text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'We consider two variational evolution problems related to Monge-Kantorovich mass transfer. These problems provide models for collapsing sandpiles and for compression molding. We prove the following connection between these problems and nonlocal geometric curvature motion: The distance functions to surfaces moving according to certain nonlocal geometric laws are solutions of the variational evolution problems. Thus we do the first step of the proof of heuristics developed in earlier works. The main techniques we use are differential equations methods in the Monge-Kantorovich theory.'
address: |
- Mathematical Sciences Research Institute\
1000 Centennial Dr.\
Berkeley CA 94720
author:
- Mikhail Feldman
title: Variational evolution problems and nonlocal geometric motion
---
[^1]
In this paper we study two models involving limits as $p \rightarrow \infty$ of solutions of $p$-Laplacian evolution problems. One is a model of collapsing sandpiles proposed by L. C. Evans, R. F. Gariepy and the author [@EFG]. Another is a model of compression molding proposed by G. Aronsson and L. C. Evans [@AEComprMot]. In both models the limits were characterized as solutions of variational evolution problems related to Monge-Kantorovich mass transfer. Solutions that have the form of the distance function to a moving boundary arise naturally in the both models. The equations of the motion of the boundary for both models were derived [*heuristically*]{} in [@EFG] and [@AEComprMot]. According to the equations the outer normal velocity of the boundary at a point depends on both the local geometry (curvatures) and the nonlocal geometry of the boundary. Thus, the motion of the boundary is a nonlocal geometric motion.
In this paper we prove rigorously the connection between geometric and variational evolution problems. Namely, assuming that a moving surface satisfying the geometric equation is given, we prove that the distance function defines a solution of the corresponding variational evolution problem. The main assumption is that the surface remains convex (or, more generally, semiconvex) during the evolution. For the collapsing sandpiles model we also prove the corresponding result for the solutions that have the form of the maximum of several distance functions (such solutions represent several sand cones interacting in the process of collapse). In the proofs we utilize the connection between the models and the Monge-Kantorovich mass transfer. This allows us to use the differential equations methods in the Monge-Kantorovich theory, which have been developed recently by L. C. Evans and W. Gangbo [@EG].
The examples suggest that solutions of the geometric equations derived in [@EFG] can develop singularities, even if the initial data is smooth. Thus we do not assume below that the moving surface is smooth. This however makes the technique more involved.
In the forthcoming paper [@Feld] we construct convexity preserving viscosity solutions of the geometric equations given convex initial data, thus finishing the proof of the heuristics developed in [@EFG] and [@AEComprMot].
This paper is organized as following. In Sections \[ColSand\] - \[Th.4\] we work on the collapsing sandpiles model. In Section \[CompMod\] we work on the compression molding model. In Appendix \[Apndx 1\] we estimate the local Lipschitz constant of the gradient of the distance function of a set.
Notation {#NotatSect}
========
Let $\Omega \in { {\bf R}}^n$ be a bounded open set. Let $\Gamma = \partial \Omega$. We will use the following two distance functions of the set $\Omega$. The first, interior distance function, is $$\label{distFunc00}
d_{\Omega}(x)=
\left\{
\begin{array}{ll}
{ \mbox{dist}}(x, \partial \Omega), & x \in \Omega, \\
0, & x\in { {\bf R}}^n \setminus \Omega.
\end{array}
\right.$$ The second, signed distance function, is $$\label{sgnDist}
d^{s}_{\Omega} (x)=\left\{
\begin{array}{cr}\mbox{dist}(x,\partial \Omega), & x\in \Omega,
\\ -{\mbox{dist}}(x,\partial \Omega), & x\notin \Omega.
\end{array}
\right.$$
Let $x \in R^n$. We denote ${\cal N}_{
\partial \Omega}(x)$ the set of all points of $\partial \Omega$ nearest to $x$, i.e., $${\cal N}_{\partial \Omega}(x) = \{
y \in \partial \Omega, \;\; \;\;
|x-y| = \inf_{z \in \partial \Omega} |x-z| \}.$$ Note that ${\cal N}_{\partial \Omega}(x)$ may have more then one point.
Let $y \in \partial \Omega$. Then $\partial \Omega$ is differentiable at $y$ if there exists $\delta>0$ such that in a suitable coordinate system $(x_1, ..., x_n)$ in ${ {\bf R}}^n$ we have $$\label{SurfDiff}
y=0,\; \;\; \Omega
\cap
B_\delta(0)=\{ x_n>\Psi(x')\} \cap
B_\delta(0) \;\;\; \mbox{ where } \;\; \Psi(x')=
o(|x'|).$$ Here $x'=(x_1, ..., x_{n-1})$, and $B_\delta(0) = B^n_\delta(0)$ is the ball in ${ {\bf R}}^n$ with radius $\delta$ and center at $0$. $\partial \Omega$ is twice differentiable at $y$ if in a suitable coordinate system in ${ {\bf R}}^{n-1}$ the function $\Psi$ has the expression $$\label{SurfTwiceDiff}
\Psi(x')=
\sum^{n-1}_{i=1} \kappa_ix^2_i+o(|x'|^2).$$ The numbers $\kappa_1,...,$ $\kappa_{n-1}$ are the principal curvatures of $\partial \Omega$ at $y$.
Let $\{\Omega_t \},\; t\in { {\bf R}}^1_+$, be a family of open bounded sets in ${ {\bf R}}^n$. We call a family of sets $\{\Omega_t \}$ Lipschitz (with respect to $t$) if there exists a constant $M>0$ such that for any $t_1, t_2$ the set $(\Omega_{t_1} \setminus\Omega_{t_2}) \cup
(\Omega_{t_2} \setminus\Omega_{t_1}) $ is subset of both $M\mid t_1 - t_2 \mid$ neighborhood (in ${ {\bf R}}^n$) of $\partial \Omega_{t_1}$ and $M\mid t_1 - t_2 \mid$ neighborhood of $\partial \Omega_{t_2}$.
Let $$\begin{aligned}
&& E=\cup_t \left(\Omega_t \times \{ t \}\right)
\subset { {\bf R}}^n \times { {\bf R}}^1_+,
\; \;\;\;
\nonumber \\
&& {\bf \Gamma} = \cup_t \left(\partial \Omega_t \times \{ t \}\right)
\subset { {\bf R}}^n \times { {\bf R}}^1_+, \;\;\;
\label{defE} \\
&& \Gamma_t = \partial \Omega_t \subset { {\bf R}}^n.
\nonumber\end{aligned}$$
Let $(y,t) \in { {\bf R}}^n \times { {\bf R}}^1_+$ be such that $y \in \Gamma_t$ and the function $(x,\tau) \rightarrow d^{s}_{\Omega_{\tau}} (x)$ is differentiable at $(y,t)$. We define the outer normal velocity of $\Gamma _t$ at $y$ as $$\label{outNormVelDef}
V(y) = {\partial \over \partial
t}d^{s}_{\Omega_t} (y).$$
Let $(y,t_0) \in {\bf \Gamma}$. The surface ${\bf \Gamma}$ is (2,1) differentiable at $(y,t_0)$ if in a suitable coordinate system $(x_1, ..., x_n)$ in ${ {\bf R}}^n$ we have $ y=0,\; \; E
\cap
B^{n+1}_\delta(0)=\{ x_n>\Psi(x', t)\} \cap
B^{n+1}_\delta(0)$ for small $\delta>0$ and $$\Psi(x',t)= v(t-t_0) + \sum^{n-1}_{i=1} \kappa_ix^2_i
+ o(\mid x'-y'\mid^2 + \mid t-t_0 \mid) .$$
${\cal L}^n$ is $n$-dimensional Lebesgue measure.
${\cal H}^n$ is $n$-dimensional Hausdorff measure.
The collapsing sandpiles model {#ColSand}
==============================
In the paper [@EFG] the $p$-Laplacian evolution problems $$\label{pHeatUnS}
\left\{
\begin{array}{ll}
\partial_t u_{p} - \mbox{div}(\mid Du_p \mid ^ {p-2} Du_p ) =0, &
\mbox{in}\;\; { {\bf R}}^n \times (0, \infty), \\
u_p=g, & \mbox{on}\;\; { {\bf R}}^n \times \{0\},
\end{array}
\right.$$ are considered in the “infinitely fast/slow diffusion limit” $p \rightarrow \infty$. $Du_p$ denotes the gradient of $u_p$ with respect to the spatial variables $x_1, ..., x_n$. The operator $$\Delta_p(u) = \mbox{div}(\mid Du_p \mid ^ {p-2} Du_p )$$ is $p$-Laplacian. The initial data $g$ is nonnegative, Lipschitz continuous function with compact support. We assume that $$\label{UnstLip}
{ \mbox{Lip}}[g] = L > 1$$ where ${ \mbox{Lip}}[g]$ is the Lipschitz constant of $g$.
The initial data $g$ satisfying (\[UnstLip\]) is unstable in the following sense. It is shown in [@EFG] that $u_p \rightarrow u$ uniformly on ${ {\bf R}}^n \times (0, \infty)$, and that $$\label{UnstLipU}
\mid Du \mid \leq 1 \;\;\; \mbox{ a.e. in } \;\; { {\bf R}}^n \times (0, \infty).$$ This limit $u$ does not depend on $t$ for $t>0$, i.e., $u=u(x)$. By (\[UnstLip\]) and (\[UnstLipU\]) we get $u \neq g$. Thus the limit solution is discontinuous at $t=0$. The transformation $g \rightarrow u$ takes place at $t=0$.
The described problem can be interpreted as a crude model of a collapsing sandpile. The function $u(x,t)$ is the height of the pile at the location $x$, time $t$. The main physical assumption is that a sandpile is stable if its slope does not exceed 1, i.e., if ${ \mbox{Lip}}[u(\cdot, t)] \leq 1$. The condition (\[UnstLip\]) implies that the initial profile $g$ is unstable. According to the model above, the initially unstable sandpile collapses instantaneously to a stable one, with the height function $u$.
The transformation $g \rightarrow u$ was studied in [@EFG] by relating it to the evolution problem governed by Monge-Kantorovich mass transfer. Introduce the convex functional $I_{\infty}: L^2({ {\bf R}}^n) \rightarrow { {\bf R}}^1 \cup \{ \infty \}$ defined by $$I_{\infty} [ v ] =
\left\{
\begin{array}{ll}
0 & \mbox{if $v \in L^2({ {\bf R}}^n),\;\;\; \mid Dv \mid \leq 1\; $ a.e.} \\
+\infty& \mbox{otherwise}.
\end{array}
\right.$$ Then $u(x)=w(x,1)$, where $w$ is the unique solution of the evolution problem $$\label{evW}
\left\{
\begin{array}{ll}
\frac{w}{t} - \partial_t w
\in \partial I_{\infty} [ w ], & \mbox{for a.e. } t \in [\frac{1}{L}, 1], \\
w=\frac{g}{L}=\hat{g}, & \mbox{at } t=\frac{1}{L},
\end{array}
\right.$$ with $L=\mbox{Lip} [g] > 1$. Note that $\mbox{Lip}[\hat{g}] =1$, and thus the evolution problem (\[evW\]) is well-posed. The connection between this evolution problem and Monge-Kantorovich mass transfer is explained in [@EFG].
In the static case the limits as $p \rightarrow \infty$ of the boundary value problems for $p$-Laplacian type equations were considered in the papers [@BDM], [@Jan] and references therein. In many cases the limits have the form of the distance function to the boundary of the domain.
A limit evolution problem that represents a stable case of the problem (\[pHeatUnS\]) was considered in the paper [@AEW] as a model of a growing stable sandpile. Let $u(x,t)$ be, as before, the height of the pile, and let $f(x,t)\geq 0$ be the sand source function which describes the rate at which sand is coming at location $x$ at time $t$. The dynamics is defined as following. Consider the nonhomogeneous evolution problem $$\label{pHeatStabl}
\left\{
\begin{array}{ll}
\partial_t u_{p} - \mbox{div}(\mid Du_p \mid ^ {p-2} Du_p ) =f, &
\mbox{in}\;\; { {\bf R}}^n \times (0, \infty), \\
u_p=0, & \mbox{on}\;\; { {\bf R}}^n \times \{0\},
\end{array}
\right.$$ and let $p \rightarrow \infty$. Then for any $T>0$ $$u_p \rightarrow u \;\;\;\;\;\;\;\; \mbox{uniformly on} \; { {\bf R}}^n \times [0, T]$$ and this limit $u$ satisfies the equation $$\label{evU}
f - \partial_t u
\in \partial I_{\infty} [ u ], \mbox{for a.e. } t \geq 0$$ with zero initial condition. In particular the case was considered when sand sources are concentrated in $m$ points, i.e., if $$f(x,t)= \sum_{k=1}^m f_k(t) \delta_{d_k}(x),$$ where $ \delta_{d_k}(\cdot)$ is the Dirac mass at the point $d_k \in { {\bf R}}^n$. The solution has the explicit form $$u=\max(0, z_1(t)-\mid x-d_1 \mid, ..., z_m(t)-\mid x-d_m \mid),$$ where the height functions $\{ z_k(t) \}_{k=1}^m$ satisfy a system of ODE. Thus the solution has the form of growing and interacting cones centered at the points $\{d_k \}_{k=1}^m $.
By analogy with the static problems and the model of growing sandpiles [@AEW], one can expect to find solutions of (\[evW\]) that have the form of distance function to the moving boundary or superposition of such, i.e., growing and interacting cones. However, unlike the situation considered in [@AEW], interacting spherical cones do not maintain their shape under the evolution defined by (\[evW\]). A more suitable class of solutions is the following class of more general “cones”. A single cone solution has the form $$\label{2}
w(x,t) = d_{\Omega_t}(x),$$ where $\{ \Omega_t \}$ is a suitable expanding family of sets. A multiple interacting cones solution has the form (for simplicity we consider the case of two cones) $$\label{2'}
w(x,t)=\max(d_{\Omega^1_t}(x),d_{\Omega^2_t}(x)),$$ where $\{ \Omega_t^1 \}$, $\{ \Omega_t^2 \}$ are suitable expanding families of sets. Thus in order to find such solutions one has to find the appropriate families of sets.
The following equations of motion of the boundaries of the sets in (\[2\]) and (\[2’\]) are derived heuristically in [@EFG]. Assume that ${\bf \Gamma}$ defined by (\[defE\]) is a smooth surface in ${ {\bf R}}^n \times { {\bf R}}^1$. Let $y \in \Gamma_t$ and let $\kappa_1, ..., \kappa_{n-1}$ be the principal curvatures of $\Gamma_t$ at $y$. Denote $\kappa = (\kappa_1, ..., \kappa_{n-1})$. Let $\gamma(y)=\gamma_{\Omega_t}(y)$ be the radius of the largest ball touching $\Gamma_t$ at $y$ from within $\Omega_t$. Let $V(y)$ be the outer normal velocity of $\Gamma _t$ at the point $y$ defined by (\[outNormVelDef\]). Then in the case of a single cone solution (\[2\]) the equation for the surface $\Gamma_t$ is: $$\begin{aligned}
\label{1}
V(y)\;\; & = & \frac{1}{t}
F(\kappa(y), \gamma(y))\;\; \mbox{ for } y \in \Gamma_t,
\nonumber \\
\mbox{where}\;\;\;\;\;\; & & \\
F(\kappa, \gamma) & = &
\frac{\int^{\gamma }_0 s\prod^{n-1}_{i=1} (1-\kappa_i
s)ds}{\int^{\gamma }_0 \prod^{n-1}_{i=1} (1-\kappa_i
s)ds}. \nonumber\end{aligned}$$
To write equations for the motion of the surfaces $\Gamma_t^k =
\partial\Omega_t^k,\; k=1,2$ in the case of the interacting cones solution (\[2’\]), we extend the function $\gamma(\cdot)=\gamma_{\Omega}(\cdot)$ of a set $\Omega \in { {\bf R}}^n$ to a function ${ {\bf R}}^n\rightarrow { {\bf R}}^1$ by setting $\gamma \equiv 0$ on ${ {\bf R}}^n\backslash \overline{\Omega}$, and $\gamma (x)=\gamma(y)$ for $x \in \Omega$ such that $y \in {\cal N}_{\partial \Omega}(x)$. Then the equations are: $$\begin{aligned}
\label{1'}
V(y)\;\;\;\;\;\;\;
& = & \frac{1}{t} F(\kappa(y), \gamma_{\Omega_t^k}(y),
\gamma_{\Omega_t^1 \cap \Omega_t^2}(y) )\;\; \mbox{ for } y \in \Gamma_t^k,
\;k=1,2,
\nonumber \\
\mbox{where}\;\;\;\;\;\;\;\;\;\;\; & & \\
F(\kappa, \gamma, \delta) & = &
\frac{\int^{\gamma}_{\delta} s\prod^{n-1}_{i=1} (1-{\kappa}_i
s)ds}{\int^{\gamma}_{\delta} \prod^{n-1}_{i=1}
(1-{\kappa}_i
s)ds}.
\nonumber\end{aligned}$$
The equations (\[1\]) and (\[1’\]) have the form of nonlocal geometric curvature motion. Geometric nonlocality is caused by the functions $\gamma(\cdot)$.
In this paper we prove that if the moving surfaces satisfy geometric equations (\[1\]) or (\[1’\]) and some regularity conditions, then the functions (\[2\]) or (\[2’\]) respectively are solutions of the evolution problem (\[evW\]). Namely, we prove the following.
\[Theorem 1\] Let $\{\Omega_t \}$ be a locally Lipschitz continuous family of open convex bounded sets, where $ t\in [a,b]$, $ b>a>0$. Suppose that the equation (\[1\]) is satisfied at every point of $(2,1)$ differentiability of the surface ${\bf \Gamma}$ defined by (\[defE\]). Then the function (\[2\]) is a solution of the evolution equation $$\label{3}
{w\over t}-\partial_t w\in \partial I_\infty [w]\;\;\; \; \mbox{for
}a.e.\;\;t\in [a,b].$$
\[Theorem 2\] Let $\{ \Omega^1_t \}$, $\{ \Omega^2_t \}$ be two Lipschitz continuous families of open bounded convex sets, where $ t\in
[a,b]$, $b>a>0$. Suppose that the equation (\[1’\]) is satisfied at every point of $(2,1)$ differentiability of the surfaces ${\bf \Gamma}^k = \cup_t(\Gamma^k_t \times \{ t \}), \;k=1,2$. Then the function (\[2’\]) is a solution of the evolution equation (\[3\]).
The convexity condition in Theorems \[Theorem 1\], \[Theorem 2\] can be relaxed to the following “semiconvexity” condition.
\[Cond.1\] [ **(lower curvature bound condition with radius $r$ ).**]{} An open set $\Omega \subset { {\bf R}}^n$ satisfies this condition if $\gamma_{ { {\bf R}}^n \setminus \overline{\Omega}}(x) \geq 2r$ for any $x \in { {\bf R}}^n \setminus \overline{\Omega }$.
It is easy to check that if a set is either convex or has $C^{1,1}$ boundary, then the set satisfies the Condition \[Cond.1\].
\[remrkCond 1\] Let the set $\Omega$ satisfy Condition \[Cond.1\]. Let $\partial \Omega$ be twice differentiable at a point $y$, i.e., (\[SurfDiff\]) and (\[SurfTwiceDiff\]) are satisfied in an appropriate coordinate system on ${ {\bf R}}^n$. Then the principal curvatures of $\partial \Omega$ at $y$ satisfy $$\kappa_i \geq -\frac{1}{2r} \;\;\; \;
\mbox{for } \; i=1,...,n-1.$$ This follows from the fact that the graph of the function $\Psi$ defined by (\[SurfDiff\]) and (\[SurfTwiceDiff\]) lies above the ball $B_{2r}^n(0, -2r)$ by Condition \[Cond.1\].
The following two theorems state that convexity can be replaced by the weaker Condition \[Cond.1\] in Theorems \[Theorem 1\] and \[Theorem 2\].
\[Theorem 3\] Let $\{\Omega_t \}$ be a locally Lipschitz continuous family of open bounded sets, where $ t\in [a,b]$, $ b>a>0$. Let for every $ t\in [a,b]$ the set $\Omega_t$ satisfy Condition \[Cond.1\] with radius $r_0 > 0$. Suppose that the equation (\[1\]) is satisfied at every point of $(2,1)$ differentiability of the surface ${\bf \Gamma}$ defined by (\[defE\]). Then the function (\[2\]) is a solution of the evolution equation (\[3\]).
\[Theorem 4\] Let $\{ \Omega^1_t \}$, $\{ \Omega^2_t \}$ be two Lipschitz continuous families of open bounded sets, where $ t\in
[a,b]$, $b>a>0$. Let for every $ t\in [a,b]$ the sets $\Omega_t^1$, $\Omega_t^2$ satisfy Condition \[Cond.1\] with radius $r_0 > 0$. Suppose that the equation (\[1’\]) is satisfied at every point of $(2,1)$ differentiability of the surfaces ${\bf \Gamma}^k = \cup_t(\Gamma^k_t \times \{ t \}), \;k=1,2$. Then the function (\[2’\]) is a solution of the evolution equation (\[3\]).
Theorems \[Theorem 1\] and \[Theorem 2\] follow from Theorems \[Theorem 3\] and \[Theorem 4\] since convex sets satisfy Condition \[Cond.1\] with any radius.
The proof of Theorems \[Theorem 3\] and \[Theorem 4\] utilizes the following relation of the equation (\[3\]) and the Monge-Kantorovich mass transfer problem. Fix $t$. Equation (\[3\]) implies that that the function $w(\cdot,t)$ is the Monge’s potential for the optimal transfer of the measure with density $\frac{w}{t}(\cdot,t)$ into the measure with density $\partial_t{w}(\cdot,t)$. Then, according to [@EG], there exists a measurable function $a(x)$, the [*mass transport density*]{}, that satisfies the following properties: $$\begin{aligned}
a \geq 0, \;\;\; & \; & \mbox{supp}(a) \subset \{x \;\mid\;\mid Dw\mid=1\},
\nonumber \\
& & \label{defMassTrDen} \\
-\mbox{div}(aDw) & = & \frac{w}{t} - \partial_tw, \nonumber\end{aligned}$$ where the last equation is understood in the weak sense and $t$ is fixed. Conversely, if there exists a function $a(x)$ satisfying (\[defMassTrDen\]), then the function $w$ is a solution of (\[3\]) for the given $t$, see Section \[Th.3\] below.
The idea of the proof of Theorem \[Theorem 3\] is as follows. Let a family of sets $\{\Omega_t\}$ satisfy (\[2\]). An open bounded set can be represented as the union of distance rays, i.e., maximal intervals that start at the boundary, on which distance to the boundary is the linear function with slope 1. The Condition \[Cond.1\] implies that the collection of distance rays possesses “nice” measure-theoretic properties. Fix $t$ and consider a distance ray $R_y$ of $\Omega_t$ starting at a smooth point $y\in \partial \Omega_t$ . Introduce the coordinate $s$ on $R_y$, the distance to the boundary. The equation (\[defMassTrDen\]) can be formally rewritten as: $$\label{frmlEq}
Da \cdot Dw + a \Delta w - \frac{w}{t} - \partial_tw = 0.$$ Since $w(\cdot, t) = { \mbox{dist}}(\cdot, \Gamma_t)$, we have on $R_y$ $$Da \cdot Dw = \frac{da}{ds}, \;\;\; w=s ,\;\;\;
-\Delta w = \sum^{n-1}_{i=1}
{{\kappa_i}\over
{1-\kappa_i s}}, \;\;\;
\partial_tw = V(y),$$ where $\kappa_1, ..., \kappa_{n-1}$ are the principal curvatures of $\Gamma_t$ at $y$, and $V(y)$ is the outer normal velocity of $\Gamma_t$ at $y$. Thus the equation (\[frmlEq\]) can be formally rewritten on the ray $R_y$ as the following ODE $$\label{TrDenODE}
a'(s)-a(s)\sum^{n-1}_{i=1}
{{\kappa_i}\over
{1-\kappa_i s}} -{s \over t}+V(y)=0,$$ This ODE has a solution $a(s)$ with zero boundary conditions at both ends of the ray $R_y$ if $V(y)$ satisfies (\[1\]). Define $a(x,t)$ by ODE (\[TrDenODE\]) with zero boundary conditions on each distance ray that starts at a smooth point of the boundary. We prove that $a(\cdot, t)$ is a measurable function that satisfies (\[defMassTrDen\]). In order to do this we examine the properties of distance function and distance rays, and use a nonsmooth (Lipschitz) change of coordinates on the set $\Omega_t$ with $n-1$ coordinates along the boundary and the $n$-th coordinate along distance rays.
We prove Theorems \[Theorem 3\] and \[Theorem 4\] in Sections \[Sets\] - \[Th.4\]. In Section \[Sets\] we examine properties of sets satisfying Condition \[Cond.1\], and in particular we describe the change of coordinates mentioned above. In Section \[SectDens\] we define the mass transport density function and examine some properties of this function. In Section \[ChngVar\] we prove that the mass balance equation (\[1\]) is satisfied. In Section \[Th.3\] we conclude the proof of Theorem \[Theorem 3\]. Finally, in Section \[Th.4\] we sketch the proof of Theorem \[Theorem 4\].
Properties of distance function and ridge sets of the sets satisfying the lower curvature bound condition {#Sets}
=========================================================================================================
Let $\Omega$ be an open set. When there is no possible confusion, we write $d(x)$ for $d_{\Omega}(x)$ defined by (\[distFunc00\]).
Let $x\in \overline{\Omega}$. Denote by $R_x$ the longest line segment through $x$ in $\overline{\Omega}$ along which $d_{\Omega}(\cdot)$ is a linear function with slope 1. We call $R_x$ the distance ray of $x$. Note that a point $x$ can have more than one distance ray, in such case denote any of them by $R_x$. If $x \in \Omega$ then all such rays $R_x$ have equal length. If $x \in \partial \Omega$ then rays $R_x$ in the case of general open set $\Omega$ can have different length. However if $\Omega$ satisfies Condition \[Cond.1\] then there exists at most one $R_x$ for $x \in
\partial \Omega$, see Proposition \[Proposition 2\] below.
For any $x \in \Omega$ one endpoint of $R_x$ lies on $\partial \Omega$ and belongs to the set ${\cal N}_{\partial \Omega}(x)$. Call this endpoint the lower end of $R_x$. Call another endpoint of $R_x$ the upper end of $R_x$ (the names “lower” and “upper” correspond to the positions of the endpoints of $R_x$ on the graph of $\;y=d_{\Omega}(x)$). Let $y, \;\; z$ be upper and lower ends of the ray $R$, and let $x=\lambda y +
(1-\lambda) z$ where $\lambda \in (0, \;1)$. Then we say that the point $x$ lies in the relative interior of the ray $R$.
\[remUnNr\] If $x \in \Omega$ lies in the relative interior of a distance ray $R_x$, then the ray $R_x$ is the unique distance ray that contains $x$, and the set ${\cal N}_{\partial \Omega}(x)$ consists of one point, see [@EH].
\[remPosReach\] It follows from Remark \[remUnNr\] that sets satisfying Condition \[Cond.1\] are sets with positive reach defined by H. Federer [@FedCMeas]. More precisely, if a set $\Omega$ satisfies Condition \[Cond.1\] with radius $r$ then reach($\Omega$)=$2r$.
The function $\gamma_{\Omega}(\cdot)$ in the equations (\[1\]), (\[1’\]) is called ridge function and can be expressed as following: $$\label{defGamma}
\gamma_{\Omega}(x) =
\left\{
\begin{array}{ll}
0 & x \in { {\bf R}}^n \setminus \overline{\Omega}, \\
\mid R_x \mid & x \in \Omega, \\
\sup \mid R_x \mid & x \in \partial \Omega.
\end{array}
\right.$$ The function $\gamma_{\Omega}(\cdot)$ is uppersemicontinuous in ${ {\bf R}}^n$, cf. [@EH], Proposition 3.2.
Define the set $${\cal R} = \{\; x \in \overline{\Omega} \;\;\; | \;\;\;
d_{\Omega}(x) = \gamma_{\Omega}(x)\; \}.$$ The set ${\cal R}$ is called [*ridge set*]{} of $\Omega$. Note that if $x \in {\cal R}$ then either $x$ is the upper end of some distance ray of $\Omega$, or otherwise $x$ lies at $ \partial \Omega$ and there are no points $y\in \Omega$ such that $x \in {\cal N}_{\partial \Omega}(y)$.
\[Proposition 2\] Let $\Omega \subset { {\bf R}}^n$ be a bounded open set satisfying Condition \[Cond.1\] with radius $r>0$. Let $\Omega_r$ be the $r$-neighborhood of $\Omega.$ Then
\(a) $\gamma_{{ {\bf R}}^n \setminus \overline{\Omega}_r }(x) \geq r$ for all $x \in { {\bf R}}^n \setminus \Omega_r$, i.e., the set $\Omega_r$ satisfies Condition \[Cond.1\] with radius $r\over 2$.
\(b) For any $y_r \in \partial \Omega_r$ there exists a unique $y \in
\partial \Omega$ such that $|y_r-y|=r$, and $|y_r-x|>r$ for all $x\in \overline{\Omega}\backslash \{ y \}$. For any $y\in \partial \Omega$ there exists $y_r \in \partial \Omega_r$ such that $|y-y_r|=r$. If $\partial \Omega$ is differentiable at $y$ then $y-y_r$ is orthogonal to $\partial \Omega$ at $y$. If there exists a ball inside $\Omega$ touching $\partial \Omega$ at $y$, then $\partial \Omega$ is differentiable at $y$. In particular, $\partial \Omega_r$ is everywhere differentiable.
\(c) For any $x\in \Omega$ we have $$\label{distRelat}
\mbox{dist}(x,\partial \Omega_r)=\mbox{dist}(x,\partial \Omega)+r.$$ If $x \in {\Omega_r} \setminus \overline{\Omega}$ then ${ \mbox{dist}}(x, \partial \Omega_r) < r$ and $x \notin {\cal R}_r$, where ${\cal R}_r $ is ridge set of $\Omega_r$.
\(d) Sets $\Omega$ and $\Omega_r$ have the same ridge set, i.e., ${\cal R}={\cal R}_r$.
\(a) Let $x \in { {\bf R}}^n \setminus \overline{\Omega}_r$. Then ${ \mbox{dist}}(x, \Omega) = R > r$. Let $y \in \partial{\Omega}$ be a point such that $\mid x-y \mid = R$. Let $y_r$ be such point of the interval connecting $x$ and $y$ that $\mid y-y_r \mid = r$. Then $y_r \in \overline{\Omega}_r$.
Consider first the case $R \geq 2r$. Then ${ \mbox{dist}}(x, \Omega_r)\geq r$. Since if this is not true then there exists such point $z_1 \in \overline{\Omega}_r$ that $\mid x-z_1 \mid < r$, and there exists $z_2 \in \overline\Omega $ such that $\mid z_1 - z_2 \mid \leq r$. Now we get $R \leq \mid x-z_2\mid \leq \mid x-z_1 \mid + \mid z_1 - z_2 \mid < 2r$, a contradiction. So $\gamma_{{ {\bf R}}^n \setminus \overline{\Omega}_r }(x)
\geq { \mbox{dist}}(x, \Omega_r)\geq r$.
Consider now the case $r<R < 2r$. Then by Condition \[Cond.1\] we have $ \gamma_{{ {\bf R}}^n \setminus
\overline{\Omega} }(x) \geq 2r$ and so $R={ \mbox{dist}}(x, \Omega) < \gamma_{{ {\bf R}}^n \setminus
\overline{\Omega} }(x)$. Thus $y$ is the unique point on $\partial \Omega$ such that $\mid x-y \mid = { \mbox{dist}}(x,\Omega)$. Let $z = y + 2r\frac{x-y}{\mid x-y \mid}$. Then from the condition \[Cond.1\] and [@EH], Lemma 3.4, it follows that for any $\tilde{x}$ of the form $tx+(1-t)z$ where $0<t<1$ the point $y$ is the unique point on $\partial \Omega$ such that ${ \mbox{dist}}(\tilde{x}, \partial \Omega)
= \mid \tilde{x} - y \mid$. Then for each such point $\tilde{x}$, the point $y_r$ is the unique point on $\partial \Omega_r$ such that ${ \mbox{dist}}(\tilde{x},
\partial \Omega_r)
= \mid \tilde{x} - y_r \mid$. For suppose on the contrary that there exists $z_1 \in
\partial \Omega_r$ such that $z_1 \neq y_r$ and $\mid \tilde{x} - z_1 \mid \leq \mid
\tilde{x} - y_r\mid $. Then there exists $z_2 \in \partial \Omega$ such that $\mid z_1 - z_2 \mid \leq r$. Consider first the case $z_2 \neq y$. Then $$\mid \tilde{x} - y \mid = \mid \tilde{x} - y_r \mid + \mid y_r - y \mid =
\mid \tilde{x} - y_r \mid + r \geq \mid \tilde{x} - z_1 \mid +
\mid z_1 - z_2 \mid \geq \mid \tilde{x} - z_2 \mid,$$ a contradiction with the fact that $y$ is the unique nearest to $\tilde{x}$ point on $\partial \Omega$. Consider now the case $z_2 = y$. Since the ball of radius $R$ and center $\tilde{x}$ lies outside $\Omega$ and $|z_1-y| \leq r$, it follows that $z_1$ does not lie on the interval connecting $\tilde{x}$ and $ y$, and so $$\mid \tilde{x} - y \mid < \mid \tilde{x} - z_1 \mid + \mid z_1 - y \mid \leq
\mid \tilde{x} - y_r \mid +r = \mid \tilde{x} - y \mid,$$ a contradiction. So, $y_r$ is the unique nearest to $\tilde{x}$ point on $\partial \Omega_r$ for any $\tilde{x}$ on the interval connecting $z$ and $y_r$, so $\gamma_{{ {\bf R}}^n \setminus \overline{\Omega}_r}(x) \geq \mid z - y_r \mid = r$.\
(b) Let $y_r \in \partial \Omega_r$. Then ${ \mbox{dist}}(y_r, \Omega) = r$. Thus there exists $y\in \partial \Omega$ such that $\mid y-y_r \mid = r$. By Condition \[Cond.1\] , $\gamma_{{ {\bf R}}^n \setminus
\overline{\Omega}}(y_r) \geq 2r
> { \mbox{dist}}(y_r, \partial \Omega)$, and thus by [@EH], Lemma 3.4, $y$ is the unique point of $\partial \Omega$ nearest to $y_r$.
Let $y\in \partial \Omega$. Let $y_i \in { {\bf R}}^n \setminus
\overline{\Omega} $ for $i=1,2,...$, and let $y_i \rightarrow y$. By condition \[Cond.1\], $\gamma_{{ {\bf R}}^n \setminus
\overline{\Omega}}(y_i) \geq 2r$, and by uppersemicontinuity of $\gamma_{{ {\bf R}}^n \setminus \overline{\Omega}}(\cdot)$ we get $\gamma_{{ {\bf R}}^n \setminus \overline{\Omega}}(y) \geq 2r$. Thus we have shown that every point of $\partial \Omega$ has an exterior tangent ball of the radius $2r$. So if a point $y$ of $\partial \Omega$ has an interior tangent ball, then $\partial \Omega$ is differentiable at $y$.
Let $y \in \partial \Omega$. Then, as we have shown, $\gamma_{{ {\bf R}}^n \setminus \overline{\Omega}}(y) \geq 2r$. Thus there exists a distance ray $R_y$ of the set ${ {\bf R}}^n
\setminus \overline{\Omega}$ starting at $y$ such that the length of $R_y$ is at least $2r$. Let $e$ be the unit vector in the direction of $R_y$, let $y_r = y + re$. Then ${\cal N}_{\partial \Omega}(y_r) = \{ y \}$ by Remark \[remUnNr\]. Since $\mid y-y_r \mid = r$ we get $y_r \in \partial \Omega_r$.
We have shown that the interval connecting $y$ and $y_r$ lies on the distance ray of the set ${ {\bf R}}^n \setminus \overline{\Omega}$ starting at $y$, and on distance ray of the set $\Omega_r$ starting at $y_r$. It follows that the vector $y-y_r$ is orthogonal to $\partial \Omega_r$ at $y_r$ and also to $\partial \Omega$ at $y$ if $\partial \Omega$ is differentiable at $y$.\
(c) Let $x\in \Omega^0, \, y\in \partial \Omega$, and $|x-y|=
{ \mbox{dist}}(x, \partial \Omega)$. Then, by (b), $$y_r := y + \frac{r}{|y-x|}(y-x) \in \partial \Omega_r.$$ Clearly, $$|x-y_r| = |x-y| + r = d_{\Omega}(x) + r.$$ Assume that there exists $z_r \in \partial \Omega_r$ such that $$|x - z_r| < d_{\Omega}(x) + r.$$ Let $z$ be the point of intersection of $\partial \Omega$ with the interval connecting $z_r$ and $x$ (such point exists since $x \in \Omega, z_r \notin \overline{\Omega}$). Then $|z-z_r| \geq r$ and so $$|x - z| < d_{\Omega}(x).$$ This contradicts the fact that $z \in \partial \Omega$. Thus $y_r \in {\cal N}_{\partial \Omega_r}(x)$ and thus (\[distRelat\]) is proved.
It remains to prove the last assertion of the statement (c).
Let $x\in {\Omega}_r\backslash
\overline{\Omega}$ . Let $y_r \in {\cal N}_{\partial \Omega_r}(x)$. Let $R_{y_r}$ be the ray orthogonal to $\partial \Omega_r$ at $y_r$ and let $y=R_{y_r} \cap \partial \Omega$. Then by (b) we have the following: $x\in R_{y_r},\;\;x$ lies on the interval connecting $y$ and $y_r$, and $|y-
y_r|=r=\mbox{dist}(y,\partial \Omega_r)$. Then it follows from [@EH], Lemma 3.4 that $x \notin {\cal R}_r$. Also, ${ \mbox{dist}}(x, \partial \Omega_r) = \mid x-y_r \mid < |y-
y_r|=r$.\
(d) We show first that ${\cal R}\subset {\cal R}_r$. If $x\in \Omega$ and $B_\rho(x)\subset \Omega$, then $B_{\rho +r}(x)\subset \overline{\Omega}_r$. Let $x\in {\cal R}\backslash \partial \Omega$. Let $y \in {\cal N}_{\partial \Omega}(x)$ and let $y_r=y+r\frac{y-x}{|y-x|}$. Then by (b),(c) we see that $y_r \in \partial \Omega_r,$ and $|x-y_r|={ \mbox{dist}}(x, \partial \Omega) +r=
\mbox{dist}(x,\partial
\Omega_r)$. Let $\tilde{x}=y+(1+\varepsilon)(x-y)$ where $\varepsilon>0$ is small enough so that $\tilde{x}\in \Omega$. Since $x\in {\cal R}$, it follows that ${ \mbox{dist}}(\tilde{x}, \partial \Omega)<|\tilde{x}-
y|=(1+\varepsilon){ \mbox{dist}}(x, \partial \Omega)$. Let $\tilde{y}\in\partial\Omega$ be such point that ${ \mbox{dist}}(\tilde{x}, \partial \Omega)=|\tilde{x}-\tilde{y}|$, let $\tilde{y}_r=\tilde{y} +r\frac{\tilde{y}-
\tilde{x}}{|\tilde{y}-\tilde{x}|}$, then $\tilde{y}_r\in
\partial \Omega_r$ and $|\tilde{x}-
\tilde{y}_r|={ \mbox{dist}}(\tilde{x}, \partial \Omega_r)$. Then $$\mbox{dist}(\tilde{x}, \partial \Omega_r)=
\mid \tilde{y}-\tilde{x} \mid + r
<(1+\varepsilon)
{ \mbox{dist}}(x, \partial \Omega)+r
=|\tilde{x}-y_r|,$$ so $\gamma_{\Omega_r}(x)= |x-y_r| = d_{\Omega_r}(x)$. Thus $x\in {\cal R}_r$.
Let $x\in{\cal R}\cap \partial \Omega$, i.e., there does not exist a ball inside $\Omega$ that touches $\partial \Omega$ at $x$. Let $B_r(y_r)$ be a ball of radius $r$ and center at such point $y_r$ that $\overline{B}_r(y_r)\cap \overline{ \Omega}=\{x\}$, such ball exists by Condition \[Cond.1\]. Then ${ \mbox{dist}}(y_r, \partial
\Omega)=r$, and by (b) $x$ is the unique point of $\partial \Omega$ such that $|x - y_r| = r$. Let $\tilde{x}=y_r+(1+\varepsilon)(x-y_r)$ for $\varepsilon\in(0, {r \over 2})$. Given $\varepsilon$, there are two possibilities: either $\tilde{x} \in \Omega$ or $\tilde{x} \notin \Omega$.
Let $\tilde{x} \in \Omega$. Then, since $x\in {\cal R}\cap \partial
\Omega,\;\;{ \mbox{dist}}(\tilde{x}, \partial \Omega)<|x-\tilde{x}|=\varepsilon r$. Let $\tilde{y}\in \partial \Omega$ be such that $|\tilde{x}-
\tilde{y}|={ \mbox{dist}}(\tilde{x}, \partial \Omega)$, let $\tilde{y}_r=\tilde{y}+
r\frac{\tilde{y}-
\tilde{x}}{|\tilde{y}-\tilde{x}|}$. Then $\tilde{y}_r \in \partial
\Omega_r$ and by (c) we get $|\tilde{x}-\tilde{y}_r|={ \mbox{dist}}(\tilde{x}, \partial \Omega)+r=
\mbox{dist}
(\tilde{x},\partial \Omega_r)$, and so $\mbox{dist}(\tilde{x}, \partial
\Omega_r)<(\varepsilon +1)r= |\tilde{x}-y_r| $, so $y_r \notin {\cal N}_{\partial \Omega_r}(\tilde{x})$.
Let now $\tilde{x} \notin \Omega$, i.e., $\tilde{x} \in
\Omega_r \setminus \overline{\Omega}$ since $\varepsilon < {r \over 2}$. Then, by (c), ${ \mbox{dist}}(\tilde{x}, \partial \Omega_r) < r$, but $|\tilde{x}-y_r| = (\varepsilon +1)r$, so $y_r \notin {\cal N}_{\partial \Omega_r}(\tilde{x})$.
Thus $y_r \notin {\cal N}_{\partial \Omega_r}(\tilde{x})$ for any small $\varepsilon>0$. This implies $x \in {\cal R}_r.$
Thus ${\cal R}\subset {\cal R}_r.$
Now we show that ${\cal R}_r \subset {\cal R} $. By (b), (c) we get ${\cal R}_r \cap (\overline{\Omega}_r \setminus \overline{\Omega} ) =
\emptyset$.
It remains to consider $x\in \overline{\Omega}\cap {\cal R}_r $. Let first $x\in \Omega
\cap {\cal R}_r$. Let as above $y \in {\cal N}_{\partial \Omega}(x)$, $ y_r =y+r\frac{y-x}{|y-x|} \in {\cal N}_{\partial \Omega_r}(x)$, $\tilde{x}=y+(1+\varepsilon)(x-y)$ where $\varepsilon>0$ is small. From $x \in {\cal R}_r$ we conclude that $\mbox{dist}(\tilde{x}, \partial \Omega_r)<|\tilde{x}-
y_r|=(1+\varepsilon){ \mbox{dist}}(x, \partial \Omega)+r$. Let $\tilde{y}_r\in
\partial
\Omega_r$ be such that $|\tilde{x}-\tilde{y}_r|=\mbox{dist}(\tilde{x},
\partial
\Omega_r)$ and let $\tilde{y}$ be the unique point of $\partial \Omega$ nearest to $\tilde{y}_r$, then $|\tilde{y}-
\tilde{y}_r|=r$ (uniqueness of $\tilde{y}$ follows from (b)). Then by (c), (b) the points $\tilde{x},\tilde{y},\tilde{y}_r$ lie on one distance ray and ${ \mbox{dist}}(\tilde{x}, \partial \Omega)=
|\tilde{x}-\tilde{y}|$, so $${ \mbox{dist}}(\tilde{x}, \partial \Omega)=|\tilde{x}-\tilde{y}|=|\tilde{y}_r-
\tilde{x}|-
|\tilde{y}_r-y|<|\tilde{x}-y_r|-r=$$ $$(1+\varepsilon){ \mbox{dist}}(x, \partial \Omega)=(1+\varepsilon)|x-
y|=|\tilde{x}-y|$$ and so $\tilde{x}\in {\cal R}$.\
The case of $x\in \partial \Omega \cap {\cal R}_r$ is similar to the case above.
\[RemSet1\] Examples of nonconvex polygons on a plane show that without assuming some condition of the type of Condition \[Cond.1\] the statements (b)-(d) of Proposition \[Proposition 2\] are not true in general.
\[Proposition 3\] Let $\Omega \in R^n$ satisfy Condition \[Cond.1\] with radius $r$. Then $\partial \Omega_{\rho}\in C^{1,1}$ for any $\rho \in (0, \frac{r}{M})$.
The assertion follows from inequality 4.8(8) of [@FedCMeas] and Proposition \[Proposition 2\] (d). Also, inequality (\[1.1\]) proved in Appendix A1 can be applied instead of 4.8(8) of [@FedCMeas].
Recall the following facts (see e.g. [@GilTr], Lemma 14.16, 14.17). Let $\Omega \subset { {\bf R}}^n $ be an open bounded set with $C^2$ boundary $\partial \Omega$. Let $\Omega_r$ be, as above, $r$-neighborhood of $\Omega$. Then $\partial \Omega_r\in C^2$ for small $r$. If $y\in \partial \Omega$ and if $y_r\in \partial
\Omega_r$ is the unique point on $\partial \Omega_r$ such that $|y-
y_r|=r$ then the principal coordinate systems of $\partial \Omega$ at $y$ and of $\partial \Omega_r$ at $y_r$ are parallel. Denote by $\kappa_i$ and $\kappa_{i,r},\;\; (i=1,...n-1)$ the principal curvatures of $\partial \Omega$ at $y$ and of $\partial \Omega_r$ at $y_r$ respectively. We have $$\label{relNbhd}
\kappa_{i,r}=\frac{\kappa_i}{1+\kappa_ir}\;\;\;
\mbox{ for }\;\;i=1,...,n-1.$$ In the next two propositions we show similar facts for sets with nonsmooth boundaries satisfying Condition \[Cond.1\] and for Lipschitz families of such sets.
\[Proposition 4\] Let $\Omega \subset { {\bf R}}^n$ be an open set satisfying Condition \[Cond.1\] with radius $r_0$. Let $0<r \leq r_0$. Fix $y\in \partial \Omega$. Let $y_r\in \partial
\Omega_r$ be such point that $|y-y_r|=r$. Then:
a\) Let $\partial \Omega$ be twice differentiable at $y$, i.e., in a suitable coordinate system in ${ {\bf R}}^n$ we have $ y=0,\; \; \Omega
\cap
B_\delta(0)=\{ x_n>\Psi(x')\} \cap
B_\delta(0)$ for small $\delta>0$ and $\Psi(x')=
\sum^{n-1}_{i=1} \kappa_ix^2_i+o(|x'|^2)$. Then the point $y_r$ is unique given the point $y$, and in the coordinate system introduced above $y_r=(0,-r)$, $\Omega_r\cap B_{\delta_1}(y_r)=\{ x_n>\Psi_r (x')\}
\cap B_{\delta_1}(y_r) $ for sufficiently small $\delta_1>0$, where $\Psi_r(x')=-r+\sum^{n-1}_{i=1}\frac{\kappa_i}
{1+\kappa_ir}x^2_i+o(|x'|^2)$, i.e., $\Omega_r$ is twice differentiable at $y_r$ and (\[relNbhd\]) holds.
b\) If $\partial \Omega_r$ is twice differentiable at $y_r$ and $\gamma_{\Omega}(y) > 0$ then $\partial \Omega$ is twice differentiable at $y$ and (\[relNbhd\]) holds (i.e., the functions $\Psi$ and $\Psi_r$ have same expansions at 0 as in a)).
a\) Uniqueness of $y_r$ and the fact that $y_r=(0,-r)$ follow from smoothness of $\partial \Omega$ at $y$. So the interval connecting $y$ and $y_r$ lies on $x_n$-axis. By Propositions \[Proposition 2\] and \[Proposition 3\] the surface $\partial \Omega_r$ is of class $C^{1,1}$, and the interval connecting $y$ and $y_r$ is orthogonal to $\partial \Omega_r$. So $\Omega_r\cap B_{\delta_1}(y_r)=\{ x_n>\Psi_r (x')\}
\cap B_{\delta_1}(y_r) $ for some $\delta_1 > 0$, where $\Psi_r$ is a $C^{1,1}$ function.
Let $\varepsilon>0$. There exists $\sigma>0$ such that $$|\Psi(x')-\sum^{n-1}_{i=1}
\kappa_ix^2_i|<\varepsilon|x'|^2 \;\; \mbox{if} \;\; |x'|< \sigma.$$ Denote $A_\varepsilon=\{(z',z_n) \in { {\bf R}}^n \;\; |\;\; z_n\geq \sum^{n-
1}_{i=1}(\kappa_i-\varepsilon)z^2_i\}$. Then we get: $$\partial \Omega \cap B_\sigma (0)\subset
A_\varepsilon\backslash A_{-\varepsilon}.$$ Since $\gamma_{ { {\bf R}}^n \setminus \overline{\Omega}}(y) \geq 2r$, the point $y_r$ lies in the relative interior of the distance ray $R_y$ of the set $ { {\bf R}}^n \setminus \overline{\Omega}$. Then there exists $\sigma_1>0$ such that for any point $x_r \in
\cap B_{\sigma_1}(y_r)$ we have $ {\cal N}_{\partial \Omega}(x_r) \subset \partial
\Omega \cap B_{\sigma}(0)$. (Proof: suppose this is false, then there exist $\sigma_1>0$ and a sequence $x_r^k \rightarrow y_r$ such that for $x^k \in {\cal N}_{\partial \Omega}(x_r^k)$ we get $|x^k-y| > \sigma_1$. Then, passing to a subsequence, we get $x^k \rightarrow x$ where $x\in \partial \Omega$ and $x \neq y$. By continuity of distance function we get $x \in {\cal N}_{\partial \Omega}(y_r)$ and $x \neq y$. By Condition \[Cond.1\] $y_r$ lies in the relative interior of the distance ray $R_y$ of the set ${ {\bf R}}^n \setminus \overline{\Omega}$. Thus $y_r$ has a unique nearest point on $\partial \Omega$. A contradiction.). So $$\label{Aepsilon}
x\in A_\varepsilon \backslash A_{-\varepsilon}\;\;\;\mbox{ if }\;\;
x \in {\cal N}_{\partial \Omega}(x_r)\;\;\; \mbox{ where }\;\;
x_r \in B_{\sigma_1}(y_r).$$
Let $N_{\varepsilon}$ be $r$-neighborhood of the set $A_{\varepsilon}$. By Remark \[remrkCond 1\] we have $\kappa_i \geq -\frac{1}{2r}$. Let $\varepsilon < \frac{1}{4r}$. Then (by [@GilTr] Lemma 14.16, 14.17), $\partial N_\varepsilon$ is the graph of a $C^\infty$ function $g_\varepsilon$, and $\frac{\partial^2g_\varepsilon}{\partial x_i\partial
x_j}(0)=\frac{\kappa_i-\varepsilon}{1+(\kappa_i-
\varepsilon)r}\delta_{ij}$. Let $$\tilde{N}_{\varepsilon,\delta}=\{(z',z_n)\;\; |\;\;
z_n\geq \sum^{n-1}_{i=1}(\frac{\kappa_i-\varepsilon}{1+(\kappa_i-
\varepsilon)r}-\delta)z^2_1-r\}.$$ If $\sigma_2>0$ is small enough then $$\label{Nvareps}
\tilde{N}_{\pm\varepsilon,-\varepsilon}\cap B_{\sigma_2}(y_r)\subset
N_{\pm\varepsilon}\cap B_{\sigma_2} (y_r) \subset \tilde{N}_{\pm\varepsilon,
\varepsilon} \cap
B_{\sigma_2} (y_r).$$ Thus if we choose $\sigma_3$ small, then we get by (\[Aepsilon\]), (\[Nvareps\]) $$\partial \Omega_r \cap B_{\sigma_3}(y_r) \subset \tilde{N}_{\varepsilon,
\varepsilon}\backslash \tilde{N}_{-\varepsilon,-\varepsilon},$$ which means that for $x_r =(x'_r, \Psi_r(x'_r))\in \partial \Omega_r \cap
B_{\sigma_3}(y_r)$ we have $$\sum^{n-1}_{i=1}(\frac{\kappa_i-\varepsilon}{1+(\kappa_i-
\varepsilon)r}-\varepsilon)x^2_{r,i}-r\leq \Psi_r(x'_r)\leq \sum^{n-
1}_{i=1}(\frac{\kappa_i+\varepsilon}{1+(\kappa_i+
\varepsilon)r}+\varepsilon)x^2_{r,i}-r.$$ Sending $\varepsilon$ to 0 we conclude the proof.
b\) By Proposition \[Proposition 2\] the condition $\gamma_{\Omega}(y) > 0$ implies that $\gamma_{\Omega_r}(y_r) > r$. Now we can perform calculation similar to the proof of assertion a).
We need similar facts for families of sets. Let $\{\Omega_t \} $ be a family of open sets in ${ {\bf R}}^n$. Let $(\Omega_t)_{r}$ be $r$-neighborhood of $\Omega_t$ in ${ {\bf R}}^n$. Define $$\label{defEr}
E_r = \cup_t [(\Omega_t)_r \times \{ t \}], \;\;\;
{\bf \Gamma}_r = \cup_t[ \partial (\Omega_t)_r \times \{ t \}].$$
\[Proposition 4t\] Let $\{\Omega_t \} $ be a locally Lipschitz continuous family of open sets in ${ {\bf R}}^n$. Let for every $ t\in { {\bf R}}^1_+$ the set $\Omega_t$ satisfy Condition \[Cond.1\] with radius $r_0 > 0$. Let $0<r \leq r_0$. Fix $y\in \partial \Omega_{t_0}$. Let $y_r\in \partial
(\Omega_{t_0})_r$ be such point that $|y-y_r|=r$. Then:
a\) Let ${\bf \Gamma}$, defined by (\[defE\]), be (2,1) differentiable at $(y, t_0)$. That is in a suitable coordinate system $(x_1, ..., x_n)$ in ${ {\bf R}}^n$ we have $$y=0,\; \; \Omega
\cap
B^{n+1}_\delta(0)=\{ x_n>\Psi(x',t)\} \cap
B^{n+1}_\delta(0)$$ for small $\delta>0$ and $$\Psi(x',t)=v(t-t_0)+
\sum^{n-1}_{i=1} \kappa_ix^2_1+o(|t-t_0|+|x'|^2).$$ Then the point $y_r$ is unique given the point $y$, and in the coordinate system introduced above $y_r=(0,-r)$, $$\Omega_r\cap B^{n+1}_{\delta_1}(y_r)=\{ x_n>\Psi_r (x',t)\}
\cap B^{n+1}_{\delta_1}(y_r)$$ for sufficiently small $\delta_1>0$, where $$\Psi_r(x',t)=-r+v(t-t_0)+\sum^{n-1}_{i=1}\frac{\kappa_i}
{1+\kappa_ir}x^2_1+o(|t-t_0|+|x'|^2).$$
b\) If ${\bf \Gamma}_r$ is (2,1) differentiable at $(y_r,t_0)$ and $\gamma_{\Omega_{t_0}}(y) > 0$ then ${\bf \Gamma}$ is (2,1) differentiable at $y$ and the functions $\psi$ and $\psi_r$ have same expansions at $(0, t_0)$ as in a).
The proof is similar to the one of Proposition \[Proposition 4\]. In particular, the sets $A_\varepsilon$, $N_\varepsilon$ and $\tilde{N}_{\varepsilon,\delta}$ are defined as following. $$A_\varepsilon=\{\;(z',z_n, t)\;\; |\;\; z_n\geq v(t-t_0) + \sum^{n-
1}_{i=1}(\kappa_i-\varepsilon)z^2_i - \varepsilon |t-t_0| \;\}.$$ $N_\varepsilon$ is defined as following: for each $t=t^*$ the set $N_\varepsilon \cap \{ t=t^* \}$ is $\varepsilon$-neighborhood in ${ {\bf R}}^n$ of the set $\{x \;\; | \;\; (x,t^*) \in A_\varepsilon\}$. Finally, $$\tilde{N}_{\varepsilon,\delta}=\{\;(z',z_n)\;\; |\;\;
z_n\geq v(t-t_0) +
\sum^{n-1}_{i=1}(\frac{\kappa_i-\varepsilon}{1+(\kappa_i-
\varepsilon)r}-\delta)z^2_1-r -( \varepsilon + \delta) |t-t_0|\; \}.$$ The rest of the argument does not change.
To derive further properties of the distance function of a set $\Omega$ satisfying the Condition \[Cond.1\], we need to integrate over such sets. To do this it is convenient to decompose the set by suitable subsets, and define a local coordinate system in each subset. The construction is carried out below and can be roughly described as following. The subsets are unions of distance rays passing through subsets of the boundary. The coordinate systems consist of variables $x_1, ..., x_{n-1}$ on the boundary, and the variable $x_n$ along the distance rays. However, since $\partial \Omega$ is not smooth enough, we have to use $\partial \Omega_r$. Now we turn to the construction.
By Proposition \[Proposition 3\] the set $\Omega_r$ for small enough $r>0$ has $C^{1,1}$ boundary. Fix such $r$. We can choose sets ${\cal U}_1,...,{\cal U}_N$, bounded and open in ${ {\bf R}}^n$, so that $\partial \Omega_r \subset \cup_{k=1}^{N}{\cal U}_k$, and for each $k=1,...,N$ in a suitable coordinate system in ${ {\bf R}}^n$ we have $$\Omega_r \cap {\cal U}_k=
\{\; (x',x_n)\;\;|\;\;x'\in \tilde{U}_k,\;x_n>\Phi_k(x')\;\}\cap
{\cal U}_k,$$ where $\tilde{U}_k =\{\;x'\;|\;(x',x_n)\in {\cal U}_k \;\}
\subset { {\bf R}}^{n-1}$ is an open set, and $\Phi_k$ is a $C^{1,1}$ function on ${ {\bf R}}^{n-1}$.
We can also choose open sets ${\cal U}_0$ and $ {\cal U}_{N+1}\subset { {\bf R}}^n$ such that ${\cal U}_0\subset \Omega_r$, ${\cal U}_{N+1}
\subset { {\bf R}}^n\backslash \Omega_r$, both ${\cal U}_0$ and ${\cal U}_{N+1}$ do not intersect $\partial \Omega$, and ${ {\bf R}}^n=\cup^{N+1}_{k=0}{\cal U}_k$. Let $\tilde{\Psi}_k,\;k=1,...,N$, be a smooth partition of unity on ${ {\bf R}}^n$ related to the sets ${\cal U} _0, {\cal U} _1,..., {\cal U} _{N+1}$, i.e., $\sum^{N+1}_{k=0}\tilde{\Psi}_k\equiv 1$ on ${ {\bf R}}^n$ and $\tilde{\Psi}_k\in C^\infty_0({\cal U}_k)$. Then by our choice of ${\cal U} _0$ and ${\cal U} _{N+1}$ we get $\sum^N_{k=1}\tilde{\Psi}_k\equiv 1$ on $\partial
\Omega_r$.
Define maps $G_k:\tilde{U}_k\times { {\bf R}}^1\rightarrow
{ {\bf R}}^n\;\;(k=1,...,N)$ by $$\label{I2}
G_k(x',x_n)=y+x_nDd^s_{\Omega_r}(y)$$ where $y=(x',\Phi_k(x'))\in \partial \Omega_r$, and $d^s_{\Omega_r}(\cdot)$ is the signed distance to $\partial \Omega_r$.
For each $z\in \partial \Omega_r$ the vector $Dd^s_{\Omega_r}(z)$ is the inner unit normal to $\partial \Omega_r$ in $z$. Since $\partial \Omega_r$ is a $C^{1,1}$ manifold, $Dd^s_{\Omega_r}$ is Lipschitz on $\partial
\Omega_r$. Then since $\Phi_k$ is $C^{1,1}$ on ${ {\bf R}}^{n-1}$, the map $G_k$ is Lipschitz on bounded subsets of $\tilde{U}_k\times { {\bf R}}^1$. In particular, the map $\tilde{U}_k \rightarrow \partial \Omega_r$ defined by $x' \rightarrow G_k(x',0)$ is Lipschitz.
From that by explicit computation we get the following
\[jacobian lemma\] Let the map $G_k:\tilde{U}_k\times { {\bf R}}^1\rightarrow
{ {\bf R}}^n$ be defined by (\[I2\]). Then the Jacobian $JG_k$ is a locally bounded measurable function. For ${\cal L}^{n-1}$ a.e. point $x' \in \tilde{U}_k$ the surface $\partial \Omega_r$ is twice differentiable at the point $G_k(x',0)$. For every such point $x'$ the map $G_k$ is differentiable at $(x', x_n)$ for every $x_n \in { {\bf R}}^1$. At every such point $(x', x_n)$ the Jacobian of $G_k$ is given by $$\label{jacobian}
J_nG_k(x',x_n)=\sqrt{1+|D\Phi_k(x')|^2}\prod^{n-
1}_{i=1}\left[1-\kappa_{r,i} x_n\right],$$ where $\kappa_{r,1},..., \kappa_{r,n-1}$ are the principal curvatures of $\partial \Omega_r$ at $G_k(x', 0)$.
Denote by $U_k$ the sets $$\label{defUk}
U_k =
\cup_{x'\in \tilde{U}_k}\{\; (x',x_n)\;|\;x_n\in
[0,\gamma_{\Omega_r}(y))\;\}\subset { {\bf R}}^n,$$ where $\gamma_{\Omega_r}:{ {\bf R}}^n\rightarrow
{ {\bf R}}^1$ is the ridge function (\[defGamma\]) of the set $\Omega_r$. Note that $\gamma_{\Omega_r}(z)\geq r$ for all $z\in
\overline{\Omega}_r$ by Proposition \[Proposition 2\] (c, d).
The map $G_k$ is one-to-one on $U_k$ (by definition of $\gamma_{\Omega_r}$ and Lemma 3.4 of [@EH]). In addition $\Omega
\backslash {\cal R}\subset \cup^N_{k=1} G_k(U_k),$ since every point of $\Omega_r \backslash {\cal R}$ lies inside some distance ray. For each $k=1, ..., N$ define a function $\Psi_k:{ {\bf R}}^n\rightarrow { {\bf R}}^1$ by extending the function $\tilde{\Psi}_k $ from the boundary inside $\Omega_r$ as a constant along distance rays and defining $\Psi_k$ to be the $0$ on the ridge of $\Omega_r$ and outside $\Omega_r$, i.e., $$\label{defPsi}
\Psi_k(z)=\left\{ \begin{array}{lcl}\tilde{\Psi}_k(y), &
\mbox{where }y\in {\cal N}_{\partial \Omega_r}(z), &\;\;\;
\mbox{if }z\in \Omega_r \setminus {\cal R} \\ 0, & & \;\;\;\mbox{otherwise.}
\end{array} \right.$$ Note that $y$ is uniquely defuned by $z$ in the first case of (\[defPsi\]). We have $$\mbox{supp} \Psi_k \subset G_k(U_k); \;\;\;\;
\sum^N_{k=1}\tilde{\Psi}_k (y(z))\equiv 1 \;\; \mbox{for all}\;\;
z\in \Omega_r \backslash {\cal R}.$$
We thus decomposed the set $\Omega_r \setminus {\cal R}$ by the sets $G_k(U_k)$ and defined a related partition of unity $\Psi_k$. Note that in general the sets $U_k$ and $G_k(U_k)$ are neither open nor closed, and the functions $\Psi_k$ are discontinuous. In the next propositions we prove that these sets and functions are measurable.
\[Proposition I1\] Define a function $\tilde{\gamma}_{r,k}:{ {\bf R}}^{n-1}\rightarrow { {\bf R}}^1$ by $$\label{I3}
\tilde{\gamma}_{r,k}(x')=\left\{
\begin{array}{ll}\gamma_{\Omega_r}(G_k(x',0)),
& \mbox{if }x'\in \mbox{cl}(\tilde{U}_k )\\ 0, & \mbox{if }x'\notin
\mbox{cl}(\tilde{U}_k)
\end{array} \right.$$ where $\mbox{cl}(\cdot)$ denotes closure of a set. Then $\tilde{\gamma}_{r,k}$ is uppersemicontinuous (usc). The sets $U_k$ are ${\cal L}^n$ measurable. The ridge set of $\Omega_r$ has ${\cal L}^n$ measure 0.
It is enough to show that for any sequence $\{x_m'\}_{m=1}^{\infty}$ such that $x_m'\in
\mbox{cl}(\tilde{U}_k)$, $x_m'\rightarrow x'$ and $\tilde{\gamma}_{r,k}(x_m')\rightarrow L$ we have $\gamma_{r,k}(x')\geq L$.
The function $\gamma_{\Omega_r}$ is usc on ${ {\bf R}}^n$. Since $x \in \mbox{cl}(\tilde{U}_k)$ and $G_k$ is Lipschitz on $\mbox{cl}(\tilde{U}_k) \times {0})$, $$\tilde{\gamma}_{r,k}(x')=\gamma_{\Omega_r}(G_k(x', 0))
\geq \lim_{m\rightarrow \infty}
\gamma_{\Omega_r}(G_k((x_m')).$$ Thus, $\tilde{\gamma}_{r,k}$ is usc.
We have $U_k=\{\; (x',x_n)\;|\;x'\in \tilde{U}_k,\;0\leq x_n<
\tilde{\gamma}_{r,k}(x')\; \}$. Thus $U_k$ is ${\cal{L}}^n$-measurable (note that $\tilde{U}_k$ is open in ${ {\bf R}}^{n-1}$).
To prove that the ridge has ${\cal{L}}^n$ measure 0, we note that $G_k^{-1}({\cal R})\cap
\overline{U}_k=\{ (x',x_n)|x'\in \tilde{U}_k,\;x_n=\tilde{\gamma}_{r,k }(x') \}$. Then ${\cal{L}}^n(G_k^{-1}({\cal R})\cap
\overline{U}_k)=0$ because $\tilde{\gamma}_{r,k }(\cdot)$ is usc. Since $G_k$ is Lipschitz on $\mbox{cl}(\tilde{U}_k) \times [0, { \mbox{diam}}{\Omega_r}]$ and since $\cup_{k=1}^N \overline{U}_k = \overline{\Omega}_r$, we have ${\cal{L}}^n({\cal R})=0$.
We will use repeatedly the following fact:
\[Lemma I1\] Let $f:A\subset { {\bf R}}^m\rightarrow { {\bf R}}^n (m\leq
n)$ be a Lipschitz function, $A$ be a ${\cal{L}}^m$-measurable set, and let $g:{ {\bf R}}^m\rightarrow { {\bf R}}^1$ be a ${\cal{L}}^m$-measurable function. Then the function $g\circ f^{-1}:{ {\bf R}}^n \rightarrow { {\bf R}}^1$ is ${\cal{L}}^n$-measurable.
We can assume that $g \equiv 0$ outside $A$. Let $V\subset {{ {\bf R}}^1}\cup \{ + \infty \}$ be an open set. Then $W=g^{-1}(V)\cap A$ is ${\cal{L}}^m$-measurable. Then $(g\circ f^{-1})^{-1}V=f(g^{-
1}(U))\cap A=f(W),$ which is a measurable set since $f$ is Lipschitz and $m\leq n$ (see [@EGar], Lemma 2 of 3.3.1).
\[Proposition I2\] Functions $\Psi_k$ and $\Psi_k\circ G_k$ are ${\cal{L}}^n$-measurable.
The functions $G_k$ are Lipschitz. Then by Lemma \[Lemma I1\] it is enough to prove that functions $\Psi_k\circ G_k$ are ${\cal{L}}^n$-measurable (since then $\Psi_k=(\Psi_k\circ G_k)\circ G^{-1}_k$). Let $(x',x_n)\in U_k.$ Then $$\Psi_k(G_k(x',x_n))=
\tilde{\Psi}_k(G_k(x',0))=\tilde{\Psi}_k(x',\Phi_k(x')).$$ Consider the function $\Theta(x',x_n)=\tilde{\Psi}_k(x',\Phi_k (x'))$ . Since $\tilde{\Psi}_k\in C^\infty({ {\bf R}}^{n})$ and $\Phi_k\in C^{1,1}({ {\bf R}}^{n-1})$, we see that $\Theta$ is a $C^{1,1}$ function ${ {\bf R}}^n\rightarrow { {\bf R}}^1$. Now since $\Psi_k\circ
G_k= \Theta$ on $U_k$, $\Psi_k\circ
G_k \equiv 0$ outside $U_k$, and by Proposition \[Proposition I1\] the set $U_k$ is measurable, the assertion follows.
\[Lemma I2\] Let $\Omega$ be a bounded open set satisfying the Condition \[Cond.1\]. Let $X\subset
\partial \Omega$ and ${\cal H}^{n-1}(X)=0$. Let $Y$ be the union of distance rays passing through $X$, i.e., $Y=\cup_{x\in X}R_x$. Then ${\cal L}^n(Y)=0$.
We can assume that $X\subset G_k(U_k)\cap \partial \Omega$ for some $k$, and then $Y\subset G_k(U_k)$. We also can assume that $$\label{canAssume}
\gamma_{\Omega}(x) > 0 \;\;\; \mbox{ for all } \; x \in X.$$ If (\[canAssume\]) is not true then we can replace $X$ by $X \cap \{ x \; | \; \gamma_{\Omega}(x) > 0 \}$ and $Y$ does not change.
Since $\tilde{\gamma}_{r,k}(x')$ is a u.s.c. function of $x'$, we get ${\cal L}^n(\{ x_n=\gamma_{r,k}(x',0)\})
=0$. We also know that $J_nG_k>0$ on the set $U_k$. The map $G_k$ is one-to-one on $U_k$, and so $N(G_k,y)=1$ for $y\in Y \subset G_k(U_k)$, where $N(f,z)$ denotes the multiplicity function, which is the number of elements of $f^{-1}(z)$. Then by area formula we obtain $$\label{arF1}
{\cal L}^n(Y) = \int_{{ {\bf R}}^n}\chi _Y(y) dy=
\int_{U_k}\chi _Y(G_k(x))J_nG_k(x)dx,$$ where $\chi_A(\cdot)$ is the characteristic function of the set $A$, i.e., $\chi_A(\cdot)$ equals 1 on $A$ and 0 outside $A$. By Lemma \[jacobian lemma\] the Jacobian $J_nG_k$ is a locally bounded measurable function. Let $x' \in \tilde{U}_k$ be such point that the surface $\partial \Omega_r$ is twice differentiable at the point $G_k(x',0)$. By Lemma \[jacobian lemma\] the Jacobian $J_nG_k$ is defined by the expression (\[jacobian\]) at $(x', x_n)$ for every $x_n \in { {\bf R}}^1$. The principal curvatures of $\partial \Omega_r$ at the point $G_k(x', 0)$ satisfy $$\label{kappa-gamma}
\kappa_{r,i}>-\frac{1}{r}, \;\;\;
\kappa_{r,i}\tilde{\gamma}_{r,k}(x')\leq 1$$ by definition of $\tilde{\gamma}_{r,k}(\cdot)$, and by Condition \[Cond.1\]. We also have $$\label{gammaBd}
\tilde{\gamma}_{r,k}(x') \leq \mbox{diam}\Omega_r.$$ Let $m \leq n$ be such that $-\frac{1}{r} \leq \kappa_{r,1},..., \kappa_{r,m} < 0$ and $\kappa_{r,m+1},..., \kappa_{r, n-1} \geq 0$. Then for $x_n \in (r,\tilde{\gamma}_{r,k}(x')) $ we calculate using (\[kappa-gamma\]), (\[gammaBd\]): $$\label{jacIneq}
0 \leq \frac{J_nG_k(x', x_n)} {J_nG_k(x', r)} \leq
\frac{\prod^{m}_{i=1}\left[1-\kappa_{r,i} x_n\right]}
{\prod^{m}_{i=1}\left[1-\kappa_{r,i} r\right]} \leq
\left[1+\frac{1}{r}\tilde{\gamma}_{r,k}(x')\right]^m
\leq C,$$ where $C$ depends on $r$, $n$ and $\mbox{diam}(\Omega)$. By Lemma \[jacobian lemma\] the inequalities (\[jacIneq\]) hold for a.e. $x \in U_k \cap \{(x',x_n) \;|\;x_n>r\}$.
Define a map $\tilde{G}_k:\tilde{U}_k\rightarrow
\partial \Omega \subset { {\bf R}}^n$ by $\tilde{G}_k(x')= G_k(x',r) $. Then by (\[jacobian\]) $$\label{JtoJ}
J_{n-1}\tilde{G}_k (x')=J_nG_k(x',r)
\;\; \mbox{for a.e.} \;x'\in\tilde{U}_k.$$ So using the fact that the relation $Y\subset \overline{\Omega}$ implies $\chi_Y(G_k(x',x_n))=0$ for $x_n<r$, we get from (\[arF1\]), (\[gammaBd\]), (\[jacIneq\]) and (\[JtoJ\]) using area formula: $$\begin{aligned}
{\cal L}^n(Y) &\leq& C\int_{U_k}\chi_Y(G_k(x))\;J_{n-
1}\tilde{G}_k(x')dx
\nonumber \\
& = & C\int_{\tilde{U}_k}J_{n-
1}\tilde{G}_k(x')[\int_r^{\tilde{\gamma}_{r,k}(x')}
\chi_Y(G_k(x',x_n))dx_n]dx'
\nonumber \\
& = & C \int_{\tilde{U}_k} \chi_X(\tilde{G}_k(x'))
[\tilde{\gamma}_{r,k}(x')-
r]\cdot J_{n-1}\tilde{G}_k(x')dx'
\nonumber \\
& \leq &
C\int_{\tilde{U}_k}\chi_X(\tilde{G}_k(x'))\;J_{n-
1}\tilde{G}(x')dx'
\nonumber \\
& = &
C\int_{\partial \Omega}\chi_X(z)N(\tilde{G}_k,z) d{\cal H}^{n-
1}(z) = C{\cal H}^{n-1}(X)=0,
\nonumber\end{aligned}$$ where we have used the fact that $N(\tilde{G}_k,z)= 1$ for $z\in X$ (which follows from (\[canAssume\]) and Proposition \[Proposition 2\] (b,c) ).
Without assuming a condition of the type of Condition \[Cond.1\] the assertion of Proposition \[Lemma I2\] is not true. For example consider a nonconvex polygon in $R^2$ and take $X$ to be one point, a vertex of a re-enterant corner.
\[InvMeasZero\] Note that it follows from the proof above that $${\cal L}^n(G_k^{-1}(Y)) =
\int_{\tilde{U}_k}J_{n-
1}\tilde{G}_k(x')[\int_r^{\tilde{\gamma}_{r,k}(x')}
\chi_Y(G_k(x',x_n))dx_n]dx' = 0.$$
The following proposition describes the structure of the boundary of a set that satisfies Condition \[Cond.1\].
\[Proposition 4.1\] Let $\Omega$ be a bounded set. Let $\Omega$ satisfy the condition \[Cond.1\] with radius $r_0$. Then $\partial \Omega$ is $({\cal H}^{n-1}, n-1)$ rectifiable subset of ${ {\bf R}}^n$. In addition, $\partial \Omega$ is twice differentiable ${\cal H}^{n-1}$ a.e. on the set $$\label{set B}
{\cal B}=\{ x \in \partial \Omega \; | \; \gamma_{\Omega}(x)>0 \}.$$
Let $r>0$ be such number that $\partial \Omega_r$ is $C^{1,1}$. Existence of such $r$ follows from Proposition \[Proposition 3\]. It follows from Proposition \[Proposition 2\] b) that the nearest point projection mapping $P: \Omega_r \rightarrow \Omega$ is well-defined and onto. Moreover, by inequality 4.8(8) of [@FedCMeas] (or by inequality (\[1.1a\]) proved in Appendix \[Apndx 1\] below) the map $P$ is Lipschitz. Since $\partial \Omega_r$ is $C^{1,1}$, we get that $\partial \Omega$ is $({\cal H}^{n-1}, n-1)$ rectifiable.
By Proposition \[Proposition 2\] every point of the set ${\cal B}$ has a unique nearest point on $\partial \Omega_r$. Since $\partial \Omega_r$ is $C^{1,1}$, it follows that $\partial \Omega_r$ is twice differentiable ${\cal H}^{n-1}$ a.e. Since the map $P$ is Lipschitz it follows that for ${\cal H}^{n-1}$ a.e. $y \in {\cal B}$ the surface $\partial \Omega_r$ is twice differentiable at the corresponding point $y_r$. Applying Proposition \[Proposition 4\] b), we conclude the proof.
Propositions \[Proposition I1\], \[Lemma I2\] and \[Proposition 4.1\] imply
\[A.E.\] Let $\Omega \subset { {\bf R}}^n$ satisfy Condition \[Cond.1\]. Then ${\cal L}^n$ a.e point of $\Omega$ lies in the relative interior of a distance ray that intersects $\partial \Omega$ at a point at which $\partial \Omega$ is twice differentiable.
We need an analogue of the Proposition \[Proposition 4.1\] for a Lipschitz family $\{\Omega_t\}$ of sets that satisfy Condition \[Cond.1\] with a radius $r_0$ independent of $t$. Let $E, \;
{\bf \Gamma} $ be defined by (\[defE\]). Let $E_r, \;
{\bf \Gamma}_r $ be defined by (\[defEr\]). Denote $$\label{defFxt}
D(x,t) = d^{s}_{\Omega_t} (x),$$ where $d^{s}_{\Omega_t} (x)$ is the signed distance function (\[sgnDist\]).
\[Proposition 4.1.t\] Let $\{\Omega_t\}$ be a locally Lipschitz family of open bounded sets that satisfy the Condition \[Cond.1\] with radius $r_0$ independent of $t$. Then
a\) ${\cal L}^{n+1}$ a.e. $(x_0, t_0) \in E$ lies in the relative interior of a distance ray of the set $\Omega_{t_0}$ that intersects $\partial \Omega_{t_0}$ at such point $y \in \partial \Omega_{t_0}$ that ${\bf \Gamma}$ is (2,1) differentiable at $(y,t_0)$;
b\) There exists a constant $C$ depending only on $r_0$ such that the function $D(x,t) - C|x|^2$ on the set $E_{r_0}$ is concave in $x$ for every $t$.
First we prove the following fact.
\[semiconcDist\] Let $\Omega$ satisfies Condition \[Cond.1\] with radius $r_0$. Then the distance function $d_{\Omega}(x)$ is semiconcave in $\Omega$, i.e. there exists a constant $C$ depending only on $r_0$ such that the function $d_{\Omega}(x) - C|x|^2$ is concave in $\Omega$.
Let $x\in \Omega$. Let $w \in { {\bf R}}^n$ satisfy $$|w| \leq d_{\Omega}(x).$$ Let $y \in {\cal N}_{\partial \Omega_{r_0}}(x)$, where $\Omega_{r_0}$ is $r_0$- neighborhood of $\Omega$. Then by Proposition \[Proposition 2\](c), $$\begin{aligned}
|x-y|& =& d_{\Omega}(x) + r_0,
\nonumber \\
|x+w-y| & \geq & d_{\Omega}(x+w) + r_0,
\nonumber \\
|x-w-y| & \geq & d_{\Omega}(x-w) + r_0.
\nonumber\end{aligned}$$ Then we have $$d_{\Omega}(x+w) - 2d_{\Omega}(x) + d_{\Omega}(x-w) \leq
|x+w-y| - 2|x-y| + |x-w-y| \leq \frac{C}{r_0}|w|^2,$$ and so $d_{\Omega} -2 \frac{C}{r_0}|x|^2$ is concave.
Now we prove Proposition \[Proposition 4.1.t\]. The function $D(x,t)$ defined in (\[defFxt\]) is Lipschitz since $\{\Omega_t\}$ is a Lipschitz family of sets. By Proposition \[Proposition 2\] the set $(\Omega_t)_{r_0}$ satisfies Condition \[Cond.1\] with radius $\frac{r_0}{2}$, and $$d^{s}_{(\Omega_t)_{r_0}} (x) = d^{s}_{\Omega_t} (x) + r_0
\;\;\;\mbox{for }\;(x,t) \in E_{r_0}.$$ Then by Lemma \[semiconcDist\] the function $D(x,t) - C|x|^2$ on the set $E_{r_0}$ is concave in $x$ for every $t$. Now by Theorem 1 of Appendix 2 of [@Krylov] the function $D(x,t)$ is (2,1) differentiable at ${\cal L}^{n+1}$ a.e. point of $E_{r_0}$. It follows from Proposition \[Proposition 4t\] that if $D(\cdot)$ is (2,1) differentiable at a point $(x_0, t_0) \in E$, then ${\bf \Gamma}$ is (2,1) differentiable at $(y,t_0)$, where $y \in {\cal N}_{\partial \Omega_{t_0}}(x_0)$. Note that $x_0$ lies in the relative interior of its distance ray if $d_{\Omega_{t_0}}$ is twice $x$-differentiable at $x_0$. The Proposition is proved.
Properties of mass transport density. {#SectDens}
=====================================
Let $\Omega$ be a bounded set that satisfies the condition \[Cond.1\] with radius $r_0$. Let $t>0$. Define a function $a=a_{\Omega, t}:{ {\bf R}}^n\rightarrow { {\bf R}}^1$ as following. Let $x\in \partial \Omega$ be such a point that the surface $\partial \Omega$ is twice differentiable at $x$ and let $\kappa_1, ..., \kappa_{n-1}$ be the principal curvatures of $\partial
\Omega$ at $x$. Introduce the coordinate $s$ on the distance ray $R_x$ by $s(z)=|z-x|$ for $z\in R_x$. Then $s$ changes in the interval $(0, \gamma(x))$. Define the function $a$ on $R_x$ as the following: $$\label{4}
a(s)={1\over t}\prod^{n-1}_{i=1}{1\over {1-{s
\kappa_i}}}\int^s_0 \prod^{n-1}_{i=1}(1-\xi \kappa_i)\left[
\frac{\int^{\gamma(x)}_0 \nu \prod^{n-1}_{i=1} (1-\kappa_i
\nu)d\nu}{\int^{\gamma(x)}_0 \prod^{n-1}_{i=1} (1-\kappa_i
\nu)d\nu}-\xi \right] d\xi$$
By Corollary \[A.E.\] the function $a$ is now defined a.e. in $\Omega$.
Define $a\equiv 0$ on ${ {\bf R}}^n \backslash \Omega$. Now $a$ is defined a.e. in ${ {\bf R}}^n$.
\[MasTrDen\] The function $a(\cdot)=a_{\Omega, t}(\cdot)$ is called mass transport density.
Define the function $V: \partial \Omega \rightarrow { {\bf R}}$ at the point of twice differentiability of $\partial \Omega$ as following. $$\label{Veloc}
V(x)=
{1\over
t}\frac{\int^{\gamma (x)}_0 s\prod^{n-1}_{i=1} (1-\kappa_i
s)ds}{\int^{\gamma (x)}_0 \prod^{n-1}_{i=1} (1-\kappa_i
s)ds}.$$
\[Proposition 1\] Let $\Omega \in { {\bf R}}^n$ be a bounded open set that satisfies Condition \[Cond.1\] with radius $r_0$. Then there exists a constant $C$ depending only on $r_0,\; n,\; { \mbox{diam}}\,{\Omega}$ such that the following is true. Let $x \in \partial \Omega$ be such point that $\partial \Omega$ be twice differentiable at $x$. Then the mass transport density on the distance ray $R_x$, defined by (\[4\]) for $s\in [0,\gamma(x)]$, satisfies:\
(a) $|a(s)|\leq C$. Derivative $a'(s)$ exists and is continuous for $s \in
(0,\gamma(x))$.\
(b) $a(s)$ satisfies the equation $$a'(s)-a(s)\sum^{n-1}_{i=1}
{{\kappa_i}\over
{1-\kappa_i s}} -{s \over t}+V(x)=0$$ on $(0,\gamma(x))$, the boundary conditions $a(0)=0$, $a(\gamma(x))=0$, and $a(s)>0$ on $(0,\gamma(x)).$\
(c) For $s \in (0,\gamma (x))$ the inequalities hold $$|a(s)|\leq C(\gamma(x)-
s),\;\;\;\;
|a(s)|\leq Cs.$$ (d) $|a'(s)|\leq C$ on $(0,\gamma(x)).$
\(a) Existence and continuity of $a'(s)$ on $(0,\gamma(x))$ are checked explicitly. The bound $|a(s)|\leq C$ will follow from $(c)$.\
(b) The equation is checked explicitly. The boundary conditions follow from $(c)$. The inequality $a(s)>0$ on $(0,\gamma(x))$ holds since either $\kappa_i < 0$, or $0\leq
\gamma(x)\leq {1\over \kappa_i}$ and then, defining $$\tilde{V}(s)=\frac{\int^s_0 \nu\prod^{n-1}_{i=1}(1-\kappa_i
\nu)d\nu}{\int^s_0 \prod^{n-1}_{i=1}(1-\kappa_i \nu)d\nu}$$ we calculate: $${d\tilde{V} \over ds}=\frac{s\prod^{n-1}_{i=1}(1-
\kappa_i s) \int^s_0 \prod^{n-1}_{i=1}(1-\kappa_i \nu)d\nu-
\prod^{n-1}_{i=1}(1-\kappa_i s)\int^s_0 \nu\prod^{n-1}_{i=1}(1-
\kappa_i \nu)d\nu}{[\int^s_0 \prod^{n-1}_{i=1}(1-\kappa_i \nu)d\nu)]^2}$$ $$=\frac{\prod^{n-1}_{i=1}(1-\kappa_i s)}{[\int^s_0 \prod^{n-1}_{i=1}(1-
\kappa_i \nu)d\nu)]^2}\int^s_0 \prod^{n-1}_{i=1}(1-\kappa_i
\nu)(s-\nu)d\nu>0\
\ \mbox{for } s\in (0,\gamma(x)),$$ and so $$a(s)={1\over t}
[\tilde{V}(\gamma(x))-\tilde{V}(s)]
\prod^{n-1}_{i=1}{1\over {1-\kappa_i
s}}\int^s_0 \prod^{n-1}_{i=1}(1-\kappa_i \nu)d\nu
>0\ \ \mbox{for }s \in
(0,\gamma(x)).$$
(c)We write $\gamma$ for $\gamma(x)$: Denote $\Phi(s)=\int^s_0 \nu\prod^{n-1}_{i=1}(1-
\kappa_i\nu)d\nu,\;\Psi(s)=\int^s_0
\prod^{n-1}_{i=1}(1-\kappa_i\nu)d\nu$. For $s \in (0,\gamma)$ we calculate: $$\begin{aligned}
ta(s) & = &
\frac{[\Phi(\gamma)\Psi(s)-\Phi(\gamma)\Psi(\gamma)]+
[\Phi(\gamma)\Psi(\gamma)-\Phi(s)\Psi(\gamma)]}{\Psi (\gamma)}
\prod^{n-1}_{i=1}{1\over {1-\kappa_is}}
\nonumber \\
& = & (\gamma-s)
\frac{-\Phi(\gamma)\Psi'(s_1)+\Phi'(s_2)\Psi(\gamma)}{\Psi
(\gamma)}\prod^{n-1}_{i=1}\frac{1}{1-\kappa_i
s},
\nonumber\end{aligned}$$ where $s_1,s_2 \in [s,\gamma]$. Note, that $\tilde{V}(\xi)={\Phi(\xi)\over {\Psi(\xi)}}$. So $$\begin{aligned}
ta(s)& = & \prod^{n-1}_{i=1}{1\over {1-\kappa_is}}\left[-
\tilde{V}(\gamma) \prod^{n-1}_{i=1}(1-\kappa_is_1)+\prod^{n-
1}_{i=1}(a-\kappa_is_2)\cdot s_2\right]\;(\gamma-s)
\nonumber \\
& = &
\left[-\tilde{V}(\gamma)\prod^{n-1}_{i=1}\frac{1-\kappa_is_1}{1-
\kappa_is}+s_2\prod^{n-1}_{i=1}\frac{1-\kappa_is_2}{1-
\kappa_is}\right]\;(\gamma-s).
\nonumber\end{aligned}$$ If $\kappa_i\geq 0$ then for $s^*\in(s,\gamma)$ we have $0\leq s
\kappa_i<s^*\kappa_i<1,$ and so $$0\leq \frac{1-\kappa_i s^*}{1-\kappa_i
s}\leq1.$$ If $\kappa_i< 0$ then it follows from the Condition \[Cond.1\] that $|\kappa_i| \leq C$, and so $$0 \leq \frac{1-\kappa_i s^*}{1-\kappa_is} =
\frac{1+|\kappa_i| s^*}{1+|\kappa_i|s} \leq 1+C\gamma
\leq 1+ C { \mbox{diam}}\Omega.$$ It remains to estimate $\tilde{V}(s)$ for $s \in [0, \gamma]$. $$\label{tildVEst}
\tilde{V}(s)\leq
\frac{\int_0^s \gamma
\prod^{n-1}_{i=1} (1-\nu\kappa_i)d\nu}{\int_0^s \prod^{n-
1}_{i=1} (1-\nu\kappa_i)d\nu}=\gamma \leq \mbox{diam}\Omega.$$ So, $ta(s)\leq C\mbox{diam}\;\Omega\;(\gamma-s),$ and, since $t>0$, $$\label{5}
0<a(s)\leq C(\gamma-s)\mbox{ for }s \in (0,\gamma(x)).$$ To estimate $a(s)$ near $s=0$, we compute using (\[tildVEst\]) $$\begin{aligned}
|ta(s)| & = &
\left|[\tilde{V}(\gamma)- \tilde{V}(s)]
\prod^{n-1}_{i=1}\frac{1}{1-\kappa_i s}\int^s_0
\prod^{n-1}_{i=1}(1-\kappa_i\nu)d\nu \right|
\nonumber \\
& \leq & \gamma \left|\prod^{n-1}_{i=1}\frac{1}{1-\kappa_i
s}\int^s_0 \prod^{n-1}_{i=1}(1-\kappa_i\nu)d\nu\right|=\gamma
s \prod^{n-1}_{i=1}\frac{1-\kappa_i s^*}{1-\kappa_i s},
\nonumber\end{aligned}$$ where $s^*\in[0,s]$. Since $\kappa_i \gamma \leq 1$, then for $s \leq {\gamma \over 2}$ we get $$0 \leq \frac{1-\kappa_i s^*}{1-\kappa_i
s}\leq 2 \;\;\mbox{ if } \kappa_i \geq 0,$$ and since $\kappa_i > -\frac{1}{r}$ we get $$0 \leq \frac{1-\kappa_i s^*}{1-\kappa_i
s}\leq C(\mbox{diam}(\Omega), r) \;\;\mbox{ if } \kappa_i < 0.$$ So $|ta(s)|\leq C \gamma s$, and, since $ t>0$, we get $$|a(s)|\leq C s.$$ If $s\geq{\gamma \over 2}$, then $s\geq\gamma-s$, and from (\[5\]) $$|a(s)|\leq C(\gamma-s)\leq Cs.$$\
(d) We prove that $|a'(s)|\leq C$ for $s \in (0,\gamma)$. $$a'(s)=a(s)\sum^{n-1}_{i=1}\frac{\kappa_i}{1-\kappa_i
s}+{s \over t}-V(x).$$ By (\[tildVEst\]), $|V(x)|\leq
\frac{\gamma}{t}$. It remains to prove that $$\label{MTrDenDerivEst}
\left|a(s)\sum^{n-1}_{i=1}\frac{\kappa_i}{1-\kappa_i
s}\right|\leq C$$ If $\kappa_i \geq 0$ then $0\leq \kappa_i \gamma\leq 1$, and since $\gamma>s$ we get: $$0\leq\frac{\kappa_i}{1-\kappa_i s}\leq\frac{\kappa_i}{\kappa_i
\gamma-\kappa_i s}=\frac{1}{\gamma-s},$$ If $\kappa_i < 0$, then $$\left|\frac{\kappa_i }{1-\kappa_i s}\right|\leq \frac{1}{s}.$$ Now (\[MTrDenDerivEst\]) follows from (c).
Mass balance equation {#ChngVar}
=====================
Let $\{\Omega_t\}$ be a Lipschitz family of open sets satisfying Condition \[Cond.1\] with radius $r_0>0$. Let $\{\Omega_t\}$ satisfy equation (\[1\]) on the time-interval $[a,b]$ in the sense described in Theorem \[Theorem 3\]. Let the function $w(x,t)$ be defined by (\[2\]). Let $a(x,t)$ be the function $a_{\Omega_t, t}(x)$ from Definition \[MasTrDen\].
The purpose of this section is to show that for a.e. $t \in [a,b]$ $$\label{I1}
\int_{{ {\bf R}}^n} aDwD\phi dz=\int_{{ {\bf R}}^n} ({w\over t}-w_t)\phi
dz$$ for any $\phi \in C^\infty({ {\bf R}}^n)$, where $w_t$ denotes $\partial_t w$. Thus we show that the function $a(\cdot, t)$ satisfies (\[defMassTrDen\]) for a.e $t$.
It follows from Proposition \[Proposition 4.1.t\] that for a. e. $t$ the following is true: ${\cal L}^n$ a. e. point $x\in \Omega_t$ lies in the relative interior of the distance ray that intersects $\partial \Omega_t$ at such point $y\in \partial \Omega_t$ that the surface ${\bf \Gamma}$ is (2,1) differentiable at $(y,t)$, where ${\bf \Gamma}$ is defined by (\[defE\]). Fix such $t$. For the rest of this section we drop $t$ in the notation, i.e., we write $\Omega$, $a(x)$, $w(x)$, $w_t(x)$ for $\Omega_{t}$, $a(x, t)$, $w(x, t)$, $w_t(x, t)$.
We use same notation $a(\cdot)$ for both mass transport density $a(x)$, a function defined on ${ {\bf R}}^n$, and for mass transport density on a ray $R_x$, the function $a(s)$ defined on $[0, \gamma(x)]$.
Denote by $\tilde{\Gamma}$ the subset of $\partial \Omega$ that consists of all points at which the surface ${\bf \Gamma}$ is $(2,1)$-differentiable. By choice of $t$ and Proposition \[Proposition 4.1\] we get $${\cal H}^{n-1}({\cal B} \setminus \tilde{\Gamma})=0,$$ where the set ${\cal B}$ is defined by (\[set B\]). From definition of ${\cal B}$ it follows that no distance rays of $\Omega$ have their lower ends in the subset $\partial \Omega \setminus {\cal B}$ of the boundary. Denote by $R(\tilde{\Gamma})$ the union of all distance rays that have lower ends in $\tilde{\Gamma}$, i.e., $R(\tilde{\Gamma})= \cup_{x \in \tilde{\Gamma}}R_x $. Then it follows from Proposition \[Lemma I2\] and Remark \[InvMeasZero\] that $$\label{MeasZeroRel}
{\cal L}^{n} ( \Omega \setminus R(\tilde{\Gamma}) ) = 0, \;\;\;
{\cal L}^{n}[ G_k^{-1} (\Omega \setminus R(\tilde{\Gamma}))] =0 \;
\mbox{for}\;k=1,...,N.$$
Let $V(\cdot)$ be the function (\[Veloc\]). It follows from (\[MeasZeroRel\]) that at a. e. point $x$ of $\Omega$ the function $w_t(x)$ is given by the expression $\frac{1}{t}V(y)$ where $y \in {\cal N}_{\partial \Omega}(x)$. In the next proposition we prove that $w_t$ is measurable.
We continue to use local coordinate systems on $\Omega_r$, described in Section \[Sets\] and notation introduced there, in particular sets $U_k$, partition of unity $\Psi_k$ and coordinate mappings $G_k$ defined by (\[defUk\]), (\[defPsi\]) and (\[I2\]) respectively.
\[Proposition I3\]
a\) For $k=1,...,n$ there exists a bounded ${\cal{L}}^n$ measurable function $v_k:U_k\rightarrow { {\bf R}}^1$ such that $$\label{I5}
v_k(x',x_n)=V(G_k(x',r))\;\;\; \mbox{if } \;\;\;
G_k(x',r)\in \tilde{\Gamma},\; x_n \in [r,\tilde{\gamma}_{r,k}(x')].$$
b\) The function $w_t$ is ${\cal L}^n$ measurable and for any bounded measurable function $\varphi:\; U_k\rightarrow { {\bf R}}^1$ $$\label{I6}
{1\over t}\int_{U_k\cap \{x_n\geq r \}}
v_k{\varphi}J_nG_kdx=\int_{G_k(U_k)}w_t(\varphi \circ G^{-1}_k)dx$$
a\) $\partial \Omega_r$ is a $C^{1,1}$ manifold, and so the second fundamental form of $\partial \Omega_r$ is defined as a differential form with $L^{\infty}$ coefficients. At all points where $\partial \Omega_r$ is twice differentiable the principal curvatures of $\partial \Omega_r$ are the roots of the characteristic polynomial of the second fundamental form. To write this in coordinates, we recall that $\partial \Omega_r \cap G_k(U_k)$ is the graph of a $C^{1,1}$ function $x_n = \Phi_k(x')$ defined on the set $\tilde{U}_k \in { {\bf R}}^{n-1}$. Then there exist $L^{\infty}$ functions $[u_{i j}(x')]_{i,j=1}^n$ such that at the points where $D^2\Phi_k$ exists, $\frac{\partial^2 \Phi_k}{\partial x_i \partial x_j}
= u_{ij}$.
Denote by $P_{x'}(s)$ the characteristic polynomial of the matrix $[u_{ij}(x')]$ i.e. $P_{x'}(s)=\mbox{det}([u_{ij}(x')]-sI_{n-1})$. Note that the coefficients of the polynomial $P_{x'}(\cdot)$ are measurable functions of $x'$ since they are compositions of polynomials and measurable functions $u_{ij}(x')$.
The eigenvalues of the matrix $[u_{ij}(x')]_{i,j=1}^n$ are principal curvatures of $\partial \Omega_r$ at the points of twice differentiability of $\Phi_k$. Let $x' $ be such a point, let $\kappa_{r,1},...,\kappa_{r,n-1}$ be principal curvatures of $\partial \Omega_r$ in $G_k(x',0)$, and let $m$ be such number that $\kappa_{r,i}\neq 0$ for $i=1,...,m$ and $\kappa_{r,i}=0$ for $i=m+1,...,n-1$. Then $P_{x'}(s)=\prod^{n-1}_{i=1}(\kappa_{r,i}-s)$, and so $$(-1)^{n-1}s^{n-1}P_{x'}({1\over s})=\prod^{n-1}_{i=1}\;\;(1-
\kappa_{r,i}s)=
\prod^m_{i=1}\kappa_{r,i} \prod^m_{i=1}
(\frac{1}{\kappa_{r,i} }-s).$$
For any $x'\in U_k$ the point $G_k(x',r)$ belongs to $\partial \Omega$ and $|G_k(x',0)-G_k(x',r)|=r$. If the surface $\partial \Omega$ is twice differentiable at the point $y=G_k(x',r)$, then, by Proposition \[Proposition 4\], the surface $\partial \Omega_r$ is twice differentiable at the point $y_r=G_k(x',0)$, and, denoting by $\{ {\kappa}_i \}_{i=1}^n$ and $\{ {\kappa}_{r,i} \}_{i=1}^n $ the principal curvatures of $\partial \Omega$ at $y$ and of $\partial \Omega_r$ at $y_r$, we get $${\kappa}_{r,i} =\frac{{\kappa}_i}{1+{\kappa}_ir},\;\mbox{so
}\frac{1}{{\kappa}_{r,i} }=\frac{1}{{\kappa}_i}+r\;\;(\mbox{if
}{\kappa}_i\neq 0).$$ Then, assuming as before that ${\kappa}_1,...,{\kappa}_m \neq 0$ and ${\kappa}_{m+1},...,{\kappa}_{n-1}=0$, we get if $m \ne 0$ $$\begin{aligned}
\prod^{n-1}_{i=1}(1-{\kappa}_is)
&=& \prod^m_{i=1}{\kappa}_i\prod
^m_{i=1}(\frac{1}{{\kappa}_i}-s)
=\prod^m_{i=1}{\kappa}_i\prod
^m_{i=1}\left[\frac{1}{{\kappa}_{r,i}}-(s+r)\right]
\nonumber \\
& = &
(-1)^{n-
1}\frac{\prod^m_{i=1}{\kappa}_i}{\prod^m_{i=
1}{\kappa}_{r,i}}(s+r)^{n-1}P_{x'}(\frac{1}{s+r}).
\nonumber\end{aligned}$$ If $m=0$ (i.e., all $\kappa_{r,i}\equiv 0$), then $\prod^{n-1}_{i=1}(1-{\kappa}_is)=(-1)^{n-1}(s+r)^{n-
1}P_{x'}(\frac{1}{s+r})=1$.
Now we see that at all $x'$ as above, the function $V$ defined in (\[Veloc\]) can be expressed at the point $ G_k(x',r) \in \partial \Omega$ as $$\label{I7}
V(G_k(x',r))=\frac{\int^{{\tilde{\gamma}_{r,k}}(x')-r
}_0
s(s+r)^{n-
1}P_{x'}(\frac{1}{s+r})ds}{\int^{{\tilde{\gamma}_{r,k}}(x')-r}_0
(s+r)^{n-1}P_{x'}(\frac{1}{s+r})ds}$$ since, by Proposition \[Proposition 2\], $\gamma_{\Omega_r}(z)=\gamma(z)_{\Omega}+r$ for $z\in
\Omega$.
Define a function $\eta=\eta(x',x_n)$ on $\tilde{U_k} \times { {\bf R}}^1 $ by the right-hand side of (\[I7\]). This function $\eta$ is ${\cal L}^n$-measurable: $\eta$ does not depend on $x_n$, the coefficients of the polynomial $P_{x'}(\cdot)$ are measurable functions of $x'$ and by Proposition \[Proposition I1\], the functions $\tilde{\gamma}_{r,k}$ are measurable functions of $x'$.
Then it follows from Proposition \[Proposition I1\] that the function $$v_k(x',x_n)=\eta (x',x_n) \chi_{U_k}(x',x_n)$$ satisfies all properties asserted in (a) (note that the right-hand side of (\[I5\]) is bounded by diam $\Omega$).
By Lemma \[Lemma I1\] the function $v_k\circ G^{-1}_k$ is ${\cal L}^n$-measurable. By (\[MeasZeroRel\]) we get $$\label{I8'}
\begin{array}{l}
{1\over t}v_k\circ G^{-1}_k=w_t\;\;\;\mbox{at}\; {\cal L}^n
\;\mbox{a.e. point of}\;\,
G_k(U_k \cap \{ \, x_n \geq r \,\}),
\\
\\
\mbox{and } \;\; w_t \equiv 0 \;\;\; on \;\;\;
G_k(U_k \cap \{ \, x_n < r \,\})
\end{array}$$ and so $w_t$ is measurable.
Since map $G_k$ is one-to-one and Lipschitz on $U_k$, and since $U_k$ is measurable, we get by Theorem 3.2.5 of [@Federer] and Lemma \[Lemma I1\] that for every bounded measurable function $\varphi$ $$\label{I8}
\int_{U_k}v_k\varphi \;J_nG_kdz=\int_{G_k(U_k)}(v_k\circ
G^{-1}_k)(x)
(\varphi \circ G^{-1}_k)(x) dx$$
Now (\[I6\]) follows from (\[I8\]), (\[I8’\]).
\[Corollary I1\] The function ${w\over t}-w_t$ is measurable, and for each $\varphi \in C^\infty({ {\bf R}}^n)$ $$\label{I9}
\int_{{ {\bf R}}^n}({w\over t}-w_t)\varphi dz={1\over
t}\sum^N_{k=1}\int_{G_k^{-1}(R(\tilde{\Gamma}) ) }
(w\circ G_k-v_k)(\varphi \circ
G_k)(\Psi_k\circ G_k)J_nG_kdx.$$
We have $${w\over t}-w_t=\sum^N_{k=1}({w\over t}-
w_t)\Psi_k.$$ Since $\mbox{supp}\;\Psi_k\subset G_k(U_k)$, the integrand on the right-hand side of (\[I9\]) is measurable by Propositions \[Proposition I3\] and \[Lemma I2\]. It is also clear that the integrand on the right-hand side is bounded.
Then from Proposition \[Proposition I3\] (b) and area formula, we get $$\int_{{ {\bf R}}^n}({w\over t}-w_t)\varphi dz={1\over
t}\sum^N_{k=1}\int_{{ {\bf R}}^n}(w\circ G_k-v_k)(\varphi \circ
G_k)(\Psi_k \circ G_k)J_nG_k dx.$$ This formula is true for each $\varphi \in C^\infty$, and, by approximation, for every bounded measurable $\varphi$. Now (\[MeasZeroRel\]) implies (\[I9\]).
Now we prove that the mass transport density $a(\cdot)$ is measurable. The proof is similar to the proof of Proposition \[Proposition I3\].
\[Proposition I4\] The function $a(\cdot)$ is ${\cal L}^n$-measurable function on ${ {\bf R}}^n$. The function $a \circ G_k^{-1}$ is ${\cal L}^n$-measurable function on $U_k$.
For $(x', s)$ such that $$G_k(x',r)\in \tilde{\Gamma},\;\;\;s\in
[r,\tilde{\gamma}_{r,k}(x',r)],$$ the function $a\circ G_k(x',s)$ is given by expression (\[4\]), where the curvatures are computed at the point $G_k(x',r)$. Calculations similar to those in Proposition \[Proposition I3\] show that this expression can be rewritten using the characteristic polynomial $P_{x'}(\cdot)$ of the matrix $D^2\Phi_k(x')$ as following: $$a\circ G_k(x',s)={1\over t}\;\frac{\int^{s-r}_0
(\nu+r)^{n-
1}P_{x'}(\frac{1}{\nu+r})d\nu}{(s+r)^{n-1}P_{x'}(\frac{1}{s+r})}\times$$ $$\label{I12}
\left[ \frac{\int^{\tilde{\gamma}_{k,r}(x')-r}_0 \nu(\nu+r)^{n-
1}P_{x'}(\frac{1}{\nu+r})d\nu}{\int^{\tilde{\gamma}_{k,r}(x')-r}_0
(\nu+r)^{n-1}P_{x'}(\frac{1}{\nu+r})d\nu}-\frac{\int^{s-r}_0 \nu(\nu+r)^{n-
1}P_{x'}(\frac{1}{\nu+r})d\nu}{\int^{s-r}_0 (\nu+r)^{n-
1}P_{x'}(\frac{1}{\nu+r})d\nu} \right]$$ By (\[MeasZeroRel\]), the function $a \circ G_k^{-1}(\cdot)$ is defined by the above formula at a.e. point of $U_k$, and is ${\cal L}^n$-measurable function of variables $x',s$ on $U_k$ since it is a rational function of variables $s$ and $\tilde{\gamma}_{k,r}(x')$ with measurable coefficients and $\tilde{\gamma}_{k,r}(x')$ is ${\cal L}^{n-1}$ measurable function of $x'$. By Lemma \[Lemma I1\], the function $a(\cdot)$ is measurable.
[**Proof of (\[I1\]).**]{} We will transform the right-hand side of (\[I9\]). From (\[defPsi\]) we get $$\label{PsiIndepOfs}
\Psi_k(G_k(x',s)) = \Psi_k(G_k(x',r)), \;\;\;\;
s \in (r, \tilde{\gamma}_{r,k}(x')).$$ Let $\tilde{G}_k: \tilde{U}_k \rightarrow \partial\Omega$ be the map defined by $$\tilde{G}_k(x') = G_k(x',r).$$ Then for $k=1,...,N,$ using definitions of $w,v_k,\Psi_k$ we compute $$\begin{aligned}
{1\over t}\int_{G_k^{-1}(R(\tilde{\Gamma}) )}
(w\circ G_k
- v_k)(\varphi \circ
G_k)(\Psi_k \circ G_k)J_nG_k dx
\hspace{8em} & &
\nonumber \\
= \int_{\tilde{G}_k^{-1}(\tilde{\Gamma}) }
\Psi_k(G_k(x',r)) \left\{
\int^{\tilde{\gamma}_{r, k}(x')}_r {1\over t}[s-r
\begin{array}{l} \\ \\ \end{array} \right. \hspace{3em} & &
\label{RHS} \\
& & \nonumber \\
\hspace{5em}
\left. \begin{array}{l} \\ \\ \end{array}
-v_k(x',s)]\varphi (G_k(x',s))J_nG_k(x',s)ds
\right\} dx' & &
\nonumber\end{aligned}$$
Let $x' \in { {\bf R}}^{n-1}$ be such that $G_k(x', r) \in \tilde{\Gamma} $. Then $\partial \Omega$ is twice differentiable at $G_k(x',r)$, and so $${1\over t}v_k(x',s)= V(G_k(x',r))\;\;\;
\mbox{ for }\;s\in [r, \tilde{\gamma}_{r,k}(x')].$$ By Proposition \[Proposition 4\], $\partial \Omega_r$ is twice differentiable at $G_k(x',0)$. Let $\kappa_1,...,\kappa_{n-1}$ be principal curvatures of $\partial \Omega$ at $G_k(x',r)$, and $\kappa_{r,1},...,\kappa_{r,n-1}$ be principal curvatures of $\partial \Omega_r$ at $G_k(x',0)$. Then $\kappa_{r,i} =\frac{\kappa_i}{1+r\kappa_i}$. Let $y=G_k(x',r)$. Then the mass transport density on the ray $R_y$ satisfies ODE from Proposition \[Proposition 1\](b). Thus we get for $s\in
[r,\tilde{\gamma}_{r,k}(x')]=[r,\gamma_{\Omega}(y)+r]$: $$\begin{aligned}
{1\over t}(s-r-v_k)
& = &
-\frac{da(s-r)}{ds}+a(s-r)\sum^{n-
1}_{i=1}\frac{\kappa_i}{1-\kappa_i(s-r)}
\nonumber \\
& = &
-\frac{da(s-r)}{ds}+a(s-
r)\sum^{n-1}_{i=1}\frac{\kappa_{r,i} }{1-\kappa_{r,i} s}
\nonumber\end{aligned}$$ where $a(s)$ is mass transport density on the ray $R_y$.
By Proposition \[Proposition 1\], $\frac{da}{ds}$ exists and is continuous and bounded on the interval $(0,\tilde{\gamma}_{r,k}(x')-r)$, and $$a(0)=a(\tilde{\gamma}_{r,k}(x')-r)=0.$$ We also see by (\[jacobian\]) that $$\label{jacDeriv}
\frac{\partial J_nG_k}{\partial s}(x',s)=J_nG_k(x',s)\sum^{n-
1}_{i=1}\frac{-\kappa_{r,i} }{1-\kappa_{r,i} s}.$$ Then we calculate, integrating by parts: $$\int^{\tilde{\gamma}_{r, k}(x')}_r {1\over t}[s-r-
v_k(x',s)]\varphi (G_k(x',s))J_nG_k(x',s)ds
=$$ $$\int^{\tilde{\gamma}_{r,k}(x')}_r\left[-\frac{da(s-r )}{ds}-a(s-
r)\frac{\frac{\partial}{\partial
s}J_nG_k(x',s)}{J_nG_k(x',s)}\right]\varphi (G_k(x',s))J_nG_k(x',s)ds$$ $$=\int^{\tilde{\gamma}_{r,k}(x')}_r a(s-
r)\frac{\partial}{\partial
s}\varphi(G_k(x',s))J_nG_k(x',s)ds$$ This allows to derive from (\[I9\]), (\[RHS\]) and (\[PsiIndepOfs\]) the following equality: $$\begin{aligned}
\label{I10}
\int_{{ {\bf R}}^n}({w\over t}-w_t)\varphi dz =
\hspace{22em}
& &
\\
\sum_{k=1}^N
\int_{\tilde{G}_k^{-1}(\tilde{\Gamma})}
\int^{\tilde{\gamma}_{r,k}(x')}_r
\Psi_k(G_k(x',s))
a(G_k(x',s))\frac{\partial}{\partial s}\varphi
(G_k(x',s)J_nG_k(x',s)ds dx' &&
\nonumber\end{aligned}$$
Let $z=G_k(x',s)$. Using the fact that distance function is twice differentiable on $R(\tilde{\Gamma})$, we get: $$\begin{aligned}
\frac{\partial}{\partial s}
(\varphi(G_k(x',s)) &=& ( D\varphi)(G_k(x',s))\cdot
\frac{\partial }{\partial s}(G_k(x',s))
\nonumber \\
&=& D\varphi (z)\cdot Dd_\Omega(z).
\label{s-deriv}\end{aligned}$$
We insert the right-hand side of (\[s-deriv\]) into the right-hand side of (\[I10\]) and use area formula and (\[MeasZeroRel\]) to change variables from $(x', s)$ to $z$ in the integral on the right-hand side of (\[I10\]). Then using that $\sum_k
\Psi_k\equiv 1$ on $\Omega^0 \backslash {\cal R}$ and that $a \equiv 0$ on ${\cal R}$ and in ${ {\bf R}}^n \setminus \Omega$, we get (\[I1\]).
Proof of Theorem \[Theorem 3\] {#Th.3}
==============================
To conclude the proof of Theorem \[Theorem 3\] we need to demonstrate the following. Let, as above, $w(x)=d_{\Omega}(x)$, and let $v\in W^{1,\infty}({ {\bf R}}^n)$ satisfy $|Dv|\leq 1$ a.e. Then $$\label{I15}
\int_{{ {\bf R}}^n}({w\over t}-w_t)w\geq \int_{{ {\bf R}}^n}({w\over t}-w_t)v.$$
We show that this follows from (\[I1\]). By approximation, (\[I1\]) is true for Lipschitz $\varphi$. Let $\varphi=w-v$. Then by (\[I1\]) that it is enough to prove that $$\int_{{ {\bf R}}^n} aDw(Dw-Dv)\geq 0.$$ But, since $a\geq 0$ in ${ {\bf R}}^n$ and $a \equiv 0$ outside $\Omega^0$ and on ${\cal R}$, and $|Dw|\equiv 1$ on $\Omega \backslash {\cal R}$, $$\int_{{ {\bf R}}^n} aDw Dw=\int_{{ {\bf R}}^n}a\geq \int_{{ {\bf R}}^n}aDw
Dv$$ since $ |DwDv|\leq\sqrt{|Dw|^2|Dv|^2}\leq 1$.
Theorem \[Theorem 3\] is proved.
Proof of Theorem \[Theorem 4\] {#Th.4}
==============================
The proof of Theorem \[Theorem 4\] is similar to the proof Theorem \[Theorem 3\]. We give a formal calculation. Each step can be justified the same way as in the proof of Theorem \[Theorem 3\].
Fix $t\in [a, b]$. Define $${\cal D}_1={\cal D}_1(t)=\{ x\in
\Omega^1_t\; |\; d_{\Omega^1_t}(x)>d_{\Omega^2_t}(x)\},$$ $${\cal D}_2={\cal D}_2(t)=\{ x\in
\Omega^2_t \; |\; d_{\Omega^2_t}(x)>d_{\Omega^1_t}(x)\}.$$ Let $r>0$, let $\Omega^1_r=\Omega^1_{t,r}$ and $\Omega^2_r=\Omega^2_{t,r}$ be $r$-neighborhoods of $\Omega^1=\Omega^1_t$ and $\Omega^2=\Omega^2_t$, and let $r$ be so small that $\partial \Omega^1_r$ and $\partial \Omega^2_r$ are $C^{1,1}$ manifolds.
Let l=1,2. Let $\{ \tilde{U}^l_k\}^{N_l}_{k=1}$ and $\{ U^l_k\}^{N_l}_{k=1}$ be the coordinate neighborhoods associated the set $\Omega^l_r$ same way as $\{ \tilde{U}_k\}$, $\{ U_k\}$ with $\Omega_r$ in the proof of Theorem \[Theorem 3\]. Let $G^l_k:U^l_k\rightarrow
{ {\bf R}}^n$ be the corresponding coordinate mappings (\[I2\]) and let $\{ \Psi^l_k\}^{N_l+1}_{k=0}$ be the corresponding partitions of unity (\[defPsi\]). Let $\tilde{\gamma}^l_{r,k}$ be the functions defined as in (\[I3\]) from the functions $\gamma_{\Omega_r^l}$ using the coordinate mappings $G^l_k$ for $k=1,...,N_l$. Define the following functions $\tilde{\delta}^l_{r,k}$ for $k=1,...,N_l$. Let $\delta (\cdot)\equiv
\gamma_{\Omega^1 \cap \Omega^2}(\cdot)$. Then $$\tilde{\delta}^l_{r,k}(x')=\left\{
\begin{array}{ll}\delta(G^l_k(x',r))+r,
& \mbox{if }x'\in \mbox{cl}(\tilde{U}^l_k )
\;\; \mbox{ and }\;\; G^l_k(x',r) \in
\partial(\Omega^1 \cap \Omega^2);
\\
r, & \mbox{otherwise}.
\end{array} \right.$$ Note that $$(G^l_k)^{-1}({\cal D}_l) \cap U^l_k = \{\; (x', s) \;\;\; |
\;\;\; x' \in \tilde{U}^l_k, \;s \in [\,\tilde{\delta}^l_{r,k}(x'),\;
\tilde{\gamma}^l_{r,k}(x')\,] \;\}.$$
The functions $\tilde{\delta}^l_{r,k}$ are uppersemicontinuous, the proof is similar to the proof of Proposition \[Proposition I1\], using the properties of sets satisfying Condition \[Cond.1\].
Define the mass transport density $a(\cdot)$. Let $a\equiv 0$ on ${ {\bf R}}^n \backslash (\Omega^1 \cup \Omega^2)$. To define $a$ in $\Omega^1 \cup \Omega^2$ it is enough to define $a$ on ${\cal D}_1 \cup {\cal D}_2$. Let $x\in {\cal D}_l\backslash {\cal R}_l$, where $l=1,2$. Let $y \in {\cal N}_{\partial \Omega_l}(x)$ and $s = |x-y|$. Then we define $$a(x)={1\over t}\prod^{n-1}_{i=1}\frac{1}{1-s
\kappa_i}\int^s_{\delta(y)} \prod^{n-1}_{j=1}(1-\kappa_j\xi)\left[
\frac{\int^{\gamma^l(y)}_{\delta(y)}\nu \prod^{n-1}_{k=1}(1-
\kappa_k\nu)d\nu}{\int^{\gamma^l(y)}_{\delta(y)} \prod^{n-1}_{k=1}(1-
\kappa_k\nu)d\nu}-\xi \right]d\xi,$$ where $\kappa_k$ are curvatures of $\partial \Omega^l_t$ at $y$. Then $a$ satisfies the equation (\[TrDenODE\]) on subintervals of distance rays of ${\Omega^l_t}$ that lie in ${\cal D}_l,$ and $a\equiv 0$ on $\partial {\cal D}_l$ and on ${\cal R}_l,\;l=1,2$. We calculate, denoting by $\kappa_{r,i} $ the principal curvatures of $\partial \Omega^l_r$ at the point $y_r \in \Omega^l_r$ nearest to $x$ $$\begin{aligned}
\int_{{ {\bf R}}^n}({w\over t}&-&w_t)\varphi dz
\nonumber \\
&=&{1\over
t}\sum^2_{l=1}\sum^{N_l}_{k=1}\int_{U^l_k\cap (G^l_k)^{-1}({\cal
D}_l)}(w\circ G^l_k-
V\circ G^l_k)(\varphi \circ G^l_k)(\Psi^l_k
\circ G^l_k)J_nG^l_k\,dx
\nonumber \\
&=&\sum^2_{l=1}\sum^{N_l}_{k=1}
\int_{\tilde{U}^l_k}
\Psi^l_k(G^l_k(x',0))
\nonumber \\
& & \times
\int^{\tilde{\gamma}^l_{r,k}(x')}_{\tilde{\delta}^l_{r,k}(x')}
{1\over t}\left[(s-r)-
V(G^l_k(x',0))\right](\varphi(G^l_k(x',s))J_nG^l_k(x',s)\, ds dx'
\nonumber \\
&=&\sum^2_{l=1}\sum^{N_l}_{k=1}
\int_{\tilde{U}^l_k}
\Psi^l_k(G^l_k(x',0))
\int^{\tilde{\gamma}^l_{r,k}(x')}_{\tilde{\delta}^l_{r,k}(x')}\left( -
\frac{d\;a(G^l_k(x',s))}{d\;s} \right.
\nonumber \\
& & \left.
+a(G^l_k(x',s))\sum^{n-
1}_{i=1}\frac{\kappa_{r,i} }{1-\kappa_{r,i} s} \right)
\varphi \, J_nG^l_k(x',s)\,ds dx'
\nonumber \\
&=& \sum^2_{l=1}\sum^{N_l}_{k=1}
\int_{\tilde{U}^l_k}
\Psi^l_k(G^l_k(x',0))
\nonumber \\
& & \times
\int^{\tilde{\gamma}^l_{r,k}(x')}_{\tilde{\delta}^l_{r,k}(x')}
a(G^l_k(x',s))\
\frac{\partial \varphi(G^l_k(x',s ))}{\partial s}JG^l_k(x',s)\,ds dx'
\nonumber \\
&=&\sum^2_{l=1}\sum^{N_l}_{k=1}\int_{G^l_k(U^l_k)\cap {\cal
D}_l}a(z)\Psi_k^l(z)\;D\varphi (z)\;Dw(z)\,dz
\nonumber \\
&=& \int_{{ {\bf R}}^n}aDwD\varphi\, dz.
\nonumber\end{aligned}$$ Thus we have proved that the mass balance equation is satisfied. This implies Theorem \[Theorem 4\].
Compression molding model {#CompMod}
=========================
Compression molding is the process of deformation of an incompressible plastic material between two horizontal plates. The following simplified mathematical model of the process was derived by G. Aronsson [@ArCompMol] based on Hele-Show model for non-Newtonian fluid. Suppose that the distance between the horizontal plates is small. Then we can assume that the region occupied by plastic at each time $t$ has the form $U_t = \Omega_t \times [0, h_t]$ where $\Omega_t \subset { {\bf R}}^2$, and that the pressure in plastic does not depend on the vertical coordinate, i.e., pressure is the function $u(x,t)$ where $x \in \Omega_t$. Evolution of rescaled $\Omega_t$ and $u(x,t)$ is described by the following free boundary problem. Given an open set $\Omega_0 \in { {\bf R}}^2$ find an expanding family of open sets $\Omega_t \in { {\bf R}}^2, t \geq 0,$ and a function $u(x,t)$ defined on $\cup_t (\Omega_t\times \{ t \})$ such that $$\begin{aligned}
- \mbox{div} (|Du|^{ p-2 } Du)=1 & \;\;\; & \mbox{in } \Omega_t,
\label{CompMoldEqP} \\
u=0 & & \mbox{on } \Gamma_t,
\label{bdCondP} \\
V= |Du|^{p-2} & & \mbox{on } \Gamma_t,
\label{fBdCondP}\end{aligned}$$ where $\Gamma_t = \partial \Omega_t$, $V$ denotes the outer normal velocity of $\Gamma_t$. Condition (\[fBdCondP\]) means that the free boundary $\Gamma_t$ moves with the velocity of the flow.
In the paper [@AEComprMot] the asymptotic limit as $p \rightarrow
\infty$ in the problem (\[CompMoldEqP\]) - (\[fBdCondP\]) was considered. This limit corresponds to the case of highly non-Newtonian fluid.
It was shown in [@AEComprMot] that formally sending to a limit in (\[CompMoldEqP\]) - (\[fBdCondP\]) one obtains the following problem. Find a family $\{ \Omega_t \}$ of open subsets of $ { {\bf R}}^2$, a function $u(x,t)$, and a mass transport density function $a(x,t)$ satisfying: $$\label{DensLimFormOm}
\left\{
\begin{array}{lll}
|Du| \leq 1 & \;\;\; & \mbox{a.e. in } \Omega_t, \\
a \geq 0 & & \mbox{a.e. in } \Omega_t, \\
\mbox{supp}(a) \in \{ |Du| = 1 \}, & & \\
- \mbox{div}(aDu)=1 & & \mbox{ in } \Omega_t,
\end{array}
\right.$$ where the last equation is understood in the weak sense, and $$\label{DensLimFormGam}
u=0, \;\; V=a \;\; \mbox{ on } \Gamma_t,$$ where $V$ is the outer normal velocity of $\Gamma_t$.
It was shown in [@AEComprMot] that solutions of (\[DensLimFormOm\]), (\[DensLimFormGam\]) have the form $$\label{DistFormula}
u(x,t)=d_{ \Omega_t}(x),$$ where the right-hand side is defined by (\[distFunc00\]), and that formally the problem (\[DensLimFormOm\]) - (\[DensLimFormGam\]) can be rewritten as following. Find $\{\Omega_t \}$ such that $$\label{CompMoldEv}
w- \partial_t w \in \partial I_{\infty} [u],$$ where $$\label{indFnct}
w(\cdot, t) = \chi_{\Omega_t}(\cdot),$$ where $ \chi_{\Omega_t}(\cdot)$ is the indicator function of $\Omega_t$ that equals 1 inside and 0 outside $\Omega_t$. Existence of a weak solution of (\[DistFormula\]) - (\[CompMoldEv\]) was proved in [@AEComprMot]. Namely, there exists a family $\{\Omega_t\}$ of sets of finite perimeter such that $\partial_t w$ is a nonnegative Radon measure and $$\begin{aligned}
\int_{{ {\bf R}}^2} w(x, t)(v(x)-u(x, t))dx \leq
\int_{{ {\bf R}}^2} (v-u(\cdot, t)) d(\partial_t w(\cdot,t))
\nonumber \\
\label{weakCompMold} \\
\;\; \mbox{for every $v$ with }
|Dv| \leq 1, \;\mbox{for a.e. } t.
\nonumber\end{aligned}$$ The following law of motion of the free boundary $\Gamma_t$ was derived in [@AEComprMot] by a formal calculation: $$\label{MotBdCompMol}
V=\gamma (1-\frac{\kappa\gamma}{2}),$$ where $$\left\{
\begin{array}{l}
V = \mbox{ outer normal velocity of } \Gamma_t, \\
\gamma = \mbox{ function $\gamma_{\Omega_t}(\cdot)$
defined by (\ref{defGamma})}, \\
\kappa = \mbox{ curvature of } \Gamma_t.
\end{array}
\right.$$
The equation (\[MotBdCompMol\]) was derived as following. Starting from (\[DensLimFormOm\]), we perform calculations similar to the ones that lead from (\[defMassTrDen\]) to (\[TrDenODE\]). Thus we deduce that equation (\[DensLimFormOm\]) can be formally rewritten on each distance ray $R_x$, where $x \in \Gamma_t$, as ODE $$\label{TrDenODECMold}
a'(s)-a(s)
\frac{\kappa}{1-\kappa s} + 1 = 0, \;\;\;
s\in(0,\gamma_{\Omega_t}(x)).$$ By the nature of mass transfer process in the compression molding model (i.e., mass transfer from within the set onto the boundary), we expect that the mass transport density equals to zero on the ridge of $\Omega_t$ and equals to the outer normal velocity at the boundary. This translates into the following boundary conditions for the ODE (\[TrDenODECMold\]) on $R_x$: $$\label{TrDenODECMoldBCond}
a(0) = V(x), \;\;\; a(\gamma(x))=0.$$ The function $a(\cdot)$ and the constant $V$ can be found from (\[TrDenODECMold\]), (\[TrDenODECMoldBCond\]). $V$ has the expression (\[MotBdCompMol\]).
In this section we prove the connection between the variational equation (\[CompMoldEv\]) and the geometric equation (\[MotBdCompMol\]).
Let $E, {\bf \Gamma}$ be defined by (\[defE\]).
\[Theorem 1-CM\] Let $\{\Omega_t \},\; t\in { {\bf R}}^1_+$, be an expanding, locally Lipschitz continuous family of open bounded sets. Let for every $ t\in { {\bf R}}^1_+$ the set $\Omega_t$ satisfy Condition \[Cond.1\] with radius $r_0 > 0$. Suppose that the equation (\[MotBdCompMol\]) is satisfied at every point of $(2,1)$ differentiability of the surface ${\bf \Gamma}$. Then the equation (\[CompMoldEv\]) is satisfied for a.e. $t\in { {\bf R}}^1$, where the functions $u(x,t), \; w(x,t)$ are defined by (\[DistFormula\]), (\[indFnct\]).
The proof goes along the lines of the proof of Theorem \[Theorem 3\]. We will sketch the proof and present some details for the parts that are different from the proof of Theorem \[Theorem 3\].
The main steps of the proof are following:\
[*Step 1*]{}. Definition and properties of mass transport density.\
[*Step 2*]{}. Show that the main mass balance equation is satisfied.\
[*Step 3*]{}. Show that the assertion of the Theorem follows from the main mass balance equation.
We discuss each step.
[*Step 1*]{}.
Let $(x, t)$ be a point of $(2,1)$ differentiability of the surface ${\bf \Gamma}$ in ${ {\bf R}}^2 \times { {\bf R}}^1$. Define mass transport density $a(y,t)= a(s)$ on the distance ray $R_x$ of $\Omega_t$ as the solution of (\[TrDenODECMold\]), (\[TrDenODECMoldBCond\]). Do this for every such point $(x, t)$. It follows from the hypothesis of the Theorem and from Proposition \[Proposition 4.1.t\] that $a(y,t)$ is now defined at ${\cal L}^{3}$ a.e. point of the set $E = \cup_{t}(\Omega_t \times \{t\} )
\subset { {\bf R}}^2 \times { {\bf R}}^1_+$. Define $a(y,t) $ by zero at all other points of ${ {\bf R}}^2 \times { {\bf R}}^1_+$.
Solving (\[TrDenODECMold\]), (\[TrDenODECMoldBCond\]) (and taking into account (\[MotBdCompMol\])) we see that on the ray $R_x$ $$\label{trDenRayCM}
a(s)=\frac{1-\kappa s}{2\kappa} -
\frac{(1-\kappa\gamma)^2}{2\kappa(1-\kappa s)}.$$ This function satisfies assertions (a) and (d) of Proposition \[Proposition 1\]. Instead of assertion (c) of Proposition \[Proposition 1\] we have $$\label{lipTrDen}
|a(s)|\leq C(\gamma(x)-
s),\;\;\;\;
|a(s)- V(x)|\leq Cs.$$ To see this, we note that from the conditions of the Theorem and (\[MotBdCompMol\]) it follows that $$V(x) \leq C(r_0, { \mbox{diam}}\Omega_t).$$ Then the assertions (a) and (d) of Proposition \[Proposition 1\] for the function $a(s)$ defined by (\[trDenRayCM\]) follow from (\[lipTrDen\]) like in the proof of Proposition \[Proposition 1\].
Let us prove (\[lipTrDen\]). We can rewrite (\[trDenRayCM\]) as $$a(s) = \frac{\gamma -s}{2}\left(1 +
\frac{1-\kappa \gamma}{1-\kappa s}\right).$$ Now, since $0 \leq s \leq \gamma \leq { \mbox{diam}}\Omega_t$ and $\kappa \gamma \leq 1$ and $\kappa \geq -\frac{1}{r_0}$, we get the estimate $$\label{est1}
0 \leq \frac{1-\kappa\gamma}{1-\kappa s} \leq C(r_0, { \mbox{diam}}\Omega_t),$$ and the first inequality of (\[lipTrDen\]) follows. To prove the second, we calculate using (\[TrDenODECMoldBCond\]) $$a(s) - V = a(s)-a(0) = \frac{s}{2}\left[-1-
\frac{(1-\kappa\gamma)^2}{1-\kappa s}\right],$$ and use (\[est1\]) and the inequality $$0 \leq 1-\kappa \gamma \leq C(r_0, { \mbox{diam}}\Omega_t)$$ to finish the proof of (\[lipTrDen\]).
[*Step 2*]{}.
The purpose of this step is to prove that for any smooth function $\varphi \in C^1_c({ {\bf R}}^2 \times (0,T))$ we have $$\label{balEqComprMold}
\int_0^T \int_{{ {\bf R}}^2} w (\varphi + \varphi_t)
-aDuD\varphi = 0.$$ This is main mass balance equation for compression molding model.
The proof follows Section \[ChngVar\].
Fix $t$. We use local coordinate systems on $(\Omega_t)_{r}$ where $0<r\leq r_0$, defined in Section \[Sets\] and notation introduced there, in particular sets $\tilde{U}_k$, $U_k$, partition of unity $\Psi_k$, coordinate mappings $G_k$ and functions $\tilde{\gamma}_{r,k}$ defined by (\[defUk\]), (\[defPsi\]), (\[I2\]) and (\[I3\]) respectively (k=1,...,N).
We use notation $E_r$, ${\bf \Gamma}_r$ introduced in (\[defEr\]). From Proposition \[Proposition 2\] it follows that for $r \in [0, r_0]$ $${\bf \Gamma}_r = \{( x ,t)\in { {\bf R}}^2 \times (0,T) \;\;
| \;\; d^s_{\Omega_t}(x) = -r \}.$$ The function $$(x,t) \rightarrow d^s_{\Omega_t}(x)$$ is Lipschitz since $\{ \Omega_t \}$ is a Lipschitz family of sets. Then by Proposition 3.2.15 of [@Federer], for a.e. $r \in [0, r_0]$ $$set \;\; {\bf \Gamma}_r\;\; is\;\; ({\cal H}^2, 2)\;\; rectifiable.
\label{assm1}$$ In addition, by Lemma \[semiconcDist\], the function $(x,t) \rightarrow d^s_{\Omega_t}(x) - C|x|^2$, where $C$ is large enough depending on $r_0$, is concave in $x$ for every $t$ in the set $E_{r_0}$. Then it follows from Theorem 1 of Appendix 2 of [@Krylov] and Propositions \[Proposition 4.1.t\](b) and \[Proposition 3\] that for a.e. $r \in [0, r_0]$ $$\begin{aligned}
& &surface\;\;{\bf \Gamma}_r\;\; is \;\; (2,1)\;\,
differentiable \;\, at \;\;{\cal H}^2\;\;
a.e. \;\, point;
\label{assm2} \\
&& \nonumber \\
& &functions \;\; \partial_t d^s_{\Omega_t}(x), \;
Dd^s_{\Omega_t}(x),\; D^2 d^s_{\Omega_t}(x)\;\, are\; \,
bounded\;\, and\;\;
\nonumber \\
& & {\cal H}^2\;
measurable\; on\; {\bf \Gamma}_r.
\label{assm3}\end{aligned}$$
Function $(x,t) \rightarrow \gamma_{\Omega_t}(x)$ is uppersemicontinuous since the family of sets $\{\Omega_t\}$ is Lipschitz. Then similar to Propositions \[Proposition I3\], \[Proposition I4\] and Corollary \[Corollary I1\] we prove that the mass transport density $a(y,t)$ defined at Step 1 is Lebesgue measurable in ${ {\bf R}}^2 \times { {\bf R}}^1_+$. In the proof we use (\[assm3\]) (i.e., we choose such $r$ in the definition of local coordinate systems that (\[assm3\]) is satisfied).
It follows from Proposition \[Proposition 4.1.t\] that for a.e. $t$ we have the following: a.e. point $x \in \Omega_t$ lies in the relative interior of the distance ray of the set $\Omega_t$ that intersects $\partial \Omega_t$ at a point of (2,1) differentiability of the surface ${\bf \Gamma}$. Fix such $t$. Let $r\in (0, r_0]$. Let $y \in \partial \Omega_t$ be a point of (2,1) differentiability of ${\bf \Gamma}$. Let $\kappa^t$ be the curvature of $\partial \Omega_t$ at $y$, let denote $\kappa_r^t$ be the curvature of $\partial( \Omega_t)_r$ at the unique point $y\in \partial (\Omega_t)_r$ nearest to $x$, then $$\frac{1}{\kappa^t_r} = \frac{1}{\kappa^t} + r.$$ We calculate using properties of the function $a$ on the rays proved in the Step 1, equation (\[TrDenODECMold\]), and Lemma \[jacobian lemma\] with $n=2$: $$\begin{aligned}
\int_{{ {\bf R}}^2}w \varphi \,dz &=&\int_{\Omega_t} \varphi \,dz
\nonumber \\
&=&
\sum^{N}_{k=1}\int_{U_k}(\varphi \circ G_k)(\Psi_k
\circ G_k)J_2G_kdx
\nonumber \\
&=& \sum^{N}_{k=1}
\int_{\tilde{U}_k}
\Psi_k(G_k(x',r))
\int^{\tilde{\gamma}_{r,k}(x')}_r
\left[\frac{da(s-r)}{ds} \right.
\nonumber \\
& & \left. +a(s-r)\frac{\kappa^t_{r} }{1-\kappa^t_{r} s}\right]
(\varphi(G_k(x',s))J_2G_k(x',s) ds dx'.
\nonumber \end{aligned}$$ In the last expression we integrate by parts and use (\[jacDeriv\]), (\[TrDenODECMoldBCond\]) and (\[s-deriv\]). Then we get
$$\begin{aligned}
\int_{{ {\bf R}}^2}w \varphi \,dz &=&
\sum^{N}_{k=1}
\int_{\tilde{U}_k}
\Psi_k(G_k(x',r))
\left[ \begin{array}{l} \\ \\ \end{array} \right.
a(0)\varphi(G_k(x',r) )J_2G_k(x',r)
\nonumber \\
& &
+ \int^{\tilde{\gamma}_{r,k}(x')}_r
a(s-r)
\frac{\partial \varphi(G_k(x',s ))}{\partial s}J_2G_k(x',s)ds
\left. \begin{array}{l} \\ \\ \end{array} \right ]
dx'.
\nonumber \\
&=&
\int_{\hat{\partial}( \Omega_t)_r }\varphi(P^t_r(y))
\partial_t d^s_{\Omega_t}(P^t_r(y)) (1-r \kappa_r(y)) d{\cal H}^1(y)
+ \int_{\Omega_t}aD \varphi Du \,dz,
\nonumber\end{aligned}$$
where $$\label{gammaBdry}
\hat{\partial}( \Omega_t)_r = \{ y\in\partial ( \Omega_t)_r \;\; |
\gamma_{( \Omega_t)_r}(y) > r \},$$ and $$\label{nearPtPr}
P^t_r : (\Omega_t)_r \rightarrow \Omega_t$$ is the nearest point projection mapping. By Proposition \[Proposition 2\] b) the mapping $P^t_r$ is well-defined and onto.
By Proposition \[Proposition 4t\], if ${\bf \Gamma}_r$ is (2,1) differentiable at $y \in \hat{\partial}( \Omega_t)_r$, then ${\bf \Gamma}$ is (2,1) differentiable at $y'=P^t_r(y)$ and $$\partial_t d^s_{\Omega_t}(P^t_r(y)) = \partial_t d^s_{\Omega_t}(y)$$ By the equality $$d^{s}_{(\Omega_t)_{r}} (x) = d^{s}_{\Omega_t} (x) + r
\;\;\;\mbox{for }\;(x,t) \in E_{r_0}, \;\; r \in (0, r_0]$$ we see that $$\partial_t d^s_{\Omega_t}(P^t_r(y)) = \partial_t d^s_{(\Omega_t)_r}(y)$$ for $y \in \hat{\partial}( \Omega_t)_r$.
Thus we have showed that for a.e. $t \in [0, T]$, every $r \in (0, r_0]$ the equality holds $$\label{aeTcomprmold}
\int_{{ {\bf R}}^2}w \varphi \,dz =
\int_{\hat{\partial} (\Omega_t)_r }
(\varphi\circ P^t_r) \, \partial_t d^s_{(\Omega_t)_r}(1- r\kappa_r^t)
d{\cal H}^1
+ \int_{\Omega_t}aD \varphi Du \,dz.$$
In the next lemma we show that in the first integral at the right-hand side of (\[aeTcomprmold\]) we can integrate over the whole boundary $\partial (\Omega_t)_r$.
\[Int0\] Let $r \in (0, r_0]$ be such that properties (\[assm1\]), (\[assm2\]), (\[assm3\]) are satisfied. Then for a.e. $t \in (0, T)$ $$\int_{\partial (\Omega_t)_r \, \cap \{\gamma_{(\Omega_t)_r} \leq r \} }
\varphi \partial_t d^s_{(\Omega_t)_r}(1- r\kappa_r^t) d{\cal H}^1 =
0$$
We will show that for a.e. $t$ $$\label{intgrnd0}
\partial_t d^s_{(\Omega_t)_r}(y)(1- r\kappa_r^t(y)) = 0
\;\;\; {\cal H}^1 \;\mbox{ a.e. on }
\partial (\Omega_t)_r \, \cap \{\gamma_{(\Omega_t)_r} \leq r \}.$$
By (\[assm2\]) for a.e. $t$ the surface ${\bf \Gamma}_r$ is (2,1) differentiable ${\cal H}^1$ a.e. on $\partial (\Omega_t)_r$. Fix such $t$. Then in (\[intgrnd0\]) we can consider only $y$ at which ${\bf \Gamma}_r$ is (2,1) differentiable. Fix such $y$.
By Proposition \[Proposition 2\], $$\gamma_{(\Omega_t)_r} \geq r \;\; \mbox{ on } \; \partial (\Omega_t)_r.$$ Thus $$\gamma_{(\Omega_t)_r}(y) = r.$$ It follows that $$\kappa_r^t(y) \leq \frac{1}{r}.$$ Let $y' = P^t_r(y)$, i.e., $y'$ is the point of $ \partial \Omega_t$ nearest to $y$. Consider 3 cases.
[**Case 1.**]{} [*$y$ is the unique point of $\partial (\Omega_t)_r$ for which $y'$ the nearest point on $\partial \Omega_t$.*]{}
Then the calculations of [@EG], Proposition 7.1, Steps 2,3 imply that $$\kappa_r^t(y) = \frac{1}{r}.$$ Thus we have (\[intgrnd0\]) in this case.
[**Case 2.**]{}[*There exists $z \in \partial (\Omega_t)_r$, $z \neq y$, such that $y' = P^t_r(z)$, and* ]{} $$B_r(y) \cap B_r(z) \neq \emptyset.$$
Note that $$|y-y'| = |z-y'| = r.$$ Then, denoting $w$ the point $\frac{y+z}{2}$, we get $$w \in B_r(y) \cup B_r(z).$$ We have $$B_r(y),\; B_r(z) \subset { {\bf R}}^2 \setminus \overline{\Omega}_t.$$ Thus $
w \in { {\bf R}}^2 \setminus \overline{\Omega}_t,
$ and $y'$ is the point of $\partial \Omega_t$ nearest to $w$. Denote $$v=y' + \frac{r}{|w-y'|}(w-y').$$ Then, by Condition \[Cond.1\], we get the following: $$v \notin \Omega_t, \;\;\;
y' \; is \; the\;unique \; point \; of\; \partial\Omega_t \;
nearest \; to \; v.$$ But $$|v-y'|=r.$$ Thus we proved that $$v \in \partial (\Omega_t)_r.$$ The points $y$ and $z$ divide the circle $\partial B_r(y')$ on two arcs, and the point $v$ is the middle point of one of these arcs. Denote this arc $C_1$. Repeating the above argument inductively and using continuity of distance function, we prove that $$C_1 \subset \partial (\Omega_t)_r.$$ But then $$\kappa_r^t(y) = \frac{1}{r}.$$ Thus we have (\[intgrnd0\]) in the Case 2.
[**Case 3.**]{}[*There exists $z \in \partial (\Omega_t)_r$, $z \neq y$, such that $y' = P^t_r(z)$, and* ]{} $$B_r(y) \cap B_r(z) = \emptyset.$$
Then $$\partial B_r(y) \cap \partial B_r(z) = {y'}.$$ Since $\{ \Omega_t \}$ is an expanding family of sets, we have $$\label{expand}
B_r(y),\; B_r(z) \subset { {\bf R}}^2 \setminus \overline{\Omega}_{\tau}
\;\;\; \mbox{ for } \; \tau < t.$$ If $$\label{subCond1}
y' \in \partial \Omega_{\tau^*} \;\;\;
\mbox{ for some } \;\; \tau^* < t,$$ then the same is true for all $\tau \in [\tau^*, t]$, and (\[expand\]) implies that $$y \in \partial (\Omega_{\tau})_r \;\;\; \mbox{ for all } \;\;
\tau \in [\tau^*, t].$$ Then $$\partial_t d^{s}_{(\Omega_t)_{r}} (y) = 0.$$ Thus we have (\[intgrnd0\]) if (\[subCond1\]) is satisfied.
The remaining case is $$\label{subCond2}
y' \notin \partial \Omega_{\tau} \;\;\;
\mbox{ for all } \;\; \tau < t.$$ Introduce a coordinate system $(x_1, x_2)$ on $R^2$ in which $$y'=(0,0),\;\; y=(0, r).$$ Then $$z=(0, -r).$$ Since $\{\Omega_t\}$ is a continuous expanding family of sets, the function $$\phi (\tau) = { \mbox{dist}}(y', \Omega_{\tau})$$ is continuous and nonincreasing, and $$\phi (t) = 0.$$ Let $w_{\tau} \in {\cal N}_{\partial \Omega_{\tau}}(y')$. Then by (\[expand\]) $$\label{wTau}
w_{\tau} \in \partial B_{\phi (\tau)}(0,0) \setminus
\left [ B_r(0,r) \cup B_r(0, -r)) \right ].$$ In particular, $$\label{convWtau}
w_{\tau} \rightarrow y'\;\;\; \mbox{as}\;\;\; \tau \rightarrow t.$$ By Condition \[Cond.1\] we get $$\label{vBall}
B_r(v_{\tau}) \subset { {\bf R}}^2 \setminus \overline{\Omega}_{\tau} \;\;\;
\mbox{ where } \;\;
v_{\tau}= w_{\tau} + \frac{r}{| y' - w_{\tau}|}(y' - w_{\tau}).$$ By (\[wTau\]), (\[convWtau\]) we see that there exists a sequence $\tau_j \rightarrow t$ such that $$v_{\tau_j} \rightarrow p, \;\; \mbox{where} \;\; p \;\;
\mbox{is either} \;\;
(-r, 0) \;\; \mbox{or} \;\; (r, 0).$$ Then by (\[vBall\]) and continuity of the family $\{ \Omega_t \}$ we conclude $$\mbox{either } \;\;B_{r} (-r, 0)
\subset { {\bf R}}^2 \setminus \overline{\Omega}_t
\;\;\mbox{ or }\;\;
B_{r}(r, 0) \subset { {\bf R}}^2 \setminus \overline{\Omega}_t.$$ Let in fact $$B_{r}(r, 0) \subset { {\bf R}}^2 \setminus \overline{\Omega}_t.$$ Then $y'=(0,0)$ is the point of $\partial \Omega_t$ nearest to the point $(r, 0)$. Thus $$(r, 0) \in \partial (\Omega_t)_r.$$ But $y=(0, r)$, and thus $$B_r(y) \cap B_r(r, 0) \neq \emptyset.$$ Thus the points $y$ and $(r, 0)$ satisfy the conditions of Case 2. Thus we get $$\kappa_r^t(y) = \frac{1}{r}.$$ Case 3 is proved. Then Lemma \[Int0\] is proved.
The properties (\[assm1\]), (\[assm2\]), (\[assm3\]) are satisfied for a.e. $r \in (0, r_0]$. Then there exists a decreasing sequence $r_i \rightarrow 0 $, where $i=1, 2, ...$ and $r_i \in (0, r_0]$, such that (\[assm1\]), (\[assm2\]), (\[assm3\]) are satisfied for each $r_i$. Note that by (\[assm3\]), the function $$(y,t) \rightarrow \kappa_{r_i}^t(y)$$ defined on ${\bf \Gamma}_{r_i}$ is ${\cal H}^2$ measurable.
Now we can integrate (\[aeTcomprmold\]) by $t$ and use Lemma \[Int0\] to get $$\begin{aligned}
\nonumber
\int_0^T \int_{{ {\bf R}}^2}w \varphi \,dzdt
& = & \int_0^T \int_{\partial (\Omega_t)_{r_i} }
(\varphi\circ P^t_{r_i}) \, \partial_t d^s_{(\Omega_t)_{r_i}}
(1-r_i \kappa_{r_i}^t)
d{\cal H}^1\,dt
\\
&& + \int_0^T \int_{\Omega_t}aD \varphi Du \,dzdt.
\label{0Tcomprmold}\end{aligned}$$
It follows that the first integral at the right-hand side does not depend on $r_i$. Thus it is enough to compute the limit as $i \rightarrow \infty$. We will prove the following: $$\label{IndFunctEqual}
\lim_{i\rightarrow\infty}
\int_0^T \int_{\partial (\Omega_t)_{r_i} }
(\varphi\circ P^t_{r_i}) \, \partial_t d^s_{(\Omega_t)_{r_i}}
(1- r_i \kappa_{r_i}^t)
d{\cal H}^1\,dt =
-\int_0^T \int_{{ {\bf R}}^2} \chi_{\Omega_t} \partial_t\varphi\,dxdt$$
We first prove that such equality is true if the boundary satisfies additional regularity assumptions.
Let $\{ \Omega_t \}$ satisfy (\[assm1\]) - (\[assm3\]) with $r=0$. Let $V(x,t)$ be the outer normal velocity of $\Gamma_t$, defined by (\[outNormVelDef\]) at every point $(x,t)$ of differentiability of ${\bf \Gamma}$ and defined as 0 at all points where ${\bf \Gamma}$ is not differentiable. Then for every $\varphi \in C^1_c({ {\bf R}}^2 \times (0,T))$ $$\label{kinetimatic}
\int_0^T \int_{\partial \Omega_t }
\varphi V d{\cal H}^{1} dt =
-\int_0^T \int_{{ {\bf R}}^2} \chi_{\Omega_t} \partial_t\varphi\,dxdt$$
Let the set $E$ be defined by (\[defE\]). Then by (\[assm1\]) $E$ has locally finite perimeter. By (\[assm2\]) $${\cal H}^2 (\partial E \setminus \partial_* E )= 0,$$ where $\partial_* E$ is the reduced boundary of $E$. Let $\Phi: { {\bf R}}^2 \times { {\bf R}}^1 \rightarrow { {\bf R}}^2 \times { {\bf R}}^1$ be defined by $$\Phi(x,t) = (0,...,0,\varphi(x,t))$$ Let $\nu_E(x,t)$ be the measure-theoretical outer normal to ${\bf \Gamma}$ at $(x,t) \in {\bf \Gamma}$. Then by Green-Gauss theorem for sets with finite perimeter ([@EGar], section 5.8) $$\int_0^T \int_{{ {\bf R}}^2} \chi_{\Omega_t} \partial_t\varphi\,dxdt =
\int_E { \mbox{div}}\Phi\,dxdt = \int_{\bf \Gamma} \Phi\, \nu_E\, d{\cal H}^2,$$ where $${ \mbox{div}}\Phi =\sum_{i=1}^{3} \frac{\partial \Phi_i}{\partial x_i},\;\;\;
\mbox{ where }\;x_{3}=t.$$ At every point $(x,t)$ of differentiability of ${\bf \Gamma}$ we have $$\nu_E(x,t) = - D_{(x,t)}d^s_E(x,t),$$ where $d^s_E$ is the signed distance to the boundary of the set $E$ in the $(x,t)$-space. Thus we get $$\int_E \partial_t\varphi\,dxdt =
- \int_{\bf \Gamma}\varphi\, \partial_t d^s_E \,d{\cal H}^2.$$ Let $f: {\bf \Gamma} \rightarrow { {\bf R}}^1$ be the mapping defined by $$f(x,t) = t$$ Then $f$ is Lipschitz, and $f$ and ${\bf \Gamma}$ are differentiable at ${\cal H}^2$ a.e. point $(x,t) \in {\bf \Gamma}$. At such point $(x,t)$ the gradient of $f$ is a linear mapping $Df(x,t):\,T_{x,t}{\bf \Gamma} \rightarrow { {\bf R}}^1$, where $T_{x,t}{\bf \Gamma}$ is the tangent to ${\bf \Gamma}$ at $(x,t)$ space. Let $e_1, e_2$ be such orthonormal basis in ${ {\bf R}}^2$ that $e_2$ is the inner normal to $\partial \Omega_t$ at $x$. Let $\tau$ be the unit vector in the $t$-direction. Then the vectors $$e_1,\;
\tilde{e}_2=-\partial_t d^s_E(x,t) e_2 + |D_x d^s_E(x,t)|\tau$$ form an orthonormal basis in $T_{x,t}{\bf \Gamma}$. We calculate: $$\begin{aligned}
Df(x,t)e_1 &=& 0,
\nonumber \\
Df(x,t)\tilde{e}_2 &=& |D_x d^s_E(x,t)|.
\nonumber\end{aligned}$$ Thus $$|Df(x,t)|= |D_x d^s_E(x,t)|.$$ Also, the following relation holds: $$D_{x,t} d^s_{\Omega_t}(x,t) = \frac{1}{|D_x d^s_E(x,t)|}D_{x,t} d^s_E(x,t).$$ Now, applying formula 3.2.22 of [@Federer] (which is applicable by (\[assm1\])), we get: $$\begin{aligned}
\int_{\bf \Gamma}\varphi\, \partial_t d^s_E \,d{\cal H}^2
& = & \int_{\bf \Gamma}\varphi \, \partial_t d^s_{\Omega_t}
|Df(x,t)| \,d{\cal H}^2
\nonumber \\
&= & \int_0^T \int_{f^{-1}({t})}\varphi \, \partial_t d^s_{\Omega_t}
\,d{\cal H}^{1}dt
\nonumber \\
&= & \int_0^T \int_{\partial \Omega_t} \varphi \, V
\,d{\cal H}^{1}dt.
\nonumber\end{aligned}$$ The lemma is proved.
Now we can prove (\[IndFunctEqual\]). Each $\{ (\Omega_t)_{r_i} \}$ satisfies (\[assm1\]) - (\[assm3\]). Thus we have $$\label{approx}
\int_0^T \int_{\partial (\Omega_t)_{r_i} }
\varphi V^{(r_i)} d{\cal H}^{1} dt =
-\int_0^T \int_{{ {\bf R}}^2} \chi_{(\Omega_t)_{r_i}} \partial_t\varphi\,dxdt,$$ where $V^{r_i}$ is velocity of $(\Omega_t)_{r_i}$. It follows from Proposition \[Proposition 4.1\] that ${\cal L}^3({\bf \Gamma}) = 0$, and thus $$\chi_{(\Omega_t)_{r_i}}(x) \rightarrow \chi_{\Omega_t}(x)
\;\;\;\; \mbox{for a.e. } (x,t) \in { {\bf R}}^n\times(0,T).$$ Then by Dominated Convergence Theorem the right-hand side of (\[approx\]) converges to the right-hand side of (\[IndFunctEqual\]) as $i \rightarrow \infty$. Thus it remains to prove that the left-hand side of (\[approx\]) and the left-hand side of (\[IndFunctEqual\]) converge to the same limit.
Let $P^t_{r, r_i }: \partial (\Omega_t)_{r} \rightarrow
\partial (\Omega_t)_{r_i}$ be the nearest point projection (well-defined by Proposition \[Proposition 2\]). $P^t_{r, r_i }$ is a Lipschitz map by Propositions \[Proposition 2\], \[Proposition 3\]. Then using Lemma \[jacobian lemma\] and identity (\[relNbhd\]), the difference between the left-hand sides of (\[approx\]) and (\[IndFunctEqual\]) can be transformed to $$\begin{aligned}
\label{lhs1}
&& \int_0^T \int_{\partial (\Omega_t)_{r} }
(\varphi\circ P^t_{r, r_i} - \varphi\circ P^t_{r}) \,
V^{r_i}\circ P^t_{r, r_i} \,
(1- r \kappa_{r}^t)
d{\cal H}^1\,dt + \\
&& r_i\int_0^T \int_{\partial (\Omega_t)_{r} }
\varphi\circ P^t_{r, r_i} \,
V^{r_i}\circ P^t_{r, r_i}\, \kappa_{r}^t\,
d{\cal H}^1\,dt = I_{1, i} + r_i I_{2, i}.
\nonumber\end{aligned}$$
We have $|V^{r_i}|<C$, where $C$ does not depend on $i$, and $|\kappa_{r}^t| \leq \frac{1}{r}$. Using the fact that $P^t_{r, r_i} \rightarrow P^t_r$ as $r_i \rightarrow 0$ and Dominated Convergence Theorem we see that $I_{1, i} \rightarrow 0$. We also have $|I_{2, i}| < C$. Thus the expression (\[lhs1\]) converges to zero. Thus (\[IndFunctEqual\]) is proved.
The equalities (\[0Tcomprmold\]) and (\[IndFunctEqual\]) imply (\[balEqComprMold\]).
[*Step 3*]{}.
The family $\{\Omega_t\}$ is expanding. Thus the left-hand side of (\[IndFunctEqual\]) defines a nonnegative linear functional of $\varphi \in C^{\infty}({ {\bf R}}^n \times [0,T])$. By [@EGar], Chapter 1.8, Corollary 1, it follows from (\[IndFunctEqual\]) that $\partial_t \chi_{\Omega_t}$ is a nonnegative Radon measure. Thus the mass balance equation (\[balEqComprMold\]) can be rewritten as $$\int_0^T \int_{{ {\bf R}}^2} w \varphi\, dxtdt +
\int_{{ {\bf R}}^2 \times (0,T)}\varphi \,dw_t
-\int_0^T \int_{{ {\bf R}}^2} aDuD\varphi\, dxtdt = 0.$$ From this equation it follows that for a.e. $t \in [0,T]$, every $\varphi \in C^1_c({ {\bf R}}^2)$ $$\int_{{ {\bf R}}^2} w(\cdot, t) \varphi\, dx +
\int_{{ {\bf R}}^2}\varphi \,dw_t(\cdot, t)
-\int_{{ {\bf R}}^2} a(\cdot, t)Du(\cdot, t)D\varphi\, dx = 0.$$ Now, repeating argument of Section \[Th.3\], we conclude the proof of Theorem \[Theorem 1-CM\].
Appendix {#Apndx 1}
========
Let $\Omega$ be an open set. We write $d(x)$ for $d_{\Omega}(x)$ below. The purpose of this section is to prove that the gradient of $d_{\Omega}$ is locally Lipschitz at $x \in \Omega \setminus {\cal R}$ and give an estimate of the Lipschitz constant in the terms of the distance between $x$ and endpoints of the distance ray $R_x$.
Proposition \[Proposition 3Ap\] should be compared with Proposition 4.1 of [@EG] and with inequality 4.8(8) of [@FedCMeas].
In the inequality (\[1.1a\]) below the only assumption regarding the point $x_1$ is that it is close enough to $x$. In particular it is possible that $x_1 \in {\cal R}$. In the inequality 4.8(8) of [@FedCMeas] the conditions on $x$ and $x_1$ are symmetric and exclude the possibility that $x_1 \in {\cal R}$.
Proposition \[Proposition 3Ap\] improves the estimates of Proposition 4.1 of [@EG] in the following. Two quantities are estimated explicitly in Proposition \[Proposition 3Ap\]: the local Lipschitz constant of $Dd(\cdot)$ at $x$ and the size of the neighborhood of $x$ in which the estimates (\[1.1a\]) and (\[1.1\]) hold.
\[Proposition 3Ap\] There exist constants $C$ and $M$ depending only on $ n$ such that the following is true. Let $\Omega \in R^n$ be an open set. Let $\varepsilon > 0$, $x\in \Omega
\backslash {\cal R}$, and let $$\label{ptsAssm}
d_{\Omega}(x) \geq M\varepsilon, \;\;\;
\gamma_{\Omega}(x) - d_{\Omega}(x) \geq M\varepsilon.$$ Then by Remark \[remUnNr\] there exists a unique ray $R_x$. Denote $y$ and $v$ the lower and upper ends of $R_x$, i.e., $y=R_x \cap \partial \Omega$ and $v=R_x \cap {\cal R}$. Then for every $x_1\in \Omega$ satisfying $\mid x-x_1 \mid <\varepsilon$, the inequality holds $$\label{1.1a}
\mid y-y_1\mid \leq C\left(1+\frac{ \mid x-y \mid}{ \mid x-v \mid}\right)
\mid x-x_1 \mid,$$ where $y_1 \in {\cal N}_{\partial\Omega}(x_1)$. If in addition the function $d(\cdot)$ is differentiable at $x_1$ then $$\label{1.1}
|Dd(x)-
Dd(x_1)|\leq \frac{C }{ \varepsilon }|x-x_1|.$$
We first prove the inequalities (\[1.1a\]), (\[1.1\]) assuming that $$\label{convCond}
\mid x- v \mid \geq
\mid x- y \mid.$$
Denote $d:=d_{\Omega}(x)$. By (\[convCond\]), we can find a point $O$ on the interval of $R_x$ between $v$ and $x$ such that $\mid x-O\mid = d$. Let $y \in {\cal N}_{\partial \Omega}(x)$, then we also have $y \in {\cal N}_{\partial \Omega}(O)$. In the calculations below $C$ will denote different constants depending only on $n$. We assume that $$\label{Mgt10}
M>10.$$
Choose $x_1 \in B_{\varepsilon}(x)$. Let $\hat{x}_1$ be projection of $x_1$ onto $R_x$, then $$|x-\hat{x}_1|\leq \varepsilon \leq {1\over 10}d$$ by (\[ptsAssm\]), (\[Mgt10\]). Thus $\hat{x}_1$ lies between $y$ and $O$ on $R_x$. Let $d_1=|\hat{x}_1-O|$, then $$\label{1.2}
{11\over 10}\geq {d\over {d_1}}\geq{9\over 10},\; \; \; \;|d-d_1|
<\varepsilon$$ Introduce the following coordinate system in $R^n$: let the point $O$ be the origin, let $e_n=\frac{x-O}{|x-O|}$ (thus $e_n$ is the unit vector along the ray $R_x$), let $e_1,...,e_{n-1}, e_n$ be an orthonormal basis in ${ {\bf R}}^n$. Then in these coordinates $$x=(0,d)_,\;\;\;\hat{x}_1=(0,d_1),\;\;\;y=(0,2d),$$ where $0 \in { {\bf R}}^{n-1}$. We also have $x_1=(x'_1,d_1),$ where $x'_1 \in { {\bf R}}^{n-1}$. Let $y_1 \in \partial
\Omega$ be such that $|x_1-y_1|=d(x_1)$. Let the coordinates of $y_1$ be $(y'_1,y_{1,n})$ where $y'_1 \in { {\bf R}}^{n-1}$, $ y_{1,n} \in
{ {\bf R}}^1$. Then $$\label{1.3}
|x_1-y_1|\leq|x_1-y|$$ or $$\label{1.4}
|x'_1-y'_1|^2+|d_1-y_{1,n} |^2\leq|x'_1|^2+|d_1-2d|^2$$
\[Claim 1\] There exists $M_1$, depending on $n$, such that for any $ \delta \in (0,2)$ the following is true: $$\label{claimCond}
\mbox{if} \;\;d>\frac{M_1}{\delta^2}\varepsilon\;\;
\mbox{then} \;\;\;|y'_1|<\delta d.$$
Note that $$\label{1.41}
|x'_1| < \varepsilon, \;\;\;\; |y'_1|<2d.$$ Indeed, the first inequality is true since $x_1 \in B_{\varepsilon}(x)$. To prove the second inequality of (\[1.41\]) we use (\[1.4\]) and (\[1.2\]) to get $$|y'_1- x'_1|^2 \leq \varepsilon^2+|d+\varepsilon|^2
<(d+\sqrt{2}\varepsilon)^2,$$ so $$\label{1.5}
|y'_1|\leq |x'_1|+d+\sqrt{2}\varepsilon<d+3\varepsilon<2d$$ since $d\geq 10\varepsilon$. Thus (\[1.41\]) is proved.
Suppose that the assertion of Claim \[Claim 1\] is false, that is $$\label{wrongIneq}
|y'_1|\geq \delta d.$$ By (\[claimCond\]) $$\label{deltaD}
\delta d > {M_1 \over \delta}\varepsilon > 2\varepsilon$$ if $M_1 > 4$. Then we get from (\[1.4\]), (\[1.41\]), (\[1.2\]), (\[wrongIneq\]) $$\begin{aligned}
(y_{1,n} -d_1)^2 & \leq &
\varepsilon^2+(2d-d_1)^2-(|y'_1|-|x'_1|)^2
\nonumber \\
& \leq &
\varepsilon^2+({11 \over 10}d)^2-(\delta d-\varepsilon)^2
\label{1.6} \\
& < & 4d^2(1-
\frac{\delta^2}{4})-2d^2+2\delta d\varepsilon
\nonumber \\
& \leq &
4d^2(1-
\frac{\delta^2}{4})
\nonumber\end{aligned}$$ The last inequality follows from $-2d^2+2\delta d\varepsilon < 0$ which holds by (\[claimCond\]) if $\delta < 2$ and $ M_1>8$.
Consider two cases:\
:$\;\; y_{1,n} \geq d_1$. Denote by $\alpha$ the angle between vectors $y_1 - x_1$ and $(x'_1,y_{1,n})-x_1$. In the triangle $x_1, y_1,\;(x'_1,y_{1,n})$ the side $(x'_1,y_{1,n})-
y_1$ is orthogonal to $x_1-(x'_1,y_{1,n})$, since $(x'_1,y_{1,n})-
y_1=(x'_1-y'_1,0)$, and $x_1-(x'_1,y_{1,n})=(0,d_1-y_{1,n})$. Thus we get $$\tan\alpha=\frac{|y'_1-x'_1|}{y_{1,n} - d_1}.$$ But $$|y'_1-x'_1|\geq||y'_1|-|x'_1||=|\delta d-
\varepsilon|=\delta d-\varepsilon$$ by (\[deltaD\]). By (\[1.6\]) we get $$\tan\alpha \geq \frac{\delta d-\varepsilon}{2d\sqrt{1-
\frac{\delta^2}{4}}}=\frac{\delta}{2\sqrt{1-
\frac{\delta^2}{4}}}(1-{1\over \delta}\cdot{\varepsilon \over d})\geq
\frac{\delta}{4\sqrt{1-\frac{\delta^2}{4}}}.$$ where the last inequality follows from (\[deltaD\]). Thus, using the condition $\delta \in (0, 2)$ we get $$\label{1.7}
|\cos\alpha|=\frac{1}{\sqrt{1+\tan^2\alpha}}\leq
\sqrt{\frac{4(4-\delta^2)}{16-
3\delta^2}} \leq 1-
{\delta^2\over {2(16 - 3\delta^2)}}.$$
Consider the triangle $x_1,y_1,\;(x'_1,0)$. The angle between the vectors $(x'_1,0) - x_1$ and $y_1-x_1$ is $\pi-
\alpha$, and so we get $$\label{ThCos}
|y_1-(x'_1,0)|^2=|x_1-(x'_1,0)|^2+|y_1-x_1|^2-2|x_1-(x'_1,0)||y_1-
x_1|\cos(\pi-\alpha).$$ Since $x_1=(x'_1, d_1)$, we have $$|x_1-(x'_1,0)|=d_1.$$ From (\[1.4\]) we get $$|y_1-x_1|\leq\sqrt{\varepsilon^2+(2d-d_1)^2}.$$ Then by (\[ThCos\]), (\[1.7\]) we get $$\begin{aligned}
|y_1-(x'_1,0)|^2
& \leq &
d^2_1 +[\varepsilon^2+(2d-
d_1)^2]+2d_1\sqrt{\varepsilon^2+(2d-d_1)^2}\,(1-{\delta^2 \over 32})
\nonumber \\
& \leq &
d^2_1+(2d-d_1)^2+2d_1(2d-d_1)(1+\frac{\varepsilon^2}{2(2d-
d_1)^2})(1-{\delta^2 \over 32})+\varepsilon^2
\nonumber \\
& \leq &
4d^2+2d_1\frac{\varepsilon^2(1-{\delta^2\over 32})}{2(2d-
d_1)}+\varepsilon^2-2d_1(2d-d_1){\delta^2 \over 32}
\nonumber \\
& \leq &
4d^2+d_1\frac{\varepsilon^2}{2d-d_1}+\varepsilon^2-{1\over
16}d_1(2d-d_1)\delta^2.
\nonumber\end{aligned}$$ Thus we see that $$\label{mainClaimEst}
|y_1-(x'_1,0)|^2 < (2d-\varepsilon)^2,$$ provided $$\label{lastInq}
d_1\frac{\varepsilon^2}{2d-d_1}-{1\over 16}d_1(2d-
d_1)\delta^2<-4d\varepsilon.$$ Using (\[1.2\]) we see that (\[lastInq\]) is satisfied if $$\label{nearContr}
d^2\delta^2>C(\varepsilon^2+d\varepsilon)$$ where $C$ is a large enough constant. In its turn, (\[nearContr\]) is true if $d>{M_1\over \delta^2}\varepsilon$, where $M_1$ is large enough. Thus, for such $d$, the inequality (\[mainClaimEst\]) holds, and so $$|y_1-(x'_1,0)|<2d-\varepsilon.$$ Then $$|y_1|\leq|x'_1|+|y_1-(x'_1,0)|<2d=|y|.$$ This contradicts the fact that $y \in {\cal N}_{\partial \Omega}(O)$. Thus (\[wrongIneq\]) is false. Thus Case 1 of Claim \[Claim 1\] is proved.
$y_{1,n} \leq d_1.$
Using this condition, we get $$\begin{aligned}
|(x'_1,0)-y_1|^2 & = &
|(x'_1,0)-x_1+x_1-y_1|^2=|-(0,d_1)+(x'_1-
y'_1,\;d_1-y_{1,n})|^2
\nonumber \\
& = & d^2_1+|x_1-y_1|^2+2(y_{1,n}
-d_1)d_1
\nonumber \\
& \leq & d^2_1+|x_1-y_1|^2\leq
d^2_1 +\varepsilon^2+(2d-d_1)^2
\nonumber \\
& = & 4d^2-2d_1(2d-d_1)+\varepsilon^2<(2d-\varepsilon)^2,
\nonumber \\
& & \mbox{ if }\ d_1(2d-d_1)>2d\varepsilon.
\nonumber \end{aligned}$$ The last inequality is in fact true by (\[1.2\]) and $d>10\varepsilon$. Then, as in the Case 1, we arrive to a contradiction. Thus Case 2 cannot happen for $d$ and $ \varepsilon$ defined in Claim \[Claim 1\]. Thus, Claim \[Claim 1\] is proved.
Now the rest of the proof of part (a) of Proposition \[Proposition 3\] follows the proof of Proposition 4.1 in [@EG].
The main step is to show that there exist constants $C$ and $ M$ depending only on $n$ such that the inequality $d>M\varepsilon$ implies $$\label{1.11}
|y-y_1|\leq C|\hat{x}_1-x_1|.$$
Let $\delta \in (0,1)$ be chosen later, and let $d>{M_1\over
\delta^2}\varepsilon$ where $M_1$ is the constant from Claim \[Claim 1\]. Then by Claim \[Claim 1\] we get $$\label{RightIneq}
|y'_1|<\delta d.$$ Since $y\in {\cal N}_{\partial \Omega}(O)$, we get $|y_1|>|y|=2d$, and so $$|y'_1|^2+|y_{1,n} |^2\geq
4d^2.$$ Thus we estimate using (\[RightIneq\]) $$\begin{aligned}
y_{1,n} & \geq & \sqrt{4d^2-|y'_1|^2} = 2d\sqrt{1-
\frac{|y'_1|^2}{4d^2}}\geq 2d(1-\frac{|y'_1|^2}{8d^2}-C{|y'_1|^4
\over {d^4}})
\nonumber \\
\label{1.899}
& \geq & 2d-\frac{|y'_1|^2}{4d}-C\frac{|y'_1|^4}{4d^3}\geq 2d-
\frac{|y'_1|^2}{4d}(1+C\delta^2),\end{aligned}$$ where all constants $C$ do not depend on $y'_1$ and $d$. So $$0<2d-d_1=y_{1,n} -d_1+2d-y_{1,n} \leq (y_{1,n} -
d_1)+\frac{|y'_1|^2}{4d}(1+C\delta^2).$$ Now we can estimate $$\begin{aligned}
(2d-d_1)^2 & \leq & (y_{1,n} -d_1)^2+(y_{1,n} -
d_1)\frac{|y'_1|^2}{2d}(1+C\delta^2)+\frac{|y'_1|^4}{16{d^2}}
(1+C\delta^2)
\nonumber \\
\label{1.9}
& \leq & (y_{1,n} -d_1)^2+(y_{1,n} -
d_1)\frac{|y'_1|^2}{2d}(1+C\delta^2)+C\delta^2 |y'_1|^2\end{aligned}$$ where we again used (\[RightIneq\]). Using (\[1.4\]), (\[1.2\]) and (\[1.41\]) we get the following estimate $$\begin{aligned}
|y_{1,n} -d_1| & \leq & \sqrt{\varepsilon^2 +(2d-d_1)^2}=(2d-
d_1)\sqrt{1+\frac{\varepsilon^2}{(2d-d_1)^2}}
\nonumber \\
& \leq & (2d-
d_1)(1+\frac{\varepsilon^2}{2(2d-d_1)^2})
\leq(2d-d_1)(1+C\frac{\varepsilon^2}{d^2})
\nonumber \\
&\leq & (2d-
d_1)(1+C\delta^4)\;\;\;\mbox{since}\;\;d\geq
\frac{M_1}{\delta^2}\varepsilon. \nonumber\end{aligned}$$ Thus, by (\[1.9\]): $$\begin{aligned}
(2d-d_1)^2 & \leq & (y_{1,n} -d_1)^2+
(2d-d_1)\frac{|y'_1|^2}{2d}(1+C\delta^4)+C|y'_1|^2\delta^2
\nonumber \\
& \leq & (y_{1,n} -
d_1)^2+\Theta|y'_1|^2,
\label{1.10}\end{aligned}$$ where $$\Theta=(1-\frac{d_1}{2d})(1+C\delta^4)+C\delta^2.$$ Using (\[1.2\]) we estimate $$\Theta \leq{11\over
20}(1+C\delta^4)+C\delta^2<1,$$ if $\delta>0$ is small enough. Fix such $\delta$. Let $M_1$ be the constant from Claim \[Claim 1\] defined by this $\delta$. Let $M=\frac{M_1}{\delta^2}$ be our choice of the constant $M$ in the inequalities (\[ptsAssm\]). Then the inequalities (\[ptsAssm\]) imply (\[1.10\]) with $\Theta<1$. Thus (\[1.10\]) and (\[1.4\]) give: $$\label{1.10.2}
|y'_1-x'_1|^2\leq |x'_1|^2+\Theta |y'_1|^2.$$ Now we estimate: $$\mid y'_1 \mid \leq \mid y'_1- x'_1 \mid + \mid x'_1 \mid$$ and from this, using (\[1.10.2\]), we get $$\begin{aligned}
\mid y'_1 \mid^2 & \leq & \mid y'_1- x'_1 \mid^2 +
2 \mid y'_1- x'_1 \mid \mid x'_1 \mid + \mid x'_1 \mid^2
\nonumber \\
& \leq & (1+\mu) \mid y'_1- x'_1 \mid^2 + (1+\frac{C}{\mu})
\mid x'_1 \mid^2
\nonumber \\
& \leq & (1+\mu)( |x'_1|^2+\Theta |y'_1|^2)+
(1+\frac{C}{\mu}) \mid x'_1 \mid^2.
\nonumber\end{aligned}$$ Choose $\mu>0$ so small that $(1+\mu)\Theta<1$, then we get $$\label{1.10.3}
\mid y'_1 \mid \leq C \mid x'_1 \mid=C\mid x_1 - \hat{x}_1\mid.$$
It remains to estimate $\mid y_{1,n} - y_n \mid$ where $y_n=2d$.
From (\[1.899\]) we get $$\label{1.10.4}
2d-y_{1,n} \leq C \frac{\mid y'_1 \mid^2}{d} \leq C\mid y'_1 \mid,$$ since $\mid y'_1 \mid < \delta d$.
Since $x_1=(x'_1, d_1)$, $d(x_1)=\mid x_1 - y_1 \mid$, and $d(\hat{x}_1)) = \mid \hat{x}_1 -y \mid = 2d-d_1$, we calculate (using the fact that $\mbox{Lip}[d(\cdot)]=1$) $$y_{1,n} -d_1 \leq \mid y_1-x_1 \mid =
2d-d_1+ d(x_1) - d(\hat{x}_1) \leq
2d-d_1+\mid x_1- \hat{x}_1 \mid$$ and so $$\label{1.10.5}
y_{1,n} - 2d \leq \mid x_1- \hat{x}_1 \mid.$$ The estimates (\[1.10.3\]), (\[1.10.4\]), and (\[1.10.5\]) imply (\[1.11\]).
Now since $$Dd(x_1)=\frac{x_1-y_1}{\mid x_1-y_1\mid}, \;\;\;
Dd(\hat{x}_1)=\frac{\hat{x}_1-y}{\mid \hat{x}_1-y\mid},$$ we get the following estimate $$|Dd(x_1)-Dd(\hat{x}_1)|\leq \frac{|x_1-\hat{x}_1|+|y_1-
y|}{|\hat{x}_1-y|},$$ then using the equality $|\hat{x}_1-y|=2d-d_1$ and the inequalities (\[1.2\]) and (\[1.10\]) we deduce that $$|Dd(x_1)-Dd(\hat{x}_1)|\leq {C\over d}|x_1-\hat{x}_1|.$$ Note that $Dd(\hat{x}_1)=Dd(x)$ since the points $x$ and $\hat{x}_1$ are on the same ray $R^1_x$. Now from $|\hat{x}_1-x_1|\leq |x-x_1|$, we get (\[1.1\]) in the case of (\[convCond\]). Note also that (\[1.11\]) and inequality $|\hat{x}_1-x_1|\leq |x-x_1|$ imply (\[1.1a\]).
In the case $\mid x- v \mid <
\mid x- y \mid $ we can make the following reduction to the case (\[convCond\]). Let $$\label{defQ}
Q=\mid x- y \mid -
\mid x- v \mid,$$ and let $\tilde \Omega$ be the set $ \{z \in \Omega \;\; \mid \;\; { \mbox{dist}}(z, \partial \Omega)
> Q \}$. Then $B_{\varepsilon}(x) \subset
\tilde{\Omega}$ since the function ${ \mbox{dist}}(\cdot, \partial \Omega)$ is Lipschitz with constant 1. For $z \in \tilde{\Omega} $ we have $$\label{ReductDist}
{ \mbox{dist}}(z, \partial {\Omega)}) = { \mbox{dist}}(z, \partial \tilde{\Omega}) + Q.$$ To see this consider $z_1 \in {\cal N}_{\partial \Omega}(z)$, and let $z_2$ be the point of intersection of the interval connecting $z$ and $z_1$ with $ \partial \tilde{\Omega} $. Then $z_1 \in {\cal N}_{\partial \Omega}(z_2)$ since $z_2$ lies in the relative interior of the distance ray $R_{z_1}$ that intersects $\partial \Omega$ in $z_1$. Thus $\mid z_1 - z_2 \mid = Q$. Now suppose that there exists a point $\hat{z}_2 \in \partial
\tilde{\Omega}$ such that $\mid z-\hat{z}_2\mid < \mid z-z_2\mid $. Let $\hat{z}_1 \in {\cal N}_{\partial \Omega}(\hat{z}_2)$, then $\mid \hat{z}_1 - \hat{z}_2 \mid = Q$. Then $$\mid z- \hat{z}_1 \mid \leq
\mid z - \hat{z}_2 \mid + \mid \hat{z}_2 - \hat{z}_1\mid < \mid z-z_2 \mid +
Q =
\mid z - z_1 \mid,$$ a contradiction to the fact that $z_1 \in {\cal N}_{\partial \Omega}(z)$. This proves (\[ReductDist\]). So the distance rays and the ridge set of the set $ \tilde{\Omega}$ are intersections of those of the set $\Omega$ with the set $ \tilde{\Omega}$.
Then in (\[1.1\]) we can consider $d(\cdot)$ as the distance to $\partial \tilde{\Omega}$, so we can consider $\tilde{\Omega}$ instead of $\Omega$, and then the inequality (\[convCond\]) is satisfied.
To show (\[1.1a\]), denote by $w$ and $w_1$ the points of intersection of $\partial \tilde{\Omega}$ with the intervals $(x,y)$ and $(x_1, y_1)$ respectively. Then it follows from (\[ReductDist\]) $w \in {\cal N}_{\partial \tilde{\Omega}}(x)$ and $w_1\in {\cal N}_{\partial \tilde{\Omega}}(x_1)$, and $$\mid w-y \mid = \mid w_1 - y_1 \mid =Q.$$ Then it follows from (\[defQ\]) that $$\mid x-v \mid = \mid x-w\mid.$$ Using (\[defQ\]) and letting $
A= \frac{ \mid x-y \mid}{ \mid x-v \mid},
$ we get $$\begin{aligned}
y &=& x+ A(w-x), \;\;\;
\nonumber \\
y_1 &=& x_1 + \left[1+(A-1)\frac{\mid x-w \mid}{\mid x_1-w_1\mid} \right]
( w_1 - x_1).
\nonumber\end{aligned}$$ The inequality (\[1.1a\]) can be applied to the set $ \tilde{\Omega}$, so $$\mid w-w_1\mid \leq C \mid x-x_1\mid.$$ Now we estimate: $$\begin{aligned}
\mid y-y_1\mid & = & \left| x-x_1 +
A\left[(w-x) - \frac{\mid x-w \mid}{\mid x_1-w_1\mid}(w_1 - x_1)\right] \right.
\nonumber \\
& & + \left.
\left(\frac{\mid x-w \mid}{\mid x_1-w_1\mid}-1\right) ( w_1 - x_1) \right|
\nonumber \\
&\leq & C(1+A)(\mid x-x_1\mid + \mid w-w_1\mid)
\nonumber \\
&\leq &
C(1+A) \mid x-x_1\mid.
\nonumber\end{aligned}$$ This implies (\[1.1a\]).
Acknowledgement {#acknowledgement .unnumbered}
===============
It is the author’s pleasure to thank L. C. Evans and R. F. Gariepy for their interest, insight and and encouragement.
[99]{}
G. Aronsson. [*Asymptotic solutions of a compression molding problem*]{}, Preprint LiTH-MATH-R-95-01, 1995, Department of Mathematics, Linkoping University, Linkoping , Sweden.
G. Aronsson, L. C. Evans. [*An asymptotic model for compression molding*]{}, Forthcoming.
G. Aronsson, L. C. Evans, Y. Wu. [*Fast/Slow diffusion and growing sandpiles*]{}, J.Diff.Equat., 131 (1996), no. 2, 304-335.
T. Bhattacharya, E. DiBenedetto, J. Manfredi. [*Limits as $p\rightarrow \infty$ of $\Delta_pu=f$ and related extremal problems*]{}, Rend. Sem. Mat. Univ. Politee, Torino, 1989.
L. C. Evans, M. Feldman, R. F. Gariepy. [*Fast/Slow diffusion and collapsing sandpiles*]{}, J.Diff.Equat., 137(1997), 166-209
L. C. Evans, W. Gangbo. [ *Differential equations methods for the Monge-Kantorovich mass transfer problem*]{}, Preprint.
L. C. Evans, R. F. Gariepy. [ *Measure theory and fine properties of functions*]{}, CRC Press, 1992.
W. D. Evans, D. J. Harris. [ *Sobolev embeddings for generalized ridged domains*]{}, Proc.London Math.Soc., 54(1987), 141-175.
H. Federer. [*Curvature measures*]{}, Trans. Am. Math. Soc., 93(1959), 418-491.
H. Federer. [*Geometric measure theory*]{}, Springer, 1969.
M. Feldman. [*Convex viscosity solutions of nonlocal geometric motion of planar curves*]{}, in preparation.
D. Gilbarg, N.Trudinger. [ *Elliptic Partial Differential Equations of second order (2nd Ed.)*]{}, Springer, 1983.
U.Janfalk. [ *On certain problems concerning the $p$-Laplace operator*]{}, Linkoping Studies in Science and Technology, Dissertation \#326, Linkoping University, Sweden, 1993.
N. Krylov. [*Nonlinear elliptic and parabolic equations of second order*]{}, Reidel, Dordrecht, 1987.
H. M. Soner. [*Motion of a set by a curvature of its boundary*]{}, J.Diff.Equat., 101(1993), 313-372.
[^1]: Research is supported in part by NSF grants DMS-9701755 (MSRI) and DMS-9623276.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Strichartz estimates are a manifestation of a dispersion phenomenon, exhibited by certain partial differential equations, which is detected by suitable Lebesgue space norms. In most cases the evolution propagator $U(t)$ is a one parameter group of unitary operators. Motivated by the importance of decay estimates in group representation theory and ergodic theory, Strichartz-type estimates seem worth investigating when $U(t)$ is replaced by a unitary representation of a non-compact Lie group, the group element playing the role of time. Since the Schrödinger group is a subgroup of the metaplectc group, the case of the metaplectic or oscillatory representation is of special interest in this connection. We prove uniform weak-type sharp estimates for matrix coefficients and Strichartz estimates for that representation. The crucial point is the choice of function spaces able to detect such a dispersive effect, which in general will depend on the given group action. The relevant function spaces here turn out to be the so-called modulation spaces from Time-frequency Analysis in Euclidean space, and Lebesgue spaces with respect to Haar measure on the metaplectic group. The proofs make use in an essential way of the covariance of the Wigner distribution with respect to the metaplectic representation.'
address: 'Dipartimento di Scienze Matematiche, Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129 Torino, Italy'
author:
- 'Alessandra Cauli, Fabio Nicola'
- Anita Tabacco
title: Strichartz estimates for the metaplectic representation
---
Introduction and statement of the main results
==============================================
Strichartz estimates represent one of the main research theme in modern Harmonic Analysis and Partial Differential Equations. The literature in this connection is growing incredibly fast, and new results are often applied to wellposedness and scattering of nonlinear PDEs, see [@Tao] and the references therein.
Maybe the simplest case is given by the free Schrödinger equation in $\mathbb{R}^n$. The corresponding propagator $U(t)=e^{it\Delta}$ is easily proved to satisfy the so-called [*dispersive estimate*]{}: $$\|U(t)\psi\|_{L^\infty}\lesssim |t|^{-n/2}\|\psi\|_{L^1}.$$ One then deduces mixed-norm estimates, known as [*Strichartz estimates*]{}, which read $$\|U(t)\psi\|_{L^q(\mathbb{R};L^r(\mathbb{R}^n))}\lesssim\|\psi\|_{L^2(\mathbb{R}^n)}$$ for $\frac{2}{q}+\frac{n}{r}=\frac{n}{2}$, $2\leq q$, $r\leq \infty$ and $(q,r,n)\not=(2,\infty,2)$.
Strichartz estimates are a manifestation of two effects: compared with the basic $L^2$-conservation law, corresponding to the pair $q=\infty,r=2$, the other pairs express
- a gain (loss) of space (time) local regularity,
- a gain (loss) of time (space) decay at infinity.
Dispersive and Strichartz estimates hold, for different ranges of exponents, and possibly with a loss of derivatives, for several classes of equations, even on manifolds, homogeneous spaces, etc. In general, the evolution propagator $U(t)$ is a strongly continuous unitary representation of the abelian group ${\mathbb{R}}$. Now, for a non-compact abelian group $G$, the [*irreducible*]{} unitary representations are one-dimensional and their matrix coefficients are just (multiples of) the group characters, with no decay at all. The above decay is in part due to a lack of “coherence” of the irreducible components of $U(t)$: frequency components move in different directions and, in some cases, with different speeds.
Motivated by the importance of decay estimates in representation theory and ergodic theory (see e.g. [@howe2; @moore] and the references therein), Strichartz-type estimates seem worth investigating for strongly continuous unitary representations $
\mu:G\to \mathcal{U}(H)
$ of a non-compact locally compact Hausdorff group $G$, where $H$ is a Hilbert space. The representation $\mu(g)$ plays now the role of the above propagator $U(t)$. Generally speaking, we are interested in estimates of the type $$\label{coeff0}
||\mu(g) \psi\|_{L^q(G; X_\theta)}\lesssim\|\psi\|_{H}$$ for some scale of Banach spaces $X_\theta$, valid for a range of pairs $(q,\theta)$.
In this note we develop this idea for the metaplectic group $G=Mp(n,{\mathbb{R}})$, that is the double covering of the symplectic group $Sp(n,{\mathbb{R}})$, and the corresponding metaplectic, or oscillatory, representation, first constructed by Segal and Shale [@26; @27] in the framework of quantum mechanics (see also van Hove [@38]) and by Weil [@39] in number theory. This is a strongly continuous unitary representation of $Mp(n,{\mathbb{R}})$ in $L^2({\mathbb{R}^n})$, which turns out to be faithful, so that we can think of $Mp(n,{\mathbb{R}})$ as a subgroup of $\mathcal{U}(L^2({\mathbb{R}^n}))$, and the representation given just by the inclusion. Following [@DeGosson] we will therefore denote by $\widehat{S}$ a metaplectic operator and by $S=\pi(\widehat{S})\in Sp(n,{\mathbb{R}})$ its projection in the symplectic group (the construction of the metaplectic representation is briefly recalled in Section \[sec2\] below).
Now, it turns out that the operator $e^{it\Delta}$ is a particular metaplectic operator, so that a natural candidate for the spaces $X_\theta$ in would seem to be the Lebesgue spaces. However, the Fourier transform is itself a metaplectic operator, and therefore we should actually look for spaces invariant with respect to the action of the Fourier transform. $U(n)$-invariance (see Section \[sec4\]) finally suggests, as right function spaces, the modulation spaces $M^p$, widely used in Time-frequency Analysis [@DeGosson; @Grochenig].
In short, for a given Schwartz function $\varphi\in\mathcal{S}({\mathbb{R}^n})\setminus\{0\}$, consider the time-frequency shifts $\varphi_{z}(y)=e^{i\xi\cdot y}\varphi(y-x)$, $z=(x,\xi)\in{\mathbb{R}^n}\times{\mathbb{R}^n}$. Then for $1\leq p\leq\infty$ we define the $M^p$ norm of $\psi\in\mathcal{S}'({\mathbb{R}^n})$, as $$\|\psi\|_{M^p}=\Big(\int_{{\mathbb{R}^{2n}}} | \langle \psi,\varphi_z\rangle |^p\, dz\Big)^{1/p}$$ (with obvious changes when $p=\infty$). Different windows $\varphi$ give equivalent norms. We have $\mathcal{S}({\mathbb{R}^n})\subset M^p\subset \mathcal{S}'({\mathbb{R}^n})$ for every $1\leq p\leq \infty$, $M^2=L^2({\mathbb{R}^n})$, $M^p\subset M^q$ if and only if $p\leq q$, $(M^p)'=M^{p'}$ if $p<\infty$. Modulation space norms measure the phase space concentration of a function; roughly speaking we can think of a function in $M^p$ as a function having $L^p$ decay at infinity and $F L^p$ local regularity. Let us also observe that modulation spaces have been recently applied in PDEs by several authors, see e.g. [@cordero3; @cordero1; @cordero2; @ruzhansky; @wang] and the references therein (some of their properties are collected in Section \[sec2\]).
We follow the usual pattern, namely we begin with a dispersive-type estimate.
\[mainthm\] The following estimate holds: $$\label{dispersiva}
\|\widehat{S}\psi\|_{M^{\infty}}\lesssim (\lambda_{1}(S)\ldots\lambda_{n}(S))^{-1/2}\|\psi\|_{M^1}$$ for $\widehat{S}\in Mp(n,\mathbb{R}),\ \psi\in\mathcal{S}({\mathbb{R}^n})$, where $\lambda_{1}(S),\ldots,\lambda_{n}(S)$ are the singular values $\geq 1$ of $S=\pi(\widehat{S})\in Sp(n,\mathbb{R})$.
The result is sharp as far as the decay is concerned (see Section \[sec4\]).
As a consequence we can obtain the following estimates on matrix coefficients.
\[mainthm2\] Let $G=Mp(n,{\mathbb{R}})$ with the Haar measure. The following estimate holds: $$\label{coeff}
\|\langle \widehat{S}\varphi_1,\varphi_2\rangle\|_{L^{4n,\infty}(G)}\lesssim\|\varphi_1\|_{M^1}\|\varphi_2\|_{M^1},$$ for $\varphi_1,\varphi_2\in\mathcal{S}({\mathbb{R}^n})$.
Here $L^{4n,\infty}$ is the weak-type $L^{4n}$ space on $G=Mp(n,\mathbb{R})$.
Corollary \[mainthm2\] refines a result by Howe [@howe], who proved that for fixed $\varphi_1,\varphi_2\in \mathcal{S}(\mathbb{R}^n)$ the matrix coefficient in is in $L^{4n+\epsilon}$ for every $\epsilon>0$ but in general not in $L^{4n}$. In fact, estimates for matrix coefficients have a long tradition in representation theory, see for example [@cowling2; @cowling; @howe; @howe2; @oh] and the references therein. Usually, dealing with a unitary representation of a group $G$ in a Hilbert space $H$, one takes $\varphi_1,\varphi_2$ in $K$-invariant finite dimensional subspaces of $H$, $K\subset G$ being a maximal compact subgroup, and the constants in the estimates will depend on the dimension of such subspaces. Sometimes this finiteness condition is replaced by taking $\varphi_1,\varphi_2$ in higher order Sobolev-type spaces, and often an $\epsilon$-loss in the decay appears, as above (see e.g. [@moore]). On the contrary, in we have the low regularity space $M^1$, and functions in $M^1$ do not need to have any differentiability, even in a fractional sense.
Weak-type estimates for matrix coefficients such as seem of great interest in their own right; for example, they could play a key role in extending Cowling’s strengthened version of the Kunze-Stein phenomenon [@cowling3] to groups of rank higher than 1.
As a consequence of the dispersive estimates we therefore obtain the following Strichartz-type estimates.
\[mainthm3\] Let $G=Mp(n,{\mathbb{R}})$ with the Haar measure. The following estimates hold: $$\|\widehat{S}\psi\|_{L^q(G;M^r)}\lesssim\|\psi\|_{L^2},$$ for $$\frac{4n}{q}+\frac{1}{r}\leq \frac{1}{2},\quad 2\leq q,r\leq \infty.$$
The range of admissible pairs $(q,r)$ in Theorem \[mainthm3\] is represented in Figure \[figura\], which also shows a comparison with the case of the Schrödinger group (as already observed, the one-parameter group $e^{it\Delta}$ is a subgroup of $Mp(n,{\mathbb{R}})$). Notice however that the exponent $r$ refers to different function spaces; in fact we have $L^r\subset M^r$ for $2\leq r\leq\infty$, with strict inclusion when $r>2$. As one can see, the admissibility condition implies $q\geq 8n$. Also, we have a whole region of admissible pairs, and not just a segment, because the modulation spaces $M^r$ are nested, unlike the Lebesgue spaces. Let us observe that, compared with the trivial estimate for $q=\infty,r=2$, again the other admissible pairs $(q,r)$ represent a gain (loss) of time (space) decay at infinity. Instead, we do no longer have any smoothing effect, as expected: among the metaplectic operators we also meet linear changes of variables, which do not produce smoothing in any reasonable space. This is in turn related to the fact that $M^1\subset M^\infty$ in .
\[figura\]
(0,0) – (5.5,0) node\[above right\] [$1/q$]{}; (0,0) – (0,4.5) node\[right\] [$1/r$]{}; (0,3) node \[left\][$\frac{1}{2}$]{} – (4,1); (0,3) – (2,0) node \[below\][$\frac{1}{8n}$]{}; (4,1) – (4,0) node \[below\][$\frac{1}{2}$]{}; (0,0) – (2,0) – (0,3); (2,2) – (4,3.2) node \[right\][Schrödinger equation]{}; (0,1) node\[left\][$\frac{n-2}{2n}$]{} – (4,1); (1,0.5) – (3.3,1.8) node \[right\][metaplectic representation]{}; at (9,4.5) [(Lebesgue/modulation space exponent)]{};
Let us observe that similar estimates seem worth investigating for other unitary representations, e.g. the oscillatory representation restricted to subgroups of ${Mp}(n,{\mathbb{R}})$ (cf. [@cordero01; @cordero02; @cordero03; @cordero04]), unitary representations of linear Lie groups such as ${SL}(n,{\mathbb{R}})$ or more general semisimple Lie groups, where the Cartan decomposition should play the role of our singular value analysis. Part of the problem is to identify low regularity spaces strictly tailored to the given representation, playing the role of the modulation spaces used here. We plan to carry on this investigation in future work.
The paper is organized as follows. In Section \[sec2\] we recall some preliminary results on time-frequency and symplectic methods used in the proofs of the main results. That material is mainly extracted from [@DeGosson]. Section \[sec3\] is devoted to the proof of the above results. Finally in Section \[sec4\] we collected some concluding remarks.
Preliminaries {#sec2}
=============
We recall here a number of definitions and results that we will use in the following. We refer to [@DeGosson; @helgason; @leray] for details.
Notation
--------
We denote by $\langle\cdot,\cdot\rangle$ the inner product in $L^2({\mathbb{R}^n})$, linear in the first argument. The notation $A\lesssim B$, for expressions $A,B\geq0$, means $A\leq C B$ for a constant $C$ depending only on the dimension $n$ and parameters which are fixed in the context. We also write $A\asymp B$ for $A\lesssim B$ and $B\lesssim A$.
The symplectic group {#sec2.1}
--------------------
The symplectic group $Sp(n,\mathbb{R})$ is the group of $2n\times 2n$ real matrices $S$ such that $S^T JS=J$, where $$J=\begin{pmatrix}
0& I\\
-I&0
\end{pmatrix}.$$ We recall that every symplectic matrix $S$ admits a unique polar decomposition $S=S_0U$ where $S_0$ is symplectic, symmetric and positive definite and $U$ is a symplectic rotation, i.e. belongs to $$U(2n,\mathbb{R}):=Sp(n,{\mathbb{R}})\cap O(2n,{\mathbb{R}})\simeq U(n).$$ A positive definite matrix can always be diagonalized using an orthogonal matrix. When this matrix is in addition symplectic we can use a symplectic rotation to perform this diagonalization: if $S$ is positive definite, there exists $U\in U(2n,\mathbb{R})$ such that $S=U^TDU$ where $$D={\rm diag}(\lambda_1,\ldots,\lambda_n,\lambda^{-1}_1,\ldots,\lambda^{-1}_n)$$ and $\lambda_1\geq \ldots\geq \lambda_n\geq \lambda_n^{-1}\geq \ldots \geq \lambda_1^{-1}>0$ are the eigenvalues of $S$.
By combining polar decomposition and this diagonalization result we see that every symplectic matrix $S$ can be written as $$S=U_1 D U_2$$ with $U_1,U_2\in U(2n,\mathbb{R})$ and $D$ diagonal as above, where $\lambda_1\geq \ldots\geq \lambda_n\geq \lambda_n^{-1}\geq \ldots \geq \lambda_1^{-1}>0$ are now the singular values of $S$.
Integration on the symplectic group {#integration-on-the-symplectic-group .unnumbered}
-----------------------------------
$Sp(n,\mathbb{R})$ turns out to be a unimodular Lie group. The following integration formula for $U(2n,\mathbb{R})$-bi-invariant functions on $Sp(n,\mathbb{R})$ will be crucial in the following.
Recall that $f:Sp(n,\mathbb{R})\to \mathbb{C}$ is called $U(2n,\mathbb{R})$-bi-invariant if $f(U_1 S U_2)=f(S)$ for every $S\in Sp(n,\mathbb{R})$, $U_1,U_2\in U(2n,{\mathbb{R}})$.
Consider the Abelian subgroup $A=\{a_t\}$ of $Sp(n,\mathbb{R})$ given by $$a_t=\left(\begin{matrix}e^{\frac{t}{2}} & 0 \\ 0 & e^{-\frac{t}{2}}\end{matrix}\right),\quad t={\rm diag}(t_1,\ldots,t_n),\ (t_1,\ldots,t_n)\in{\mathbb{R}^n}.$$ If $f$ is a $U(2n,{\mathbb{R}})$-bi-invariant function on $Sp(n,\mathbb{R})$, its integral with respect to the Haar measure is given by $$\begin{gathered}
\label{integrale}
\int_{Sp(n,\mathbb{R})} f(S)dS\\
=C\int_{t_1\geq\ldots \geq t_n\geq 0}f(a_t)\prod_{i<j}\sinh\frac{t_i-t_j}{2}\prod_{i\leq j}\sinh\frac{t_i+t_j}{2}\,dt_1\ldots dt_n
\end{gathered}$$ for some constant $C>0$.
Weak-type Young inequality on unimodular groups {#weak-type-young-inequality-on-unimodular-groups .unnumbered}
-----------------------------------------------
We will also need the Young inequality for weak type spaces, which reads as follows.
On a measure space $X$, for $0< p<\infty$ the weak-type Lebesgue space $L^{p,\infty}(X)$ is defined as the space of measurable functions $f:X\to\mathbb{C}$ such that $$\|f\|_{L^{p,\infty}}:=\sup\limits_{\lambda>0} \{\lambda \cdot ({\rm meas}\{x:\,|f(x)|\geq \lambda\})^{1/p}\}<\infty.$$ Let now $G$ be a unimodular locally compact Hausdorff group. Let $$1<p,q,r<\infty,\quad
\frac{1}{p}+\frac{1}{r}=\frac{1}{q}+1.$$ Then there exists a constant $C_{p,q,r}>0$ such that for all $f$ in $L^p(G)$ and $g$ in $L^{r,\infty}(G)$ we have $$\label{wyoung}
\|f\ast g\|_{L^{q}(G)}\leq C_{p,q,r}\|g\|_{L^{r,\infty}(G)}\|f\|_{L^p(G)}.$$
The metaplectic group {#sec2.2}
---------------------
There are many construction of the metaplectic group $Mp(n,\mathbb{R})$, i.e. the double covering of the symplectic group $Sp(n,\mathbb{R})$, and the metaplectic representation $Mp(n,\mathbb{R})\rightarrow \mathcal{U}(L^2({\mathbb{R}^n}))$. Since it turns out to be a faithful representation, we can in fact think of group elements as unitary operators themselves. This is the point of view of the following construction, where $Mp(n,\mathbb{R})$ is defined as a subgroup of the unitary group $\mathcal{U}(L^2({\mathbb{R}^n}))$ and the corresponding representation is just the inclusion. The difficult point is to prove the existence of a projection $$\pi: Mp(n,\mathbb{R})\to Sp(n,\mathbb{R})$$ which makes $Mp(n,\mathbb{R})$ the double covering of $Sp(n,\mathbb{R})$.
We recall here the main points of the construction, and we refer to [@DeGosson] and [@leray] for details.
It can be proved that the symplectic group $Sp(n,\mathbb{R})$ is generated by the so-called free symplectic matrices $$S=\left(\begin{matrix}A & B \\ C & D\end{matrix}\right)\in Sp(n,\mathbb{R}),\quad \det B\neq 0.$$ To each such matrix we associate the generating function $$W(x,x')=\frac{1}{2}DB^{-1}x\cdot x-B^{-1}x\cdot x'+\frac{1}{2}B^{-1}Ax'\cdot x'.$$ Conversely, to every polynomial of the type $$W(x,x')=\frac{1}{2}Px\cdot x-Lx\cdot x'+\frac{1}{2}Qx'\cdot x'$$ with $$P=P^T, Q=Q^T$$ and $$\det L\neq 0$$ we can associate a free symplectic matrix, namely $$S_W=\left(\begin{matrix}L^{-1}Q & L^{-1} \\ PL^{-1}Q-L^T & PL^{-1}\end{matrix}\right).$$ Now, given $S_W$ as above and $m\in\mathbb{Z}$ such that $$m\pi \equiv {\rm arg}\,\det L\quad {\rm mod}\, 2\pi,$$ we define the operator $\widehat{S}_{W,m}$ by setting, for $\psi \in \mathcal{S}(\mathbb{R}^n)$, $$\widehat{S}_{W,m}\psi(x)=\frac{1}{(2\pi i)^{n/2}}\Delta(W)\int_{\mathbb{R}^n}e^{iW(x,x')}\psi(x')dx'$$ (to be clear, $(2\pi i)^{n/2}=(2\pi)^{n/2}e^{i\pi n/4}$) where $$\Delta(W)=i^m\sqrt{|\det L|}.$$ The operator $\widehat{S}_{W,m}$ is called a [*quadratic Fourier transform*]{} associated to the free symplectic matrix $S_W$. The class modulo 4 of the integer $m$ is called [*Maslov index*]{} of $\widehat{S}_{W,m}$. Observe that if $m$ is one choice of Maslov index, then $m+2$ is another equally good choice: hence to each function $W$ we associate two operators, namely $\widehat{S}_{W,m}$ and $\widehat{S}_{W,m+2}=-\widehat{S}_{W,m}$.
The quadratic Fourier transform corresponding to the choices $S_W=J$ and $m=0$ is denoted by $\widehat{J}$. The generating function of $J$ being simply $W(x,x')=-x\cdot x'$, it follows that $$\widehat{J}\psi(x)=\frac{1}{(2\pi i)^{n/2}}\int_{\mathbb{R}^n}e^{-ix\cdot x'}\psi(x')dx'=\frac{1}{i^{n/2}}F\psi(x)$$ for $\psi\in \mathcal{S}(\mathbb{R}^n)$, where $F$ is the usual unitary Fourier transform.
The quadratic Fourier transforms $\widehat{S}_{W,m}$ form a subset of the group $\mathcal{U}(L^2(\mathbb{R}^n))$ of unitary operators acting on $L^2(\mathbb{R}^n)$, which is closed under the operation of inversion and they generate a subgroup of $\mathcal{U}(L^2(\mathbb{R}^n))$ which is, by definition, the metaplectic group $Mp(n,\mathbb{R})$. The elements of $Mp(n,\mathbb{R})$ are called metaplectic operators.
Every $\widehat{S}\in Mp(n,\mathbb{R})$ is thus, by definition, a product $$\widehat{S}_{W_1,m_1}\ldots\widehat{S}_{W_k,m_k}$$ of metaplectic operators associated to free symplectic matrices.
In fact, it can be proved that every $\widehat{S}\in Mp(n,\mathbb{R})$ can be written as a product of exactly two quadratic Fourier transforms: $\widehat{S}=\widehat{S}_{W,m}\widehat{S}_{W',m'}$. Now, it can be proved that the map $$\widehat{S}_{W,m}\longmapsto S_W$$ extends to a group homomorphism $$\pi: Mp(n,\mathbb{R})\to Sp(n,\mathbb{R}),$$ which is in fact a double covering.
We also observe that each metaplectic operator is, by construction, a unitary operator in $L^2({\mathbb{R}^n})$, but also an authomorphism of $\mathcal{S}({\mathbb{R}^n})$ and of $\mathcal{S}'({\mathbb{R}^n})$.
Modulation spaces {#sec2.3}
-----------------
Fix a window function $\varphi\in \mathcal{S}(\mathbb{R}^n)\setminus\{0\}$. The short-time Fourier transform (STFT) of a function/temperate distribution $\psi\in \mathcal{S'}(\mathbb{R}^n)$ with respect to $\varphi$ is defined by $$V_\varphi \psi(x,\xi)=(2\pi)^{-n}\int_{\mathbb{R}^n}e^{-i\xi \cdot y}\psi(y)\overline{\varphi(y-x)}dy, \quad x,\xi \in \mathbb{R}^n.$$ For $1\leq p, q\leq \infty$ and a Schwartz function $\varphi\in \mathcal{S}(\mathbb{R}^n)\setminus\{0\}$, the modulation space $M^{p,q}(\mathbb{R}^n)$ is defined as the space of $\psi\in \mathcal{S'}(\mathbb{R}^n)$ such that $$\|\psi\|_{M^{p,q}}:=\Big(\int_{\mathbb{R}^{n}}\Big(\int_{\mathbb{R}^{n}}|V_\varphi \psi(x,\xi)|^p dx\Big)^{q/p} d\xi \Big)^{1/q} < \infty,$$ with obvious changes if $p=\infty$ or $q=\infty$.
If $p=q$, then we write $M^{p}$ instead of $M^{p,p}$.
We will also need a variant, sometimes called Wiener amalgam norm in the literature, defined by $$\|\psi\|_{W(FL^p, L^q)}:=(\int_{\mathbb{R}^{n}}\Big(\int_{\mathbb{R}^{n}}|V_\varphi \psi(x,\xi)|^p d\xi\Big)^{q/p} dx \Big)^{1/q},$$ where the Lebesgue norms appear in the inverse order. Both these norms provide a measure of the time-frequency concentration of a function and are widely used in Time-frequency Analysis [@DeGosson; @Grochenig].
We have $M^{p_1,q_1}\subseteq M^{p_2,q_2}$ if and only if $p_1\leq p_2$ and $q_1\leq q_2$. Similarly $W(F L^{p_1},L^{q_1})\subseteq W(F L^{p_2},L^{q_2})$ if and only if $p_1\leq p_2$ and $q_1\leq q_2$.
The duality goes as expected: $$(M^{p,q})'= M^{p',q'},\quad 1\leq p,q<\infty,$$ and in particular $$\label{dualita}
|\langle f,g\rangle|\lesssim \|f\|_{M^p}\|g\|_{M^{p'}}.$$ In the dispersive estimates we meet, in particular, the Gelfand triple $$M^1\subset L^2({\mathbb{R}^n})\subset M^\infty.$$ We observe that $$\mathcal{S}({\mathbb{R}^n})\subset M^1\subset L^2({\mathbb{R}^n})$$ with dense and strict inclusions. For atomic characterizations of the space $M^1$ we refer to [@DeGosson; @Grochenig].
We will also use the complex interpolation theory for modulation spaces, which reads as follows: for $1\leq p,q,p_i,q_i\leq \infty$, $i=0,1$, $0\leq \theta\leq 1$, $$\frac{1}{p}=\frac{1-\vartheta}{p_0}+\frac{\vartheta}{p_1},\quad \frac{1}{q}=\frac{1-\vartheta}{q_0}+\frac{\vartheta}{q_1},$$ we have $$(M^{p_0,q_0},M^{p_1,q_1})_{\vartheta}=M^{p,q}.$$
The Wigner distribution {#sec2.4}
-----------------------
We now introduce a quadratic time-frequency distribution which will play a key role in the following. Again it represents a basic tool in the analysis of signals [@Grochenig] and in phase space Quantum Mechanics [@DeGosson; @DeGosson2]. We refer to [@DeGosson; @DeGosson2] for details.
The cross-Wigner distribution $W(\psi,\varphi)$ of functions $\psi,\varphi\in L^2(\mathbb{R}^n)$ is defined to be $$W(\psi,\varphi)(x,\xi)=(2\pi)^{-n}\int_{\mathbb{R}^n}e^{- i\xi \cdot y}\psi\Big(x+\frac{y}{2}\Big)\overline{\varphi\Big(x-\frac{y}{2}\Big)}dy.$$ We also set $W\psi=W(\psi,\psi)$.
We recall the important [*Moyal identity*]{} (see e.g. [@DeGosson Theorem 182]): $$\label{moyal}
\langle W\psi,W\varphi\rangle_{L^2(\mathbb{R}^{2n})}=(2\pi)^{-n} |\langle \psi,\varphi\rangle|^2.$$ We will also need the following estimates.
\[pro0\] We have $$\label{eq0}
\|W(\psi,\varphi)\|_{L^1({\mathbb{R}^{2n}})}\lesssim \|\varphi\|_{M^1} \|\psi\|_{M^1},$$ $$\label{eq1}
\int_{{\mathbb{R}^n}} \sup_{x\in{\mathbb{R}^n}} |W(\psi,\varphi)(x,\xi)|d\xi\lesssim \|\varphi\|_{M^1}\|\psi\|_{M^{\infty,1}}$$ and $$\label{eq2}
\int_{{\mathbb{R}^n}} \sup_{\xi\in{\mathbb{R}^n}} |W(\psi,\varphi)(x,\xi)|dx\lesssim \|\varphi\|_{M^1}\|\psi\|_{W(F L^\infty, L^1)}.$$
Formula is proved in [@DeGosson Proposition 3.6.5].
Let us prove and . It is easy to see that $$\label{eq00}
|W(\psi,\varphi)(x,\xi)|=2^n|V_\varphi \psi(2x,2\xi)|$$ so that it is sufficient to prove similar estimates with $W(\psi,\varphi)(x,\xi)$ replaced by $V_\varphi \psi(x,\xi)$. To this end we recall from [@Grochenig Lemma 11.3.3] that, for $\varphi,\varphi_0 \in \mathcal{S}(\mathbb{R}^n)$ such that $\|\varphi_0\| \neq 0$ and $\psi\in \mathcal{S'}(\mathbb{R}^n)$ we have $$\label{conv}
|V_{\varphi}\psi(x,\xi)|\lesssim \frac{1}{\|\varphi_0\|^2}(|V_{\varphi_0} \psi|*|V_{\varphi}\varphi_0 |)(x,\xi),$$ for all $(x,\xi)\in \mathbb{R}^{2n}$.
Now, we apply this inequality with a fixed Schwartz window $\varphi_0$ and we also observe that $$|V_{\varphi}\varphi_0(x,\xi)|=|V_{\varphi_0}\varphi(-x,-p)|.$$ The desired estimates for $V_{\varphi}\psi(x,\xi)$ then follow by applying the Young inequality for mixed-norm Lebesgue spaces in .
One of the most important property of the cross–Wigner distribution is its covariance with respect to the action of metaplectic operators. In fact we have (see e.g. [@DeGosson Corollary 2.17]) $$\label{covarianza}
W(\widehat{S}\psi, \widehat{S}\varphi)(z)= W(\psi,\varphi)(S^{-1}z),\quad z\in{\mathbb{R}^n}\times{\mathbb{R}^n}.$$ for every $\widehat{S}\in Mp(n,\mathbb{R})$, with projection ${S}\in Sp(n,\mathbb{R})$.
Proof of the main results {#sec3}
=========================
In this section we prove Theorems \[mainthm\], Corollary \[mainthm2\] and Theorem \[mainthm3\].
By duality it is equivalent to prove that $$| \langle \widehat{S}\psi,\varphi\rangle|\lesssim (\lambda_1(S)\ldots\lambda_n(S))^{-1/2}
\|\psi\|_{M^1}\|\varphi\|_{M^1}.$$ By the Moyal identity and the covariance property , we have $$\begin{aligned}
|\langle \widehat{S}\psi,\varphi\rangle|^2&=(2\pi)^n \langle W(\widehat{S}\psi),W\varphi\rangle_{L^2(\mathbb{R}^{2n})}
\\
&=(2\pi)^n\langle W\psi(S^{-1}\cdot),W\varphi\rangle_{L^2(\mathbb{R}^{2n})}.\end{aligned}$$ We now can write $S^{-1}=S_1U_1$ with $S_1\in Sp(n,\mathbb{R})$ positive definite and $U_1\in U(2n,\mathbb{R})$. Hence, by an orthogonal change of variable we obtain $$|\langle \widehat{S}\psi,\varphi\rangle|^2
=(2\pi)^n \langle W\psi(S_1\cdot),W\varphi(U_1^T\cdot)\rangle_{L^2({\mathbb{R}^{2n}})}.$$ We now diagonalize $S_1$, $S_1=U_2^TDU_2$ (see Section \[sec2.1\]) where $$D={\rm diag}(\lambda_1,\ldots,\lambda_n,\lambda^{-1}_1,\ldots,\lambda^{-1}_n)$$ with $\lambda_1\geq \ldots\geq \lambda_n\geq \lambda_n^{-1}\geq \ldots \geq \lambda_1^{-1}>0$ and $U_2\in U(2n,\mathbb{R})$. With a further change of variable we obtain$$|\langle \widehat{S}\psi,\varphi\rangle|^2
=(2\pi)^n\langle W\psi(U_2^TD\,\cdot),W\varphi(U_1^TU_2^T\cdot)\rangle _{L^2(\mathbb{R}^{2n})}.$$ Let $$F_1=W\psi(U_2^T\cdot)=W(\widehat{U}_2\psi),$$ and $$F_2=W\varphi(U_1^TU_2^T\cdot)=W(\widehat{U}_2\widehat{U}_1\varphi).$$ We estimate $$\begin{aligned}
\langle &W\psi(U_2^TD\,\cdot),W\varphi(U_1^TU_2^T\cdot)\rangle_{L^2({\mathbb{R}^{2n}})}
\\ &=
\int_{\mathbb{R}^{2n}} F_1(\lambda_1 x_1,\ldots,\lambda_n x_n,\lambda_1^{-1}\xi_1,\ldots,\lambda_n^{-1}\xi_n) \overline{F_2(x,\xi)} dxd\xi
\\
&\leq \int_{\mathbb{R}^{2n}} \sup_{\xi\in \mathbb{R}^n}|F_1(\lambda_1x_1,\ldots,\lambda_n x_n,\xi_1,\ldots,\xi_n)|\sup_{x\in \mathbb{R}^n}|F_2(x,\xi)|dxd\xi
\\
&=\lambda_1^{-1}\ldots \lambda_n^{-1} \int_{\mathbb{R}^{n}} \sup_{\xi\in \mathbb{R}^n}|F_1(x,\xi)|dx \int_{\mathbb{R}^n}\sup_{x\in \mathbb{R}^n}|F_2(x,\xi)|d\xi\leq
\\
&\lesssim \lambda_1^{-1}\ldots \lambda_n^{-1}\|\widehat{U}_2\psi\|_{M^1} \|\widehat{U}_2\psi\|_{ W(\mathcal{F}L^\infty,L^1) } \|\widehat{U}_2\widehat{U}_1\varphi\|_{M^1}\|\widehat{U}_2\widehat{U}_1\varphi\|_{M^{\infty,1} }
,
\end{aligned}$$ where we used, in the last line, Proposition \[pro0\].
Using the inclusions $$M^1=M^{1,1}\hookrightarrow M^{\infty,1},\quad M^1=W(\mathcal{F}L^1,L^1) \hookrightarrow W(\mathcal{F}L^\infty,L^1)$$ we continue the above estimate as $$\lesssim \lambda_1^{-1}\ldots \lambda_n^{-1}\|\widehat{U}_2\psi\|_{M^1}^2\|\widehat{U}_2\widehat{U}_1\varphi\|_{M^1}^2.$$ It is then sufficient to show that $$\label{3.2} \|\widehat{U}_2\psi\|_{M^1}\leq C\|\psi\|_{M^1}$$ and $$\label{3.1}\|\widehat{U}_2\widehat{U}_1\varphi\|_{M^1}\leq C \|\varphi\|_{M^1}$$ for a constant $C>0$ independent of $\widehat{U}_1$, $\widehat{U}_2$.
Let us verify , which implies too. By the definition of the $M^1$ norm and we have $$\|\widehat{U}_2\psi\|_{M^1}\asymp \|W(\widehat{U}_2\psi,\varphi)\|_{L^1({\mathbb{R}^{2n}})}$$ for some fixed $\varphi\in \mathcal{S}(\mathbb{R}^n)\setminus\{0\}$, which by covariance is equal to $$\|W(\psi,\widehat{U}_2^{-1}\varphi)(U_2^{-1}\cdot)\|_{L^1({\mathbb{R}^{2n}})}=\|W(\psi,\widehat{U}_2^{-1}\varphi)\|_{L^1({\mathbb{R}^{2n}})}.$$ Hence, using and the continuous embedding $\mathcal{S}(\mathbb{R}^n)\hookrightarrow M^1$ it is sufficient to prove that if $\varphi\in \mathcal{S}(\mathbb{R}^n)$ then $$\{\hat{U}\varphi:\,U\in U(2n,\mathbb{R})\}$$ is a bounded subset of $\mathcal{S}(\mathbb{R}^n)$, that is, every Schwartz seminorm is bounded on it. Since $U(2n,\mathbb{R})$ is compact it is sufficient to show that every seminorm is locally bounded, i.e. we can limit ourselves to take $U$ in a sufficiently small neighborhood for any fixed $U_0\in U(2n,\mathbb{R})$. Equivalently, we can consider $U$ of the form $U=U_1J^{-1}U_0$ where $U_1$ belongs to a sufficiently small neighborhood of $J$ in $U(2n,{\mathbb{R}})$. Now, $$\begin{aligned}
\hat{U}\varphi(x)&=\pm \widehat{U}_1[\widehat{J}^{-1}\widehat{U}_0\varphi](x)\\
&=c \sqrt{|\det L|}\int_{\mathbb{R}^n}e^{\frac{i}{2}Px\cdot x-iLx\cdot y+\frac{i}{2}Qy\cdot y}[\underbrace{\widehat{J}^{-1}\widehat{U}_0\varphi]}_{\in \mathcal{S}(\mathbb{R}^n)}(y)dy\end{aligned}$$ where $|c|=1$ and, say, $\|P\|<\epsilon$, $\|Q\|<\epsilon$, $\|L-I\|<\epsilon$. If $\epsilon<1$, it is clear that $\hat{U}\varphi$ belongs to a bounded subset of $\mathcal{S}(\mathbb{R}^n)$, as one can verify by direct inspection.
In order to prove Corollary \[mainthm2\] we need the following preliminary result.
\[pro1\] Let $\alpha>0$, $\beta>0$. Consider the function $$h(S)=(\lambda_1(S)\ldots\lambda_n(S))^{-\alpha}$$ on $Sp(n,{\mathbb{R}})$, where $\lambda_1(S)\ldots\lambda_n(S)$ are the singular values $\geq1$ of the symplectic matrix $S$.
We have $h\in L^{\beta,\infty}$ on $Sp(n,{\mathbb{R}})$, with respect to the Haar measure, if $$\alpha\beta\geq 2n.$$
We have to estimate the measure of the set $$\mathcal{D}_\lambda=\{S\in Sp(n,{\mathbb{R}}):\ h(S)\geq \lambda\}, \quad \lambda>0$$ or equivalently $$\int_{Sp(n,{\mathbb{R}})} \chi_{\mathcal{D}_\lambda} dS,$$ where $\chi_{\mathcal{D}_\lambda}$ is the indicator function of $\mathcal{D}_\lambda$. Observe that $\mathcal{D}_\lambda=\emptyset$ if $\lambda>1$ so that we can suppose $0<\lambda\leq 1$.
We use formula with $f=\chi_{\mathcal{D}_\lambda}$ since $h$, and therefore $f$, is $U(2n,{\mathbb{R}})$-bi-invariant. With the notation in we have $$h(a_t)=e^{-\alpha(t_1+\ldots +t_n)/2},$$ where $t=(t_1,\ldots,t_n)$ and $t_1, \ldots, t_n\geq0$. Hence $h(a_t)\geq\lambda$ if and only if $$t_1+t_2+\ldots+t_n\leq A_\lambda:=-2\log \lambda/\alpha.$$ By , $${\rm meas}\,\mathcal{D}_\lambda=C
\int_{t_1\geq t_2\geq...\geq t_n\geq 0\atop t_1+...+t_n\leq A_\lambda} \prod_{i<j} \sinh\frac{t_i-t_j}{2}\, \prod_{i\leq j} \sinh\frac{t_i+t_j}{2}\, dt_n\ldots dt_1.$$
Now we have $$\begin{aligned}
\prod_{i<j} \sinh &\frac{t_i-t_j}{2}\, \prod_{i\leq j} \sinh\frac{t_i+t_j}{2}\\
&\leq \exp\Big({\sum_{i<j}\Big(\frac{t_i-t_j}{2}+\frac{t_i+t_j}{2}\Big)+t_1+...+t_n}\Big)\\
&=\exp\big({(n-1)t_1+(n-2)t_2+...+t_{n-1}+t_1+...+t_n}\big)\\
&=e^{t_n}e^{2t_{n-1}}e^{3t_{n-2}}\ldots e^{nt_1}.\end{aligned}$$ By first integrating with respect to the variable $t_n$ from $t_n=0$ to $t_n=A_\lambda-t_{n-1}-\ldots-t_1$, we obtain $${\rm meas}\,\mathcal{D}_\lambda\leq C \int_{t_1\geq t_2\geq...\geq t_{n-1}\geq 0\atop t_{n-1}\leq A_\lambda-t_{n-2}-...-t_1}e^{A_\lambda}e^{t_{n-1}}\ldots e^{(n-1)t_1}\, dt_{n-1}\ldots dt_1.$$ Now we can repeat the same argument for $t_{n-1}$ and so on. We obtain $${\rm meas}\,\mathcal{D}_\lambda\leq Ce^{nA_\lambda}=C \lambda^{-2n/\alpha},\quad 0<\lambda\leq 1.$$ Hence ${\rm meas}\,\mathcal{D}_\lambda\leq C'\lambda^{-\beta}$ if $2n/\alpha\leq \beta$, which is the desired result.
Using and we have $$|\langle \widehat{S}\varphi_1,\varphi_2\rangle|\lesssim (\lambda_1(S)\ldots\lambda_n(S))^{-1/2}\|\varphi_1\|_{M^1}\|\varphi_2\|_{M^1}.$$ Hence it is sufficient to prove that the function $$\widehat{S}\longmapsto (\lambda_1(S)\ldots\lambda_n(S))^{-1/2}$$ is in $L^{4n,\infty}$ on $Mp(n,{\mathbb{R}})$ with respect to the Haar measure. Since this function factorizes through $Sp(n,{\mathbb{R}})$ it is enough to prove that the function $$h(S):= (\lambda_1(S)\ldots\lambda_n(S))^{-1/2}$$ is in $L^{4n,\infty}$ on $Sp(n,{\mathbb{R}})$. This follows from Proposition \[pro1\] with $\alpha=1/2$ and $\beta=4n$.
We are now ready to prove the Strichartz estimates for the metaplectic representation.
We know that $$\label{elle2}
\|\widehat{S}\psi\|_{L^2}=\|\psi\|_{L^2}$$ for $\psi\in L^2({\mathbb{R}^n})$, which gives the desired Strichartz estimate for $q=\infty,\ r=2$, because $M^2=L^2$, and also for $q=\infty,\ 2\leq r\leq\infty$, because $L^2\hookrightarrow M^r$ for $r\geq 2$. Hence from now on we can suppose $
q<\infty.
$
Now by Theorem \[mainthm\], $$\|\widehat{S}\psi\|_{M^\infty}\lesssim (\lambda_1(S)\ldots\lambda_n(S))^{-1/2}\|\psi\|_{M^1}.$$ By interpolation with we obtain, for every $2\leq r\leq\infty$, $$\label{interp}
\|\widehat{S}\psi \|_{M^r}\lesssim(\lambda_1(S) \ldots\lambda_n(S))^{-(\frac{1}{2}-\frac{1}{r})}\|\psi\|_{M^{r'}}.$$ Let now $G=Mp(n,{\mathbb{R}})$, as in the statement. We apply the usual $ TT^\ast$ method (see [@Tao page 75]) to the operator $T\psi=\widehat{S}\psi$. To prove that $$T:L^2\to L^q(G;M^r)$$ continuously, we will verify that $$TT^\ast :L^{q'}(G;M^{r'})\to L^q(G;M^r)$$ continuously.
We have
$$T^*F(\cdot)=\int_G\widehat{S}^{-1}F(\widehat{S},\cdot)d\widehat{S}$$ if $F(\widehat{S},x)$ is, say, a continuos function on $G\times {\mathbb{R}^n}$ with compact support.
Hence $$[TT^*F](\widehat{S},\cdot)=\int_G \widehat{S}\widehat{S'}^{-1}F(\widehat{S'},\cdot)d\widehat{S'}.$$ Now, using we can estimate this expression, for every $2\leq r\leq\infty$, $1\leq q\leq\infty$, as $$\begin{aligned}
\|TT^*F\|_{L^q(G;M^r)}&\leq \|\int_G \|\widehat{S}\widehat{S'}^{-1}F(\widehat{S'},\cdot)\|_{M^r}d\widehat{S'}\|_{L^q(G)}\nonumber\\
&\leq \|\int_G (h \circ \pi)(\widehat{S}\widehat{S'}^{-1})\|F(\widehat{S'},\cdot)\|_{M^{r'}}d\widehat{S'}\|_{L^q(G)},\label{ultima}\end{aligned}$$ where we set $$h(S)=(\lambda_1(S) \ldots\lambda_n(S))^{-(\frac{1}{2}-\frac{1}{r})}$$ as a function on $Sp(n,{\mathbb{R}})$ and $\pi:G=Mp(n,{\mathbb{R}})\to Sp(n,{\mathbb{R}})$ is the projection.
Now suppose that the pair $(q,r)$ satisfies $2\leq q,r\leq\infty$ and $\frac{4n}{q}+\frac{1}{r}\leq \frac{1}{2}$; see Figure \[figura\]. Observe that this implies $q>2$ and we are also supposing $q<\infty$, which implies $r>2$. Choose $$\alpha=\frac{1}{2}-\frac{1}{r},\quad \frac{1}{\beta}= \frac{2}{q}.$$ We see that $\alpha,\beta>0$ and $\alpha\beta\geq 2n$ so that by Proposition \[pro1\] we have $h\in L^{\beta,\infty}$ in $Sp(n,{\mathbb{R}})$ and $h \circ \pi\in L^{\beta,\infty}$ on $G$. We moreover have $$\frac{1}{q}+1=\frac{1}{q'}+\frac{1}{\beta},\quad 1<q,q',\beta<\infty.$$ Hence we can apply the weak-type Young inequality on $G$ to the last expression in , and we see that it is therefore dominated by $
\|F\|_{L^{q'}(G;M^{r'})}.
$
This concludes the proof.
Concluding remarks {#sec4}
==================
The motivation for modulation spaces
------------------------------------
Let us point out the main elements which led us to consider the modulation space $M^1$ and its dual $M^\infty$ as natural candidates for the dispersive estimate .
Estimate clearly does not hold with $M^1$ and $M^\infty$ replaced by $L^1$ and $L^\infty$, respectively, because, for example, the pointwise multiplication by $e^{it|x|^2}$ is a metaplectic operator but Lebesgue norms do not detect any decay as $|t|\to+\infty$. Hence we focused on a space which controls $L^1$ decay in space [*and*]{} $L^1$ decay in momentum, as $M^1$ does indeed.
But in the course of the proof of Theorem \[mainthm\] we also used in an essential way another property of $M^1$, namely that the set of operators $\widehat{U}$ are uniformly bounded on $M^1$ when $U=\pi(\widehat{U})$ varies in $U(2n,{\mathbb{R}})$, as proved in .
Motivated by these issues, it would be very interesting to get characterizations of function spaces, in particular modulation spaces, in terms of symplectic invariance.
Sharpness of the results
------------------------
It is easy to see that the exponent $-1/2$ in is sharp. In fact, one can apply that estimate to a Gaussian function $\psi$ and the metaplectic operator $\widehat{S}\psi(x)=c \sqrt{\det L}\,\psi (Lx)$ (for suitable $c\in\mathbb{C}$, $|c|=1$), with $L={\rm diag}(\lambda_1,\ldots,\lambda_n)$, $\lambda_1,\ldots,\lambda_n\geq 1$. We have $S=(\lambda_1^{-1},\ldots,\lambda_n^{-1},\lambda_1,\ldots,\lambda_n)$ (cf. [@DeGosson Proposition 116]) and $$\|\widehat{S}\psi\|_{M^\infty}\asymp (\lambda_1\ldots\lambda_n)^{-1/2},$$ as proved in [@cordero2 Lemma 3.2] (and in [@toft Lemma 1.8] in the case $\lambda_1=\ldots=\lambda_n$).
Let us observe that the exponent $4n$ in is sharp as well; in fact Howe [@howe] proved that for fixed $\varphi_1,\varphi_2\in \mathcal{S}(\mathbb{R}^n)$ the matrix coefficients in general do not belong to $L^{4n}$.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors are very indebted to Professors Elena Cordero, Michael Cowling, Jaques Faraut and Vladimir Uspenskiy for discussions and remarks which improved the paper in an essential way.
[99]{}
E. Cordero, F. De Mari, K. Nowak, A. Tabacco, [*Analytic features of reproducing groups for the metaplectic representation*]{}, J. Fourier Anal. Appl. 12, no. 2, 157–180, (2006).
E. Cordero, F. De Mari, K. Nowak, A. Tabacco, [*Reproducing groups for the metaplectic representation*]{}, Pseudo-differential operators and related topics, 227–244, Oper. Theory Adv. Appl., 164, Birkhäuser, Basel, (2006).
E. Cordero, F. De Mari, K. Nowak, A. Tabacco, [*Dimensional upper bounds for admissible subgroups for the metaplectic representation*]{}, Math. Nachr. 283, no. 7, 982–993 (2010).
E. Cordero, K. Gröchenig, F. Nicola, L. Rodino, [*Generalized metaplectic operators and the Schrödinger equation with a potential in the Sjöstrand class*]{}, J. Math. Phys. 55, no. 8, 081506, (2014).
E. Cordero, F. Nicola, [*Strichartz estimates in Wiener amalgam spaces for the Schrödinger equation*]{}, Math. Nachr. 281, no. 1, 25–41 (2008).
E. Cordero, F. Nicola, [*Metaplectic representation on Wiener amalgam spaces and applications to the Schrödinger equation*]{}, J. Funct. Anal. 254, no. 2, 506–534, (2008).
E. Cordero, A. Tabacco, [*Triangular subgroups of $Sp(d,{\mathbb{R}})$ and reproducing formulae*]{}, J. Funct. Anal. 264, no. 9, 2034–2058, (2013).
M. Cowling, [*Sur les coefficients des représentations unitaires des groupes de Lie simples.*]{} Analyse harmonique sur les groupes de Lie (Nancy-Strasbourg, France, 1976–78), II, Lecture Notes in Math. 739, Springer, Berlin, 132–178, (1979).
M. Cowling, [*Herz’s “principe de majoration" and the Kunze-Stein phenomenon*]{}. Harmonic analysis and number theory (Montreal, PQ, 1996), 73–88, CMS Conf. Proc., 21, Amer. Math. Soc., (1997).
M. Cowling, U. Haagerup, R. Howe, [*Almost $L^2$ matrix coefficient*]{}, J. Reine Angew. Math. 387, 97–110 (1988).
M. A. de Gosson, [*Symplectic Methods in Harmonic Analysis and in Mathematical Physics*]{}, Birkhäuser Basel, (2011).
M. A. de Gosson, [*The Wigner Transform*]{}, Advanced Textbooks in Mathematics, World Scientific, (2017).
M. Goto, [*Absolutely closed Lie groups*]{}, Math. Ann. 204, 337–341, (1973).
K. Gröchenig, [*Foundations of Time-Frequency Analysis*]{}, Birkhäuser Basel, (2001).
S. Helgason, [*Groups and Geometric Analysis: Integral Geometry, Invariant Differential Operators, and Spherical Functions*]{}, American Mathematical Society, Mathematical Surveys and Monographs 83, (1984).
R. Howe, [*On a notion of rank for unitary representations of the classical groups*]{}, Harmonic analysis and group representations, 223–331, Liguori, Naples, (1982).
R. Howe, E-C. Tan, [*Nonabelian harmonic analysis. Applications of $SL(2,{\mathbb{R}})$*]{}, Universitext. Springer-Verlag, New York, (1992).
Leray, Jean [*Lagrangian analysis and quantum mechanics. A mathematical structure related to asymptotic expansions and the Maslov index*]{}, translated from the French by Carolyn Schroeder, MIT Press, Cambridge, Mass.-London, (1981).
C. C. Moore, [*Exponential decay of correlation coefficients for geodesic flows*]{}, Group representations, ergodic theory, operator algebras, and mathematical physics (Berkeley, Calif., 1984), 163–181, Math. Sci. Res. Inst. Publ., 6, Springer, New York, (1987).
H. Oh, [*Uniform pointwise bounds for matrix coefficients of unitary representations and applications to Kazhdan constants*]{}, Duke Math. J. 113, no. 1, 133–192, (2002).
M. Ruzhansky, M. Sugimoto, B. Wang, [*Modulation spaces and nonlinear evolution equations. Evolution equations of hyperbolic and Schrödinger type*]{}, 267–283, Progr. Math., 301, Birkhäuser/Springer Basel AG, Basel, (2012).
I. E. Segal, [*Foundations of the theory of dynamical systems of infinitely many degrees of freedom I*]{}, Mat.-Fys. Medd. Dansk. Vid. Selsk., 31(12), (1959). D. Shale, [*Linear symmetries of free Boson fields*]{}, Trans. Amer. Math. Soc., 103:149-167, (1962).
T. Tao, [*Nonlinear dispersive equations: local and global analysis*]{}, American Mathematical Society, (2006).
J. Toft, [*Continuity properties for modulation spaces, with applications to pseudo-differential calculus. I.*]{}, J. Funct. Anal., 207(2):399–429, (2004).
L. van Hove, [*Sur certaines groupes representations unitaires d’un groupe infini de transformations*]{}, Mem. Acad. Roy. de Belgique, Classe des Sci., 25(6), (1951).
B. Wang, Z. Huo, C. Hao, Z. Guo, [*Harmonic analysis method for nonlinear evolution equations. I.*]{}, World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, (2011).
A. Weil, [*Sur certaines groupes d’operateurs unitaires*]{}, Acta Math., 111:143-211, (1964).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study an equation structured by age and a phenotypic trait describing the growth process of a population subject to aging, competition between individuals, and mutations. This leads to a renewal equation which occurs in many evolutionary biology problems. We aim to describe precisely the asymptotic behavior of the solution, to infer properties that illustrate the concentration and adaptive dynamics of such a population. This work is a continuation of [@Nordmann2018a] where the case without mutations is considered. When mutations are taken into account, it is necessary to control the corrector which is the main novelty of the present paper.
Our approach consists in defining, by the Hopf transform, a Hamilton-Jacobi equation with an effective Hamiltonian as in homogenization problems. Its solution carries the singular part of the limiting density (typically Dirac masses) and the corrector defines the weights. The main new result of this paper is to prove that the corrector is uniformly bounded, using only the global Lipschitz and semi-convexity estimates for the viscosity solution of the Hamilton-Jacobi equation. We also establish the limiting equation satisfied by the corrector. To the best of our knowledge, this is the first example where such bounds can be proved in such a context.
author:
- 'Samuel Nordmann[^1]'
- 'Beno\^ it Perthame[^2] [^3]'
bibliography:
- 'library.bib'
title: 'Dynamics of concentration in a population structured by age and a phenotypic trait with mutations. Convergence of the corrector'
---
[**Key-words:**]{} Adaptive evolution; Asymptotic behaviour; Dirac concentration; Hamilton-Jacobi equations; Mathematical biology; Renewal equation; Viscosity solutions; correctors.\
[**AMS Class. No:**]{} 35B40, 35F21, 35Q92, 49L25.
Introduction
============
Main results
------------
We study a mathematical model describing the growth process of a population structured by age and a phenotypic trait, subject to aging, competition between individuals and mutations. Our goal is to describe the asymptotic behavior of the solution, in particular the selection of the fittest traits and the adaptative dynamics of such traits. For ${\varepsilon}>0$, we choose $m_{\varepsilon}(t,x,y)$ to represent the population density of individuals who, at time $t\geq0$, have age $x\geq0$ and a quantitative phenotipic trait $y\in{\mathbb{R}}^n$, solution of a renewal type equation $$\label{equa}
\left\{\begin{array}{ll}
{\varepsilon}{\partial}_t m_{\varepsilon}+ {\partial}_x\left[A(x,y)m_{\varepsilon}\right] +\left(\rho_{\varepsilon}(t)+d(x,y)\right)m_{\varepsilon}=0,
\\[2mm]
A(0,y)m_{\varepsilon}\left(t,0,y\right)=\frac{1}{{\varepsilon}^n}\int_{{\mathbb{R}}^n}\int_{{\mathbb{R}}_+}{M(\frac{y'-y}{{\varepsilon}})b(x',y')m_{\varepsilon}(t,x',y')dx'dy'},
\\[2mm]
\rho_{\varepsilon}(t)=\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}^n}{m_{\varepsilon}(t,x,y)dx dy},
\\[2mm]
m_{\varepsilon}(t=0,x,y)=m_{\varepsilon}^0(x,y)>0.
\end{array}\right.$$ The parameter ${\varepsilon}>0$ stands for a hyperbolic rescaling $(t,y)\leftrightarrow({\varepsilon}^{-1}t,{\varepsilon}^{-1}y)$ in the mutation term. Our main concern is to study the asymptotics of $m_{\varepsilon}$ when ${\varepsilon}\to0$.
This work is the continuation of the study begun in [@Nordmann2018a], where the model without mutations is studied and where we proved that there is a measure $\mu$ (typically a Dirac mass) and a bounded profile ${\mathcal Q}$ such that $$m_{\varepsilon}(t,x,y) \rightharpoonup {\mathcal Q}(t,x,y)\, \mu(t,y),$$ in other words, the asymptotic singularity is carried in the variable $y$ only. In the present work, the mutation term in the second line of adds a significant difficulty because the profile ${\mathcal Q}$ turns out to be strongly related to the limit of the corrector in the spectral problem defining the underlying effective Hamiltonian, i.e., the *effective fitness* in our context, arising in a Hamilton-Jacobi equation which defines the singular part.
While the classical approach consists in studying the asymptotics of $v_{\varepsilon}(t,x,y):={\varepsilon}\ln (m_{\varepsilon}(t,x,y))$, here, and following [@BP:PS-dispersal; @Nordmann2018a], we define a priori some kind of variable separation setting $$m_{\varepsilon}(t,x,y):= p_{\varepsilon}(t,x,y) e^\frac{U_{\varepsilon}(t,y)- \int_0^t \rho_{\varepsilon}(\cdot)}{{\varepsilon}}$$ where the exponential term carries the singular part of the limiting population density and $U_{\varepsilon}$ is defined autonomously through a Hamilton-Jacobi equation (see ) involving the *effective fitness* $\Lambda$ (defined in ). The main finding of this paper is the proof that, with our choice of $U_{\varepsilon}$, the corrector $p_{\varepsilon}$ is uniformly bounded. We also show that $p_{\varepsilon}$ converges to a multiple of the principal eigenfunction $Q(t,x,y)$ of the formal limiting operator (defined in ). It justifies that the ansatz $U_{\varepsilon}$ is appropriate and carries all the information on the singular behavior of $m_{\varepsilon}$ when ${\varepsilon}\to0$. The formal idea of the method is explained with more details in section \[SecFormalApproach\] and consists in keeping as simple as possible the Hamilton-Jacobi equation which defines $U_{\varepsilon}$ while all the functional analytic difficulties are carried by the corrector $p_{\varepsilon}$ which satisfies a linear equation.
The first step is to define $U_{\varepsilon}$ and study its properties.
\[th:MainResultsConvergenceU\] Under the assumptions of section \[SecAssumptions\], when ${\varepsilon}\to0$, the ansatz $U_{\varepsilon}(t,y)$ is well defined through and converges in $W_{loc}^{1,r}$, $1 \leq r < \infty$, to some $U\in W^{1,\infty}_{loc}$ which is semi-convex and is a viscosity solution of the Hamilton-Jacobi equation .
The second step is to prove that the corrector is uniformly bounded and converges.
\[th:MainResultsConvergenceP\] Under the assumptions of section \[SecAssumptions\], for any fixed $T>0$, $p_{\varepsilon}(t,x,y)$ and $\int_{x>0}p_{\varepsilon}(t,x,y)dx$ are bounded from above and below, uniformly in ${\varepsilon}>0$, $t\in[0,T]$, $y\in{\mathbb{R}}^n$.
In addition, for $t\in[0,T]$, $x\in[0,\overline{x}]$, $y\in{\mathbb{R}}^n$, up to extraction of a subsequence ${\varepsilon}\to0$, $p_{\varepsilon}$ converges in $L^\infty$-weak-$^\star$ to $\gamma(t,y) Q(t,x,y)$, where $Q$ is the eigenfunction defined in and $\gamma\in L^\infty$ formally satisfies .
With these informations, it is standard that the *concentration points* of the population density $m_{\varepsilon}(t,x,y)$ are carried by the set $$\label{DefSetS}
\mathcal{S}:=\left\{t\geq0,\ y\in{\mathbb{R}}^n:U(t,y)=\sup\limits_{y'\in{\mathbb{R}}^n}U(t,y')\right\}.$$
\[MainResults\] Under the assumptions of section \[SecAssumptions\], when ${\varepsilon}\to0$\
1. The total population $\rho_{\varepsilon}$ converges weakly to some positive $\rho\in L^\infty$ and $$\forall t>0,\quad \int_0^t \rho = \sup\limits_{y\in{\mathbb{R}}^n} U(t,y).$$ 2. The population $m_{\varepsilon}$ vanishes locally uniformly outside the set $\mathcal{S}$.\
3. Under further assumptions on the initial conditions, and for small times $t\in[0,T]$, we have $$\mathcal{S}=\{(t,\bar y(t))\},$$ where $\bar y(t)$ follows the Canonical Equation $$\left\{\begin{aligned}
&\frac{\mathrm{d}}{\mathrm{d}t}\bar y(t)=\left(D_y^2U(t,\bar y(t))\right)^{-1}\cdot\nabla_y\Lambda(\bar y(t),1)+{\partial}_\eta\Lambda(\bar y(t),1)\int_{{\mathbb{R}}^n}M(z)z dz,\\
&\bar y(0)=\bar y^0.
\end{aligned}\right.$$
Notice that we establish convergence of the full families $U_{\varepsilon}$ and $\rho_{\varepsilon}$, without using the famous uniqueness result of [@CaLa], because of the simple dependency on the unknown $\rho_{\varepsilon}$ in our setting.
We point out that the main restriction on the coefficients is the assumption that the transport in $x$ outwards the support of $b(\cdot,y)$ occurs in finite time, see . The other assumptions, detailed in section \[SecAssumptions\], are formulated directly on the limiting eigenproblem and are quite general.
The model
---------
Some possible biological interpretations of the model are as follows. The function $A(x,y)$ is the speed of aging of individuals with age $x$ and phenotypic trait $y$. The total size of the population at time $t$ is denoted with $\rho_{\varepsilon}(t)$. In the mortality term, $d(x,y)>0$ represents intrinsic death rate and the nonlocal term $\rho_{\varepsilon}(t)$ represents competition. The condition at the boundary $x=0$ describes the birth of newborns that happens with rate $b(x,y)>0$ and with the probability kernel $M$ for mutations.
The terminology of *renewal equation* comes from this boundary condition. It is related to the McKendrick-von Foerster equation which is only structured in age (see [@perthame] for a study of the linear equation). This model has been extended with other structuring variables, as size [@DiekPhys; @Mis-Pe-Ry], DNA content, maturation, etc., in the context of cell divisions [@Doumic-Gabriel], or proliferative and quiescent states of tumor cells [@Adimy-Cr-Ru; @GyllWebb]. Space structured problems have also been extensively studied [@Jabin2016; @Mi-patch; @Mi-Pe-spatial; @BP:PS-dispersal].
To keep the model quite general, we have allowed the progression speed $A$ to depend on $x$. Thus, although the variable $x$ is referred to as *age*, it can represent other biological quantities that evolve throughout the individual lifespan such as, for instance, the size of individuals, a physiological age, a parasite load, etc.
The rescaling parameter ${\varepsilon}> 0$ comes from a hyperbolic rescaling of $t$ and $y$. Accordingly, the dynamics are considered in two different time scales. The first one is the individual lifetime scale ${\varepsilon}t$, i.e., the characteristic time for the population to reach the dynamical equilibrium for a fixed $y$. The second one is the evolutionary time scale $t$, corresponding to the evolution of the population distribution with respect to the variable $y$. Formally, at the limit when ${\varepsilon}\to0$, the time scales are completely separated. This rescaling is a classical way to give a continuous formulation of the adaptive evolution of a phenotypically structured population (see [@CFBA; @OD; @Di-Ja-Mi-Pe; @AL.SM.BP]). Note that the mutation kernel is supposed to be *thin-tailed*, i.e., it decreases faster than any exponential. A fat-tailed kernel needs a different rescaling, see [@Bouin].
From the modelization point of view, is a form of mathematical formulation of Natural Selection and Evolution. On an ecological time scale, only the phenotype $\bar y(t)$ which maximizes the ecological fitness $U(t,\cdot)$ can survive. On an evolutionary time scale, we observe the dynamics of $\bar y(t)$. Similar results as have been obtained for various models with parabolic equations [@GB.BP:07; @GB.SM.BP:09; @AL.SM.BP] and integrodifferential equations [@Ba-Pe; @DJMR; @lorenzi-2013]. More generally, convergence to positive measures in selection-mutation models has been studied by many authors [@ackleh_F_T; @calsina-al-2013; @Busse2016]. The special case of age-structured populations are also considered in [@Meleard2009; @Tran2008; @VC:PG:AMG].
A formal presentation of the method {#SecFormalApproach}
-----------------------------------
To analyze the singular perturbation problem at hand, the usual approach relies on the WKB change of unknown ([@GB.BP:07; @Di-Ja-Mi-Pe]), which consists in the change of variable $m_{\varepsilon}(t,x,y)=e^{\frac{v_{\varepsilon}(t,x,y)}{{\varepsilon}}}$. In the context of concentrations, this form is motivated by the heuristics that a Dirac mass is nothing but a narrow Gaussian. Indeed, in a weak sense $\left(\pi{\varepsilon}\right)^{\frac{-n}{2}} e^{-\frac{\Vert y-\bar y\Vert ^2}{{\varepsilon}}} \rightharpoonup \delta_{\bar y=y}$ as $\epsilon\to0$. This approach has been extensively used in works on a similar issue, see for instance [@BG-LCE-PES; @combustion; @Champagnat2011a; @Evans1989]. With this change of unknown, at the limit ${\varepsilon}\to0$, the function $v_{\varepsilon}(t,x,y)$ satisfies a constrained Hamilton-Jacobi equation, on which estimates can be difficult to prove because it carries all the difficulties in the asymptotic analysis, concentration effect and profile defined by the corrector. For that reason, the perturbed test function method has been invented [@Evans_perturbed] and widely used, which avoids computing the corrector.
Here, we propose a variant of the method. The principle can be viewed as a Taylor expansion $v_{\varepsilon}(t,x,y)= v^1_{\varepsilon}(t,y)+{\varepsilon}v_{\varepsilon}^2(t,x,y)$, to choose only some convenient terms to define the Hamilton-Jacobi equation for $v_{\varepsilon}^1$, and then to prove that the corrector $v^2_{\varepsilon}$ is bounded. With a slight rewriting, we proceed to the change of variable $$\label{factorization}
m_{\varepsilon}(t,x,y)=p_{\varepsilon}(t,x,y) \, \exp{\frac{U_{\varepsilon}(t,y)- \int_0^t \rho_{\varepsilon}(s)ds }{{\varepsilon}}}$$ where $U_{\varepsilon}(t,y)$ is defined *ad hoc* through a standard Hamilton-Jacobi equation for which classical regularity properties can be proved. Then,the new and difficult step is to prove estimates on the corrector $p_{\varepsilon}(t,x,y)$. Note that $p_{\varepsilon}$ satisfies a linear equation rather than a constrained Hamilton-Jacobi equation: it makes thus possible to use classical comparison principles and ideas issued from the General Relative Entropy method (see [@PMSMBP1]).\
We are now left with the task of finding a good candidate for $U_{\varepsilon}(t,y)$ and formally identifying $p_{\varepsilon}(t,x,y)$. Injecting in , we find $$\label{equaFactor}
\left\{\begin{array}{ll}
{\varepsilon}{\partial}_t p_{\varepsilon}+ {\partial}_x\left[A(x,y)p_{\varepsilon}\right] +d(x,y)p_{\varepsilon}+ {\partial}_t U_{\varepsilon}(t,y) p_{\varepsilon}=0,
\\[2mm]
A(0,y)p_{\varepsilon}\left(t,0,y\right)=\frac{1}{{\varepsilon}^n}\int_{{\mathbb{R}}^n}\int_{{\mathbb{R}}_+}{M(\frac{y'-y}{{\varepsilon}})e^{\frac{U_{\varepsilon}(t,y')-U_{\varepsilon}(t,y)}{{\varepsilon}}}b(x',y')p_{\varepsilon}(t,x',y')dx'dy'}.
\end{array}\right.$$ With the change of variable $z=\frac{y'-y}{{\varepsilon}}$, the renewal term becomes $$\label{boundarycondeps}
A(0,y)p_{\varepsilon}(t,0,y)=\int_{{\mathbb{R}}^n}\int_{{\mathbb{R}}_+}{M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{{\varepsilon}}}b(x',y+{\varepsilon}z)p_{\varepsilon}(t,x',y+{\varepsilon}z)dx'dz}.$$ When ${\varepsilon}$ is small, we can formally approximate $$\label{Approximation}
A(0,y)p_{\varepsilon}(t,0,y)\approx\eta_{\varepsilon}(t,y)\int_{{\mathbb{R}}_+}b(x',y)p_{\varepsilon}(t,x',y)dx',$$ where $$\eta_{\varepsilon}(t,y):=\int_{{\mathbb{R}}^n}{M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{{\varepsilon}}}dz}.$$ Then, formally putting ${\varepsilon}{\partial}_tp_{\varepsilon}=O({\varepsilon})$ in the first line of , we end up with the following approximate problem $$\label{equaFactorApprox}
\left\{\begin{array}{ll}
{\partial}_x\left[A(x,y)p_{\varepsilon}\right] +d(x,y)p_{\varepsilon}+ {\partial}_t U_{\varepsilon}(t,y) p_{\varepsilon}=O({\varepsilon}),
\\[2mm]
A(0,y)p_{\varepsilon}\left(t,0,y\right)=\eta_{\varepsilon}(t,y)\int_{{\mathbb{R}}_+}b(x',y)p(t,x',y)dx' + O({\varepsilon}).
\end{array}\right.$$
Considering $\eta_{\varepsilon}(t,y)$ as a parameter, we introduce the following eigenproblem: for fixed $(y,\eta)\in{\mathbb{R}}^n\times(0,+\infty)$, find $(\Lambda(y,\eta),Q(x,y,\eta))$, solution of $$\label{EigenProblemFinal}
\left\{\begin{array}{ll}
{\partial}_x\left[A(x,y)Q\right]+d(x,y)Q-\Lambda(y,\eta)Q=0,\quad\forall x>0,\\[1mm]
A(0,y)Q(0,y,\eta)=\eta,\\[1mm]
Q> 0, \quad \int_{{\mathbb{R}}_+} b(x,y) Q(x,y,\eta)dx=1.
\end{array}\right.$$ The third line corresponds to a normalization of the eigenfunction which is convenient for future calculations. Formally, $\Lambda$ corresponds to the *effective fitness*, and $Q$ to the age profile at equilibrium in an environment characterized by the parameters $(y,\eta)$.\
In an attempt to indentify $p_{\varepsilon}$ with $Q$, this formal approach suggests to define $U_{\varepsilon}$ as a solution of the Hamilton-Jacobi equation $$\label{DefU_eps}\left\{
\begin{aligned}
&{\partial}_t U_{\varepsilon}(t,y)=-\Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{{\varepsilon}}}dz\right) &\quad \forall t\geq0, \, y\in{\mathbb{R}}^n,
\\
&U_{\varepsilon}(0,y)=U_{\varepsilon}^0(y) &\forall y\in{\mathbb{R}}^n,
\end{aligned}\right.$$ for some initial conditions $U^0_{\varepsilon}$.
We stress out that, when ${\varepsilon}$ vanishes, the full term $\exp{\frac{U_{\varepsilon}(t,y)- \int_0^t \rho_{\varepsilon}(s)ds }{{\varepsilon}}}$ represents a bounded measure, e.g., a Dirac mass as in the example of Gaussian concentration, and thus $$\sup_{y \in {\mathbb{R}}^n} U_{\varepsilon}(t,y)- \int_0^t \rho_{\varepsilon}(s)ds =0, \qquad \forall t\geq 0,$$ which explains the first formula in Theorem \[MainResults\].
However, because of the non-local term $\eta_{\varepsilon}$, proving uniform in ${\varepsilon}>0$ estimates on $U_{\varepsilon}$ is quite technical, a fact that can also be seen on the limiting equation when ${\varepsilon}\to0$. Taking for granted that $U_{\varepsilon}$ converges to a function $U$ locally uniformly, $U$ turns out to be a viscosity solution of the Hamilton-Jacobi equation $$\label{DefU}\left\{
\begin{aligned}
&{\partial}_t U(t,y)= H(y,\nabla_y U) &\forall t\geq0,\ \forall y\in{\mathbb{R}}^n,\\
&U(0,y)=U^0(y) &\forall y\in{\mathbb{R}}^n.
\end{aligned}\right.$$ with the Hamiltonian $H(y,p)$ defined by $$\label{DefHamiltonian}
H(y,p):= -\Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{p\cdot z}dz\right).$$ From this equation, it is classical to prove uniform a priori bounds on ${\partial}_t U$ (indeed, ${\partial}_t U$ satisfies a transport equation). We deduce that $H(y,\nabla_y U)$ is bounded, then that $$\eta(t,y):= \int_{{\mathbb{R}}^n}M(z)e^{\nabla_y U(t,y)\cdot z}dz$$ is bounded. Besides, since $p\mapsto H(y,p)$ is convex, we deduce that $U$ is semi-convex, namely in space-time, the Hessian $D^2 U$ is bounded from below. Since ${\partial}_t U$ is bounded, we have ${\partial}^2_t U \in L_{loc}^1$. Then, using ${\partial}^2_t U= -{\partial}_t \eta \, {\partial}_\eta\Lambda$, we infer that ${\partial}_t\eta\in L^1_{loc}$. In the sequel, all these estimates are proved on $U_{\varepsilon}$, uniformly in ${\varepsilon}>0$. These are the classical (and sharp) general estimates for Hamilton-Jacobi equations with convex Hamiltonians.\
With these optimal estimates on $U_{\varepsilon}$ in hands, we can bound the corrector $p_{\varepsilon}$. We set $$Q_{\varepsilon}(t,x,y):= Q(x,y,\eta_{\varepsilon}(t,y)), \qquad \Lambda_{\varepsilon}(t,y):=\Lambda(y,\eta_{\varepsilon}(t,y))$$ and find that $$\label{EquationOnQepsFormelle}
\left\{\begin{array}{ll}
{\varepsilon}{\partial}_t Q_{\varepsilon}+ {\partial}_x\left[A(x,y)Q_{\varepsilon}\right] +d(x,y)Q_{\varepsilon}+ {\partial}_t U_{\varepsilon}(t,y) Q_{\varepsilon}={\varepsilon}{\partial}_t Q_{\varepsilon},
\\[2mm]
A(0,y)Q_{\varepsilon}\left(t,0,y\right)=\frac{1}{{\varepsilon}^n}\int_{{\mathbb{R}}^n}\int_{{\mathbb{R}}_+}{M(\frac{y'-y}{{\varepsilon}})e^{\frac{U_{\varepsilon}(t,y')-U_{\varepsilon}(t,y)}{{\varepsilon}}}b(x',y')Q_{\varepsilon}(t,x',y')dx'dy'}.
\end{array}\right.$$ Note that the boundary term at $x=0$ is obtained using the definition of $\eta_{\varepsilon}(t,y)$ and the normalization $\int_{{\mathbb{R}}_+} b(x,y) Q(x,y,\eta)dx=1$ for all $(y, \eta)$.
The right hand side of the first line can be controlled with the available a priori bounds on $U_{\varepsilon}$. Indeed, integrating equation to obtain $AQ$, we compute $${\partial}_t Q_{\varepsilon}= {\partial}_t\eta_{\varepsilon}(t,y) {\partial}_\eta Q(x,y,\eta_{\varepsilon}(t,y))={\partial}_t\eta_{\varepsilon}\left(\frac{1}{\eta_{\varepsilon}}+{\partial}_\eta\Lambda(y,\eta_{\varepsilon})\int_0^x\frac{1}{A(\cdot,y)}\right)Q_{\varepsilon}.$$ Except from this term, we see that $Q_{\varepsilon}$ and $p_{\varepsilon}$ satisfy the same equation , which is linear and admits a comparison principle. Under assumptions on the initial conditions, we deduce that $p_{\varepsilon}$ is bounded from above and below by multiples of $Q_{\varepsilon}$. Passing to the limit in the equation, we prove that $p_{\varepsilon}$ converges weakly to a multiple of $Q$. It justifies our approach, especially the approximation , and formally proves .\
Now, having in hand that $U_{\varepsilon}$ converges () and that $p_{\varepsilon}$ is uniformly bounded (), can be understood and formally justified as follows. On the one hand, the saturation term “$\rho_{\varepsilon}(t)$” in implies the total population $\rho_{\varepsilon}$ to be bounded, uniformly in ${\varepsilon}>0$. On the other hand, from , the asymptotics of $m_{\varepsilon}(t,\cdot,\cdot)$ when ${\varepsilon}$ vanishes are driven by the points $y$ where $U(t,\cdot)$ is maximal. In other words, when ${\varepsilon}\to0$, $m_{\varepsilon}$ vanishes outside the set $\mathcal{S}$ defined in . Then, we can study the evolutionary dynamics through the dynamics of the (unique) critical point of $U(t,\cdot)$.
Outline of the paper
--------------------
Section \[SecU\] is devoted to the definition of the eigenelements $(Q,\Lambda)$, the definition of the ansatz $U_{\varepsilon}$, the statement of a priori estimates, and the asymptotics of $U_{\varepsilon}$ when ${\varepsilon}\to0$. The proofs are postponed to section \[sec:appendixConcentrationWithMutations\]. In section \[sec:StudyOfP\], we study the corrector $p_{\varepsilon}$ and prove . Next, we study the asymptotics of the population density, and we prove in section \[sec:Mutation\_Proof\]. Finally, some longer or more technical proofs are gathered in section \[sec:appendixConcentrationWithMutations\].
Assumptions {#SecAssumptions}
-----------
Most of our assumptions are formulated directly on the solution $(\Lambda(y,\eta),Q(x,y,\eta))$ of the limiting eigenproblem and, therefore, may seem abstract to the reader. However, we think that, besides being quite general, this formulation gives a better insight into the nature of our assumptions and the spirit of our approach.
### Example {#RemarksOnAssumptions}
Before stating the general assumptions, we first give for the reader’s convenience a concrete set of assumptions on the coefficients $A$, $b$, $d$ which are sufficient to fulfil the general assumptions. Note that, besides, we need the initial conditions to be “well prepared,” which is not detailed in this example.
To avoid any difficulty when $\vert y\vert\to+\infty$, we can assume for example that $A$, $b$ and $d$ have a compact dependence on $y$. Namely, if $\psi$ is a globally smooth diffeomorphism from ${\mathbb{R}}^n$ into the unit ball, we assume $A(x,y):= A_\star(x,\psi(y))$, $b(x,y):=b_\star(x,\psi(y))$, $d(x,y):=d_\star(x,\psi(y))$, where $A_\star$, $b_\star$ and $d_\star$ are defined on the closed unit ball. This way, the $y$ space ${\mathbb{R}}^n$ is compactified. Then, the coefficients can be chosen to fullfill the following conditions, for all $x\geq0$ and uniformly in $y\in\overline{B}_1$, $$\left\{\begin{gathered}
A_\star(\cdot,\cdot),\ b_\star(x,\cdot) \geq0,\ d_\star(x,\cdot) \geq0 \text{ are }C^1,\\
A_\star(x,y)\geq\underline{A}>0, \qquad \frac{1}{A_\star(\cdot,y)}\; \text{is integrable on the support of } b_\star(\cdot,y),\\
b_\star(x,y)\leq Ke^{Kx}, \qquad \underline\eta b_\star(x,y)-d_\star(x,y)\geq \underline{r},\\
M(\cdot)\text{ is a Gaussian probability kernel},
\end{gathered}\right.$$ for some constants $K>0$, $\underline{r}>0$ and where $\underline \eta$ is determined from the initial conditions, see .
These assumptions can be substantially generalized. For instance, using the formula (obtained by integrating the first line in ) $$\Lambda(y,\eta)=\frac{\int_{{\mathbb{R}}_+}\left(d(x,y)-\eta b(x,y)\right) Q(x,y,\eta)dx}{\int_{{\mathbb{R}}_+}Q(x,y,\eta)dx},$$ the relation between $b_\star$ and $d_\star$ is only used to ensure the inequality $$\Lambda(y,\eta)\leq -\underline{r}<0,\quad \forall \eta\geq\underline{\eta}$$ which is sufficient in the sequel, see .
### Assumptions on the coefficients {#section:AssCoef}
We assume that for some $\underline{A}>0$ $$\label{AssumptionCoefPositifs}
b\geq0,\quad d\geq0,\quad A\geq\underline{A}>0\text{ are continuous functions},$$ $$A(\cdot,\cdot),\ b(x,\cdot),\ d(x,\cdot)\text{ are }C^1,$$ $$\label{AssumptionANoDegenerate}
d(\cdot,y), \ A(\cdot,y) \text{ are bounded from above and below in some interval of ${\mathbb{R}}_+$, uniformly in $y$}.$$ which can be viewed as a non-degeneracy condition. Regarding the mutation kernel, we assume that $$M(\cdot) \text{ is a probability kernel and vanishes faster than any exponential.}$$ For instance, $M(\cdot)$ can be a Gaussian distribution or have a compact support. Note that the case without mutations corresponds to $M=\delta_0$ and has been already treated in [@Nordmann2018a].
In addition, we assume $$\exists K>0\text{ such that } \sup_{ y\in{\mathbb{R}}^n} \int_0^{\overline x}\frac{1}{A(x',y)}dx'\leq ,K\label{AssumptionIntegralACombined}$$ where $\overline x$ defines the largest support of $b(x,y)$, $$\overline x:=\sup\left\{x\geq0 : \exists y\in{\mathbb{R}}^n, \ b(x,y)>0\right\}\in[0,+\infty].$$ $$\exists K>0\text{ such that }\sup_{ y\in{\mathbb{R}}^n} \int_0^{+\infty}\frac{1}{A(x',y)}dx'\leq K,
\label{AssumptionIntegralA}$$ or $$\exists \overline{x}>0\text{ such that }\forall x\geq\overline x,\ \forall y\in{\mathbb{R}}^n,\quad b(x,y)=0.\label{AssumptionBCompact}$$ This assumption means that the transport outwards the support of $b$ occurs in finite time (it can be seen on the characteristics). It is a necessary and sufficient condition for the ratio $\frac{{\partial}_\eta Q}{Q}$ to be bounded on $[0,\overline{x}]$, which is used in section \[SecP\] to prove an estimate on $p_{\varepsilon}$. Our approach would also work if the support of $A(\cdot,y)$ is compact, but we omit this case for simplicity.
We also need the eigenvalue $\Lambda$ of to be well defined and differentiable (w.r.t $\eta$). As we see it in ,defining, $$\label{DefinitionF}
F(y,\lambda):=\int_{{\mathbb{R}}_+}\frac{b(x,y)}{A(x,y)}\mathrm{exp}\left(\int_0^x\frac{\lambda-d(x',y)}{A(x',y)}dx'\right)dx, \qquad \lambda\in{\mathbb{R}},$$ the eigenvalue $\Lambda(y,\eta)$ is defined through the relation $$F(y,\Lambda(y,\eta))=\frac{1}{\eta}.$$ We assume there exists $\overline \Lambda<0$ such that $$\label{AssumptionFormelle}
F( y,\overline\Lambda),\ {\partial}_\lambda F( y,\overline\Lambda) <+\infty\quad\text{for all }y\in{\mathbb{R}}^n.$$ This assumption is not very restrictive, it is satisfied if $b(x,y)\leq K'e^{Kx}$ choosing $ \overline{\Lambda } \leq -K$.
### Assumptions on the initial conditions.
We need the population $m_{\varepsilon}(t,x,y)$ to be “well prepared” for concentration. We write $m_{\varepsilon}^0(x,y)=p_{\varepsilon}^0(x,y)e^{\frac{U_{\varepsilon}^0(y)}{{\varepsilon}}}$ according to , and we assume that $$U_{\varepsilon}^0 \text{ smoothly converges to a function }U^0\text{ when ${\varepsilon}$ vanishes},$$ $$\label{initial_Lipschitz}
\exists k^0>0\text{ such that }\forall{\varepsilon}>0,\ \forall y\in{\mathbb{R}}^n,\quad \vert\nabla_y U_{\varepsilon}^0(y)\vert\leq k^0,$$ $$\label{AssumptionInitialSemiConvexity}
\exists C>0\text{ such that }\forall{\varepsilon}>0,\ \forall y\in{\mathbb{R}}^n,\quad {\partial}^2_{y_i} U_{\varepsilon}^0(y)\geq -C,$$ $$\label{initial_integrability}
\underline{J}^0\leq\int_{{\mathbb{R}}^n}e^{\frac{U^0_{\varepsilon}(y)}{{\varepsilon}}}dy\leq\overline{J}^0,\text{ for some }\underline{J}^0,\overline{J}^0>0,$$ and $p^0_{\varepsilon}$ such that, for some $\underline{\gamma}^0, \overline{\gamma}^0>0$: $$\underline{\gamma}^0\leq \frac{p^0_{\varepsilon}(x,y)}{Q(x,y,\eta_{\varepsilon}^0(y))}\leq \overline {\gamma}^0,\label{AdditionalAssumption2Pre}$$ where $Q$ is defined through and we define $$\label{DefinitionEtaLambda0}
\eta_{\varepsilon}^0(y):=\int_{z\in{\mathbb{R}}^n}M(z)e^{\frac{U^0_{\varepsilon}(y+{\varepsilon}z)-U^0_{\varepsilon}(y)}{{\varepsilon}}}dz, \qquad \Lambda_{\varepsilon}^0(y):=\Lambda(y,\eta_{\varepsilon}^0(y)).$$ Another condition is also required on $\eta_{\varepsilon}^0$, see . Note that and ensures that $\rho_{\varepsilon}^0:=\iint_{{\mathbb{R}}_+\times{\mathbb{R}}^n} m_{\varepsilon}^0$ is uniformly bounded. Note also that assumption implies $\sup_{{\mathbb{R}}^n} U^0=0.$
### Assumptions on the distribution of phenotypes
The following assumptions only deal with the dependence of the coefficients on $y\in{\mathbb{R}}^n$. In particular, if all the coefficients have a compact dependence on $y$, then all the following assumptions are automatically satisfied.
First, we need a condition on the initial data, namely that, from , $\eta^0_{\varepsilon}(y)$ is bounded and $\Lambda_{\varepsilon}^0(y)$ is bounded and negative. More precisely, in accordance to section \[section:AssCoef\], we assume that there are two negative constants $\underline\Lambda\leq\overline\Lambda<0$ and two positive constants $0<\underline\eta\leq\overline\eta$, such that for all $y\in{\mathbb{R}}^n$, $$\label{AssumptionFWith}
\frac{1}{\overline{\eta}}\leq F(y,\underline{\Lambda})\leq F(y,\Lambda_{\varepsilon}^0(y)):=\frac{1}{\eta_{\varepsilon}^0(y)} \leq F(y,\overline{\Lambda})\leq\frac{1}{\underline{\eta}}.$$ This assumption implies $\underline{\eta}\leq\eta_{\varepsilon}^0(y)\leq\overline\eta$ and $\underline{\Lambda}\leq \Lambda^0_{\varepsilon}(y)\leq \overline{\Lambda}$ (since $\lambda\mapsto F(y,\lambda)$ is increasing). We will see in that those two inequalities hold for all times, namely $\underline{\eta}\leq\eta_{\varepsilon}(t,y)\leq\overline\eta$ and $\underline{\Lambda}\leq \Lambda(y,\eta_{\varepsilon}(t,y))\leq\overline{\Lambda}$. Note that with the notations $$\label{DefinitionBornes}
\frac{1}{\underline\eta(y)}:=F(y,\overline\Lambda),\quad
\frac{1}{\overline\eta(y)}:=F(y,\underline\Lambda),$$ assumption can be written $$\underline{\eta}\leq\underline\eta(y)\leq\eta_{\varepsilon}^0(y)\leq\overline\eta(y)\leq\overline{\eta}.$$ We stress out that this assumption implies $-\Lambda(y,\eta_{\varepsilon}(t,y))>0$. It means that every phenotype has a positive fitness, and is thus able to survive in absence of other phenotypes. This assumption is somehow restrictive, but it is not irrealistic, and allows to avoid some technicalities.\
Finally, the next assumptions deal with the derivatives of $\Lambda$. We need ${\partial}_\eta\Lambda$, $\nabla_y\Lambda$ and $\nabla_y{\partial}_\eta\Lambda$ to be bounded, and $\Lambda$ to be semi-convex. According to , we assume that there exists two constants $l,L>0$ such that for all $y\in{\mathbb{R}}^n$, $\lambda\in[\underline{\Lambda},\overline{\Lambda}]$, $$\label{AssumptionDerF}
l\leq {\partial}_\lambda F(y,\lambda)\leq L,$$ $$\label{AssumptionDerFy2}
\vert\nabla_yF(y,\lambda)\vert,\vert\nabla_y{\partial}_\lambda F(y,\lambda)\vert\leq L,$$ $$\label{AssumptionSemiConvexity}
\forall i\in\{1,\dots,n\},\quad{\partial}^2_{y_i} F(y,\lambda)\geq -L.$$
Definition and properties of $U_{\varepsilon}$ {#SecU}
==============================================
To make sense to the above heuristic, we first give a rigorous definition of the eigenelements $(\Lambda,Q)$, which only uses classical arguments. Then, we define $U_{\varepsilon}$, formally introduced in section \[SecFormalApproach\], and state some a priori estimates. We use those results to derive estimates on $\eta_{\varepsilon}$ and $\Lambda_{\varepsilon}$. Note that the study of $U_{\varepsilon}$ is autonomous and can be done separately from the analysis of the corrector. Finally, we pass to the limit, as ${\varepsilon}\to 0$, in the quantities $U_{\varepsilon}$, $\eta_{\varepsilon}$ and $\Lambda_{\varepsilon}$ to recover a viscosity solution of the Hamilton-Jacobi equation. The longer or more technical proofs are postponed to section \[sec:appendixConcentrationWithMutations\].
The eigenproblem and effective Hamiltonian
------------------------------------------
We consider the limiting problem and prove the existence of the eigenelements $(\Lambda,Q)$, along with some properties. The proof only uses elementary arguments and is postponned to section \[sec:AppendixEigenElements\]. introduced in section \[SecFormalApproach\], that we recall here: for fixed $(y,\eta)\in{\mathbb{R}}^n\times(0,+\infty)$, find $(\Lambda(y,\eta),Q(x,y,\eta))$, solution of $$\label{valeurpropre}
\left\{\begin{array}{ll}
{\partial}_x\left[A(x,y)Q(x,y,\eta)\right]+d(x,y)Q(x,y,\eta)=\Lambda(y,\eta)Q(x,y,\eta),\\[1mm]
A(0,y)Q(0,y,\eta)=\eta\int_{{\mathbb{R}}_+}{b(x',y)Q(x',y,\eta)dx'},\\[1mm]
Q(x,y,\eta)> 0, \quad \int_{{\mathbb{R}}_+} b(x,y) Q(x,y,\eta)dx=1,\ \forall (y,\eta)\in{\mathbb{R}}^n\times(0,+\infty).
\end{array}\right.$$
\[ThEigenElements\] Under the assumptions of section \[SecAssumptions\], for fixed $y\in{\mathbb{R}}^n$ and $\eta\in (\underline\eta(y),\overline\eta(y))$ (from ), there exists a unique couple $(\Lambda(y,\eta),Q(x,y,\eta))$ which satisfies .
Moreover, with $F$ defined in , $\underline{\Lambda}$ and $\overline{\Lambda}$ defined in , the eigenvalue $\Lambda(y,\eta)$ is continuously differentiable and it holds, for all $y\in{\mathbb{R}}^n,\; \eta\in (\underline\eta(y),\overline\eta(y))$, $$\label{FormuleImplicite}
F(y,\Lambda(y,\eta))=\frac{1}{\eta},$$ $$\label{Prerequis:BoundsOnLambda}
\underline{\Lambda}\leq\Lambda(y,\eta)\leq\overline{\Lambda}<0,$$ $$\label{DefinitionQ}
Q(x,y,\eta)= \eta\frac{1}{A(x,y)}\mathrm{exp}\left(\int_0^{x}{\frac{\Lambda(y,\eta)-d(x',y)}{A(x',y)}d x'}\right).$$
See section \[sec:AppendixEigenElements\].
Differentiating the relation , we immediately obtain the derivatives of $\Lambda$ $$\label{FormuleDeriveeLambda}
{\partial}_\eta \Lambda(y,\eta)=\frac{-1}{\eta^2{\partial}_\lambda F(y,\Lambda(y,\eta))} \leq \frac{-1}{\eta^2L} ,\qquad
\nabla_y \Lambda(y,\eta)=\frac{-\nabla_y F(y,\Lambda(y,\eta))}{{\partial}_\lambda F(y,\Lambda(y,\eta))}.$$ In particular, the property ${\partial}_\eta\Lambda<0$, turns out to be fundamental in the following section.
We also have some kind of concavity property for $\Lambda$. For all $y\in{\mathbb{R}}^n$, $\eta\in (\underline{\eta}(y),\overline{\eta}(y))$, we have $$\label{IntermediateInequality}
{\partial}_\eta^2\Lambda(y,\eta)+\frac{{\partial}_\eta\Lambda(y,\eta)}{\eta}\leq 0.$$
We use the short notations $F:=F(y,\Lambda(y,\eta))$ (where $F$ is defined in ) and $\Lambda:=\Lambda(y,\eta)$. Differentiationg the first relation in with respect to $\eta$, we find $${\partial}_\lambda^2F ({\partial}_\eta\Lambda)^2+{\partial}_\lambda F{\partial}_\eta^2 \Lambda=\frac{2}{\eta^3}, \qquad {\partial}_\eta^2 \Lambda=\frac{2}{\eta^3{\partial}_\lambda F}-\frac{1}{\eta^4({\partial}_\lambda F)^3}{\partial}_\lambda^2 F.$$ Combining it with the expression of ${\partial}_\eta \Lambda(y,\eta)$ in and , we deduce $${\partial}_\eta^2\Lambda+\frac{{\partial}_\eta\Lambda}{\eta}= \frac{1}{({\partial}_\lambda F)^3\eta^3}\left(({\partial}_\lambda F)^2-F{\partial}_\lambda^2F\right).$$ Using the Cauchy-Schwarz inequality and the definition of $F$ in , we have $$\label{Inequality2}
({\partial}_\lambda F)^2-F{\partial}_\lambda^2F\leq 0$$ and the proof of is completed.
Construction of $U_{\varepsilon}$ and a priori estimates
--------------------------------------------------------
We give a rigourous definition of $U_{\varepsilon}$, formally introduced in .
\[TheoremUeps\] Under the assumptions of section \[SecAssumptions\], for all ${\varepsilon}>0$ there exists a solution $U_{\varepsilon}(t,y)$ of . In addition, it satisfies, with $\underline{\Lambda},\overline{\Lambda}<0$ defined in , $$-\overline{\Lambda}\leq {\partial}_t U_{\varepsilon}(t,y)\leq-\underline{\Lambda}, \qquad \forall {\varepsilon}>0, \; t\geq0\; y\in{\mathbb{R}}^n.$$
See section \[sec:ConstructionU\].
In fact, $U_{\varepsilon}$ is the unique solution of with a locally bounded time derivative.
As a direct consequence of , we deduce the following useful corollary.
\[CorolaireImportant\] With $\underline{\Lambda},$ $\overline{\Lambda}$, $\underline{\eta}$, $\overline{\eta}$ defined in , and setting $$\label{DefinitionEtaLambda}
\eta_{\varepsilon}(t,y):=\int_{{\mathbb{R}}^n}M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{{\varepsilon}}}dz, \qquad
\Lambda_{\varepsilon}(t,y):=\Lambda(y,\eta_{\varepsilon}(t,y)),$$ we have $$\label{BoundOnEta}
\underline\eta\leq\eta_{\varepsilon}(t,y)\leq\overline{\eta} ,
\qquad
\underline\Lambda \leq\Lambda_{\varepsilon}(t,y)\leq\overline\Lambda<0 .$$
Simply use , $\Lambda_{\varepsilon}=-{\partial}_t U_{\varepsilon}$ and assumption .
Further estimates
-----------------
\[UepsLipschitz\] Under the assumptions of section \[SecAssumptions\], with $k^0$ defined in , $L,l>0$ in and $\underline{\eta}$ in , we have $$\vert \nabla_y U_{\varepsilon}(t,y)\vert \leq k^0+ \frac{L}{l\underline{\eta}^2}t, \qquad \forall {\varepsilon}>0, \; t\geq0, \; y\in{\mathbb{R}}^n.$$
See section \[sec:LipschitzEstimateEps\].
Note that the coefficient $\frac{L}{l\underline{\eta}^2}$ comes from a bound on $\vert \nabla_y\Lambda (y,\eta_{\varepsilon})\vert$. We will see that, at the limit when ${\varepsilon}\to0$, we can prove Lipschitz continuity globally in time.
We also need the following control of second order derivatives.
\[ThmSemiConvexity\] The function $U_{\varepsilon}$ is semi-convex in $(t,y)$, that is, for all the $\nu\in\mathbb{S}:=\{(t,y)\in{\mathbb{R}}^{n+1}: t^2+\vert y\vert^2=1\}$, ${\partial}^2_{\nu \nu}U_{\varepsilon}$ are bounded from below, uniformly in ${\varepsilon}>0$, $y\in{\mathbb{R}}^n$, locally uniformly in $t\geq0$. Therefore $U_{\varepsilon}$ belongs to $W^{2,1}_{loc}$ in $(t,y)$, uniformly in ${\varepsilon}>0$.
The idea is to use that the Hamiltonian has properties closely related to convexity, namely and to use the Lipschitz bounds. See section \[sec:W11estimate\].
The following corollary is essential when studying the corrector in section \[sec:StudyOfP\].
\[PropoW1estimate\] We have $\eta_{\varepsilon}\in W^{1,1}_{loc}$, uniformly in ${\varepsilon}>0$. In addition, denoting $\overline \eta(t):=\sup\limits_{y\in{\mathbb{R}}^n}\eta_{\varepsilon}(t,y)$, we have $$\int_0^T \left\vert {\partial}_t\overline \eta_{\varepsilon}(t) \right\vert dt\text{ is bounded uniformly in ${\varepsilon}>0$}, \qquad \forall T\geq0.$$
See section \[sec:ProofCorollary\].
Asymptotics
-----------
With these regularity properties, we are ready to establish the asymptotics of $U_{\varepsilon}$ when ${\varepsilon}$ vanishes.
\[TheoremU\] Under the assumptions of section \[SecAssumptions\], when ${\varepsilon}$ vanishes, $U_{\varepsilon}$ converges locally uniformly (and in $W_{loc}^{1,r}$, $1 \leq r < \infty$) to a function $U(t,y)\in W^{1,\infty}_{loc}$ which is a semi-convex viscosity solution of $$\label{HJU}
\left\{\begin{aligned}
&{\partial}_t U(t,y)=-\Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{\nabla_y U(t,y)\cdot z}dz\right),&\forall t>0,\ \forall y\in{\mathbb{R}}^n,\\
&U(0,y)=U^0(y), &\forall y\in{\mathbb{R}}^n.
\end{aligned}\right.$$ Moreover, under the assumption that $M$ is not degenerate (i.e., $M(\cdot)>0$ in a neighborhood of $0$), the function $U(t,y)$ is globally Lipschitzian.
See section \[sec:Asymptotics\_of\_U\] and section \[sec:APosterioriLipschitz\].
We also point out that, from Proposition 4.3 in [@Nordmann2018a], $$p\mapsto-\Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{p\cdot z}dz\right)\text{ is a convex mapping, }\forall y\in{\mathbb{R}}^n.$$ This class of Hamiltonian has been widely studied, and numerous results on regularity as well as representation formula are available [@Bianchini2012; @Fleming1975].
As a direct consequence of the $L^r_{loc}$ convergence of $\nabla U_{\varepsilon}$ to $\nabla U$, we have the following corollary.
\[CoroStrongCVEta\] When ${\varepsilon}\to0$, $\eta_{\varepsilon}$ converges in $L^r_{loc}$, $1 \leq r < \infty$, to $$\eta(t,y):=\int_{{\mathbb{R}}^n}M(z)e^{\nabla_yU\cdot z}dz.$$ Consequently, $\Lambda_{\varepsilon}(t,y):=\Lambda(t,\eta_{\varepsilon}(t,y))$ converges to $\Lambda(y,\eta(t,y))$ and $Q_{\varepsilon}(t,x,y):=Q(x,y,\eta_{\varepsilon}(t,y))$ to $Q(x,y,\eta(t,y))$ in $L^r_{loc}$, $1 \leq r < \infty$.
Asymptotics of the corrector - Proof of {#sec:StudyOfP}
========================================
We now turn to the main new results of the paper. These are boundedness from above and below and the asymptotics of the corrector $p_{\varepsilon}(t,x,y)$ defined through the factorisation , according to the definition of $U_{\varepsilon}$ in .
Estimates on $p_{\varepsilon}$ {#SecP}
------------------------------
This section is devoted to the proof of the first statement of , that is, $p_{\varepsilon}(t,x,y)$ and $\int_{x>0} p_{\varepsilon}(t,x,y)$ are bounded uniformly in ${\varepsilon}>0$, $x\geq0$, $y\in{\mathbb{R}}^n$, locally uniformly in $t\geq0$.
Our first result states a control of $p_{\varepsilon}$, from above and below, for $x\in[0,\overline{x}]$ (with $\overline x$ from ).
\[LemmaStrongBoundsOnP\] Under the assumptions of section \[SecAssumptions\], for any fixed $T>0$, there exists two constants $\underline{\gamma},\overline{\gamma}>0$ such that $$\underline{\gamma}\ Q_{\varepsilon}(t,x,y)\leq p_{\varepsilon}(t,x,y)\leq \overline{\gamma}\ Q_{\varepsilon}(t,x,y),$$ for all ${\varepsilon}>0$, $t\in[0,T]$, $y\in{\mathbb{R}}^n$, $x\in [0,\overline{x}]$, where $$Q_{\varepsilon}(t,x,y):=Q(x,y,\eta_{\varepsilon}(t,y)).$$
The function $p_{\varepsilon}(t,x,y)$ satisfies the following equation, for ${\varepsilon}>0,\ t>0,\ x>0,\ y\in{\mathbb{R}}^n$, $$\label{EquationOnp}
\left\{\begin{aligned}
&{\varepsilon}{\partial}_tp_{\varepsilon}+{\partial}_x\left[A(x,y)p_{\varepsilon}\right]+\left(d(x,y)-\Lambda_{\varepsilon}\right)p_{\varepsilon}=0,\\
&A(0,y)p_{\varepsilon}(t,0,y)=\iint\limits_{\substack{x>0,\ z\in{\mathbb{R}}^n}}M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{\epsilon}}b(x,y+{\varepsilon}z)p_{\varepsilon}(t,x,y+{\varepsilon}z)dx dz.
\end{aligned}\right.$$ and $Q_{\varepsilon}$ satisfies, for $(t, y)$ as parameters, the equation in the variable $x$ $$\left\{\begin{aligned}
&{\partial}_x[A(x,y)Q_{\varepsilon}]-(d(x,y)-\Lambda_{\varepsilon}(t,y) )Q_{\varepsilon}= 0,\\
&A(0,y)Q_{\varepsilon}(t,0,y)=\eta_{\varepsilon}(t,y).
\end{aligned}\right.$$ Setting $$\gamma_{\varepsilon}(t,x,y):=\frac{p_{\varepsilon}(t,x,y)}{Q_{\varepsilon}(t,x,y)},$$ we have $$\label{EquationOnV}
\left\{\begin{aligned}
&{\partial}_t \gamma_{\varepsilon}+\frac{A(x,y)}{{\varepsilon}}{\partial}_x \gamma_{\varepsilon}= -\frac{{\partial}_tQ_{\varepsilon}}{Q_{\varepsilon}} \gamma_{\varepsilon},\\
&\gamma_{\varepsilon}(t,0,y)=\iint\limits_{\substack{x>0,\ z\in{\mathbb{R}}^n}} J_{\varepsilon}(t,x,y,z) \gamma_{\varepsilon}(t,x,y+{\varepsilon}z)dz,
\end{aligned}\right.$$ where $$\label{DefinitionProbaKernelJ}
J_{\varepsilon}(t,x,y,z):= \frac{1}{\eta_{\varepsilon}}
M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{\epsilon}}b(x,y+{\varepsilon}z)Q_{\varepsilon}(t,x,y+{\varepsilon}z).$$ Our goal is to infer some bounds on $\gamma_{\varepsilon}$.
First, from the definition of $\eta_{\varepsilon}$ and the normalization $\int_{x>0}b(x,y)Q(x,y,\eta)dx=1$, we see that $J_{\varepsilon}$ is a probability kernel, $$\iint_{\substack{x>0,\ z\in{\mathbb{R}}^n}} J_{\varepsilon}(t,x,y,z) dxdz=1, \qquad \forall t \geq0, \; y \in {\mathbb{R}}^n.$$
We need to estimate $ \frac{{\partial}_tQ_{\varepsilon}}{Q_{\varepsilon}}.$ We compute, using the representation formula , $$\begin{aligned}
\frac{{\partial}_t Q_{\varepsilon}}{Q_{\varepsilon}}(t,x,y)
&= {\partial}_t\eta_{\varepsilon}(t,y) \frac{{\partial}_\eta Q}{Q}(x,y,\eta_{\varepsilon}(t,y))\\
&= {\partial}_t\eta_{\varepsilon}(t,y)\left(\frac{1}{\eta_{\varepsilon}(t,y)}+{\partial}_\eta\Lambda (y,\eta_{\varepsilon}(t,y))\int_0^x \frac{1}{A(x',y)}dx'\right).\end{aligned}$$ Because $\eta_{\varepsilon}$ is bounded from below, see , and ${\partial}_\eta\Lambda$ is bounded, see and assumption , we have $$\label{StepIntermediaire}
\left\vert \frac{{\partial}_t Q_{\varepsilon}}{Q_{\varepsilon}}(t,x,y)\right\vert \leq K\vert {\partial}_t\eta_{\varepsilon}(t,y)\vert\left(1+\int_0^x \frac{1}{A(x',y)}dx'\right),$$ for some constant $K>0$. Then, using assumption , we have, for $x\in[0,\overline{x}]$, $$\left\vert \frac{{\partial}_t Q_{\varepsilon}}{Q_{\varepsilon}}(t,x,y)\right\vert \leq K\vert {\partial}_t\eta_{\varepsilon}(t,y)\vert$$ for some constant still denoted by $K$. Setting $\overline \eta_{\varepsilon}(t):=\sup\limits_{y\in{\mathbb{R}}^n}\eta_{\varepsilon}(t,y)$, we define $$\underline \gamma_{\varepsilon}(t,x,y):= \gamma_{\varepsilon}(t,x,y) \exp\left(-K\int_0^t\vert {\partial}_t\overline\eta_{\varepsilon}(t')\vert dt'\right),$$ so that $\underline \gamma_{\varepsilon}$ is a subsolution to , namely $$\left\{\begin{aligned}
&{\partial}_t \underline \gamma_{\varepsilon}+\frac{A(x,y)}{{\varepsilon}}{\partial}_x\underline \gamma_{\varepsilon}\leq 0\\
&\underline \gamma_{\varepsilon}(t,0,y)\leq \iint\limits_{\substack{x>0,\ z\in{\mathbb{R}}^n}} J_{\varepsilon}(t,x,y,z) \underline \gamma_{\varepsilon}(t,x,y+{\varepsilon}z)dzdx
\end{aligned}\right.$$ (in fact, equality holds in the second line). From the comparison principle, we deduce $$\underline \gamma_{\varepsilon}(t,x,y)\leq \sup_{\substack{x\in[0,\overline x]\\ y\in{\mathbb{R}}^n}} \underline \gamma_{\varepsilon}(0,x,y)\leq \overline{\gamma}^0,$$ where $\overline \gamma^0$ comes from assumption . This gives a control from above on $ \gamma_{\varepsilon}(t,x,y) $ which implies $$p_{\varepsilon}(t,x,y)\leq \overline{\gamma}^0 Q_{\varepsilon}(t,x,y)\exp\left(K\int_0^t\vert \overline {\partial}_t\eta_{\varepsilon}(t')\vert dt'\right).$$ By , $\int_0^t \vert {\partial}_t\overline \eta_{\varepsilon}\vert$ is bounded uniformly in ${\varepsilon}>0$, therefore, for some constant $\overline{\gamma}$, $$p_{\varepsilon}(t,x,y)\leq \overline{\gamma} Q_{\varepsilon}(t,x,y).$$
Identically, we infer the bound from below, and the proof of is completed.
With the previous lemma in hand, we now estimate $p_{\varepsilon}$ for all $x\geq0$ (which is useless if $\overline x=+\infty$). We set $$\begin{gathered}
\overline{Q}(x,y):=\overline{\gamma}\frac{\overline\eta}{\underline{\eta}(y)}Q(x,y,\underline{\eta}(y))=\overline{\gamma}\frac{\overline{\eta}}{A(x,y)}\exp\left(\int_0^x\frac{\overline{\Lambda}-d(x',y)}{A(x',y)}dx'\right),\\
\underline{Q}(x,y):=\underline{\gamma}\frac{\underline\eta}{\overline{\eta}(y)}Q(x,y,\overline{\eta}(y))=\underline\gamma\frac{\underline{\eta}}{A(x,y)}\exp\left(\int_0^x\frac{\underline{\Lambda}-d(x',y)}{A(x',y)}dx'\right),\end{gathered}$$ where $\underline{\gamma},\overline{\gamma}$ are given by the previous lemma.
\[LemmaBoundOnP\] Under the same condition as in the previous lemma, we have $$\underline{Q}(x,y)\leq p_{\varepsilon}(t,x,y)\leq\overline{Q}(x,y), \qquad \forall {\varepsilon}>0, \; t\in[0,T], \; x\geq 0, \; y\in{\mathbb{R}}^n.$$
Note that, when restricted to $[0, \overline x]$ these bounds are weaker than in since $\frac{1}{\underline{\gamma}}\underline {Q}\leq Q_{\varepsilon}\leq \frac{1}{\overline{\gamma}}\overline{Q}$.
From , we deduce $$\underline{\gamma}\ \underline{\eta}\leq A(0,y)p_{\varepsilon}(t,0,y)\leq \overline{\gamma}\ \overline \eta.$$ Hence, on the one hand, we have $$\label{EquationOnpBIS}
\left\{\begin{aligned}
&{\varepsilon}{\partial}_tp_{\varepsilon}+{\partial}_x\left[A(x,y)p_{\varepsilon}\right]+\left(d(x,y)-\overline\Lambda\right)p_{\varepsilon}\leq 0,\\
&A(0,y)p_{\varepsilon}(t,0,y)\leq \overline{\gamma}\ \overline{\eta},
\end{aligned}\right.$$ for ${\varepsilon}>0,\ t\in[0,T],\ x>0,\ y\in{\mathbb{R}}^n$. On the other hand, $$\left\{\begin{aligned}
&{\varepsilon}{\partial}_t{\overline Q}+{\partial}_x\left[A(x,y){\overline Q}\right]+\left(d(x,y)-\overline\Lambda\right){\overline Q}= 0,\\
&A(0,y){\overline Q}(0,y)= \overline{\gamma}\ \overline{\eta}.
\end{aligned}\right.$$ From the comparison principle, we deduce $p_{\varepsilon}\leq {\overline Q}$. The lower bound can be proved similarily.
We are now ready to prove our first main result
\[thmCVWith\] Under the assumptions of section \[SecAssumptions\] and for any fixed $T>0$, $p_{\varepsilon}(t,x,y)$ and $\int_{x>0}p_{\varepsilon}(t,x,y)dx$ are bounded from above and below, uniformly in ${\varepsilon}>0$, $t\in[0,T]$, $x\geq0$, $y\in{\mathbb{R}}^n$.
The first point is deduced from and $\overline{Q}\leq \overline{\gamma}\frac{\overline{\eta}}{\underline{A}}$, with $\underline{A}$ from assumption .
To prove the second point, we only need to estimate the integrals of $\underline{Q}$ and $\overline{Q}$. Recalling $\overline\Lambda<0$, we compute $$\begin{aligned}
\int_0^{+\infty} \overline{Q}(x,y)dx
&\leq \int_0^{+\infty}\overline{\gamma} \frac{\overline{\eta}}{A(x,y)}\exp\left(\int_0^x\frac{\overline{\Lambda}}{A(x',y)}dx'\right)\\
&=\left[\overline{\gamma}\frac{\overline{\eta}}{\overline{\Lambda}}\exp\left(\int_0^x\frac{\overline{\Lambda}}{A(x',y)}dx'\right)\right]_{x=0}^{+\infty}\\
&=\overline{\gamma}\frac{\overline\eta}{-\overline \Lambda}\left[1-\exp\left(\int_0^{+\infty}\frac{\overline{\Lambda}}{A(x',y)}dx'\right)\right]
\leq \overline{\gamma}\frac{\overline \eta}{-\overline \Lambda},\end{aligned}$$ which proves the bound from above.
For the other inequality, we use the non-degeneracy assumption which implies $\int_{{\mathbb{R}}_+}Q(x,y,\eta)dx>\alpha$, for some $\alpha>0$.
Asymptotics of $p_{\varepsilon}$ {#sec:Asymptotics_of_P}
--------------------------------
We complete the proof of by showing that, when ${\varepsilon}\to0$, the mutations affect the equilibrium distribution $Q(x, y, \eta)$ only through a multiplicative factor $\gamma(t,y)$,
\[PropoAsymptoticsP\] For $t\in[0,T]$ (with $T>0$ fixed), $x\in[0,\overline{x}]$, $y\in{\mathbb{R}}^n$, and up to extraction of a subsequence when ${\varepsilon}\to0$, $p_{\varepsilon}$ converges in $L^\infty$ weak-$^\star$ to $\gamma (t,y) Q(x,y,\eta(t,y))$ with $\gamma(t,y)\in L^\infty$ which formally satisfies the equation $$\label{EquationVFinale}
\left\{\begin{aligned}
&{\partial}_t \gamma +{\partial}_\eta\Lambda \left(\int_{z\in{\mathbb{R}}^n} M(z) e^{\nabla U\cdot z} z dz\right)\cdot \nabla_y \gamma +{\partial}_t\eta \frac{{\partial}^2_\eta\Lambda}{2{\partial}_\eta\Lambda} \gamma=0,\\
&\gamma(t=0)=\gamma^0.
\end{aligned}\right.$$
The difficulty in stating rigorously equation is that ${\partial}_t\eta $ and $\int_{z\in{\mathbb{R}}^n} M(z) e^{\nabla U\cdot z} z dz$ are nothing more than bounded measures which is not smooth enough since $\gamma$ is just $L^\infty$. Nevertheless, we can prove establish the convergence of the full sequence $p_{\varepsilon}$ to $\gamma Q$ if $T$ is small enough: we use the regularity of the initial condition to rigourously pass to the limiting equation , and then use a standard uniqueness result on this equation.
Note also that, according to , the population concentrates on $\mathcal{S}$ where $U(t,y)$ achieved its maximum. From [@GB.BP:07], $U$ is differentiable at points of $\mathcal{S}$ and one has $\partial_t U= \nabla_y U = 0$ and $\eta =1$. Thus $p_{\varepsilon}$ converges to a multiple of $Q(x,y,1)$ on $\mathcal{S}$. In addition, if $M(\cdot)$ is even, then the drift term vanishes and the equation can be written $${\partial}_t [{\partial}_\eta\Lambda(y, \eta(t,y))^{1/2} \gamma ]=0.$$\
To prove , and establish we use the dual problem associated with ; for fixed $(y,\eta)\in{\mathbb{R}}^n\times(0,+\infty)$ consider $\Phi(x,y,\eta)$ the unique solution of $$\label{EquationDualeBis}
\left\{\begin{aligned}
&A(x,y){\partial}_x\Phi+\left[\Lambda(y,\eta)-d(x,y)\right]\Phi=-\eta b(x,y)\Phi(0,y,\eta),\quad\forall x>0,\\
&\int_{x>0}Q(x,y,\eta)\Phi(x,y,\eta)dx=1.
\end{aligned}\right.$$ We can solve this ordinary differential equation and find $$\label{FormulePhi}
\Phi(x,y,\eta)=-{\partial}_\eta\Lambda(y,\eta)\eta \int_x^{+\infty}\frac{b(x',y)}{A(x',y)}\exp\left(\int_x^{x'}\frac{\Lambda(y,\eta)-d(x'',y)}{A(x'',y)}d x''\right)dx'.$$ Also, multiplying equation by $\partial_\eta Q$ and integrating by parts, we get $$\label{FormuleIntegralePhi1}
\Phi(0,y,\eta)= -{\partial}_\eta\Lambda(y,\eta),$$ and, multiplying equation by $\partial^2_{\eta \eta} Q$, we end up with $$\label{FormuleIntegralePhi2}
\int_0^{+\infty} Q(x,y,\eta){\partial}_\eta \Phi(x,y,\eta)dx=-\frac{{\partial}^2_\eta\Lambda}{2{\partial}_\eta\Lambda}.$$
We fix $T>0$ and, throughout the proof, choose $t\in[0,T]$, $x\in[0,\overline{x}]$, $y\in{\mathbb{R}}^n$. We define $\gamma_{\varepsilon}:=\frac{p_{\varepsilon}}{Q_{\varepsilon}}$ and recall equation . From , we know that $\gamma_{\varepsilon}$ is bounded and thus converges (up to extraction of a subsequence) to some $\gamma$ in $L^\infty$ weak-$^\star$. Passing to the limit in , we deduce ${\partial}_x \gamma \equiv0$ (in the sense of distributions), and thus $\gamma$ only depends on $t$ and $y$. Since $Q_{\varepsilon}$ (strongly) converges to some $Q$ (see ), we deduce that $p_{\varepsilon}(t,x,y)$ converges (up to extraction of a subsequence) to $\gamma(t,y)Q(t,x,y)$ in $L^\infty$ weak-$^\star$.
We are now left with the task of identifying $\gamma$. To do so, we set $$E_{\varepsilon}(t,y):= \int_0^{+\infty} \gamma_{\varepsilon}(t,x,y)Q_{\varepsilon}(t,x,y)\Phi_{\varepsilon}(t,x,y) dx, \qquad
\Phi_{\varepsilon}(t,x,y):=\Phi(x,y,\eta_{\varepsilon}(t,y)).$$ Note that $\Phi_{\varepsilon}(t,x,y) \equiv 0$ for $x>\overline x$ and thus $E_{\varepsilon}(t,y):= \int_0^{\overline x} \gamma_{\varepsilon}Q_{\varepsilon}\Phi_{\varepsilon}$. From , and , we can write $$\begin{aligned}
{\varepsilon}{\partial}_t E_{\varepsilon}&=\int_{x>0} {\varepsilon}{\partial}_t \gamma_{\varepsilon}\; Q_{\varepsilon}\Phi_{\varepsilon}+{\varepsilon}\int_{x>0} \gamma_{\varepsilon}{\partial}_t[Q_{\varepsilon}\Phi_{\varepsilon}]\\
&= -\int_{x>0} A{\partial}_x \gamma_{\varepsilon}\; Q_{\varepsilon}\Phi_{\varepsilon}-{\varepsilon}\int_{x>0} \gamma_{\varepsilon}{\partial}_t Q_{\varepsilon}\Phi_{\varepsilon}+{\varepsilon}\int_{x>0} \gamma_{\varepsilon}{\partial}_t[Q_{\varepsilon}\Phi_{\varepsilon}]\\
&=[A \gamma_{\varepsilon}Q_{\varepsilon}\Phi_{\varepsilon}](x=0) + \int_{x>0} \gamma_{\varepsilon}{\partial}_x[AQ_{\varepsilon}\Phi_{\varepsilon}]+{\varepsilon}\int_{x>0} \gamma_{\varepsilon}Q_{\varepsilon}{\partial}_t\Phi_{\varepsilon}\\
\shortintertext{(from ${\partial}_x[AQ_{\varepsilon}\Phi_{\varepsilon}]=-\eta_{\varepsilon}b Q_{\varepsilon}\Phi_{\varepsilon}(x=0)$, and with the probability kernel $J_{\varepsilon}$ defined in~\eqref{DefinitionProbaKernelJ})}
&= \eta_{\varepsilon}\Phi_{\varepsilon}(x=0) \iint\limits_{x>0,\ z\in{\mathbb{R}}^n} J_{\varepsilon}(t,x,y,z)\left(\gamma_{\varepsilon}(t,x,y+{\varepsilon}z)-\gamma_{\varepsilon}(t,x,y)\right)dzdx+{\varepsilon}\int_{x>0} \gamma_{\varepsilon}Q_{\varepsilon}{\partial}_t\Phi_{\varepsilon}\end{aligned}$$ Recalling , we know that $\Phi_{\varepsilon}$ converges to some $\Phi$ when ${\varepsilon}\to0$. Dividing by ${\varepsilon}$ and passing to the limit ${\varepsilon}\to0$ (after extracting a subsequence, in the sense of distributions), we deduce $$\label{EntropyLimite}
{\partial}_t\left[ \gamma\int_{x>0} Q\Phi\right]={\partial}_t\gamma=\Phi(t,0,y)\left(\int_{z\in{\mathbb{R}}^n} M(z) e^{\nabla U\cdot z} z\right)\cdot \nabla_y \gamma+ \gamma\int_{x>0}Q{\partial}_t\Phi.$$
Injecting ${\partial}_t\Phi_{\varepsilon}(t,x,y) ={\partial}_t\eta_{\varepsilon}{\partial}_\eta\Phi(x,y,\eta_{\varepsilon})$ and - in , we end up with equation . From classical uniqueness results, we deduce that $\gamma_{\varepsilon}$ converges weakly to $\gamma$ for the whole sequence ${\varepsilon}\to0$ (and not for an extracted subsequence).
Concentration of the population density - proof of {#sec:Mutation_Proof}
===================================================
We now conclude on the consequences of our study on $U_{\varepsilon}$ and $p_{\varepsilon}$ with the concentration effect for the population density $m_{\varepsilon}$ itself.
Selection of the fittest phenotypes {#SecConcentration}
-----------------------------------
The following result states that the total population $\rho_{\varepsilon}$ is uniformly bounded and converges when ${\varepsilon}\to0$. Recalling , the two first statements of are direct consequences of the following proposition and the uniform $L^1$ estimate on $p_{\varepsilon}$ ().
\[ThmConcentration\] Under the assumptions of section \[SecAssumptions\],\
1. There exist two positive constants $\underline{\rho},\overline{\rho}>0$ such that $$\underline\rho\leq\rho_{\varepsilon}(t)\leq\overline{\rho},\quad\forall{\varepsilon}>0,\ t\geq 0.$$ In addition, $\rho_{\varepsilon}$ converges to some $\rho$ in the $L^\infty$-weak$^*$ topology.\
2. The integral $\int_{{\mathbb{R}}^n}e^{\frac{U_{\varepsilon}(t,y)- \int_0^t\rho_{\varepsilon}}{{\varepsilon}}}dy$ is bounded away from $0$, uniformly in ${\varepsilon}>0$, $t\geq0$.
Consequently, when ${\varepsilon}$ vanishes, we have $$\label{eqHJ_constrained}
\int_0^t\rho = \sup\limits_{y\in{\mathbb{R}}^n}U(t,y), \quad \forall t \geq 0 .$$
In addition, with $\mathcal{S}$ defined in , we have $$\begin{aligned}
\mathcal{S}
&:=\left\{t\geq0,\ y\in{\mathbb{R}}^n: U(t,y)=\sup_{{\mathbb{R}}^n} U(t,\cdot)\right\} =\left\{t\geq0,\ y\in{\mathbb{R}}^n: U(t,y)=\int_0^t\rho\right\} .\end{aligned}$$
T,he proof of becomes much simpler if we assume that there are $\underline{r},\overline{r}>0$ such that $$\underline{r}\leq b(x,y)-d(x,y)\leq\overline{r},\quad \forall x\geq0,\ y\in{\mathbb{R}}^n,$$ Indeed, integrating and using an integration by parts, we obtain $$\begin{aligned}
{\varepsilon}\frac{{\mathrm{d}}}{{\mathrm{d}}t}\rho_{\varepsilon}(t) &=\iint\limits_{{\mathbb{R}}^n\times{\mathbb{R}}_+}\left[\left(\frac{1}{{\varepsilon}^n}\int_{{\mathbb{R}}^n}{M(\frac{y'-y}{{\varepsilon}})d y}\right)b(x,y') -d(x,y')\right]m_{\varepsilon}(t,x,y')dxdy' -\rho_{\varepsilon}^2(t)\\
&\displaystyle\leq \overline{r}\rho_{\varepsilon}(t) -\rho_{\varepsilon}^2(t). \phantom{\int}\end{aligned}$$ which implies $0\leq\rho_{\varepsilon}(t)\leq\max\left({\overline{r}},\rho_{\varepsilon}^0\right)$ and provides us with an a priori upper bound on $\rho_{\varepsilon}$. With the same method, we also infer a positive lower bound on $\rho_{\varepsilon}$.
Then, using the uniform $L^1$ estimate on $p_{\varepsilon}$ from , we directly deduce .
We recall that $$\label{DefinitionRho}
\rho_{\varepsilon}(t)=\iint\limits_{x,y}p_{\varepsilon}(t,x,y)e^{\frac{U_{\varepsilon}(t,y)-\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}dxdy.$$ Multiplying by $e^{\int_0^t \frac{\rho_{\varepsilon}}{{\varepsilon}}}$ we have $$\rho_{\varepsilon}(t)e^{\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}=\int_{{\mathbb{R}}^n} e^{\frac{U_{\varepsilon}(t,y)}{{\varepsilon}}}\int_{{\mathbb{R}}_+}p_{\varepsilon}(t,x,y)dxdy,$$ and integrating over $(0,t)$, we deduce $${\varepsilon}\left(e^{\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}-1\right)=\int_0^t\int_{{\mathbb{R}}^n} e^{\frac{U_{\varepsilon}(t,y)}{{\varepsilon}}}\int_{{\mathbb{R}}_+} p_{\varepsilon}(t,x,y)dxdydt.$$ From $0< -\overline{\Lambda}\leq {\partial}_t U_{\varepsilon}\leq -\underline{\Lambda}$ () and the $L^1(dx)$ estimate on $p_{\varepsilon}$ (), we have $$\begin{aligned}
{\varepsilon}\left(e^{\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}-1\right)
&\geq \underline{I}\int_0^t\int_{{\mathbb{R}}^n} \frac{{\varepsilon}}{-\underline{\Lambda}}\frac{{\partial}_tU_{\varepsilon}(t,y)}{{\varepsilon}} e^{\frac{U_{\varepsilon}(t,y)}{{\varepsilon}}}dy\ dt\\
&\geq \frac{{\varepsilon}\underline{I}}{-\underline\Lambda}\int_{{\mathbb{R}}^n}e^{\frac{U_{\varepsilon}(t,y)}{{\varepsilon}}}-e^\frac{U^0_{\varepsilon}(y)}{{\varepsilon}} dy.\end{aligned}$$ Dividing by ${\varepsilon}e^{\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}$ on both sides we find $$\left(1-e^{-\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}\right)\geq \frac{ \underline{I}}{-\underline\Lambda}\int_{{\mathbb{R}}^n}\left(e^{\frac{U_{\varepsilon}(t,y)}{{\varepsilon}}}-e^\frac{U^0_{\varepsilon}(y)}{{\varepsilon}} \right)e^{-\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}dy,$$ that we rewrite, with $u_{\varepsilon}(t,y) = U_{\varepsilon}(t,y) - \int_0^t\rho_{\varepsilon}$ $$\int_{{\mathbb{R}}^n}e^{\frac{u_{\varepsilon}(t,y)}{{\varepsilon}}}dy\leq \frac{-\underline{\Lambda}}{\underline I}\left(1-e^{-\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}\right)+e^{-\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}\int_{{\mathbb{R}}^n}e^\frac{U^0_{\varepsilon}(y)}{{\varepsilon}} dy.$$ Then, from assumption , we deduce $$\int_{{\mathbb{R}}^n}e^{\frac{u_{\varepsilon}(t,y)}{{\varepsilon}}}dy\leq \frac{-\underline{\Lambda}}{\underline I}\left(1-e^{-\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}}\right)+\overline{J}^0e^{-\frac{\int_0^t\rho_{\varepsilon}}{{\varepsilon}}},$$ and $$\int_{{\mathbb{R}}^n}e^{\frac{u_{\varepsilon}(t,y)}{{\varepsilon}}}dy\leq \overline{K}:=\max\left(\frac{-\underline{\Lambda}}{\underline I}, \overline{J}^0\right).$$ Identically, we infer $$\int_{{\mathbb{R}}^n}e^{\frac{u_{\varepsilon}(t,y)}{{\varepsilon}}}dy\geq \underline{K}:= \min\left(\frac{-\overline{\Lambda}}{\overline I},\underline{J}^0\right)>0.$$
Now, from the definition of $\rho_{\varepsilon}$ (recalled in ) and the $L^1$ estimate on $p_{\varepsilon}$ () we find $$\underline I\ \underline K\leq \rho_{\varepsilon}(t)\leq \overline I\ \overline K.$$
Since $\rho_{\varepsilon}(t)$ is uniformly bounded, there exists a sequence ${\varepsilon}_k\to0$ such that $\rho_{{\varepsilon}_k}$ converges to some $\rho$ in the $\mathcal{L}^\infty$-weak$^*$ topology when $k\to+\infty$. Since, in addition, $U_{\varepsilon}$ converges locally uniformly to $U$ (), we deduce that $u_{{\varepsilon}_k}$ converges locally uniformly to some $u$.
Now, from $\underline{K}\leq\int_{{\mathbb{R}}^n}e^{\frac{u_{\varepsilon}(t,\cdot)}{{\varepsilon}}}\leq\overline{K}$, at the limit $k\to+\infty$, we have $$\label{LimsupNulle}
\forall t>0,\ \sup\limits_{y\in{\mathbb{R}}^n}u(t,y)= 0.$$ We deduce $$\int_0^t\rho=\sup_{y\in{\mathbb{R}}^n}U(t,y).$$ Therefore, $\int_0^t\rho$ does not depend on the extracted subsequence, and the convergence occurs for the whole sequence ${\varepsilon}\to0$, which achieves the proof.
Notice that the bound on $\rho(s)$ can be made more precise in the Cesaro sense $$\forall t>0,\quad -\overline{\Lambda}\leq\frac{1}{t}\int_0^t \rho(s)ds\leq-\underline{\Lambda}$$ since, from we know that $\int_0^t \rho=\sup_{{\mathbb{R}}^n} U(t,\cdot)\leq \sup_{{\mathbb{R}}^n} U^0(\cdot)-\underline{\Lambda}t=-\underline{\Lambda}t$ and the bound from below is similar.
Adaptive dynamics {#SecAdaptativeDynamics}
-----------------
We now prove the third statement of . We need further assumptions on the initial conditions, $$\label{initialu}
\exists! \ \bar y^0_{\varepsilon}\in{\mathbb{R}}^n,\quad U^0_{\varepsilon}(\bar y^0_{\varepsilon})=\max\limits_{y\in{\mathbb{R}}^n} U_{\varepsilon}^0(y)=0,$$ $$\label{initial_convergence_y}
\bar y^0_{\varepsilon}\text{ converges to some }\bar y^0\in{\mathbb{R}}^n\text{ when ${\varepsilon}$ vanishes},$$ $$\label{initial_concavity}
\nabla_y^2 U^0(\bar y^0)<0.$$
\[TheoremAD\] Under the assumptions of section \[SecAssumptions\] and –, for a short time interval $[0,T]$, there exists a unique $\bar y(t)\in{\mathbb{R}}^n$ on which $U(t,\cdot)$ reaches its maximum. Moreover, $t\mapsto\bar y(t)\in\mathcal{C}^1$ and satisfies the Canonical Equation $$\label{CanonicalEquation}
\left\{\begin{aligned}
&\frac{\mathrm{d}}{\mathrm{d}t}\bar y(t)=\left(\nabla_y^2U(t,\bar y(t))\right)^{-1}\cdot\nabla_y\Lambda(\bar y(t),1)+{\partial}_\eta\Lambda(\bar y(t),1)\int_{{\mathbb{R}}^n}M(z)z dz,\\
&\bar y(0)=\bar y^0.
\end{aligned}\right.$$.
Note that features a drift term ${\partial}_\eta\Lambda\int_{{\mathbb{R}}^n}M(z)zdz$. If the mutation kernel $M(\cdot)$ is even, this term vanishes and we recover the classical Canonical Equation. Let us also recall that uniqueness and regularity of a unique concentration point (monomorphism) is a hard questions with few progresses, see [@GB.BP:07; @MiRo2016; @CaLa].
Since $U$ (defined in ) satisfies with smooth initial datum, it is uniformly $C^2$ in the $y$ variable for short times $t\in[0,T]$, $T>0$ (this can be proved with the method of the characteristics, see Chapter 3.2 in [@LE:98]).
Consider such a time interval $[0,T]$, and $V\subset{\mathbb{R}}^n$ a neighborhood of $y^0$. We are interested in the solutions $(t,\bar y)\in [0,T]\times V$ of $$\label{EquationBut}
\nabla_y U(t,\bar y)=0.$$ From we know that at initial time there exists a unique solution $\bar y^0$ of . Besides, $\nabla_y^2 U^0(\bar y^0)$ is invertible. From the implicit functions theorem, there exists a unique $\bar y(t)\in{\mathbb{R}}^n$ satisfying , for $t$ in a certain time interval still denoted $[0,T]$. We can again choose a smaller $T$ to ensure that $\bar y(t)$ remains in $V$.
From , we can also choose $T$ and $V$ small enough to guarantee, $\forall t\in[0,T]$, $$\begin{gathered}
\bar y(t)\in V,\quad U(t,\cdot) \text{ is striclty concave in }V,\\
\max_{y\in V} U(t,y)=\max_{y\in{\mathbb{R}}^n} U(t,y).\end{gathered}$$ Hence, for all $t\in[0,T]$, the solution $\bar y(t)$ of must satisfy $$\label{PropMaximum}
U(t,\bar y(t))=\max_{y\in V} U(t,y)=\max_{y\in{\mathbb{R}}^n} U(t,y)=\int_0^t \rho,$$ which proves the first part of the proposition and that $t\mapsto \bar y (t)$ is $\mathcal{C}^1$.
For the Canonical Equation, we differentiate with respect to $t$, and noting that $$\eta(t,\bar y(t))=\int_{{\mathbb{R}}^n}M(z)e^{\nabla_yU(t,\bar y(t))\cdot z}dz =1,$$ we obtain $$\begin{aligned}
0
&=\frac{\mathrm{d}}{\mathrm{d}t}\left[\nabla_y U(t,\bar y (t))\right]\\
&=-\nabla_y\Lambda(\bar y (t),1)-\nabla_y^2 U(t,\bar y (t))\cdot{\partial}_\eta\Lambda(\bar y(t),1)\int_{{\mathbb{R}}^n}M(z)zdz+\nabla_y^2 U(t,\bar y (t))\cdot \dot{\bar y}(t),
\end{aligned}$$ and follows, for $t\in[0,T]$.
Under the same assumptions, $t\mapsto \rho(t)\in C^1([0,T])$ and for all $t\in[0,T]$, $$\rho(t)=-\Lambda(\bar y(t),1)$$ and $$\frac{\mathrm{d}}{\mathrm{d}t}\rho(t)=-\nabla_y\Lambda\cdot \nabla_y^2 U\cdot \nabla_y\Lambda -{\partial}_\eta\Lambda \left(\int_{{\mathbb{R}}^n} M(z) z dz\right)\cdot \left(\nabla_y^2 U\right)^2\cdot \nabla_y \Lambda.$$ where $\nabla_y^2 U$ is evaluated in $(t,\bar y(t))$, and the derivatives of $\Lambda$ in $(\bar y(t),1)$.
In particular, if $M(\cdot)$ is even, then $\frac{\mathrm{d}}{\mathrm{d}t}\rho(t)\geq 0$ and $-\overline{\Lambda}\leq \rho\leq -\underline{\Lambda}$.
Follows from ${\partial}_t U(t,\bar y (t))=\rho(t)$ and $\frac{\mathrm{d}}{\mathrm{d}t}\left[{\partial}_t u(t,\bar y (t))\right]=\frac{\mathrm{d}}{\mathrm{d}t}\rho(t).$
The limitation of to a short time interval is merely due to three independant phenomena. First, the possible loss of concavity, or apparition of singularities for $U$, coming from the Hamilton-Jacobi equation . Secondly, the possible “jump” of the point where $U$ reaches its maximum, contradicting $\max_{y\in V} U(t,y)=\max_{y\in{\mathbb{R}}^n} U(t,y)$ in . Finally, the possible blow-up in finite time of $\bar y(t)$ from the dynamics of the Canonical Equation . Regarding the last point, we point out that $\Lambda$ can sometimes be used as a Lyapunov function. Indeed, we have $$\begin{aligned}
&\frac{{\mathrm{d}}}{{\mathrm{d}}t}\left[\Lambda(\bar y(t),1)\right]=\nabla_y\Lambda(\bar y(t),1)\cdot\dot{\bar y}(t)\\
&=\nabla_y\Lambda(\bar y(t),1)\cdot\nabla_y^2U^{-1}\cdot\nabla_y\Lambda(\bar y(t),1)+{\partial}_\eta\Lambda(\bar y(t),1)\nabla_y\Lambda(\bar y(t),1)\cdot\int_{{\mathbb{R}}^n}M(z)z dz\\
&\leq {\partial}_\eta\Lambda(\bar y(t),1)\nabla_y\Lambda(\bar y(t),1)\cdot\int_{{\mathbb{R}}^n}M(z)z dz.
$$ In particular, if $M(\cdot)$ is even, then $\frac{{\mathrm{d}}}{{\mathrm{d}}t}\left[\Lambda(\bar y(t),1)\right]\leq0$. Thus, if $\bar y^0$ belongs to a “well” of $\Lambda$, then $\bar y(t)$ remains “trapped”, which prevents from an blow-up in finite time. It also implies, at least formally, that $\bar y(t)$ converges to a local minimum of $\Lambda(\cdot,1)$ when $t\to+\infty$.
Construction, estimates, and asymptotics of $U_{\varepsilon}$ - Proof of {#sec:appendixConcentrationWithMutations}
=========================================================================
The eigenproblem - proof of {#sec:AppendixEigenElements}
----------------------------
An immediate calculation on gives the explicit solution for $Q$ in terms of $\Lambda$. Multiplying by $b(x,y)$ and integrating in $x$, we obtain formula . Next, from we have, for $y\in{\mathbb{R}}^n,\ \forall\eta\in (\underline\eta(y),\overline\eta(y))$ $$F(y,\underline\Lambda)\leq\frac{1}{\eta}\leq F(y,\overline\Lambda).$$ As ${\partial}_\lambda F>0$, we conclude the existence and uniqueness of $\Lambda(y,\eta)$ as the unique solution of . Now, using , we obtain existence and uniqueness of $Q$.
Finally, the bounds follow from ${\partial}_\lambda F>0$ and $$F(y,\underline\Lambda)\leq F(y,\Lambda(y,\eta))\leq F(y,\overline\Lambda).$$
For later purpose, we also neeed the following stronger version of .
\[IntermediateInequalitySTRONG\] There exists $\delta>0$ such that, for all $y\in{\mathbb{R}}^n$, $\eta\in (\underline{\eta}(y),\overline{\eta}(y))$, we have $${\partial}_\eta^2\Lambda(y,\eta)+\frac{{\partial}_\eta\Lambda(y,\eta)}{\eta}\leq -\delta.$$
From the proof of , we have $${\partial}_\eta^2\Lambda+\frac{{\partial}_\eta\Lambda}{\eta}= \frac{1}{({\partial}_\lambda F)^3\eta^3}\left(({\partial}_\lambda F)^2-F{\partial}_\lambda^2F\right).$$ Since $\frac{1}{({\partial}_\lambda F)^3\eta^3}\geq \frac{1}{L^3\overline{\eta}^3}$, our goal is to show that $$({\partial}_\lambda F)^2-F{\partial}_\lambda^2F\leq -\delta,$$ for all $y\in{\mathbb{R}}^n$, $\eta\in (\underline{\eta}(y),\overline{\eta}(y))$.
We set, for all $x\geq0$, $y\in{\mathbb{R}}^n$, $\lambda\in{\mathbb{R}}$, $$f(x,y,\lambda):=\frac{b(x,y)}{A(x,y)}\exp\left(\int_0^x\frac{\lambda-d(x',y)}{A(x',y)}dx'\right).$$ According to , we have $F(y,\lambda)=\int_{{\mathbb{R}}_+}f(x,y,\lambda)dx$. We also define the probability measure $\tilde f(x,y,\lambda):=\frac{f(x,y,\lambda)}{F(y,\lambda)}.$ Now, setting $\mathcal{A}(x,y):=\int_0^x\frac{1}{A(x',y)}dx'$, we have $${\partial}_\lambda F(y,\lambda)= \int_{{\mathbb{R}}_+}\mathcal{A}(x,y)f(x,y,\lambda)dx,\quad {\partial}^2_\lambda F(y,\lambda)= \int_{{\mathbb{R}}_+}\mathcal{A}(x,y)^2f(x,y,\lambda)dx.$$ It gives, $$({\partial}_\lambda F)^2-F{\partial}_\lambda^2F=-F^2\int_{{\mathbb{R}}_+}\left(\mathcal{A}(x,y)-\int_{{\mathbb{R}}_+}\mathcal{A}\tilde f dx\right)^2\tilde{f}dx$$ (the function are evaluated on $y$ and $\lambda=\Lambda(y,\eta)$). We have $F^2\leq \frac{1}{\underline{\eta}^2}$, and assumption implies that the above term is negative uniformly in $y\in{\mathbb{R}}^n$, $\eta\in (\underline{\eta}(y),\overline{\eta}(y))$.
Construction of $U_{\varepsilon}$ - proof of {#sec:ConstructionU}
---------------------------------------------
We present a proof of existence based on a regularization argument. For the constrained Hamilton-Jacobi equation, a fixed point method has also been proposed in [@KimACAP2019]. Our proof is divided into three parts. First, we construct $U_{\varepsilon}$ on a truncated problem. Then, we prove a uniform a priori estimate on ${\partial}_t U_{\varepsilon}$, which allows finally to remove the truncation.
#### Extending $\Lambda$.
defines $\Lambda(y,\eta)$ only for $y\in{\mathbb{R}}^n$ and $\eta\in\left(\underline{\eta}(y),\overline{\eta}(y)\right)$. We first need to artificially extend $\Lambda(y,\eta)$ for $\eta\in(0,+\infty)$. For $y\in{\mathbb{R}}^n$, we set $$\tilde\Lambda(y,\eta)=
\begin{cases}
\underline{\Lambda}-\underline{B_y}(\eta) &if $\eta<\underline\eta(y)$,\\
\Lambda(y,\eta)&if $\underline\eta(y)\leq\eta\leq\overline\eta(y)$,\\
\overline{\Lambda}+\overline{B_y}(\eta) &if $\eta>\overline\eta(y)$,
\end{cases}$$ where $\underline{B_y}$ and $\overline{B_y}$ are chosen to be positive, increasing, bounded by $1$, and such that $\tilde\Lambda$ is smooth. Note that the extension of $\Lambda$ is completely arbitrary, but we will show, a posteriori, that $\eta_{\varepsilon}\in\left(\underline{\eta}(y),\overline{\eta}(y)\right)$. We consider the following problem $$\label{TruncatedProblem}
\left\{
\begin{aligned}
&{\partial}_t \tilde U_{\varepsilon}(t,y)=-\tilde\Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{\frac{\tilde U_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U_{\varepsilon}(t,y)}{{\varepsilon}}}dz\right), &\forall t\geq0,\ \forall y\in{\mathbb{R}}^n,\\
&\tilde U_{\varepsilon}(0,y)=u_{\varepsilon}^0(y), &\forall y\in{\mathbb{R}}^n.
\end{aligned}\right.$$
#### Solution for the truncated problem.
For a fixed $R>0$, we consider a truncation function $\phi_R:{\mathbb{R}}\to{\mathbb{R}}$ which is smooth, increasing and satisfies the following conditions:
- $\phi_R(r)= r \text{ for }r \in[-\frac{R}{2},\frac{R}{2}]$,
- $\phi_R(r)=R \text{ for }r \geq 2R$,
- $\phi_R(r)= -R \text{ for } r \leq -2R$,
- $\phi_R'\geq 0$ is uniformly bounded.
For ${\varepsilon}>0$, we consider the Cauchy problem $$\label{eqRter}
\left\{
\begin{aligned}
&{\partial}_t \tilde U^R_{\varepsilon}(t,y)=\phi_R\left(-\tilde\Lambda\left(y,\int_{{\mathbb{R}}^n}{M(z)e^{\frac{\tilde U^R_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U^R_{\varepsilon}(t,y)}{{\varepsilon}}}d z}\right)\right),\\
&\tilde U^R_{\varepsilon}(0,\cdot)=U^0_{\varepsilon}.
\end{aligned}\right.$$ for which the classical Cauchy-Lipschitz theorem provides existence and uniqueness of a solution $\tilde U^R_{\varepsilon}$, defined globally in time.
#### Estimate on the time derivative.
\[BoundOnD\_tUBIS\] We have, with $
{\partial}_t U^0_{\varepsilon}:=-\Lambda\left(y,\eta_{\varepsilon}^0(y)\right)
$ and $\eta_{\varepsilon}^0$ is defined in , $$\inf\limits_{y\in{\mathbb{R}}^n} {\partial}_t U^0_{\varepsilon}(y)\leq{\partial}_t \tilde U^R_{\varepsilon}(t,y)\leq \sup\limits_{y\in{\mathbb{R}}^n} {\partial}_t U^0_{\varepsilon}(y),\quad \forall {\varepsilon}>0,\ t>0,\ y\in{\mathbb{R}}^n .$$
The full proof of this statement, which is technical, can be found in [@Nordmann2018a] (proof of Proposition 4.7, Appendix D). We give the formal idea of the method.
Let us fixe ${\varepsilon}>0$, $R>0$ and set $V(t,y):={\partial}_t \tilde U^R_{\varepsilon}(t,y).$ Differentiating with respect to $t$, we obtain $$\label{formal_timebis}
{\partial}_tV(t,y)=\int_{{\mathbb{R}}^n} K (t,y,z) \left(\frac{V(t,y+{\varepsilon}z)-V(t,y)}{{\varepsilon}}\right) dz,$$ where $K(t,y,z):=-\phi'_R\ {\partial}_\eta\tilde\Lambda\ M(z)e^{\frac{\tilde U^R_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U^R_{\varepsilon}(t,y)}{{\varepsilon}}} $. Since ${\partial}_\eta\Lambda<0$, we have $K\geq0$. Then, if for some $t>0$, $V(t, \cdot)$ reaches its maximum at $\bar y\in{\mathbb{R}}^n$, we obtain the inequality $${\partial}_t V(t, \bar y) =\int_{{\mathbb{R}}^n} K (t,\bar y,z) \left(\frac{V(t,\bar y+{\varepsilon}z)-V(t,\bar y)}{{\varepsilon}}\right) dz \leq 0.$$ Formally, it shows that the maximum value of $V$ is decreasing with time, that is, $$\sup_y V(t, y) \leq \sup_y V(0, y)= \sup_{y}{\partial}_tU^{0}_{\varepsilon}.$$ With the same method we show ${\inf_yV\geq\inf_y{\partial}_tU^{0}_{\varepsilon}}$, which conclude the proof of .\
Hereafter, from assumption and ${\partial}_\lambda F>0$, we have $$\label{BoundD_tU^0}
-\overline{\Lambda}\leq {\partial}_tU^0_{\varepsilon}(y)\leq -\underline{\Lambda}.$$ Using , we infer $$\label{EstimateIntermediaire}
-\overline{\Lambda}\leq\phi_R\left(-\tilde \Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{\frac{\tilde U^R_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U^R_{\varepsilon}(t,y)}{\epsilon}}\right)\right)\leq-\underline{\Lambda}.$$
#### Removing the truncation.
From and the choice of $\phi_R$, for $R$ large enough, we have $$-\phi_R\left(\tilde \Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{\frac{\tilde U^R_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U^R_{\varepsilon}(t,y)}{\epsilon}}\right)\right)=\tilde \Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{\frac{\tilde U^R_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U^R_{\varepsilon}(t,y)}{\epsilon}}\right).$$ Besides, since ${\partial}_\eta\tilde\Lambda<0$, we have $$\label{EstimateIntermediaire2}
\underline{\eta}(y)\leq \int_{{\mathbb{R}}^n}M(z)e^{\frac{\tilde U^R_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U^R_{\varepsilon}(t,y)}{\epsilon}}\leq \overline{\eta}(y),$$ for all $R$ large enough, ${\varepsilon}>0$, $t\geq0,\ y\in{\mathbb{R}}^n$. Thus, from the definition of $\tilde{\Lambda}$ in , we have $$-\tilde\Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{\frac{\tilde U^R_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U^R_{\varepsilon}(t,y)}{{\varepsilon}}}dz\right)=-\Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{\frac{\tilde U^R_{\varepsilon}(t,y+{\varepsilon}z)-\tilde U^R_{\varepsilon}(t,y)}{{\varepsilon}}}dz\right),$$ that is, $\tilde U_{\varepsilon}^R$ is a solution of . The proof is thereby achieved.
A priori Lipschitz estimate - proof of {#sec:LipschitzEstimateEps}
---------------------------------------
We follow the same idea as for . However, there are some technical difficulties. First, we have to deal with a “source term” $\nabla_y\Lambda\left(y,\eta_{\varepsilon}(t,y)\right)$ which is bounded by a constant $\frac{L}{l\underline{\eta}^2}$, using , , and . In addition, we first need to prove the estimate on a truncated function, then to remove the truncation. We fix $i\in\{1,\dots,n\}$, $T>0$, and we set $$\begin{aligned}
W_{\varepsilon}(t,y):= {\partial}_{y_i}{U}_{\varepsilon}(t,y) .
\end{aligned}$$ Differentiating , we obtain $$\label{eqRDbis}
\begin{aligned}
&{\partial}_t{W}_{\varepsilon}(t,y)\\
&=-{\partial}_{y_i}\Lambda(y,\eta_{\varepsilon})-{\partial}_\eta\Lambda(y,\eta_{\varepsilon})\left(\int M(z)e^{\frac{{U}_{\varepsilon}(t,y+{\varepsilon}z)-{U}_{\varepsilon}(t,y)}{{\varepsilon}}}\left[\frac{{W}_{\varepsilon}(t,y+{\varepsilon}z)-{W}_{\varepsilon}(t,y)}{{\varepsilon}}\right]dz\right)\\
&:=\mathcal{F}(t,y,W_{\varepsilon}(t,\cdot)).
\end{aligned}$$ We formally define a truncated problem, for $R>0$, and its solution $W^R_{\varepsilon}$ satisfying $$\label{troncatureS}
\begin{aligned}
&W^R_{\varepsilon}(t,y)=\phi_R\left({\partial}_{y_i}U^0_{\varepsilon}(y)+\int_0^t \mathcal{F}(s,y,{W}_{\varepsilon}^R(s,\cdot))ds\right),\\
\end{aligned}$$ where $\mathcal{F}$ is defined above and $\phi_R$ is a truncation function as in . We can prove existence and uniqueness of a global solution of by a direct application of the Cauchy-Lipschitz theorem.\
We set $\bar W^R_{\varepsilon}:=W^R_{\varepsilon}-Ct$ with $C:=\frac{L}{l\underline{\eta}^2}$. Our goal is to show $$\label{objectifbis}
\begin{aligned}
& \forall t\in[0,T],\forall y\in{\mathbb{R}}^n, & \bar W^R_{\varepsilon}(t,y)\leq\sup {\partial}_{y_i}U_{\varepsilon}^0.
\end{aligned}$$ By contradiction, assume does not hold, i.e., there exist ${y_0\in{\mathbb{R}}^n},{t_0\in[0,T]}$ such that $$\label{absurd2'}
\bar W^R_{\varepsilon}(t_0,y_0)-\sup {\partial}_{y_i}U_{\varepsilon}^{0}>0.$$ For $\beta>0$, $\alpha>0$ small enough, $t \in [0,T], y\in{\mathbb{R}}^n$ we introduce $$\varphi_{\alpha,\beta}(t,y):=\bar W^R_{\varepsilon}(t,y)- \alpha t-\beta\vert y-y_0\vert.$$ As $\bar W^R_{\varepsilon}$ is bounded, $\varphi_{\alpha,\beta}$ reaches its maximum on $[0,T]\times{\mathbb{R}}^n$ at a point $(\bar t, \bar y).$ We have $$\forall z\in{\mathbb{R}}^n,\ \varphi_{\alpha,\beta}(\bar t,\bar y+ {\varepsilon}z)\leq\varphi_{\alpha,\beta}(\bar t,\bar y).$$ Then, we obtain the inequality $$\label{accroissement}
\forall z\in{\mathbb{R}}^n,\ \frac{\bar W^R_{\varepsilon}( \bar t, \bar y+{\varepsilon}z)-\bar W^R_{\varepsilon}( \bar t, \bar y)}{{\varepsilon}}\leq\beta\frac{\vert \bar y +z-y_0\vert-\vert \bar y -y_0\vert}{{\varepsilon}}\leq\beta \vert z\vert.$$
We choose $\alpha$ small enough so that $\varphi_{\alpha,\beta}(t_0,y_0)>\varphi_{\alpha,\beta}(0,y_0)= {\partial}_{y_i}U_{\varepsilon}^{0}(y_0),$ which is possible thanks to . It implies $\bar t>0$. Hence ${{\partial}_t\varphi_{\alpha,\beta}(\bar t,\bar y)\geq0}$, i.e. ${{\partial}_t\bar W^R( \bar t, \bar y)\geq\alpha}$ (if $\bar t=T$, then ${\partial}_t$ stands for the left derivative). Differentiating at $(\bar t,\bar y)$, we have $$\begin{aligned}
\alpha&\leq{\partial}_t\bar W^R_{\varepsilon}(\bar t,\bar y)\\
&\leq -\sup\phi'_R\times{\partial}_{y_i}\Lambda(y,\eta_{\varepsilon}(t,y))- C\\
&\quad +\sup\phi'_R\times\left(-{\partial}_\eta\Lambda(y,\eta)\right)\int M(z)e^{\frac{ U_{\varepsilon}(\bar t,\bar y+{\varepsilon}z)- U(\bar t,\bar y)}{{\varepsilon}}}\left[\frac{W^R_{\varepsilon}(\bar t,\bar y+{\varepsilon}z)-W^R_{\varepsilon}(\bar t,\bar y)}{{\varepsilon}}\right]dz.
\end{aligned}$$ Now, from $\vert{\partial}_{y_i}\Lambda(y,\eta_{\varepsilon}(t,y))\vert\leq\frac{L}{l\underline{\eta}^2}= C$ and $0\leq-{\partial}_\eta\Lambda(y,\eta)\leq\frac{L}{\underline{\eta}^2}$, we have $$\begin{aligned}
\alpha&\leq \frac{L}{\underline{\eta}^2}\left(\int M(z)e^{\frac{ U_{\varepsilon}(\bar t,\bar y+{\varepsilon}z)- U(\bar t,\bar y)}{{\varepsilon}}}\vert z\vert{\mathrm{d}}z\right)\times \beta\\
&\leq \frac{L}{\underline{\eta}^2} \left(\int{M(z)e^{\frac{-2\underline\Lambda T}{{\varepsilon}}+k_0\vert z\vert}\vert z\vert{\mathrm{d}}z}\right)\times \beta.
\end{aligned}$$ Then, passing to the limit $\beta\to 0$ we obtain ${\alpha\leq0}$: contradiction. Thus, we have $$\bar W_{\varepsilon}^{R}\leq\sup\vert {\partial}_{y_i}U_{\varepsilon}^{0}\vert=k^0.$$ We proceed similarily to obtain the reverse inequality $\bar W_{\varepsilon}^{R}\geq k^0$. We have, for all $R>0,\ {\varepsilon}>0,\ t\in[0,T],\ y\in{\mathbb{R}}^n$ $$\vert W^R_{\varepsilon}(t,y)\vert \leq k^0+Ct.$$
Finally, the bound on $W^R_{\varepsilon}$ is uniform in $R$ so we can remove the truncation, as detailed in section \[sec:ConstructionU\]. Thus, $W^R_{\varepsilon}=W_{\varepsilon}$ for $R$ large enough and $$\vert {\partial}_{y_i} U_{\varepsilon}(t,y)\vert \leq k^0+Ct.$$
Semi-convexity - proof of {#sec:W11estimate}
--------------------------
For convex Hamiltonian, the semi-convexity of the solution is a classical matter, [@GB:94; @Cannarsa2004]. Here, we have to deal with a nonlocal operator which features a difference rather than a gradient
#### Semi-convexity in $t$.
For shorter formulas, we need some notations $$V_{\varepsilon}:= {\partial}_t^2U_{\varepsilon}, \qquad J_{\varepsilon}(t,y,z):=M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{{\varepsilon}}}.$$ We begin with two results that are used later and express some properties usually connected to the convexity of the Hamiltonian. Firstly we observe that
\[LemmaIntermediaire2\] For all ${\varepsilon}>0$, $t>0$, $y\in{\mathbb{R}}^n$, we have $$({\partial}_t\eta_{\varepsilon})^2\leq \eta_{\varepsilon}\int_{{\mathbb{R}}^n}\left(\frac{{\partial}_tU_{\varepsilon}(t,y+{\varepsilon}z)-{\partial}_t U_{\varepsilon}(t,y)}{{\varepsilon}}\right)^2J_{\varepsilon}dz.$$
Use Jensen’s inequality and the definition of $\eta_{\varepsilon}$ in .
Next, we prove that $$\label{InequationV}
{\partial}_t V_{\varepsilon}\geq -{\partial}_\eta\Lambda(y,\eta_{\varepsilon}) \int_{{\mathbb{R}}^n}\frac{V_{\varepsilon}(t,y+{\varepsilon}z)-V_{\varepsilon}(t,y)}{{\varepsilon}}J_{\varepsilon}dz.$$
Differentiating twice, we find $$\label{EquationOnD_T2U}
\begin{aligned}
{\partial}_t V_{\varepsilon}=
&-{\partial}_\eta\Lambda \int_{{\mathbb{R}}^n}\frac{V_{\varepsilon}(t,y+{\varepsilon}z)-V_{\varepsilon}(t,y)}{{\varepsilon}}J_{\varepsilon}dz\\
&-{\partial}^2_{\eta}\Lambda\left({\partial}_t\eta_{\varepsilon}\right)^2
-{\partial}_\eta\Lambda\int_{{\mathbb{R}}^n}\left(\frac{{\partial}_tU_{\varepsilon}(t,y+{\varepsilon}z)-{\partial}_t U_{\varepsilon}(t,y)}{{\varepsilon}}\right)^2J_{\varepsilon}dz.
\end{aligned}$$ Next, combining with , we find $$\label{LemmaHamiltonienConvex}
{\partial}^2_{\eta}\Lambda(y,\eta_{\varepsilon})\left({\partial}_t\eta_{\varepsilon}\right)^2
+{\partial}_\eta\Lambda(y,\eta_{\varepsilon})\int_{{\mathbb{R}}^n}\left(\frac{{\partial}_tU_{\varepsilon}(t,y+{\varepsilon}z)-{\partial}_t U_{\varepsilon}(t,y)}{{\varepsilon}}\right)^2J_{\varepsilon}dz
\leq0,$$ Using the above inequality and , we find .
From inequality , we deduce
\[lemmaSemiConvex\] $V_{\varepsilon}$ is uniformly bounded from below and more precisely, with $V_{\varepsilon}^0(y):=V_{\varepsilon}(t=0,y)$, we have $$\label{objectif}
V_{\varepsilon}(t,y)\geq\inf_{y\in {\mathbb{R}}^n}V^0_{\varepsilon}(y)> -\infty,\quad \forall {\varepsilon}>0,\ \forall t>0,\ \forall y \in {\mathbb{R}}^n.$$
The proof follows closely the method of section \[sec:LipschitzEstimateEps\]. The formal idea is the following. If, for some $t>0$, $V_{\varepsilon}(t, \cdot)$ reaches its minimum at $\bar y\in{\mathbb{R}}^n$, from we obtain ${\partial}_t V_{\varepsilon}(t, \bar y)\geq0$. Formally, it shows that the minimum value of $V_{\varepsilon}$ is increasing with time, that is, $\inf_y V_{\varepsilon}(t, y) \geq \inf_y V_{\varepsilon}(0, y)$. Then, we conclude with the fact that $\inf_y V_{\varepsilon}(0, y)$ is bounded, uniformly in ${\varepsilon}>0$.
Differentiating in $t$, we obtain $$\label{LinkVEtaExplicite}
V_{\varepsilon}(t,y)=-{\partial}_\eta\Lambda(y,\eta_{\varepsilon})\int_{{\mathbb{R}}^n}\frac{\Lambda_{\varepsilon}(t,y+{\varepsilon}z)-\Lambda_{\varepsilon}(t,y)}{{\varepsilon}} J_{\varepsilon}dz.$$ In particular, our assumptions imply $\inf_{y\in {\mathbb{R}}^n}V^0_{\varepsilon}>-\infty$ uniformly in ${\varepsilon}>0$, thus implies that $V_{\varepsilon}$ is bounded from below, uniformly in ${\varepsilon}$.\
We prove by contradiction. We assume that there exists $(T,y_0)\in(0,+\infty)\times{\mathbb{R}}^n$ such that $$\label{absurd2}
V_{\varepsilon}(T,y_0)-\inf_{y\in {\mathbb{R}}^n} V_{\varepsilon}^0(y)<0.$$ For $\beta>0$, $\alpha>0$ small and for $t\in[0,T], y\in{\mathbb{R}}^n,$ we also introduce $$\varphi_{\alpha,\beta}(t,y):=V_{\varepsilon}(t,y)+ \alpha t+\beta\vert y-y_0\vert.$$
From and for a fixed ${\varepsilon}>0$, we have $V_{\varepsilon}(t,y)$ is bounded from below uniformly in $t\in[0,T]$, $y\in{\mathbb{R}}^n$. Therefore, $\varphi_{\alpha,\beta}$ goes to $+\infty$ as $\vert y\vert \to +\infty$ and reaches its minimum on ${[0,T]\times{\mathbb{R}}^n}$ at a point $(\bar t, \bar y)$. We have $$\varphi_{\alpha,\beta}(\bar t,\bar y+{\varepsilon}z)\geq\varphi_{\alpha,\beta}(\bar t,\bar y),\quad\forall z\in{\mathbb{R}}^n,$$ thus $$\label{accroissement1}
\frac{V_{\varepsilon}(\bar t,\bar y+{\varepsilon}z)-V_{\varepsilon}(\bar t,\bar y)}{{\varepsilon}}\geq\beta\frac{\vert\bar y-y_0\vert-\vert\bar y-y_0+{\varepsilon}z\vert}{{\varepsilon}}\geq-\beta \vert z\vert,\quad \forall z\in{\mathbb{R}}^n.$$
We choose $\alpha$ small enough to ensure $\varphi_{\alpha,\beta}(T,y_0)<\varphi_{\alpha,\beta}(0,y_0),$ which is possible thanks to assumption . It implies $\bar t>0$. Hence ${{\partial}_t\varphi_{\alpha,\beta}(\bar t,\bar y)\leq0}$, that is ${{\partial}_tV_{\varepsilon}(\bar t,\bar y)\leq-\alpha}$ (if $\bar t=T$ then ${\partial}_t V_{\varepsilon}^R(\bar t,\bar y)$ stands for the left-derivative). From at $(\bar t,\bar y)$, using , $$\begin{aligned}
-\alpha&\geq{\partial}_tV_{\varepsilon}(\bar t,\bar y)=-{\partial}_\eta\Lambda \int_{{\mathbb{R}}^n}M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{{\varepsilon}}}\frac{V_{\varepsilon}(t,y+{\varepsilon}z)-V_{\varepsilon}(t,y)}{{\varepsilon}}dz\\
&\geq\beta \inf\left[-{\partial}_\eta\Lambda\right]\int{M(z)e^{\frac{U_{\varepsilon}(\bar t,\bar y+{\varepsilon}z)-U_{\varepsilon}(\bar t,\bar y)}{{\varepsilon}}}\vert z\vert d z}\\
&\geq \beta \frac{1}{L\overline{\eta}}\int{M(z)e^{\left(k_0+\frac{L}{l\underline{\eta}^2}T\right)\vert z\vert}\vert z\vert d z}
\end{aligned}$$ where, in the last step, we used and $-{\partial}_\eta\Lambda\geq \frac{1}{L\overline{\eta}}$. As $\beta$ goes to $0$, we obtain $\alpha\leq0$, which is absurd. The proof is thereby achieved.
#### Semi-convexity in $y$.
Let us show how the method can be adapted to prove that ${\partial}^2_i U_{\varepsilon}$ (the second derivative w.r.t. $y_i$) is bounded from below. We set $
W_{\varepsilon}:= {\partial}_i^2U_{\varepsilon}.
$ Differentiating twice, we find $$\label{EquationOnD_y2U}
\begin{aligned}
{\partial}_t W_{\varepsilon}=
&-{\partial}_\eta\Lambda \int_{{\mathbb{R}}^n}\frac{W_{\varepsilon}(t,y+{\varepsilon}z)-W_{\varepsilon}(t,y)}{{\varepsilon}}J_{\varepsilon}dz\\
&-{\partial}^2_{\eta}\Lambda\left({\partial}_i\eta_{\varepsilon}\right)^2
-{\partial}_\eta\Lambda\int_{{\mathbb{R}}^n}\left(\frac{{\partial}_iU_{\varepsilon}(t,y+{\varepsilon}z)-{\partial}_i U_{\varepsilon}(t,y)}{{\varepsilon}}\right)^2J_{\varepsilon}dz\\
&-{\partial}^2_i\Lambda - 2{\partial}^2_{i,\eta}\Lambda {\partial}_i\eta_{\varepsilon}.
\end{aligned}$$ In contrast with the equation on ${\partial}^2_t U_{\varepsilon}$, we need to deal with a source term and a linear term in the last line.
For any constant $K>0$, Young’s inequality implies $$\label{YoungInequality}
- 2{\partial}^2_{i,\eta}\Lambda {\partial}_i\eta_{\varepsilon}\geq -K^2\vert {\partial}^2_{i,\eta}\Lambda\vert ^2-\frac{1}{K^2}({\partial}_i\eta_{\varepsilon})^2.$$ Applying this inequality and (replacing ${\partial}_t$ by ${\partial}_i$) in , we obtain $$\label{EquationOnD_y2UBIS}
\begin{aligned}
{\partial}_t W_{\varepsilon}\geq
&-{\partial}_\eta\Lambda \int_{{\mathbb{R}}^n}\frac{W_{\varepsilon}(t,y+{\varepsilon}z)-W_{\varepsilon}(t,y)}{{\varepsilon}}J_{\varepsilon}dz\\
&-\left({\partial}^2_{\eta}\Lambda+\frac{{\partial}_\eta\Lambda}{\eta_{\varepsilon}}+\frac{1}{K^2}\right)\left({\partial}_i\eta_{\varepsilon}\right)^2\\
&-{\partial}^2_i\Lambda -K^2\vert {\partial}^2_{i,\eta}\Lambda\vert ^2.
\end{aligned}$$
Using lemma , and choosing $K\geq\frac{1}{\sqrt\delta}$, we have $$\begin{aligned}
{\partial}_t W_{\varepsilon}\geq
&-{\partial}_\eta\Lambda \int_{{\mathbb{R}}^n}\frac{W_{\varepsilon}(t,y+{\varepsilon}z)-W_{\varepsilon}(t,y)}{{\varepsilon}}J_{\varepsilon}dz\\
&-{\partial}^2_i\Lambda -K^2\vert {\partial}^2_{i,\eta}\Lambda\vert ^2.\end{aligned}$$ From assumptions -, the source term $-{\partial}^2_i\Lambda -K^2\vert {\partial}^2_{i,\eta}\Lambda\vert ^2$ is bounded from below by some constant $-K'<0$. We end up with $${\partial}_t W_{\varepsilon}\geq
-{\partial}_\eta\Lambda \int_{{\mathbb{R}}^n}\frac{W_{\varepsilon}(t,y+{\varepsilon}z)-W_{\varepsilon}(t,y)}{{\varepsilon}}J_{\varepsilon}dz -K'.$$ Then, applying the same method as in the proof of (see also the proof of ), we show $$W_{\varepsilon}(t,y) \geq \inf_{y\in{\mathbb{R}}^n} W_{\varepsilon}(t=0,y)-K't,\quad \forall{\varepsilon}>0,\ t\geq 0,\ y\in{\mathbb{R}}^n.$$ Finally, we conclude the lower bound since $W_{\varepsilon}(t=0,y)\geq -C$ (from assumption ), we have $$W_{\varepsilon}(t,y) \geq -C-K't,\quad \forall{\varepsilon}>0,\ t\geq 0,\ y\in{\mathbb{R}}^n.$$
#### Other directional derivatives and conclusion.
Lower bounds on other second order derivatives in directions $\nu\in\mathbb{S}:=\{(t,y)\in{\mathbb{R}}^{n+1}: t^2+\vert y\vert^2=1\}$ can be obtained by a slight adaptation of the previous steps show. We deduce that $U_{\varepsilon}$ is semi-convex and that $\nabla U_{\varepsilon}$ is uniformly in $BV_{loc}$ (see Proposition 1.1.3 and Theorem 2.3.1 in [@Cannarsa2004]). We obtain that $U_{\varepsilon}$ is uniformly bounded in $W_{loc}^{2,1}$.
We recall that, from and , $U_{\varepsilon}$ is Lipschitz continuous, uniformly in ${\varepsilon}>0$, $y\in{\mathbb{R}}^n$, locally uniformly in $t$. Besides, from and for a fixed $T>0$, there exists a constant $K>0$ such that $({\partial}^2_tU_{\varepsilon}(t,y))^-\leq K$ for all $(t,y)\in[0,T]\times{\mathbb{R}}^n$ (where $^-$ denotes the negative part). We deduce $$\begin{aligned}
\int_0^T \vert{\partial}^2_tU_{\varepsilon}(t,y)\vert dt=\int_0^T {\partial}^2_tU_{\varepsilon}(t,y) dt+2\int_0^T ({\partial}^2_tU_{\varepsilon}(t,y))^- dt
&\leq {\partial}_tU_{\varepsilon}(t,y)+2Kt\\
&\leq-\underline{\Lambda}+2Kt,\end{aligned}$$ which proves the claim. With the same method, we show ${\partial}^2_{i}U_{\varepsilon}\in L^1_{loc}$.
Proof of {#sec:ProofCorollary}
---------
We recall the definition of $\eta_{\varepsilon}$ in , and note that differentiating , we obtain $$\label{LinkVEta}
{\partial}^2_t U_{\varepsilon}= -{\partial}_t\eta_{\varepsilon}(t,y){\partial}_\eta\Lambda(t,\eta_{\varepsilon}(t,y)), \qquad {\partial}^2_{t,x_i} U_{\varepsilon}= -{\partial}_{x_i}\eta_{\varepsilon}(t,y){\partial}_\eta\Lambda(t,\eta_{\varepsilon}(t,y)).$$ Using that $U_{\varepsilon}$ is uniformly bounded in $W^{2,1}$ () and that $-{\partial}_\eta\Lambda$ is positively bounded , we deduce that $\eta_{\varepsilon}$ is bounded in ${W^{1,1}}$, which proves the first part of .
Let us now fix $T>0$ and prove the second part. We recall that $U_{\varepsilon}$ is semi-concave (). Thus, there exists a constant $K>0$ such that ${\partial}_t\eta_{\varepsilon}(t,y)\geq -K$ for all ${\varepsilon}>0$, $(t,y)\in[0,T]\times{\mathbb{R}}^n$. Denoting $\overline{\eta}(t):=\sup_{y\in{\mathbb{R}}^n}\eta(t,y)$, we deduce ${\partial}_t\overline\eta_{\varepsilon}(t)\geq -K$ (indeed, the mapping $\theta_y: t\mapsto \eta_{\varepsilon}(t,y)+Kt$ is nondecreasing, and so is $\sup_{y\in{\mathbb{R}}^n}\theta_y$). The last inequality should be understood in the sense of the distributions.
We deduce $$\int_0^T \vert{\partial}_t\overline\eta_{\varepsilon}(t)\vert dt
=\int_0^T {\partial}_t\overline\eta_{\varepsilon}(t) dt+2\int_0^T ({\partial}_t\overline\eta_{\varepsilon}(t))^- dt
\leq \overline\eta_{\varepsilon}(t)+2Kt
\leq\overline\eta+2Kt,$$ where $^-$ denotes the negative part, and the last inequality comes from . The proof is complete.
Asymptotics of $U_{\varepsilon}$ - proof of {#sec:Asymptotics_of_U}
--------------------------------------------
#### Extraction of a subsequence
From the a priori estimate of and Ascoli’s theorem, we know that $U_{\varepsilon}$ converges locally uniformly to some $U$, up to extraction of a subsequence. Incidentally, this convergence also occurs in $W^{1,1}$, from the $W^{2,1}$ estimate in and a classical compact embedding. In addition, we know from and that $U_{\varepsilon}$ is locally Lipschitz continuous, uniformly in ${\varepsilon}>0$: for all $t\geq t'\geq0$ and $y,y'\in{\mathbb{R}}^n$ $$\label{LipschitzTbis}
\vert {U}(t,y)-U(t',y')\vert\leq-\underline{\Lambda}(t-t')+ \left(k^0+\frac{L}{l\underline\eta^2} t\right)\vert y-y'\vert .$$ Thus, the convergence occurs in $W_{loc}^{1,r}$, $1 \leq r < \infty$. Also notice that implies that $U$ is semi-convex, uniformly in $y\in{\mathbb{R}}^n$, locally uniformly in $t$.
#### Viscosity solution
We are going to show that $U$ is a viscosity solution of , i.e., $U$ satisfies $$\label{HJEBIS}
{\partial}_t U=H(y,\nabla_y U),\quad U(0,y)=U^0(y),$$ with $$H(y,p):=-\Lambda\left(y,\int_{{\mathbb{R}}^n}M(z)e^{p\cdot z}dz\right).$$ The proof is adapted from classical stability results for viscosity solutions of Hamilton-Jacobi equations (see [@GB:94]). However, this case is not completely standard because of the nonlocal term $\int_{{\mathbb{R}}^n}M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{{\varepsilon}}}dz$.
\[subsuperBIS\] The function $U$ is a viscosity solution of in $(0,\infty)\times{\mathbb{R}}^n$. Also, for all $T>0$, the viscosity inequalities stand for ${t\in (0,T]}$.
We are going to prove that $U$ is a subsolution of . Let us consider a test function $\varphi$ and a point $(t_0,y_0)$ such that $
{U}-\varphi$ reaches a global maximum at $(t_0,y_0)$. From classical results, there exists $(t_{\varepsilon},y_{\varepsilon})$ such that $$\left\{\begin{aligned}
&(t_{\varepsilon},y_{\varepsilon})\underset{{\varepsilon}\to 0}{\longrightarrow}(t_0,y_0),\\
&\max_{t,y} U_{\varepsilon}-\varphi =(U_{\varepsilon}-\varphi)(t_{\varepsilon},y_{\varepsilon}).
\end{aligned}\right.$$ For all $z\in{\mathbb{R}}^n$, ${\varphi(t_{\varepsilon},y_{\varepsilon}+{\varepsilon}z)-U_{\varepsilon}(t_{\varepsilon},y_{\varepsilon}+{\varepsilon}z)\geq \varphi(t_{\varepsilon},y_{\varepsilon})-U_{\varepsilon}(t_{\varepsilon},y_{\varepsilon})}$, thus we have $$\frac{\varphi(t_{\varepsilon},y_{\varepsilon}+{\varepsilon}z)-\varphi(t_{\varepsilon},y_{\varepsilon})}{{\varepsilon}}\geq\frac{U_{\varepsilon}(t_{\varepsilon},y_{\varepsilon}+{\varepsilon}z)- U_{\varepsilon}(t_{\varepsilon},y_{\varepsilon})}{{\varepsilon}}.$$ Since ${\partial}_\eta \Lambda<0$, equation gives $${\partial}_t\varphi(t_{\varepsilon},y_{\varepsilon})=-\Lambda\left(y_{\varepsilon},\int_{{\mathbb{R}}^n}
M(z)e^{\frac{U_{\varepsilon}(t_{\varepsilon},y_{\varepsilon}+{\varepsilon}z)-U_{\varepsilon}(t_{\varepsilon},y_{\varepsilon})}{{\varepsilon}}}d
z\right)\\
\leq -\Lambda\left(y_{\varepsilon},\int_{{\mathbb{R}}^n}
M(z)e^{\frac{\varphi(t_{\varepsilon},y_{\varepsilon}+{\varepsilon}z)-\varphi(t_{\varepsilon},y_{\varepsilon})}{{\varepsilon}}}d
z\right).$$ As ${\varepsilon}$ goes to $0$, $${\partial}_t \varphi(t_0,y_0)\leq-\Lambda\left(y_0,\int_{{\mathbb{R}}^n} M(z)e^{\nabla_y \varphi(t_0,y_0)\cdot z}\right)=H(y_0,\nabla_y \varphi(t_0,y_0)),$$ then $U$ is a viscosity subsolution of . With the same method, we prove that $ U$ is also a viscosity supersolution. It completes the first part of the proof. The second part of the statement is a well-known result, and proof can be found in [@GB:94].
#### Uniqueness
We point out that the Hamiltonian $H$ is Lipschitz continuous in the $y$ variable. We introduce a truncated Hamiltonian $$\tilde H (y,p):=
\left\{\begin{aligned}
&H(y,p) &&\text{if }\int_{{\mathbb{R}}^n}M(z)e^{p\cdot z}dz\in[\underline{\eta},\overline{\eta}],\\
&0 &&\text{otherwise}.
\end{aligned}\right.$$ Since $\underline{\eta}\leq\eta_{\varepsilon}(t,y)\leq\overline{\eta}$ (from ), we have $${\partial}_t U=\tilde H(y,\nabla_y U).$$ For this equation, a classical uniqueness result is in order (see e.g [@GB:94; @Nordmann2018a]). We deduce that $U_{\varepsilon}$ converges to $U$ for the whole sequence ${\varepsilon}\to0$ (and not for an extracted subsequence).
A posteriori Lipschitz estimate - proof of the global Lipschitz regularity in {#sec:APosterioriLipschitz}
------------------------------------------------------------------------------
From and , we know that $U$ is Lipschitz, globally in $t$ and locally in $y$. Our goal is to show that $U$ is globally Lipschitz continuous, i.e., that there exists a constant $C>0$ such that $$\label{GoalAPosterioriLipschitz}
\forall t\geq0,\ \forall (y,y')\in({\mathbb{R}}^n)^2, \quad U(t,y)-U(t,y') \leq C\vert y-y'\vert.$$ Let us fix $t\geq0$ and $(y,y')\in({\mathbb{R}}^n)^2$.
With $\eta_{\varepsilon}$ defined in and the bound $\eta_{\varepsilon}\leq\overline{\eta}$ from , we have, for all ${\varepsilon}>0$, $z\in{\mathbb{R}}^n$ $$\int_{{\mathbb{R}}^n}M(z)e^{\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U(t,y)}{{\varepsilon}}}dz\leq \eta_{\varepsilon}(t,y)\leq \overline{\eta}.$$ From the assumption that $M(\cdot)$ is not degenerate, we deduce that, for some $r_0>0$, and for all $z\in{\mathbb{R}}^n$ such that $\vert z\vert =r_0$, then $$\frac{U_{\varepsilon}(t,y+{\varepsilon}z)-U_{\varepsilon}(t,y)}{{\varepsilon}}\leq C$$ for some constant $C$ (independant of ${\varepsilon}$, $t$, $y$ and $z$). Then, chosing $z$ and ${\varepsilon}$ such that $y-y'={\varepsilon}z,$ we have $U_{\varepsilon}(t,y)-U_{\varepsilon}(t,y') \leq C\vert y-y'\vert.$ As ${\varepsilon}\to0$, we prove the goal .
[^1]: École des Hautes Études en Sciences Sociales, CAMS, 54 boulevard Raspail, 75006 Paris, France. Email: [email protected]
[^2]: Sorbonne Université, CNRS, Université de Paris, Inria, Laboratoire Jacques-Louis Lions, 75005 Paris, France. Email: [email protected]
[^3]: **Acknowledgment.** BP has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 740623)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
Wayne State University, Detroit MI 48201\
E-mail:
title: 'Measurements of the jet internal sub-structure and its relevance to parton shower evolution in p+p and Au+Au collisions at STAR'
---
Introduction
============
Relativistic ion collisions produce copious amounts of jets due to the hard scatterings between quarks and gluons of the colliding nuclei. Recent measurements at both RHIC and LHC along with theoretical advancements have shown the importance of studying and measuring the properties of these jets both in p+p and in heavy ion collisions (reviews of jet studies can be found here [@expreview]). There are two natural scales that characterize a jet and its evolution: the momentum and the angular scales. First generation measurements have measured jet quenching in an inclusive manner via the momentum asymmetry in di-jet events and have further extensively studied the momentum dependence via nuclear modification factors and fragmentation functions. Jet-medium interaction could further be dependent on the resolution scale or the coherence length of the medium which sees the jet as a singular radiating object or a multi-prong object [@cohlength].
Jet sub-structure in p+p Collisions
===================================
In order to differentially study energy loss in the medium, the jet structure in vacuum has to be first understood i.e., the DGLAP splitting functions that govern parton evolution. In recent literature, SoftDrop [@SoftDrop] has gained fame as an algorithm which can extract such scales experimentally with its procedure of walking backwards in the Cambridge/Aachen clustering tree until two sub-jets satisfy $z = \frac{min(p_{T,1}, p_{T,2})}{p_{T,1} + p_{T,2}} > z_{cut} (\Delta R/R)^{\beta}$ where [$z_g$ ]{}and [$R_g$ ]{}are the $z$ and $\Delta R$ upon termination of the algorithm with $z_{cut} = 0.1$ and $\beta = 0$. It was shown that for such choices of $z_{cut}, \beta$ the SoftDrop $z_{g}$ distribution converges to the vacuum DGLAP splitting functions for $z>z_{cut}$ in a “Sudakov-safe” manner [@Larkoski:2015lea].
![Fully unfolded measurement of the SoftDrop groomed sub-jet shared momentum fraction ($z_{g}$) and the groomed jet radius ($R_{g}$) in p+p collisions at $\sqrt{s} =$ 200 GeV. The markers and lines are described in the text.[]{data-label="fig:ppzgrg"}](star_zg_pp200GeV_hp2018.png "fig:"){width="49.00000%"} ![Fully unfolded measurement of the SoftDrop groomed sub-jet shared momentum fraction ($z_{g}$) and the groomed jet radius ($R_{g}$) in p+p collisions at $\sqrt{s} =$ 200 GeV. The markers and lines are described in the text.[]{data-label="fig:ppzgrg"}](star_rg_pp200GeV_hp2018.png "fig:"){width="47.00000%"}
The p+p data for the groomed jet structure measurements were collected with the STAR detector [@star] during the 2012 run at $\sqrt{s_{NN}} = 200$ GeV. Jets are reconstructed from charged tracks in the Time Projection Chamber (TPC) and energy depositions in the Barrel ElectroMagnetic Calorimeter (BEMC) using the anti-$k_{t}$ algorithm as implemented in the FastJet package [@FastJet], hereafter referenced as Ch+Ne jets. For additional details regarding event/track/tower quality selections, please refer to [@starprl; @nicktalk]. For the inclusive p+p analysis, events are selected by an online jet patch trigger which is a $1\times1$ patch in $\eta-\phi$ with the total sum $E_{T}$ in the patch to be greater than 7.3 GeV. Two-dimensional unfolding in $p_T$ and [$z_g$ ]{}or $R_{g}$ respectively, was done using Bayesian unfolding as implemented in the RooUnfold package [@roounfold; @dgostini] with four iterations. The response matrix is created from a PYTHIA-6 (Perugia Tune, slightly adapted to STAR data) prior and a GEANT-3 simulation of the STAR detector. The systematic uncertainties for data are taken as a quadrature sum resulting from the following sources: tracking efficiency ($4\%$), tower gain calibration ($3.8\%$), hadronic correction to the tower energy scale (described in [@starprl]) and unfolding related sources including varying the iteration parameter from 2 to 6 and the prior in the response matrix.
Figure \[fig:ppzgrg\] shows the fully unfolded [$z_g$ ]{}and [$R_g$ ]{}distributions for two jet [$p_{T}$]{} selections, 10-15 and 30-40 GeV/$c$, respectively. The STAR data are shown in the red filled star markers with the red shaded region corresponding to the overall systematic uncertainty. Leading order Monte Carlo (MC) generators such as PYTHIA-6 Perugia tune, PYTHIA-8 Monash, and Herwig-7 EE4C UE tune in the blue, black and magenta lines are also plotted for comparison to the data. For the [$z_g$ ]{}observable, we also provide the symmetrized DGLAP splitting functions (noted as AP Q-Jet in the figure) at leading order in the red dashed lines for quark jets (with the splitting being similar for quark and gluon initiated jets). All the models studied reproduce the general trends seen in the data, particularly the dependence on the jet momenta leading to a steeper [$z_g$ ]{}distribution and a narrower $R_{g}$.
Jet Angular Scale in Au+Au Collisions
=====================================
The Au+Au data used in these proceedings were collected during the 2007 run with its corresponding reference p+p run in 2006 at $\sqrt{s_{NN}} = 200$ GeV. Since the jet patch trigger is saturated in an Au+Au event, we employ a high tower (HT) trigger, requiring at least one BEMC tower with $E_{T}$ > 5.4 GeV. Event centrality in Au+Au is determined by the raw charged track multiplicity in the TPC within $|\eta| < 0.5$ and we show only events in the 0-20% centrality range. In Au+Au events, we have two separate jet collections given by the HardCore selection [@starprl], where jets are clustered with objects (tracks/towers) with $p_{T}> 2$ GeV/$c$, and Matched jets which are clustered from the constituent-subtracted [@cs] event with our nominal $p_{T} > 0.2$ GeV/$c$ for constituents, and are geometrically matched to the HardCore jets ($\Delta R < 0.4$). Further event selection criteria include a minimum [$p_{T}$]{} requirement for HardCore di-jets ($p^{\rm{Lead}}_{T} > 16, p^{\rm{SubLead}}_{T} > 8$ GeV/$c$) and an azimuthal angle ($|\Delta \phi (\rm{Lead, SubLead})| > 2\pi/3$) selection to focus on back-to-back di-jets.
![TwoSubJet observables ($z_{SJ}$ - left and $\theta_{SJ}$ - right) comparing Au+Au (solid circles) and p+p embedded in Au+Au (open circles including the systematic uncertainties in the shaded region) for recoil Matched jets (anti-$k_{t}$ $R=0.4$ jets with $R=0.1$ sub-jets) in the $p_{T}$ range 10-20 GeV$/c$. []{data-label="fig:twoSubJet"}](star_twosubjet_z_theta_auau7_hp2018.png){width="70.00000%"}
In our studies, we found the groomed jet radii (R$_{g}$) to be highly sensitive to the fluctuating underlying event found in Au+Au collisions and therefore we devised a new observable involving sub-jets of a smaller radius reconstructed within the original jet (see here [@SubJetliliana] for a recent theoretical article demonstrating similar classes of observables). For our nominal anti-$k_{t}$ jets of $R=0.4$, we reconstruct an inclusive set of anti-$k_{t}$ sub-jets with $R=0.1$ from the original jet’s constituents. An absolute minimum sub-jet [$p_{T}$]{} requirement of $2.97$ GeV/$c$ is enforced in central Au+Au collisions to reduce sensitivity to the background fluctuations. The two observables related to the momentum and angular scales are then defined as follows $z_{SJ} = \frac{min(p^{SJ1}_{T}, p^{SJ2}_{T})}{p^{SJ1}_{T} + p^{SJ2}_{T}}, ~ ~ ~ \theta_{SJ} = \Delta R (SJ1, SJ2) $ where $SJ1, SJ2$ are the leading and sub-leading sub-jets, respectively.
![HardCore and Matched (left figure) di-jet asymmetry ($|A_{J}|$) and the Matched recoil jet yield (right figure) along with the ratios shown in the bottom right panels. Markers are described in the text. []{data-label="fig:aj"}](star_aj_hc_mat_thetabins_auau7_hp2018.png "fig:"){width="55.00000%"} ![HardCore and Matched (left figure) di-jet asymmetry ($|A_{J}|$) and the Matched recoil jet yield (right figure) along with the ratios shown in the bottom right panels. Markers are described in the text. []{data-label="fig:aj"}](star_mat_recoiljetptspectra_thetabins_auau7_hp2018.png "fig:"){width="38.00000%"}
For a meaningful comparison between Au+Au and a p+p reference, the effects of background fluctuations and detector inefficiencies must be taken into account. To achieve this, HT-triggered p+p data from 2006 is embedded into minimum bias Au+Au data (p+p $\oplus$ Au+Au) from 2007, in the same centrality range (0-20%). During embedding, we account for the relative tracking efficiency (90%$\pm$7%) and relative tower energy scale (100%$\pm$2%), with a one sigma variation taken as systematic uncertainties. Figure \[fig:twoSubJet\] shows the detector level TwoSubJet $z_{SJ}$ (left) and $\theta_{SJ}$ (right) distributions for constituent-subtracted Matched recoil jets with $R=0.1$ sub-jets (henceforth denoted as $SJ-0.1$ in the figures) recoiling off the trigger (selected with a $|\Delta \phi(\rm{jet, HT})| > 2\pi/3$), in the [$p_{T}$]{} range 10-20 GeV/$c$. Results are observed to be similar in both Au+Au and p+p $\oplus$ Au+Au. We also observe a remarkable difference in the shape of $z_{SJ}$ when compared to that of the SoftDrop $z_{g}$, which is caused by selecting the core of the jet. The $\theta_{SJ}$ for jets within the considered $p_{T}$ range peaks at small values and includes a natural lower cutoff at the sub-jet radius and we now select jets based on this distribution.
Di-jet asymmetry for both HardCore and Matched jets (left panels) are shown in Figure \[fig:aj\]. The right panels of Figure \[fig:aj\] show the yield of Matched recoil jets normalized per di-jet for the different $\theta_{SJ}$ selections and the ratios of Au+Au/p+p in the bottom panels. The black, blue and red markers represent recoil jets with selections on $\theta_{SJ}$ $[0.1, 0.4], [0.1, 0.2]$ and $[0.2, 0.3]$, respectively, for inclusive, narrow and wide jets. We observe a clear di-jet imbalance indicating jet quenching effects in the $|A_{J}|$ distributions (comparing Au+Au to p+p$\oplus$Au+Au) for all HardCore jets including the wide angle jets. The Matched jets on the other hand are balanced at RHIC energies, as evident by ratios in the bottom right panels consistent with unity. This is consistent with our earlier measurement [@starprl]. We also note that wide angle jets are still balanced indicating no apparent distinction between wide and narrow jets by the medium in our selection. Further detailed differential analyses are required with the high statistics 2014 data set to extract the medium resolution scale or the coherence length and the effect on standard jet quenching observables at RHIC energies.
Conclusions
===========
STAR has presented the first fully unfolded SoftDrop [$z_g$ ]{}and [$R_g$ ]{}measurements of inclusive jets with varying transverse momentum in p+p collisions at $\sqrt{s} = $200 GeV. The measurements are overall reproduced by current leading order Monte Carlo event generators for both [$z_g$ ]{}and [$R_g$ ]{}for jets in our kinematic acceptance and reflect the momentum dependent narrowing of jet structure. Due to the sensitivity of the SoftDrop observables to the Au+Au underlying event, we introduce and measure the TwoSubJet observables, $z_{SJ}$ and $\theta_{SJ}$ for R=0.1 anti-$k_{t}$ sub-jets as representing the momentum and angular scales of a jet in a heavy ion environment. We measure the di-jet momentum asymmetry and the recoil jet yield with the special di-jet selection at STAR and find that HardCore di-jets are imbalanced and Matched di-jets are balanced for jets of varying angular scales. We find no significant difference in the quenching phenomenon for both wide and narrow jets leading to the conclusion that these special jets do not undergo significantly different jet-medium interactions due to their varying angular scales.
[99]{} M. Connors, R. Reed, S. Salur, C. Nattrass, Rev. Mod. Phys. 90, (2018) 025005 Y, Mehtar Tani, K, Tywoniuk, Phys. Rev. D 98 (2018) 051501(R) A. J. Larkoski, S. Marzani, G. Soyez, J. Thaler, Soft Drop, JHEP 05 (2014) 146 A. J. Larkoski, S. Marzani, J. Thaler, Phys. Rev. D 91 (11) (2015) 111501 K. H. Ackermann et. al., Nucl. Instrum. Methods Phys. Res., Sect. A 499, (2003) 624 M. Cacciari, G. P. Salam, and G. Soyez, J. High Energy Phys. 04 (2008) 005. L. Adamczyk, et al., Phys. Rev. Lett. 119 (6) (2017) 062301 Nick Elsey for the STAR Collaboration, these proceedings T. Adye et. al. [hepunx.rl.ac.uk/\~adye/software/unfold/RooUnfold.html](hepunx.rl.ac.uk/~adye/software/unfold/RooUnfold.html) G. D’Agostini, arXiv:1010.0632 L. Apolinário, J. G. Milhano, M. Ploskon, X. Zhang, Eur. Phys. J. C78 (2018) no.6, 529 P. Berta, M. Spousta, D. W. Miller, R. Leitner, J. High Energy Phys. 1406 (2014) 092
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The magneto-optical imaging technique is used to visualize the penetration of the magnetic induction in YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films during surface resistance measurements. The in-situ surface resistance measurements were performed at 7 GHz using the dielectric resonator method. When only the microwave magnetic field $H_{rf}$ is applied to the superconductor, no $H_{rf}$-induced vortex penetration is observed, even at high rf power. In contrast, in the presence of a constant magnetic field superimposed on $H_{rf}$ we observe a progression of the flux front as $H_{rf}$ is increased. A local thermometry method based on the measurement of the resonant frequency of the dielectric resonator placed on the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin film shows that the $H_{rf}$–induced flux penetration is due to the increase of the film temperature.'
author:
- Julien Kermorvant
- 'Jean-Claude Mage'
- Bruno Marcilhac
- Yves Lemaître
- 'Jean-François Bobo'
- Cornelis Jacominus van der Beek
date:
-
-
title: 'Microwave heating-induced DC magnetic flux penetration in YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ superconducting thin films'
---
Introduction
=============
High Temperature Superconductor (HTS) thin films are now recognized as particularly suitable for high frequency signal processing. Due to their very low surface resistance $R_{s}$, as compared to normal metals, they allow for very efficient microwave signal filtering and detection. However, in the region of high power $P_{rf}$ of the incident microwave field, their application is limited by the strong dependence of $R_{s}$ on the $P_{rf}$–magnitude. A nonlinear increase of the surface resistance with the input rf power is commonly observed,[@Oates92; @Oates95; @Samoilova95; @Hampel1996; @Wosik97; @Anlage99; @Lahl2005; @Kermorvant2009] leading to a detrimental decrease of the $Q$-factor of the devices.[@Hein97] The origin of the nonlinear microwave losses in high $T_{c}$ superconductors has been studied by many groups. Among the cited causes, there are intrinsic phenomena such as the excitation of quasiparticles when the current density induced by the microwave magnetic field (of magnitude $H_{rf}$) becomes of the order of the pair-breaking current density,[@Dahm99] but also the limitation of the supercurrent in grain boundaries, and vortex motion [@Lahl2005; @Coffey1992; @Coffey1993; @Coffey1993ii] in the superconductor induced by the rf field. The interplay of the superposed ac and dc magnetic fields in superconducting thin films is extensively described in Refs. . Recent work has, however, clearly demonstrated that it is simply local Joule heating of the superconducting film by the microwave field that leads to the nonlinear behavior,[@Hampel1996; @Wosik97; @Kermorvant2009] and that the dissipation at the origin of the heating is due to the linear electromagnetic response of the films.
In Ref. , we have introduced a valuable tool for local thermometry of superconducting films studied using a dielectric resonator. The variation of the resonator frequency as a function of the temperature depends essentially on the dielectric constant $\epsilon$ of the resonator; its calibration turns the latter into a precise local thermometer. This method has allowed us to measure the temperature of YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ films as a function of the rf input power (at 10 GHz) under nominally isothermal conditions. The observed temperature increase of the resonator and the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films as function of rf power unambiguously showed that the usually observed increase in the surface resistance is due to Joule heating, with a linear response dissipation mechanism. Candidate mechanisms at the origin of the heating can therefore be limited to quasi-particle dissipation [@Dahm99; @Mattis1958] and flux-flow losses.[@Coffey1992; @Coffey1993; @Coffey1993ii; @Bardeen1965; @Brandt1991; @Coffey1991]
![(Color online) Principle of MOI of superconductors. Thin drawn lines depict the magnetic flux as this traverses the superconductor and the MOL, thick black lines show the optical path of the impinging and reflected light, and the circled arrows illustrate the linear polarization direction of the light.[]{data-label="fig2"}](Figure1.pdf){width="1.01\linewidth"}
![Complex permeability of the (Lu,Bi)$_{3}$(Fe,Ga)$_{5}$O$_{12}$ magneto-optically active layer, measured at room temperature.[]{data-label="fig1"}](Figure2.pdf){width="1.01\linewidth"}
Here, we study the influence of an rf magnetic field generated by a rutile dielectric resonator on the DC flux distribution in YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films. For this, we have developed a dedicated set-up, that allows for the simultaneous measurement of the film surface resistance and the visualization of the magnetic flux distribution using the Magneto-Optical Imaging (MOI) Technique.[@Dorosinskii1992; @Jooss2002] MOI is based on the Faraday effect in which the rotation of the polarization plane of incident linearly polarized light is proportional to the magnetic induction component parallel to the wavevector of the incoming light. Since superconductors do not present a significant Faraday effect, one has to use a magneto-optical layer (MOL, with a strong Faraday effect) placed on top of a superconductor, as depicted in Fig. \[fig2\]. In the experiments described below, we use, as a MOL, Lu- and Bi-doped Yttrium-Iron Garnet thick films[@Uehara2009] with a ferromagnetic resonance at 0.9 GHz (see Fig. \[fig1\]). The application of a magnetic field of a few dozen mT will increase the ferromagnetic resonance frequency somewhat. However, at the experimental microwave field frequency of 7.0 GHz, the magnetization rotation leading to the Faraday effect will always be strongly overdamped, inhibiting the direct visualization of the microwave magnetic field. The MOI technique does permit the visualization of the modification of static flux structures in YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films following microwave field application. Note that the development of a similar imaging system was reported in Ref. , however, the influence of the microwave magnetic field on flux penetration was not reported there.
![(Color online) Schematic view of the experimental assembly. Panel (a) shows a side view, while panel (b) depicts a top view for different lid apertures. The left-hand side cover allows the imaging of the central region of the sample, while the right-hand cover allows for the observation of the edge region. []{data-label="fig3"}](Figure3.pdf){width="0.9\linewidth"}
Experimental details
====================
YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ films
--------------------------------------
All YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ films under study in this work were cut from the same wafer, designated SY211 in Ref. . The YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ was deposited on a 500 $\mu$m-thick MgO substrate using cylindrical hollow cathode DC sputtering, and had a thickness $d =400$ nm. Its critical temperature is $T_{c} \simeq 86$ K and its critical current density $j_{c} = 5\times 10^{10}$ Am$^{-2}$ at $T = 77$ K.
Magneto-Optical Imaging
-----------------------
The MOL is a 5 $\mu$m-thick (Lu,Bi)$_{3}$(Fe,Ga)$_{5}$O$_{12}$ ferrimagnetic garnet film with in-plane magnetic anisotropy, grown on a 500 $\mu$m-thick Gd$_{3}$Gd$_{5}$O$_{12}$ (GGG) substrate.[@Uehara2009] The MOL is covered by a 100 nm-thick Al mirror layer and a 10 nm-thick TiO$_{2}$ protective layer. It is placed face-down on the superconducting film (Fig. \[fig2\]), whence it is observed through the transparent substrate using a polarized microscope with nearly crossed polarizers. In this configuration, the reflected light intensity increases as function of the local magnetic induction perpendicular to the garnet. Bright regions in the MO image correspond to regions of high magnetic flux density, while dark areas correspond to small or zero induction. This allows for the direct observation of magnetic flux penetration into the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ films. The calibration of the luminous intensity in the absence of the superconductor, or measured at a point that is sufficiently far removed from the superconductor, allows one to convert the spatially resolved intensity maps to maps of the absolute value of the magnetic induction.
Simultaneous MOI and Surface Resistance measurements
----------------------------------------------------
In order to measure the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film’s surface resistance during MOI, a number of specific modifications are required with respect to standard MOI and standard $R_{s}$ measurements. These are sketched in Fig. \[fig3\]. First of all, we have chosen to use the dielectric resonator technique with a rutile-phase TiO$_{2}$ resonator (of diameter 7 mm and height 3 mm) operating at 10 GHz for the $R_{s}$ measurements.[@Kermorvant2009] However, the presence of the MOL, in close contact with the superconducting film, prohibits one from placing the dielectric resonator directly on the film. The resonator was therefore installed on the side of the MgO substrate. The resonator is pressed to the substrate by the bottom lid of the square Cu cavity (of width 30 mm and height 7 mm) in which the whole assembly is placed. The presence of the substrate between the resonator and the film shifts the resonant frequency from 10 GHz to 7 GHz. The MgO substrate exhibits a low microwave loss tangent, $\tan \delta = 9 \times 10^{-6}$, and high thermal conductivity, $\kappa(90\,\,{\mathrm K}) = 290$ Wm$^{-1}$K$^{-1}$.[@Cahill] Hence, it does not introduce additional losses, and negligible measurement error on both the temperature and the resonant frequency when the temperature of the superconductor rises. The microwave field is excited using the sweeper of a HP 8510C vector network analyzer through a coupling loop in the Cu cavity, the position of which can be adjusted at low temperature using an XYZ stage.
![(Color online) (a) MOI of the flux distribution in the square YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film at $H_{a} = 350$ Oe and $T = 50$ K. (b) Profiles of the magnetic induction $B$ along the horizontal line in (a), for successive applied fields between 0 and 250 Oe. (c) MOI of the flux distribution in the film at $T = 70$ K and $H_{a} =150$ Oe, applied after zero field cooling to 30 K and successive warming. (d) $B$–profiles taken along the horizontal line in (c), for different temperatures during warming to 70 K. []{data-label="fig4"}](Figure4.pdf){width="0.99\linewidth"}
In usual MOI, the object to be imaged is cooled via direct thermal contact with a Cu sample holder attached to the cold head of the cryostat. For simultaneous $R_{s}$/MOI measurements, the intercalation of the dielectric resonator, with its low thermal conductivity,[@Thurber1965] prohibits this configuration. Thus, thermal contact is made via the top of the imaged specimen, through the top lid of the Cu cavity. This imposes a reduction of the magneto-optically imaged area of the sample under study. Imaging is performed through a circular aperture of diameter 8 mm in the cavity top lid. The presence of the Al mirror layer on the MOL does not pose additional problems, since it is shielded from the TiO$_{2}$ resonator by the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film.
Results
=======
MOI in dc magnetic field
-------------------------
Figures \[fig4\](a,c) present standard MO images obtained on a square-shaped $5\times5$ mm$^{2}$ YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film. The panels (a,c) illustrate the distribution of the magnetic induction in the superconducting film during the application of a constant magnetic field $H_{a} = 350$ Oe perpendicular to the film plane at the cavity temperature of 50 K, while panels (b,d) show the effect of increasing the cavity temperature after zero field cooling to 30 K, the application of $H_{a} = 150$ Oe and subsequent warming to 70 K. In all cases, the flux distribution accurately corresponds to the predictions of the Bean model,[@Bean1962; @Brandt1993; @Zeldov1994] which has that in the flux-penetrated areas of the film, the screening current density can only take on the value $\pm j_{c}$. The magnetic flux penetrates the superconductor from the edge of the sample, and is distributed according to the characteristic pillow-like shape expected for the divergence-free flow of the critical current in the thin film.[@Brandt1995] The flux distribution only depends on the parameter $H_{a}/j_{c}$, so that increasing either $H_{a}$ or temperature (with the concomitant decrease of $j_{c}$) both lead to a progression of the flux front to the film center. The position of the flux front, $x_{f}$, verifies the relation $$x_{f} = w \left[ 1 - \frac{1}{\cosh \left( \pi H_{a} /j_{c} d \right)} \right]$$ in which $w$ is the half-width of the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film.
![ (a) Surface resistance at 7 GHz as function of the input microwave power at 60 K, for zero applied field, and $H_{a} = 130$ Oe. (b) Resonant frequency as a function of the input microwave power at 60 K. []{data-label="fig5"}](Figure5.pdf){width="0.9\linewidth"}
![(Color online) (a) Schematic representation of the position of the obtained image relatively to the sample edge. The open green contour denotes the outline of the SC film. (b) MO image obtained after the sample was zero field cooled to 60 K, and a microwave field with input power of 22 dBm was applied. []{data-label="fig6"}](Figure6.pdf){width="1.0\linewidth"}
![(Color online) (a) Schematic representation of the position of the obtained image relative to the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin film outline (open green contour). (b) MO image obtained after the sample was zero-field cooled to 60 K, and a dc field of 130 Oe applied.[]{data-label="fig7"}](Figure7.pdf){width="1.0\linewidth"}
MOI with applied microwave magnetic field
-----------------------------------------
In a first experiment, we have applied a microwave field to the superconductor, in the absence of a dc field. Since any magnetic flux should penetrate the superconducting film from the edge, we have used the right-hand Cu cavity cover of Fig. \[fig3\]. The film was cooled in Earth’s magnetic field, down to 60 K. After stabilization of the target temperature, the $TE_{011}$–mode of the dielectric resonator was excited in the low-power regime (-10 dBm applied to the coupling loop). In this mode, the rf electric field is parallel to the plane of the superconducting film, while the rf magnetic field is perpendicular to it. Next, the microwave power was increased from -10 dBm to +24 dBm in steps of 3 dBm. At each input power level, the resonator frequency and $Q$-factor were measured and the corresponding MO image recorded.
Figure \[fig5\](a) renders the microwave power dependence of the surface resistance $R_{s}$, while Fig. \[fig5\](b) shows the power dependence of the resonant frequency $f_{0}$. No qualitative difference was observed with respect to the previously performed measurements at 10 GHz, with the resonator placed in direct contact with the superconducting film.[@Kermorvant2009] Both $R_{s}$ and the resonant frequency increase with rf power, indicating the nonlinear increase of microwave losses in the high microwave power regime.
Figure \[fig6\](b) shows the MO image obtained at an input power level of 24 dBm. According to the $R_{s}$($P_{rf}$) characteristics in Fig. \[fig5\](a), the superconductor is clearly in the lossy regime. However, MOI does not reveal any vortex penetration into the superconductor, even in the high power regime.
![(a) Increase of the resonance frequency of the TiO$_{2}$ resonator as function of temperature, measured in the absence of the superconducting film. (b) Temperature increase of the superconducting film during the swept-power experiments of Fig. \[fig5\], as deduced from (a). []{data-label="fig8"}](Figure8.pdf){width="0.9\linewidth"}
: 1. No microwave field; 2. 10 dBm; 3. 14 dBm; 4. 17 dBm; 5. 20 dBm; 6. 24 dBm.[]{data-label="fig9"}](Figure9.pdf){width="1.0\linewidth"}
![(Color online) Identical flux magnetic flux penetration into the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin film exposed to a dc field of 130 Oe (a) at the sample holder temperature $T = 60$ K, with applied microwave power $P_{rf}$ = 17 dBm; (b) at the higher temperature of the sample holder $T = 62.586 $ K, chosen such that $\Delta f = 33.245$ MHz as in (a), but with $P_{rf}$ = 10 dBm. []{data-label="fig9"}](Figure10.pdf){width="1.0\linewidth"}
MOI with microwave and dc magnetic field
----------------------------------------
A second series of experiments was performed by adding a constant magnetic field to the rf field. The superconducting sample was again zero-field cooled to 60 K, following which a dc magnetic field of 130 Oe was applied perpendicularly to the film surface. This leads to the magnetic flux penetration illustrated in Fig. \[fig7\]. Next, the microwave input was switched on in order to excite the dielectric resonator. The nominal temperature of the cold head and the dc magnetic field remain constant during the experiment. The microwave power was gradually increased in steps of approximately 3 dBm. At each input microwave power level the $Q$-factor and the resonant frequency of the resonator was measured and the corresponding MO image recorded. The obtained results concerning the power dependence of the surface resistance and the resonant frequency are again plotted in Figure \[fig5\]. We observe the usual power dependence of the surface resistance of YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films, [*i.e.*]{}, a nonlinear increase of $R_{s}$ as function of $P_{rf}$. As shown by temperature dependence of $f_{0}$ depicted in Fig. \[fig8\], the simultaneous increase of the resonant frequency of the dielectric, here, by up to $\delta f = 33.245$ MHz for the highest injected power, can be understood as the result of local heating of the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film.[@Kermorvant2009]
The simultaneously acquired MO images show no progression of the flux front in the linear regime of $R_{s}( P_{rf})$, but pronounced enhanced dc magnetic flux penetration in the nonlinear regime. A mechanism for ac magnetic field–induced dc flux penetration into thin superconducting films, the so-called vortex lattice shaking, was presented in Refs. . The conditions for vortex shaking to be effective are that the ac field strength be sufficient to drive the film into the critical state, which is the case here, and that the rf screening currents be of opposite polarity on the top and bottom film surfaces. In the present experimental configuration, the latter condition is not satisfied: the use of the $TE_{011}$ cavity mode means that the rf magnetic field to which the film is subjected induces screening currents of the same polarity on the top and bottom film surface, opposite in direction to the electric field in the rutile cavity.
We therefore surmise that the progression of the flux front is due to the decrease of the critical current density $j_{c}$ associated with the increase in temperature of the YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ film, a conjecture that is checked by increasing the nominal temperature of the cold head and the sample holder using the incorporated heater, all the while maintaining a low microwave input power of 10 dBm. The temperature is adjusted so as to precisely yield a frequency change of the resonator $\delta f = 33.245$ MHz. We find that this corresponds to a temperature increase of $\Delta T = 2.586$ K. The MOI (Fig. \[fig9\]) shows that the flux penetration precisely corresponds to that previously obtained by increasing the microwave power, showing that local heating is indeed at the origin of the enhanced flux penetration and the nonlinear dependence $R_{s}$ ( $P_{rf}$ ).
Summary and Conclusions
=======================
We have developed a experimental set-up that allows for the simultaneous imaging of magnetic flux penetration into superconducting samples using the magneto-optical technique, and the measurement of their surface resistance in the range 1 – 10 GHz. The method was applied to YBa$_{2}$Cu$_{3}$O$_{7-\delta}$ thin films. No signature of the microwave magnetic field could be observed in magneto-optics. However, the application of a high-power microwave magnetic field significantly enhances dc flux penetration due to local heating of the superconducting film.
This work was partially funded by the French National Research Agency ANR, under contract number ANR-07-BLAN-0242 ÔÔSURFÕÕ.
[99]{} D. Oates, P.P. Nguyen, G. Dresselhaus, M.S. Dresselhaus, C.W. Lam, S.M. Ali, J. Superconductivity [**5**]{}, 361 (1992). Daniel E. Oates, Paul P. Nguyen, Gene Dresselhaus, Mildred S. Dresselhaus, Gad Koren and Emil Polturak, J. Superconductivity [**8**]{}, 725 (1995). T. B. Samoilova, Supercon. Sci. Technol. [**8**]{}, 259 (1995). G. Hampel, P. Kolodner, P.L. Gammel, P.A. Polakos, E. de Obaldia, P.M. Mankiewich, A. Anderson, R. Slattery, D. Zhang, G.C. Liang, C.F. Shi, Applied Physics Letters [**69**]{}, 571 (1996). Jaroslaw Wosik, Lei-ming Xie, Krzysztof Nesteruk, Dawei Li, John H Miller, Stuart A Long, J. Superconductivity [**10**]{}, 97 (1997). Steven M. Anlage, Wensheng Hu, C. P. Vlahacos, David Steinhauer, B. J. Feenstra, Sudeep K. Dutta, Ashfaq Thanawalla, and F. C. Wellstood, J. Supercond. [**12**]{}, 3 53-362 (1999) . J. Kermorvant, C.J. van der Beek, J.-C Mage, B. Marcilhac, Y. Lemaître, J. Briatico, R. Bernard, J. Villegas, J. Appl. Phys. [**106**]{}, 023912 (2009). P. Lahl and R. Wördenweber, J. Appl. Phys. [**97**]{}, 113911 (2005). T. Dahm and D.J. Scalapino, Phys. Rev. B [**60**]{}, 13125 (1999). M. Hein, W. Diete, M. Getta, S. Hensen, T. Kaiser, G. Müller, H. Piel, and H. Schilk, IEEE Trans. Appl. Supercond. [**7**]{}, 1264 (1997). D. C. Mattis and J. Bardeen, Phys. Rev. [**111**]{}, 412Ð417 (1958). John Bardeen and M. J. Stephen, Phys. Rev. [**140**]{}, A1197 (1965). Ernst Helmut Brandt, Phys. Rev. Lett. [**67**]{}, 2219 (1991) Mark W. Coffey and John R. Clem, Phys. Rev. Lett. [**67**]{}, 386 (1991). Mark W. Coffey and John R. Clem, Phys. Rev. B [**46**]{}, 11757 (1992). Mark W. Coffey and John R. Clem, Phys. Rev. B [**48**]{}, 342 (1993). Mark W. Coffey, Phys. Rev. B [**47**]{}, 15298 (1993). Ernst Helmut Brandt and Grigorii P. Mikitik, Phys. Rev. Lett. [**89**]{}, 027002-1(2002). Grigorii P. Mikitik and Ernst Helmut Brandt, Phys. Rev. B [**67**]{}, 134521 (2003). Grigorii P. Mikitik and Ernst Helmut Brandt, Phys. Rev. B [**69**]{}, 104511 (2004). G. P. Mikitik and E. H. Brandt, Journal of Low Temperature Physics [**139**]{}, 221 (2005). J. Wosik, C. Wang, L.-M. Xie, T.H. Johansen, Q.Y. Chen, and W.-K. Chu, Applied Superconductivity Conf., Aug. 4-9, 2002, Houston, TX; 15 L.A. Dorosinskii, M.V. Indenbom, V.I. Nikitenko, Yu.A. OssipÕyan, A.A. Polyanskii, and V.K. Vlasko-Vlasov, Physica C [**203**]{}, 149 (1992). Ch. Jooss, J Albrecht, H Kuhn, S Leonhardt and H Kronmüller, Rep. Prog. Phys. [**65**]{}, 651Ð788 (2002). M. Uehara, C.J. van der Beek, J. Gattacecca, V. Skidanov, Y. Quesnel, Geochemistry, Geophysics, Geosystems [**11**]{}, Q05Y09 (2009). See http://users.mrl.uiuc.edu/cahill/tcdata/tcdata.html W.R. Thurber, A.J. Mante, Physical Review [**139**]{}, 1655 (1965). C.P. Bean, Phys. Rev. Lett. [**8**]{}, 6 (1962). E.H. Brandt, M.V. Indenbom and A. Forkl, Europhys. Lett. [**22**]{}, 735 (1993). E. Zeldov, J.R. Clem, M. McElfresh and M. Darwin, Phys. Rev. B [**49**]{}, 9802 (1994). Ernst Helmut Brandt, Phys. Rev. Lett. [**74**]{}, 3025 (1995).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
**A. Lesfari**\
Département de Mathématiques\
Faculté des Sciences\
Université Chouaïb Doukkali\
B.P. 20, El-Jadida, Maroc.\
E-mail : [email protected] , [email protected]
title: '**ÉTUDE DES SOLUTIONS MÉROMORPHES D’ÉQUATIONS DIFFÉRENTIELLES**'
---
Abstract. In this paper we shall study differential equations in the complex domain. The method of indeterminate coefficients and the majorant method lead to a proof of the existence and uniqueness of meromorphic solution of differential equations. We discuss their connection with the concept of algebraic integrability systems.\
AMS Subject Classification : 34M05, 34M45, 70H06.
Position du problème
====================
Dans ce travail, nous envisagerons l’étude des équations différentielles dans le domaine complexe. Soit le système d’équations différentielles non-linéaires $$\begin{aligned}
\frac {dw_{1}}{dz}&=&f_{1}\left( z,w_{1},...,w_{n}\right)
,\nonumber\\
&\vdots&\\
\frac {dw_{n}}{dz}&=&f_{n}\left(
z,w_{1},...,w_{n}\right),\nonumber\end{aligned}$$ où $f_{1},...,f_{n}$ sont des fonctions de $n+1$ variables complexes $z,w_{1},...,w_{n}$ et qui appliquent un domaine de $\mathbb{C}^{n+1}$ dans $\mathbb{C}$. Le problème de Cauchy consiste en la recherche d’une solution $\left( w_{1}\left(
z\right) ,...,w_{n}\left( z\right) \right) $ dans un voisinage d’un point $z_{0},$ passant par le point donné $\left(
z_{0},w_{1}^{0},...,w_{n}^{0}\right) $ c’est-à-dire satisfaisant aux conditions initiales $$\begin{aligned}
w_{1}(z_{0})&=&w_{1}^{0},\nonumber\\
&\vdots&\nonumber\\
w_{n}(z_{0})&=&w_{n}^{0}.\nonumber\end{aligned}$$ Notons que le système $\left( 1\right) $ peut s’écrire sous forme vectorielle dans $\mathbb{C}^{n}$ $$\frac{dw}{dz}=f(z,w(z)) ,$$ en posant $w=$ $(w_{1},...,w_{n})$ et $f=(f_{1},...,f_{n}).$ Dans ce cas, le problème de Cauchy consistera à déterminer la solution $w(z)$ telle que $$w(z_{0})
=w_{0}=(w_{1}^{0},...,w_{n}^{0}).$$
Commençons tout d’abord par décrire quelques résultats connus. On sait que lorsque les fonctions $f_{1},...,f_{n}$ sont holomorphes au voisinage du point $\left( z_{0},w_{1}^{0},...,w_{n}^{0}\right)
$ alors le problème de Cauchy admet une solution holomorphe et une seule. Une question naturelle se pose: le problème de Cauchy peut-il admettre quelque solution non holomorphe au voisinage du point $\left( z_{0},w_{1}^{0},...,w_{n}^{0}\right) ?$ Lorsque les fonctions $f_{1},...,f_{n}$ sont holomorphes, la réponse est négative. D’autres circonstances peuvent se produire pour le problème de Cauchy relatif au système d’équations différentielles $\left( 1\right) ,$ lorsque l’hypothèse d’holomorphie relative aux fonctions $f_{1},...,f_{n}$ n’est plus satisfaite au voisinage d’un point. On constate dans une telle éventualité que les comportements des solutions peuvent revêtir les aspects les plus divers. En général, les singularités des solutions sont de deux types: mobiles ou fixes, suivant qu’elles dépendent ou non des conditions initiales. Des résultats importants ont été obtenus par Painlevé $[14]$. Supposons par exemple que le système $(1)$ s’écrit sous la forme $$\begin{aligned}
\frac{dw_{1}}{dz}&=&\frac{P_{1}(z,w_{1}w_{n})}{Q_{1}(z,w_{1},...,w_{n})},\nonumber\\
&\vdots&\nonumber\\
\frac{dw_{n}}{dz}&=&\frac{P_{n}(
z,w_{1},...,w_{n})}{Q_{n}(z,w_{1},...,w_{n})},\nonumber\end{aligned}$$ avec $$P_{k}\left( z,w_{1},...,w_{n}\right) =\sum_{0\leq
i_{1},\ldots ,i_{n}\leq p}A_{i_{1},\ldots ,i_{n}}^{\left( k\right)
}\left( z\right) w_{1}^{i_{1}}...w_{n}^{i_{n}},\text{ }1\leq k\leq
n,$$$$Q_{k}\left( z,w_{1},...,w_{n}\right) =\sum_{0\leq
j_{1},\ldots ,j_{n}\leq q}B_{j_{1},\ldots ,j_{n}}^{\left( k\right)
}\left( z\right) w_{1}^{j_{1}}...w_{n}^{j_{n}},\text{ }1\leq k\leq
n,$$ des polynômes à plusieurs indéterminées $w_{1},...,w_{n}$ et à coefficients algébriques en $z.$ On sait\
0.7cm $\left( i\right) $ que les singularités fixes sont constituées par quatre ensembles de points. Le premier est l’ensemble des points singuliers des coefficients $A_{i_{1},\ldots ,i_{n}}^{\left( k\right) }\left( z\right) ,$ $B_{j_{1},\ldots ,j_{n}}^{\left( k\right) }\left( z\right) $ intervenant dans les polynômes $P_{k}\left(
z,w_{1},...,w_{n}\right) $ et $Q_{k}\left(
z,w_{1},...,w_{n}\right) .$ En général cet ensemble contient le point $z=\infty .$ Le second ensemble est constitué des points $\alpha _{j}$ tels que $$Q_{k}\left(
z,w_{1},...,w_{n}\right) =0,$$ circonstance qui se produit si les coefficients $B_{j_{1},\ldots ,j_{n}}^{\left( k\right) }\left(
z\right) $ s’annulent tous pour $z=\alpha _{j}.$ Le troisième est l’ensemble des points $\beta _{l}$ tels que pour certaines valeurs $\left( w_{1^{\prime }},...,w_{n^{\prime }}\right) $ de $\left( w_{1},...,w_{n}\right) ,$ on ait $$P_{k}\left( \beta
_{l},w_{1^{\prime }},...,w_{n^{\prime }}\right) =Q_{k}\left( \beta
_{l},w_{1^{\prime }},...,w_{n^{\prime }}\right) =0.$$ Dès lors les seconds membres du système ci dessus se présentent sous la forme indéterminée $\frac{0}{0}$ aux points $\left( \beta _{l},w_{1^{\prime }},...,w_{n^{\prime }}\right) .$ Enfin, l’ensemble des points $\gamma _{m}$ tels qu’il existe des valeurs $u_{1},...,u_{n},$ pour lesquelles $$R_{k}\left( \gamma
_{m},u_{1},...,u_{n}\right) =S_{k}\left( \gamma
_{m},u_{1},...,u_{n}\right) =0,$$ où $R_{k}$ et $S_{k}$ sont des polynômes en $u_{1},...,u_{n}$ obtenus à partir de $P_{k}$ et $Q_{k}$ en posant $$w_{1}=\frac{1}{u_{1}},\ldots
,w_{n}=\frac{1}{u_{n}}.$$ Chacun de ces ensembles ne comporte qu’un nombre fini d’éléments. Les singularités fixes du système en question sont en nombre fini.\
0.7cm $\left( ii\right) $ que les singularités mobiles de solutions de ce système sont des singularités mobiles algébriques: pôles et (ou) points critiques algébriques. Il n’y a pas de points singuliers essentiels pour la solution $\left( w_{1},...,w_{n}\right) .$\
Considérant le système d’équations différentielles $\left( 1\right) ,$ peut-on trouver des conditions suffisantes d’existence et d’unicité de solutions méromorphes? Nous établirons un théorème d’existence et d’unicité pour la solution du problème de Cauchy relatif au système d’équations différentielles $\left( 1\right) ,$ en faisant appel à la méthode des coefficients indéterminés. La solution sera explicitée sous la forme d’une série de Laurent. Il se posera dès lors le problème de la convergence. Celui-ci sera résolu par la méthode des fonctions majorantes (pour cette notion voir par exemple \[3, 7, 9\]). Nombreux sont les problèmes, aussi bien théorique que pratique, ou apparaissent des équations différentielles dont le second membre n’est pas holomorphe. Nous verrons, dans la dernière section, que les solutions méromorphes dépendant d’un nombre suffisant de paramètres libres jouent un rôle crucial dans l’étude des équations différentielles dites algébriquement intégrables.
Existence et unicité de solutions méromorphes
=============================================
Dans ce qui suivra, nous envisagerons le problème de Cauchy relatif au système normal $\left( 1\right) $ dans l’hypothèse où $f_{1},...,f_{n}$ ne dépendent pas explicitement de $z$ c’est-à-dire $$\begin{aligned}
\frac {dw_{1}}{dz}&=&f_{1}\left(w_{1},...,w_{n}\right)
,\nonumber\\
&\vdots&\\
\frac {dw_{n}}{dz}&=&f_{n}\left(w_{1},...,w_{n}\right),\nonumber\end{aligned}$$ On suppose que $f_{1},...,f_{n}$ sont des fonctions rationnelles en $w_{1},...,w_{n}$ et que le système $\left( 2\right) $ est quasi-homogène, c’est-à-dire ils existent des entiers positifs $s_{1},...,s_{n}$ telles que $$f_{i}(\alpha ^{s_{1}}w_{1},...,\alpha ^{s_{n}}w_{n})=\alpha
^{s_{i}+1}f_{i}(w_{1},...,w_{n}) ,\quad 1\leq i\leq n,$$ pour chaque constante non nulle $\alpha $. Autrement dit, le système $(2) $ est invariant par la transformation $$z\longrightarrow \alpha ^{-1}z,\text{ }w_{1}\longrightarrow\alpha
^{s_{1}}w_{1},\ldots ,\text{ }w_{n}\longrightarrow \alpha
^{s_{n}}w_{n}.$$ Notons que si le déterminant $$\label{eqn:euler}
\Delta \equiv \det (w_{j}\frac{\partial f_{i}}{\partial
w_{j}}-\delta _{ij}f_{i}) _{1\leq i,j\leq n},$$ est non identiquement nul, alors le choix des nombres $s_{1},...,s_{n}$ est unique.\
Dans tout ce qui va suivre, nous supposerons, pour simplifier les notations que $z_0=w_0=0,$ ce qui n’affecte pas la généralité des résultats.
Supposons que $$\label{eqn:euler}
w_{i}=\frac{1}{z^{s_{i}}}\sum_{k=0}^{\infty }c_{i}^{(k)
}z^{k},\quad 1\leq i\leq n,$$ où $c^{(0)}\neq 0$, soit la solution formelle en séries de Laurent, obtenue par la méthode des coefficients indéteminés, du système quasi-homogène $(2)$. Alors, les coefficients $c_{i}^{(0)}$ satisfont aux équations non-linéaires $$\label{eqn:euler}
s_{i}c_{i}^{(0)}+f_{i}(c_{1}^{(0)},...,c_{n}^{(0)})=0,$$ où $1\leq i\leq n,$ tandis que $c_{i}^{(1)},c_{i}^{(2)},...$ satisfont chacun à un système d’équations linéaires de la forme $$\label{eqn:euler}
(\mathcal{L}-k\mathcal{I}) c^{(k)}=\mbox{polyn\^{o}me en}\quad
c_{i}^{( 0)},...,c_{i}^{(k-1)},\quad 1\leq i\leq n,\quad k\geq 1,$$ où $c^{(k)}=(c_1^{(k)},...,c_n^{(k)})^\top$ et $$\mathcal{L}\equiv(\frac{\partial f_{i}}{\partial
w_{j}}(c_{1}^{(0)},...,c_{n}^{(0)}) +\delta _{ij}s_{i}) _{1\leq
i,j\leq n},$$ est la matrice jacobienne de (5). En outre, la série $(4) $ est convergente.
*Démonstration*: En substituant $\left(4\right) $ dans $\left( 2\right) $, tout en tenant compte de la quasi-homogénéité du système, on obtient $$\begin{aligned}
\sum_{k=0}^{\infty }(k-s_{i}) c_{i}^{(k) }z^{k-s_{i}-1}&=&f_{i}(
\sum_{k=0}^{\infty }c_{1}^{(k) }z^{k-s_{1}},...,\sum_{k=0}^{\infty
}c_{n}^{( k)}z^{k-s_{n}})
,\nonumber\\
&=&f_{i}(z^{-s_{1}}(c_{1}^{(0) }+\sum_{k=1}^{\infty }c_{1}^{(
k)}z^{k}) ,...,z^{-s_{n}}(c_{n}^{(0)}+\sum_{k=1}^{\infty
}c_{n}^{(k)}z^{k})) ,\nonumber\\
&=&z^{-s_{i}-1}f_{i}(c_{1}^{(0)}+\sum_{k=1}^{\infty }c_{1}^{(k)
}z^{k},...,c_{n}^{(0)}+\sum_{k=1}^{\infty }c_{n}^{(k)
}z^{k}).\nonumber\end{aligned}$$ Ensuite, on développe le second membre comme suit $$\begin{aligned}
\sum_{k=0}^{\infty}(k-s_{i})
c_{i}^{(k)}z^{k}&=&f_{i}(c_{1}^{(0)},...,c_{n}^{(0)})+
\sum_{j=1}^{n}\frac{\partial f_{i}}{\partial
w_{j}}(c_{1}^{(0)},...,c_{n}^{(0)})
\sum_{k=1}^{\infty }c_{j}^{(k)}z^{k}\nonumber\\
&&+\sum_{k=2}^\infty z^k\sum_{(\alpha, \tau)\in
D_k}\frac{1}{\alpha !}\frac{\partial ^{\alpha }f_{i}}{\partial
w^{\alpha }}(c_{1}^{(0) },...,c_{n}^{( 0)})
\prod_{j=1}^{n}(c_{j}^{(\tau _{j})}) ^{\alpha _{j}},\nonumber\end{aligned}$$ où $\alpha =\left( \alpha _{1},...,\alpha _{n}\right),$ $\tau=\left( \tau _{1},...,\tau _{n}\right) ,$ $$\left| \alpha \right| =\sum_{j=1}^{n}\alpha _{j},\quad
\alpha !=\prod_{j=1}^{n}\alpha _{j}!,$$ $$D_k=\{(\alpha, \tau):\tau_j>0, \forall j, \left|\alpha \right|>2, \sum_{j=1}^{n}\alpha
_{j}\tau _{j}=k\}.$$ En identifiant les termes ayant même puissance au premier et au second membre, on obtient successivement pour $k=0$ l’expression $\left(5\right),$ pour $k=1,$ $$\left( \mathcal{L}-\mathcal{I}\right)c^{\left(1\right)
}=0,$$ et pour $k\geq 2,$ $$\label{eqn:euler}
((\mathcal{L}-k\mathcal{I})c^{(k)})_i=-\sum_{(\alpha, \tau)\in
D_k}\frac{1}{\alpha !}\frac{\partial ^{\alpha }f_{i}}{\partial
w^{\alpha }}(c_{1}^{(0)},...,c_{n}^{(0)})
\prod_{j=1}^{n}(c_{j}^{(\tau _{j})}) ^{\alpha _{j}},$$ où $\tau _{j}> 0,$ $\displaystyle{\sum_{j=1}^{n}\alpha _{j}\tau
_{j}=k},$ ce qui conduit aux expressions (explicites) $\left(
6\right) .$ La solution obtenue par la méthode des coefficients indéterminés est formelle du fait que nous l’obtenons en effectuant sur des séries, que nous supposons a priori convergentes, diverses opérations dont la validité reste à justifier. Le théorème se trouvera donc établi dès que nous aurons vérifié que ces séries sont convergentes. On utilise à cette fin la méthode des fonctions majorantes ainsi que les travaux de M. Adler-P. van Moerbeke \[1\] et J.P. Françoise \[4\]. Notons tout d’abord que des paramètres libres apparaissent soit dans le système $\left( 5\right) $ de $n$ équations à $n$ inconnues, lorsque celui-ci admet un ensemble continue de solutions, soit par le fait que $\lambda _{i}\equiv k\in \mathbb{N}^{*},1\leq i\leq
n,$ est une valeur propre de la matrice $\mathcal{L}.$ Dès lors, les coefficients peuvent être vus comme étant des fonctions rationnelles sur une variété affine $V,$ de fibre le lieu $$\bigcap_{i=1}^{n}\left\{ s_{i}c_{i}^{\left( 0\right) }+
f_{i}\left(c_{1}^{\left( 0\right) },...,c_{n}^{\left( 0\right)
}\right) =0\right\}.$$ Soit $n_0\in V$ et soit $K$ un sous-ensemble compact de $V,$ contenant un voisinage ouvert de $n_0.$ Notons que $K$ peut-être muni de la topologie du plan complexe. Posons $$A=1+\max \left\{\left| c_{1}^{\left(\tau _{1}\right)}(n_0)\right|,
\left| c_{2}^{\left( \tau _{2}\right) }(n_0)\right| ,...,\left|
c_{n}^{\left( \tau _{n}\right)}\right|(n_0) \right\}, 1\leq i\leq
n, 1\leq \tau _{i}\leq \lambda _{n},$$ où $\lambda_n$ désigne la plus grande valeur propre de la matrice $\mathcal{L}.$ Soient $B$ et $C$ deux constantes avec $C>A$ telles que dans le compact $K$ on ait $$\left| \frac{\partial ^{\alpha }f_{i}}{\partial w^{\alpha }}
(n_0) \right| \leq \alpha !B^{\left| \alpha \right| },$$$$\left|
\left( \mathcal{L}(n_0)-k\mathcal{I}_n\right) ^{-1}\right| \leq
C,\quad k\geq \lambda _{n}+1.$$ De $\left(7\right) $ on déduit que $$\left| c_{i}^{\left( k\right)}(n_0)\right| \leq C\sum_{(\alpha, \tau)\in D}
B^{\left| \alpha \right| }\prod_{j=1}^{n}\left| c_{j}^{\left( \tau
_{j}\right) }\right| ^{\alpha _{j}},\quad k\geq \lambda _{n}+1.$$ Considérons maintenant la série $$\Phi\left( z\right) =Az+\sum_{k=2}^{\infty }\beta _{k}z^{k},$$ où $\beta _{k}$ sont des nombres réels définis inductivement par $\beta_1\equiv A$ et $$\beta _{k}\equiv C\sum_{(\alpha, \tau)\in D}B^{\left| \alpha
\right| }\prod_{j=1}^{n}\beta ^{\alpha _{j}}_{\tau _{j}},\quad
k\geq 2.$$ On vérifie aisément par récurrence que la série $\Phi\left( z\right)$ est une majorante pour $$\sum_{k=1}^{\infty }c_{i}^{\left( k\right) }z^{k},\quad 1\leq
i\leq n.$$ En effet, on a $\left| c_{i}^{\left( 1\right) }\right|
\leq A.$ Supposons que $\left| c_{i}^{\left( j\right) }\right|
\leq \beta _{j}, j < k, \forall i.$ Alors $$\begin{aligned}
\left| c_{i}^{\left( k\right) }(n_0)\right|& \leq &
C\sum_{(\alpha, \tau)\in D}B^{\left| \alpha \right|
}\prod_{j=1}^{n}\left| c_{j}^{\left( \tau _{j}\right) }\right|
^{\alpha _{j}},\text{ }k\geq \lambda _{n}+1,\nonumber\\
&\leq &C\sum_{(\alpha, \tau)\in D}B^{\left| \alpha \right|
}\prod_{j=1}^{n}\left| \beta ^{\alpha _{j}}_{\tau _{j}}\right|,\nonumber\\
&=&\beta _{k}.\nonumber\end{aligned}$$ D’autre part, il résulte de la définition des nombres $\beta
_{k}$ que $$\Phi \left( z\right) =Az+CB^{2}\frac{\left( n\Phi
\left( z\right) \right) ^{2}}{1-Bn\Phi \left( z\right) }.$$La racine $$\Phi (z)=\frac{1+nABz-\sqrt{(1-2nAB(1+2nBC)
z+n^{2}A^{2}B^{2}z^{2})}}{2nB(1+nBC)},$$ fournit la majorante cherchée. D’où la possibilité d’un développement en série entière au voisinage de $z=0.$ Ceci achève la démonstration. $\square$\
La série (4) est l’unique solution méromorphe dans le sens où cette solution résulte de ce que les coefficients$\quad$$c_{i}^{\left( k\right) }$ se trouvent déterminés de façon univoque avec la méthode de calcul adopté.
Le résultat du théorème précédent s’applique à l’équation différentielle quasi-homogène d’ordre $n$ suivante : $$\begin{aligned}
\label{eqn:euler}
\frac{d^{n}w}{dz^{n}}=f\left(w,\frac{dw}{dz},...,\frac{d^{n-1}w}{dz^{n-1}}\right).\end{aligned}$$ $f$ étant une fonction rationnelle en $w,\frac{dw}{dz},...,\frac{d^{n-1}w}{dz^{n-1}}$ et $$\begin{aligned}
w(z_{0})&=&w_{1}^{0},\nonumber\\
\frac{dw}{dz}(z_{0})&=&w_{2}^{0},\nonumber\\
&\vdots&\nonumber\\
\frac{d^{n-1}w}{dz^{n-1}}(z_{0})&=&w_{n}^{0}.\nonumber\end{aligned}$$ En effet, l’équation $(8)$ se ramène à un système de $n$ équations du premier ordre en posant $$\begin{aligned}
w\left( z\right)&=&w_{1}\left( z\right),\nonumber\\
\frac{dw}{dz}\left( z\right)&=&w_{2}\left(
z\right),\nonumber\\
&\vdots&\nonumber\\
\frac{d^{n-1}w}{dz^{n-1}}\left( z\right)&=&w_{n}\left(
z\right).\nonumber\end{aligned}$$ On obtient ainsi $$\begin{aligned}
\frac{dw_{1}}{dz}&=&w_{2},\nonumber\\
\frac{dw_{2}}{dz}&=&w_{3},\nonumber\\
&\vdots&\nonumber\\
\frac{dw_{n-1}}{dz}&=&w_{n},\nonumber\\
\frac{dw_{n}}{dz}&=&f\left(w_{1},w_{2},...,w_{n}\right).\nonumber\end{aligned}$$ Un tel système constitue un cas particulier du système normal $\left(2\right)$.
Connection avec l’intégrabilité algébrique des systèmes hamiltoniens
====================================================================
Considérons un sytème hamiltonien $$\label{eqn:euler}
\frac{dw}{dz}=J\left(w\right) \frac{\partial H}{\partial w},\quad
w\in \mathbb{R}^{n},\quad n=2m+k,$$ où $H$ est l’hamiltonien et $J\left( w\right) $ est une matrice réelle antisymétrique telle que les crochets de Poisson correspondants vérifient l’identité de Jacobi:$$\left\{
\left\{ H,F\right\} ,G\right\} +\left\{ \left\{ F,G\right\}
,H\right\} +\left\{ \left\{ G,H\right\} ,F\right\} =0, \forall H,
F, G\in \mathcal{C}^\infty (\mathbb{R}^n),$$ où $$\left\{ H,F\right\}=\sum_{i,j}J_{ij}\frac{\partial H}{\partial
w_{i}}\frac{\partial F}{\partial w_{j}}.$$ On suppose que le système $(9)$ est complètement intégrable c’est-à-dire qu’il admet $m+k$ intégrales premières $H_{1}=H,H_{2},...,H_{m+k}$ fonctionnellement indépendantes dont $m$ intégrales sont en involution (i.e.,$\{ H_{i},H_{j}\} =0,$ $1\leq i,j\leq m,$), $k$ intégrales sont des fonctions de Casimir (i.e.,$J\frac{\partial H_{m+i}}{\partial w}=0,$ $1\leq i\leq k$) et telles que pour presque tous les $c_{i}\in \mathbb{R}$ les variétés invariantes $$\label{eqn:euler}
\overset{m+k}{\underset{i=1}{\bigcap}}\{w\in
\mathbb{R}^{n}:H_{i}(w) =c_{i}\} ,$$ sont compactes et connexes. D’après le théorème d’Arnold-Liouville $\left[ 10\right]$, les variétés $\left( 10\right) $ sont difféomorphes aux tores réels $T_{\mathbb{R}}^{m}=\mathbb{R}^{m}/r\acute{e}seau.$ En outre les flots définis par les champs de vecteurs $W_{H_{i}},1\leq
i\leq m,$ sont des mouvements rectilignes sur ce tore et les équations du problème sont intégrables par quadratures.\
Soient maintenant $w\in \mathbb{C}^{n}$, $z\in \mathbb{C}$ et $\Delta \subset $ $\mathbb{C}^{n}$ un ouvert non vide de Zariski. Comme $H_{1},...,H_{m+k}$ sont fonctionnellement indépendantes, alors l’application $$\varphi =\left( H_{1},...,H_{m+k}\right) :\mathbb{C}^{n}\longrightarrow
\mathbb{C}^{m+k},$$est une submersion générique sur $\Delta .$ Soit $ \textbf{I}=\varphi (\mathbb{C}^{n}\backslash
\Delta),$ le lieu critique de $\varphi $ et désignons par $adh
\textbf{I}$ l’adhérence (ou fermeture) de Zariski de $\textbf{I}$ dans $\mathbb{C}^{m+k}.$
Le système différentiel $\left(9\right)$ dont le côté droit est polynomial est algébriquement complètement intégrable si pour $c=(c_1,...,c_{m+k})\in \mathbb{C}^{m+k}$ $\backslash $ $adh \textbf{I},$ la fibre $$\label{eqn:euler}
M_c\equiv\varphi^{-1}\left( c\right) =\bigcap_{i=1}^{m+k}\left\{
w\in \mathbb{C}^{n}:H_{i}\left( w\right) =c_{i}\right\} ,$$ est la partie affine d’une variété abélienne (i.e., un tore complexe $T_{\mathbb{C}}^{m}\simeq
\mathbb{C}^{m}/r\acute{e}seau$ qui possède un plongement dans un espace projectif ). En outre, les flots $g_{W_{i}}^{z}\left(
w\right) ,$ $w\in M_c, $ $z\in \mathbb{C},$ définies par les champs de vecteurs $W_{H_{1}},...,W_{H_{m}}$ sont des lignes droites sur $T_{\mathbb{C}}^{m}$ c’est-à-dire $$\left[
g_{W_{i}}^{z}\left( w\right) \right] _{j}=f_{j}\left(
p+z(k_{1}^{i},...,k_{n}^{i})\right) ,$$où $f_{j}\left(
z_{1},...,z_{m}\right) $ sont des fonctions abéliennes (méromorphes) sur le tore $T_{\mathbb{C}}^{m},$ $f_{j}(p)=w_{j},$ $1\leq j\leq n.$
Soit $\overline{M_c}$ la fermeture projective de $M_c$ dans l’espace projectif complexe $\mathbb{CP}^{n}$ de dimension $n.$ Alors $\overline{M_c}$ n’est pas une variété abélienne puisque cette dernière n’est pas simplement connexe et ne peut donc en général être une intersection complète projective. Dès lors, pour que $M_c$ soit la partie affine d’une variété abélienne, la variété $\overline{M_c}$ doit être singulière à l’infini. En éclatant la singularité le long du lieu atteint par le flot et en implosant la partie du lieu qui n’est pas atteint par le flot, on montre que la variété $\overline{M_c}$ se transforme en une variété abélienne $\widetilde{M_c}$ et le lieu à l’infini se transforme en une ou plusieurs sous-variétés de codimension $1.$ Et c’est là où le théorème 1, va jouer un rôle crucial. On procède comme suit : soit $w_{i}\longrightarrow u_{i}$ une transformation birationnelle telle qu’au voisinage de $z=0,$ on ait : $$\begin{aligned}
u_{i}&=&\alpha _{i}+\circ\left( z\right) ,\text{ }1\leq i\leq
n-1,\nonumber\\
u_{n}&=&z+\circ\left( z^{2}\right) ,\nonumber\end{aligned}$$ où $\alpha _{1},...,\alpha _{n-1}$ sont des paramètres libres. Les nouvelles variables $u_{i}$ ont pour effet d’éclater la singularité de la variété projective $\overline{M_c}$ le long du lieu à l’infini atteint par le flot. Exprimées dans ces nouvelles variables $u_{1},...,u_{n},$ les équations différentielles sont régulières et holomorphes au voisinage de $u_{n}=0$ tandis que les équations définissant la fibre $M_c$ s’écrivent sous la forme : $$F_{i}\left( u_{1}\left( z\right) ,...,u_{n-1}\left( z\right)
,u_{n}\left( z\right) \right) =c_{i},\text{ }1\leq i\leq m+k,$$ où $F_{1},...,F_{m+k}$ sont des polynômes en $w$. Pour $z=0,$ on obtient $$F_{i}\left( \alpha _{1},...,\alpha
_{n-1},0\right) =a_{i},\text{ }1\leq i\leq m+k,$$ et ces relations algébriques entre les paramètres libres $\alpha
_{1},...,\alpha _{n-1}$ fournissent les équations d’une sous-variété $\mathcal{D}$ qui jouera, entre autres, un rôle important dans la compactification de la fibre $M_c.$ En fait, les paramètres libres $\alpha _{1},...,\alpha _{n-1}$ et la sous variété $\mathcal{D}$ peuvent s’obtenir directement de la manière suivante: d’abord l’on montre l’existence de solutions $w=(w_{1},w_{2},\ldots ,w_{n})$ du système $\left( 9\right) $ sous la forme de séries de Laurent $\left( 4\right) $ dépendant de $n-1$ paramètres libres $\alpha _{1},...,\alpha _{n-1}.$ En substituant ces développements dans le système $\left( 9\right) $, on voit (d’après le théorème 1) que les coefficients $c^{\left( 0\right) },c^{\left( 1\right) },...,$ satisfont aux équations $\left( 5\right) $ et $\left( 6\right).$ L’étape suivante consiste à considérer la fermeture $\mathcal{D}$ des composantes continues de l’ensemble des séries de Laurent de $w\left( z\right) $ tels que: $H_{1}\left( w\right)
=a_{1},\ldots ,$ $H_{m+k}\left( w\right) =a_{m+k}.$ Plus précisement,$$\mathcal{D}=\bigcap_{i=1}^{m+k}\left\{
\text{coefficient de }z^{0}\text{ dans }H_{i}\left( w\left(
z\right) \right) =a_{i}\right\} .$$ C’est une sous-variété (un diviseur) de codimension $1.$ Ensuite on procède à la compactification de la fibre $M_c$ $\left( 11\right) $ en une variété abélienne $\widetilde{M_c}$. Cette compactification s’obtient par l’adjonction à $M_c$ de ce diviseur $\mathcal{D}$.
Rotation d’un corps solide autour d’un point fixe
-------------------------------------------------
Les équations du mouvement d’un corps solide autour d’un point fixe s’écrivent sous la forme $$\begin{aligned}
\frac{dM}{dt}&=&M\wedge \Omega +\mu g\text{ }\Gamma \wedge L,\\
\frac{d\Gamma }{dt}&=&\Gamma \wedge \Omega ,\nonumber\end{aligned}$$ où $\wedge $ est le produit vectoriel dans $\mathbb{R}^{3},$ $M=\left( m_{1},m_{2},m_{3}\right) $ le moment angulaire du solide$,\Omega =\left( m_{1}/I_{1},m_{2}/I_{2},m_{3}/I_{3}\right)
$ la vitesse angulaire, $I_{1},I_{2}$ et $I_{3},$ les moments d’inertie, $\Gamma =\left( \gamma _{1},\gamma _{2},\gamma
_{3}\right) $ le vecteur vertical unitaire, $\mu $ la masse du solide, $g$ l’accélération de la pesanteur, et enfin, $L=\left( l_{1},l_{2},l_{3}\right) $ le vecteur unitaire ayant pour origine le point fixe et dirigé vers le centre de gravité; tous ces vecteurs sont considérés dans un système mobile dont les coordonnées sont fixées aux axes principaux d’inertie. L’espace de configuration d’un solide avec un point fixe est le groupe des rotations: $$SO\left( 3\right) =\left\{ U\text{ matrice d'ordre trois}:\text{ }U^{-1}=
U^{\top },\text{ }\det U=1\right\}.$$ C’est le groupe des matrices orthogonales d’ordre trois et le mouvement de ce solide est décrit par une courbe sur ce groupe. L’espace des vitesses angulaires de toutes les rotations est l’algèbre de Lie du groupe $SO\left( 3\right) ;$ c’est l’algèbre $$so(3)=\left\{ A\text{ matrice d'ordre trois}:U^{\top }+U=0\right\}
,$$ des matrices antisymétriques d’ordre trois$.$ Cette algèbre est engendrée comme espace vectoriel par les matrices $${e_{1}}=\left(\begin{array}{ccc}
0&0&0\\
0&0&-1\\
0&1&0
\end{array}\right),\qquad
{e_{2}}=\left(\begin{array}{ccc}
0&0&1\\
0&0&0\\
-1&0&0
\end{array}\right),\qquad
{e_{1}}=\left(\begin{array}{ccc}
0&-1&0\\
1&0&0\\
0&0&0
\end{array}\right),$$ qui vérifient les relations de commutation $$\left[
e_{1},e_{2}\right] =e_{3},\quad\left[ e_{2},e_{3}\right]
=e_{1},\quad\left[ e_{3},e_{1}\right] =e_{2}.$$ On utilisera dans la suite le fait que si l’on identifie $so\left( 3\right) $ à $\mathbb{R}^{3}$ en envoyant $\left( e_{1},e_{2},e_{3}\right) $ sur la base canonique de $\mathbb{R}^{3},$ le crochet de $so\left(
3\right) $ correspond au produit vectoriel. En d’autres termes, considérons l’application $$\mathbb{R}^{3}\longrightarrow so(3),\text{ }a=
\left( a_{1},a_{2},a_{3}\right) \longmapsto
{A}=\left(\begin{array}{ccc}
0&-a_{3}&a_{2}\\
a_{3}&0&-a_{1}\\
-a_{2}&a_{1}&0
\end{array}\right),$$ laquelle définit un isomorphisme entre les algèbres de Lie $\left( \mathbb{R}^{3},\wedge \right) $ et $\left( so(3),\left[
,\right] \right) $ où $$a\wedge b\longmapsto \left[ A,B\right] =AB-BA.$$ En utilisant cet isomorphisme, on peut réecrire le système $\left(12\right) $ sous la forme $$\begin{aligned}
\frac{dM}{dt}&=&\left[ M,\Omega \right] +\mu g\text{ }\left[
\Gamma,L\right] ,\nonumber\\
\frac{d\Gamma }{dt}&=&\left[ \Gamma ,\Omega \right] ,\nonumber\end{aligned}$$ où $$M=\left( M_{ij}\right) _{1\leq i,j\leq 3}\equiv
\sum_{i=1}^{3}m_{i}e_{i}\equiv \left(\begin{array}{ccc}
0&-m_{3}&m_{2}\\
m_{3}&0&-m_{1}\\
-m_{2}&m_{1}&0
\end{array}\right)\in so\left( 3\right) ,$$ $$\Omega =\left( \Omega _{ij}\right) _{1\leq i,j\leq 3}\equiv
\sum_{i=1}^{3}\omega _{i}e_{i}\equiv \left(\begin{array}{ccc}
0&-\omega_{3}&\omega_{2}\\
\omega_{3}&0&-\omega_{1}\\
-\omega_{2}&\omega_{1}&0
\end{array}\right)\in so\left( 3\right) ,$$ $$\Gamma =\left( \gamma _{ij}\right) _{1\leq i,j\leq 3}\equiv
\sum_{i=1}^{3}\gamma _{i}e_{i}\equiv \left(\begin{array}{ccc}
0&-\gamma_{3}&\gamma_{2}\\
\gamma_{3}&0&-\gamma_{1}\\
-\gamma_{2}&\gamma_{1}&0
\end{array}\right)\in so\left( 3\right) ,$$ et $$L = \left(\begin{array}{ccc}
0&-l_{3}&l_{2}\\
l_{3}&0&-l_{1}\\
-l_{2}&l_{1}&0
\end{array}\right)\in so\left( 3\right) ,$$ En tenant compte du fait que $M=I\Omega ,$ alors les équations pécédentes deviennent $$\begin{aligned}
\frac{dM}{dt}&=&\left[ M,\Lambda M\right] +\mu g\text{ }\left[
\Gamma ,L\right] ,\\
\frac{d\Gamma }{dt}&=&\left[ \Gamma ,\Lambda M\right] ,\nonumber\end{aligned}$$ où $$\Lambda M=\left( \Lambda _{ij}M_{ij}\right) _{1\leq i,j\leq 3}\equiv
\sum_{i=1}^{3}\lambda _{i}m_{i}e_{i}\equiv
\left(\begin{array}{ccc}
0&-\lambda _{3}m_{3}&\lambda _{2}m_{2}\\
\lambda _{3}m_{3}&0&-\lambda _{1}m_{1}\\
-\lambda _{2}m_{2}&\lambda _{1}m_{1}&0
\end{array}\right)\in so\left( 3\right),$$ avec $\lambda _{i}\equiv I_{i}^{-1}.$ Le système $\left(
13\right) $ est complètement intégrable seulement dans les cas suivants :
$a)$ : Dans ce cas, on a $$l_{1}=l_{2}=l_{3}=0,$$ c’est-à-dire le point fixe est son centre de gravité. Autrement dit, Les équations d’Euler (On parle aussi de mouvement d’Euler-Poinsot du solide) du mouvement de rotation d’un solide autour d’un point fixe, pris comme origine du repère lié au solide, lorsqu’aucune force extérieure n’est appliquée au système, peuvent s’écrire forme explicite $$\begin{aligned}
\frac{dm_{1}}{dt}&=&\left( \lambda _{3}-\lambda _{2}\right)
m_{2}m_{3},\nonumber\\
\frac{dm_{2}}{dt}&=&\left( \lambda _{1}-\lambda _{3}\right) m_{1}m_{3},\\
\frac{dm_{3}}{dt}&=&\left( \lambda _{2}-\lambda _{1}\right)
m_{1}m_{2}.\nonumber\end{aligned}$$ Ces équations forment un champ de vecteurs hamiltonien de la forme (9) avec $n=3$, $m=k=1$, $z=t$, $w= \left(
m_{1},m_{2},m_{3}\right)^{\intercal}$, $$H=\frac{1}{2}\left( \lambda _{1}m_{1}^{2}+\lambda
_{2}m_{2}^{2}+\lambda _{3}m_{3}^{2}\right),$$ l’hamiltonien et $$J=\left(\begin{array}{ccc}
0&-m_{3}&m_{2}\\
m_{3}&0&-m_{1}\\
-m_{2}&m_{1}&0
\end{array}\right)\in so\left( 3\right).$$ Ces équations admettent deux intégrales premières quadratiques : $H_1=H$ et $$H_2=\frac{1}{2}\left( m_{1}^{2}+m_{2}^{2}+m_{3}^{2}\right).$$ Ces intégrales sont fonctionnellement indépendantes, en involution et le système en question est complètement intégrable. La résolution explicite des équations d’Euler est délicate dans le cas général où $\lambda _{1}$, $\lambda _{2}$ et $\lambda _{3}$ sont tous différents; les solutions s’expriment à l’aide de fonctions elliptiques de Jacobi comme suit (pour le détail voir par exemple \[15\]) : $$\label{eqn:euler}
\left\{\begin{array}{rl}
m_1=\sqrt{\frac{2H_1-H_2\lambda_{3}}{\lambda_1-\lambda_3}}&\mathbf{cn}(t\sqrt{(\lambda
_{2}-\lambda _{3})(H_2\lambda
_{1}-2H_1)}),\\
m_2=\sqrt{\frac{2H_1-H_2\lambda_{3}}{\lambda_2-\lambda_3}}&\mathbf{sn}(t\sqrt{(\lambda
_{2}-\lambda _{3})(H_2\lambda
_{1}-2H_1)}),\\
m_3=\sqrt{\frac{H_2\lambda_{1}-2H_1}{\lambda_1-\lambda_3}}&\mathbf{dn}(t\sqrt{(\lambda
_{2}-\lambda _{3})(H_2\lambda _{1}-2H_1)}).
\end{array}\right.$$ Le mouvement de $\left( m_{1},m_{2},m_{3}\right) $ s’effectue sur l’intersection $M_c$ d’un ellipsoide avec une sphère. Les deux cercles définies par $M_c,$ forme la partie réelle d’un tore complexe de dimension $1,$ définie par la courbe elliptique $\mathcal{E}$ : $$\mathcal{E}: y^2=(1- s^{2})(1-k^2 s^{2}).$$ où $$k^2\equiv\frac{(\lambda _{1}-\lambda
_{2})(2H_1-H_2\lambda _{3})}{(\lambda_2-\lambda_3)(H_2\lambda
_{1}-2H_1)}.$$ et $$s\equiv m_2\sqrt{\frac{\lambda _{2}-\lambda _{3}}{2H_1-H_2\lambda
_{3}}}$$ L’intersection complexe $\left( \subset \mathbb{C}^{3}\right) $ est la partie affine d’une courbe elliptique $\overline{M_c}\in
\mathbb{CP}^{3}$. On montre que $\overline{M_c}$ est isomorphe à la courbe elliptique $\mathcal{E}$. En outre l’intersection réelle $\left( \subset \mathbb{R}^{3}\right) $ s’étend au tore complexe $\mathbb{C}/r\acute{e}seau$ et le flot se linéarise sur ce tore. Si $p\left( t\right) =\left( m_{1}\left( t\right)
,m_{2}\left( t\right) ,m_{3}\left( t\right) \right)$, est une solution de $\left( 14\right)$, la loi reliant $p\left(
t_{1}+t_{2}\right)$ à $p\left( t_{1}\right) $ et $p\left(
t_{2}\right)$ est la loi d’addition sur la courbe elliptique (voir \[15\]). D’après les équations $\left( 14\right) $, l’unique différentielle holomorphe sur $\overline{M_c}$ est donnée par $$\omega =\frac{dm_{1}}{\left( \lambda _{3}-\lambda
_{2}\right) m_{2}m_{3}}=\frac{dm_{2}}{\left( \lambda _{1}-\lambda
_{3}\right) m_{1}m_{3}}=\frac{dm_{3}}{\left( \lambda _{2}-\lambda
_{1}\right) m_{1}m_{2}},$$d’où$$t=\int_{p\left( 0\right)
}^{p\left( t\right) }\omega ,\text{ \quad }p\left( 0\right) \in
\overline{M_c}.$$ Le système $\left( 14\right) $ est invariant par les transformations $$t\longrightarrow\alpha ^{-1}t,\text{
}m_{1}\longrightarrow \alpha m_{1},\text{ }m_{2}\longrightarrow
\alpha m_{2},\text{ }m_{3}\longrightarrow \alpha m_{3}.$$ Celles-ci sont uniques puisque le déterminant $\left( 3\right) $ est égal à $$\begin{aligned}
\Delta&=&\left|
\begin{array}{ccc}
-\left( \lambda _{3}-\lambda _{2}\right) m_{2}m_{3}&\left( \lambda
_{3}-\lambda _{2}\right) m_{2}m_{3}&\left( \lambda _{3}-\lambda
_{2}\right) m_{2}m_{3}\\
\left( \lambda _{1}-\lambda _{3}\right) m_{1}m_{3}&-\left( \lambda
_{1}-\lambda _{3}\right) m_{1}m_{3}&\left( \lambda _{1}-\lambda
_{3}\right) m_{1}m_{3}\\
\left( \lambda _{2}-\lambda _{1}\right) m_{1}m_{2}&\left( \lambda
_{2}-\lambda _{1}\right) m_{1}m_{2}&-\left( \lambda _{2}-\lambda
_{1}\right) m_{1}m_{2}
\end{array}
\right|,\nonumber\\
&=&4\left( \lambda _{3}-\lambda _{2}\right) \left( \lambda
_{1}-\lambda _{3}\right) \left( \lambda _{2}-\lambda _{1}\right)
m_{1}^{2}m_{2}^{2}m_{3}^{2},\nonumber\\
&\neq &0.\nonumber\end{aligned}$$ On peut donc chercher des solutions du système $\left(
14\right)$ sous la forme de séries de Laurent $$\begin{aligned}
m_{1}&=&\frac{1}{t}\left( a_{0}+a_{1}t+a_{2}t^{2}+\cdots
\right),\nonumber\\
m_{2}&=&\frac{1}{t}\left( b_{0}+b_{1}t+b_{2}t^{2}+\cdots \right),\nonumber\\
m_{3}&=&\frac{1}{t}\left( c_{0}+c_{1}t+c_{2}t^{2}+\cdots
\right),\nonumber\end{aligned}$$ dépendant de $\dim (\mbox{espace de phase})-1=2$ paramètres libres. En substituant ces équations dans le système $\left( 14\right) $, on voit que :
1\) les coefficients $a_0,b_0,c_0,$ satisfont aux équations $$\begin{aligned}
a_{0}+\left( \lambda _{3}-\lambda _{2}\right) b_{0}c_{0}&=&0,\nonumber\\
b_{0}+\left( \lambda _{1}-\lambda _{3}\right) a_{0}c_{0}&=&0,\nonumber\\
c_{0}+\left(\lambda_{2}-\lambda_{1}\right)a_{0}b_{0}&=&0,\nonumber\end{aligned}$$ dont les solutions sont\
: $a_{0}=\frac{-1}{\sqrt{\left( \lambda
_{2}-\lambda _{1}\right) \left( \lambda _{1}-\lambda _{3}\right)
}},\quad b_{0}=\frac{1}{\sqrt{\left( \lambda _{2}-\lambda
_{1}\right) \left( \lambda _{3}-\lambda _{2}\right) }},\quad
c_{0}=\frac{1}{\sqrt{\left( \lambda _{1}-\lambda _{3}\right)
\left( \lambda _{3}-\lambda _{2}\right) }}.$\
: $a_{0}=\frac{1}{\sqrt{\left( \lambda
_{2}-\lambda _{1}\right) \left( \lambda _{1}-\lambda _{3}\right)
}},\quad b_{0}=\frac{1}{\sqrt{\left( \lambda _{2}-\lambda
_{1}\right) \left( \lambda _{3}-\lambda _{2}\right) }},\quad
c_{0}=\frac{-1}{\sqrt{\left( \lambda _{1}-\lambda
_{3}\right) \left( \lambda _{3}-\lambda _{2}\right) }}.$\
: $a_{0}=\frac{1}{\sqrt{\left( \lambda
_{2}-\lambda _{1}\right) \left( \lambda _{1}-\lambda _{3}\right)
}},\quad b_{0}=\frac{-1}{\sqrt{\left( \lambda _{2}-\lambda
_{1}\right) \left( \lambda _{3}-\lambda _{2}\right) }},\quad
c_{0}=\frac{1}{\sqrt{\left( \lambda _{1}-\lambda _{3}\right)
\left( \lambda _{3}-\lambda _{2}\right) }}.$\
: $a_{0}=\frac{-1}{\sqrt{\left( \lambda
_{2}-\lambda _{1}\right) \left( \lambda _{1}-\lambda _{3}\right)
}},\quad b_{0}=\frac{-1}{\sqrt{\left( \lambda _{2}-\lambda
_{1}\right) \left( \lambda _{3}-\lambda _{2}\right) }},\quad
c_{0}=\frac{-1}{\sqrt{\left( \lambda _{1}-\lambda _{3}\right)
\left( \lambda _{3}-\lambda _{2}\right) }}$.
2\) les coefficients $a_1,b_1,c_1,$ satisfont aux équations $$\begin{aligned}
\left( \lambda _{3}-\allowbreak \lambda _{2}\right)
b_{0}c_{1}+\left( \lambda _{3}
-\lambda _{2}\right) b_{1}c_{0}&=&0,\nonumber\\
\left( \lambda _{1}-\allowbreak \lambda _{3}\right)
a_{0}c_{1}+\left( \lambda _{1}
-\lambda _{3}\right) a_{1}c_{0}&=&0,\nonumber\\
\left( \lambda _{2}-\allowbreak \lambda _{1}\right)
a_{0}b_{1}+\left( \lambda _{2}-\lambda _{1}\right)
a_{1}b_{0}&=&0,\nonumber\end{aligned}$$ dont les solutions sont dant tous les cas : $a_{1}=b_{1}=c_{1}=0$.
3\) les coefficients $a_2,b_2,c_2,$ satisfont aux équations $$\begin{aligned}
&&a_{2}-\lambda _{3}b_{0}c_{2}-\lambda _{3}b_{1}c_{1}-\lambda
_{3}b_{2}c_{0} +\allowbreak \lambda _{2}b_{0}c_{2}+\lambda
_{2}b_{1}c_{1}
+\allowbreak \lambda _{2}b_{2}c_{0}=0,\nonumber\\
&&b_{2}-\lambda _{1}a_{0}c_{2}-\lambda _{1}a_{1}c_{1}-\lambda
_{1}a_{2}c_{0} +\allowbreak \lambda _{3}a_{0}c_{2}+\lambda
_{3}a_{1}c_{1}
+\allowbreak \lambda _{3}a_{2}c_{0}=0,\nonumber\\
&&c_{2}-\lambda _{2}a_{0}b_{2}-\lambda _{2}a_{1}b_{1}-\lambda
_{2}a_{2}b_{0}+\allowbreak \lambda _{1}a_{0}b_{2}+\lambda
_{1}a_{1}b_{1}+\allowbreak \lambda _{1}a_{2}b_{0}=0,\nonumber\end{aligned}$$ dont les solutions qui correspondent aux différents cas sont respectivement :\
: $a_{2}=\frac{\sqrt{\left( \lambda
_{3}-\lambda _{2}\right) }}{\sqrt{\left( \lambda _{1}-\lambda
_{3}\right) }}b_{2}+\frac{\sqrt{\left( \lambda _{3}-\lambda
_{2}\right) }}{\sqrt{\left( \lambda _{2}-\lambda _{1}\right)
}}c_{2}$.\
: $a_{2}=-\frac{\sqrt{\lambda
_{3}-\lambda _{2}}}{\sqrt{\lambda _{1}-\lambda
_{3}}}b_{2}+\frac{\sqrt{\lambda _{3}-\lambda _{2}}}{\sqrt{\lambda
_{2}-\lambda _{1}}}c_{2}$.\
: $a_{2}=\frac{\sqrt{\lambda
_{3}-\lambda _{2}}}{\sqrt{\lambda _{1}-\lambda
_{3}}}b_{2}-\frac{\sqrt{\lambda _{3}-\lambda _{2}}}{\sqrt{\lambda
_{2}-\lambda _{1}}}c_{2}$.\
: $a_{2}=-\frac{\sqrt{\lambda
_{3}-\lambda _{2}}}{\sqrt{\lambda _{1}-\lambda
_{3}}}b_{2}-\frac{\sqrt{\lambda _{3}-\lambda _{2}}}{\sqrt{\lambda
_{2}-\lambda _{1}}}c_{2}$.\
où $b_2$ et $c_2$ sont deux paramètres libres.\
Par conséquent, pour le premier cas on a $$\begin{aligned}
m_{1}&=&\frac{-1}{t\sqrt{\left( \lambda _{2}-\lambda _{1}\right)
\left( \lambda _{1}-\lambda _{3}\right) }}+\left(
\frac{\sqrt{\left( \lambda _{3}-\lambda _{2}\right)
}}{\sqrt{\left( \lambda _{1}-\lambda _{3}\right)
}}b_{2}+\frac{\sqrt{\left( \lambda _{3}-\lambda _{2}\right)
}}{\sqrt{\left( \lambda _{2}-\lambda _{1}\right) }}c_{2}\right)
t+\cdots ,\nonumber\\
m_{2}&=&\frac{1}{t\sqrt{\left( \lambda _{2}-\lambda _{1}\right)
\left( \lambda _{3}-\lambda _{2}\right)
}}+b_{2}t+\cdots ,\nonumber\\
m_{3}&=&\frac{1}{t\sqrt{\left( \lambda _{1}-\lambda _{3}\right)
\left( \lambda _{3}-\lambda _{2}\right) }}+c_{2}t+\cdots.\nonumber\end{aligned}$$ En substituant ces développements dans les intégrales premières $H_1$ et $H_2$, on obtient $$\begin{aligned}
H_{1}&=&2\frac{\sqrt{\lambda _{3}-\lambda _{2}}}{\sqrt{\lambda
_{2}-\lambda _{1}}}\left( \frac{1}{\lambda _{3}-\lambda
_{2}}-\frac{1}{\lambda _{1}-\lambda _{3}}\right)
b_{2}+2\frac{\sqrt{\lambda _{3}-\lambda _{2}}}{\sqrt{\lambda
_{1}-\lambda _{3}}}\left( \frac{1}{\lambda _{3}-\lambda
_{2}}-\allowbreak \frac{1}{\lambda _{2}-\lambda _{1}}\right)
c_{2},\nonumber\\
H_{2}&=&2\frac{\sqrt{\lambda _{3}-\lambda _{2}}}{\sqrt{\lambda
_{2}-\lambda _{1}}}\left( \frac{\lambda _{2}}{\lambda _{3}-\lambda
_{2}}-\frac{\lambda _{1}}{\lambda _{1}-\lambda _{3}}\right) b_{2}
+2\frac{\sqrt{\lambda _{3}-\lambda _{2}}}{\sqrt{\lambda
_{1}-\lambda _{3}}}\left( \frac{\lambda _{3}}{\lambda _{3}-\lambda
_{2}}-\allowbreak \frac{\lambda _{1}}{\lambda _{2}-\lambda
_{1}}\right) c_{2},\nonumber\end{aligned}$$ et on en déduit les relations $$\begin{aligned}
c_{2}&=&\frac{1}{6\sqrt{\left( \lambda _{1}-\lambda _{3}\right)
\left( \lambda _{3}-\lambda _{2}\right) }}\left( \left( \lambda
_{3}-\lambda _{2}\right) \left( \lambda _{1}H_{1}-\allowbreak
H_{2}\right) -\left( \lambda _{1}-\lambda _{3}\right) \left(
\lambda _{2}H_{1}-H_{2}\right) \right),\nonumber\\
b_{2}&=&\frac{1}{6\sqrt{\left( \lambda _{2}-\lambda _{1}\right)
\left( \lambda _{3}-\lambda _{2}\right) }}\left( \left( \lambda
_{2}-\lambda _{1}\right) \left( \lambda _{3}H_{1}-H_{2}\right)
-\allowbreak \left( \lambda _{3}-\lambda _{2}\right) \left(
\lambda _{1}H_{1}-H_{2}\right) \right).\nonumber\end{aligned}$$ On obtient évidemment des expressions similaires pour les autres cas. Il serait intéressant de comparer les solutions obtenues sous forme de séries de Laurent avec les solutions obtenues à l’aide des fonctions elliptiques de Jacobi (15) ainsi qu’avec celles obtenues par la méthode des déformations isospectrales (voir \[10,14\]).
$b)$ : Dans ce cas, on a $$I_{1}=I_{2},\quad l_{1}=l_{2}=0.$$ Il n’est pas difficile de montrer que dans ce cas aussi, l’intégration s’effectue à l’aide de fonctions elliptiques.
$c)$ : Dans ce cas, on a $$I_{1}=I_{2}=2I_3,\quad l_{3}=0.$$ L’étude de ce cas est compliquée. Le système différentiel $\left(13\right),$ s’écrit explicitement sous la forme $$\begin{aligned}
\overset{.}{m}_{1}&=&m_{2}m_{3},\nonumber\\
\overset{.}{m}_{2}&=&-m_{1}m_{3}+2\gamma_{3},\nonumber\\
\overset{.}{m}_{3}&=&-2\gamma_{2},\\
\overset{.}{\gamma}_{1}&=&2m_{3}\gamma _{2}-m_{2}\gamma _{3},\nonumber\\
\overset{.}{\gamma }_{2}&=&m_{1}\gamma
_{3}-2m_{3}\gamma _{1},\nonumber\\
\overset{.}{\gamma }_{3}&=&m_{2}\gamma _{1}-m_{1}\gamma
_{2},\nonumber\end{aligned}$$ où, sans restreindre la généralité, nous avons choisi $l_{2}=0,$ $\mu gl_{1}=1$, $I_{3}=1$ et nous avons utilisé la substitution $t\rightarrow 2t$. Ces équations forment un champ de vecteurs hamiltonien de la forme (9) avec $n=6$, $m=k=2$, $z=t$, $w= \left( m_{1},m_{2},m_{3}, \gamma_{1},
\gamma_{2}, \gamma_{3}\right)^{\intercal}$, $$H=\frac{1}{2}\left( m_{1}^{2}+m_{2}^{2}\right) +m_{3}^{2}+2\gamma
_{1},$$ l’hamiltonien et $$J=\left(\begin{array}{cccccc}
0&-m_{3}&m_{2}&0&-\gamma_{3}&\gamma _{2}\\
m_{3}&0&-m_{1}&\gamma_{3}&0&-\gamma _{1}\\
-m_{2}&m_{1}&0&-\gamma _{2}&\gamma _{1}&0\\
0&-\gamma_{3}&\gamma_{2}&0&0&0\\
\gamma_{3}&0&-\gamma_{1}&0&0&0\\
-\gamma _{2}&\gamma_{1}&0&0&0&0
\end{array}\right)$$ En plus des trois intégrales premières $$\begin{aligned}
H_{1}&\equiv& H,\nonumber\\
H_{2}&=&m_{1}\gamma _{1}+m_{2}\gamma _{2}+m_{3}\gamma _{3},\nonumber\\
H_{3}&=&\gamma _{1}^{2}+\gamma _{2}^{2}+\gamma _{3}^{2},\nonumber\end{aligned}$$ le système ci-dessus admet une quatrième intégrale première quartique obtenue par Kowalewski $$H_{4}=\left( \left( \frac{m_{1}+im_{2}}{2}\right) ^{2}-
\left( \gamma _{1}+i\gamma _{2}\right) \right) \left( \left(
\frac{m_{1}-im_{2}}{2}\right) ^{2}- \left( \gamma _{1}-i\gamma
_{2}\right) \right).$$ Les intégrales premières $H_{1}$ et $H_{4}$ sont en involution, tandis que $H_{1}$ et $H_{3}$ sont triviaux. Soit $$M_c=\bigcap_{k=1}^{4}\left\{ x:H_{k}\left( x\right) =c_{k}\right\}
,$$ la variété affine définie par l’intersection des quatre constantes du mouvenent. Soit $$( m_{1},m_{2},m_{3},\gamma
_{1},\gamma _{2},\gamma _{3}) \longmapsto (
x_{1},x_{2},m_{3},y_{1},y_{2},\gamma _{3}),$$ une transformation biunivoque de $M_c$ où $$2x_{1}=m_{1}+im_{2},\qquad y_{1}=x_{1}^{2}-\left( \gamma _{1}+i\gamma _{2}\right)
,$$ $$2x_{2}=m_{1}-im_{2},\qquad y_{2}=x_{2}^{2}-\left( \gamma _{1}-i\gamma _{2}\right)
.$$ Alors, le quotient $M_c/\sigma $ par l’involution $$\label{eqn:euler}
\sigma :M_c\longrightarrow M_c\text{ }\left(
x_{1},x_{2},m_{3},y_{1},y_{2},\gamma _{3}\right) \longmapsto
\left( x_{1},x_{2},-m_{3},y_{1},y_{2},-\gamma _{3}\right) ,$$ est une surface $S$ (de Kummer) $$\begin{aligned}
&&y_{1}y_{2}=c_{4},\nonumber\\
&&y_{1}R(x_{2})+y_{2}R(x_{1})+R_{1}\left( x_{1},x_{2}\right)
+c_{4}(x_{1}-x_{2})^{2}=0,\nonumber\end{aligned}$$ où $$R(x)=-x^{4}+c_{1}x^{2}-2c_{2}x+1-c_{4},$$ est un polynôme de degré $4$ en $x$ et $$R_{1}(x_{1},x_{2})=-c_{1}x_{1}^{2}x_{2}^{2}-
c_{2}\left( c_{2}-2x_{1}x_{2}\left( x_{1}+x_{2}\right) \right)
+\left( 1-c_{4}\right) \left( c_{1}-\left( x_{1}+x_{2}\right)
^{2}\right),$$ un autre polynôme de degré $2$ en $x_{1},x_{2}.$ Les points de ramification de $M_c$ sur $S$ sont donnés par les points fixes de l’involution $\sigma $ $\left(
17\right) $ et sont en nombre de $8.$ La surface $S$ est un revêtement double du plan $(x_{1},x_{2}),$ ramifié le long de deux courbes elliptiques se coupant exactement aux $8$ points fixes de l’involution $\sigma .$ Ces courbes donnent lieu à l’équation différentielle d’Euler $$\frac{dx_{1}}{\sqrt{R(x_{1})}}\pm
\frac{dx_{2}}{\sqrt{R(x_{2})}}=0,$$ à laquelle se trouvent liés les fameuses variables de Kowalewski $$s_{1}=\frac{R_{1}\left( x_{1},x_{2}\right)
-\sqrt{R(x_{1})}\sqrt{R(x_{2})}}{\left( x_{1}-x_{2}\right)
^{2}}+\frac{c_{1}}{2},$$ $$s_{2}=\frac{R_{1}\left( x_{1},x_{2}\right)
+\sqrt{R(x_{1})}\sqrt{R(x_{2})}}{\left( x_{1}-x_{2}\right)
^{2}}+\frac{c_{1}}{2},$$ et peuvent être vues comme étant des formules d’addition pour la fonction elliptique de Weierstrass (voir \[8\]). En termes des variables $s_{1}$ et $s_{2},$ le système $\left(16\right) $ devient $$\frac{ds_{1}}{\sqrt{P_{5}(s_{1})}}\pm
\frac{ds_{2}}{\sqrt{P_{5}(s_{2})}}=0,$$ $$\frac{s_{1}ds_{1}}{\sqrt{P_{5}(s_{1})}}\pm
\frac{s_{2}ds_{2}}{\sqrt{P_{5}(s_{2})}}=dt,$$ où $P_{5}(s)$est un polynôme de cinquième degré et l’intégration s’effectue au moyen des fonctions hyperelliptiques de genre $2.$
En utilisant le théorème 1, on montre que le système $\left(
13\right)$ dans le cas de Kowalewski, admet deux familles de solutions en séries de Laurent méromorphes $$M=\sum_{k=0}^{\infty }M^{\left( k\right) }t^{k-1},\quad \Gamma
=\sum_{k=0}^{\infty }\Gamma ^{\left( k\right)}t^{k-2},$$ dépendant de cinq paramètres libres tels que : les coefficients $M^{\left( 0\right) }$ et $\Gamma ^{\left( 0\right)
}$ satisfont au système non-linéaire $$\begin{aligned}
&&M^{\left( 0\right) }+\left[ M^{\left( 0\right) },\Lambda
M^{\left( 0\right) }\right] +\left[ \Gamma ^{\left( 0\right)
},L\right] =0,\\
&&2\Gamma ^{\left( 0\right) }+\left[ \Gamma ^{\left( 0\right)
},\Lambda M^{\left( 0\right) }\right] =0,\nonumber\end{aligned}$$ dépendant d’une variable libre $\alpha $ et définissant deux droites. Tandis que $M^{\left( k\right) }$ et $\Gamma
^{\left( k\right) }$ satisfont aux systèmes linéaires $$\left( L-kI\right)\left(\begin{array}{c}
M^{\left( 1\right) }\\
\Gamma ^{\left( 1\right) }
\end{array}\right)=0,$$ $$\left( L-kI\right)\left(\begin{array}{c}
M^{\left( k\right) }\\
\Gamma ^{\left( k\right) }
\end{array}\right)=
\left\{\begin{array}{rl} -\sum_{i=1}^{k-1}\left[ M^{\left(
i\right) },\Lambda M^{\left( k-i\right) }\right] \\
-\sum_{i=1}^{k-1}\left[ \Gamma ^{\left( i\right) },\Lambda
M^{\left( k-i\right) }\right]
\end{array}\right.
,\quad \text{ pour }k\geq 2,$$ où $L$ est la matrice jacobienne de $\left(18\right) .$ Ces systèmes fournissent une variable libre à chacun des niveaux $k=1,2,3$ et $4.$ Explicitement, on a\
$\left( *\right) $ $1^{\grave{e}re}$ famille de solutions en séries de Laurent méromorphes : $$\begin{aligned}
m_{1}\left( t\right) &=&\frac{\alpha }{t}+i\left( \alpha ^{2}-2\right) \beta +\circ\left( t\right) ,\nonumber\\
m_{2}\left( t\right)& =&\frac{i\alpha }{t}-\alpha ^{2}\beta +\circ\left( t\right) ,\nonumber\\
m_{3}\left( t\right) &=&\frac{i}{t}+\alpha \beta +\circ\left( t\right) ,\nonumber\\
\gamma _{1}\left( t\right)& =&\frac{1}{2t^{2}}+\circ\left( t\right) ,\nonumber\\
\gamma _{2}\left( t\right)& =&\frac{i}{2t^{2}}+\circ\left( t\right) ,\nonumber\\
\gamma _{3}\left( t\right) &=&\frac{\beta }{t}+\circ\left(
t\right) .\nonumber\end{aligned}$$ $\left( **\right) $ $2^{\acute{e}me}$ famille de solutions en séries de Laurent méromorphes : $$\begin{aligned}
m_{1}\left( t\right) &=&\frac{\alpha }{t}-i\left( \alpha ^{2}-2\right) \beta +\circ\left( t\right) ,\nonumber\\
m_{2}\left( t\right)& =&-\frac{i\alpha }{t}-\alpha ^{2}\beta +\circ\left( t\right) ,\nonumber\\
m_{3}\left( t\right)& =&-\frac{i}{t}+\alpha \beta +\circ\left( t\right) ,\nonumber\\
\gamma _{1}\left( t\right)& =&\frac{1}{2t^{2}}+\circ\left( t\right) ,\nonumber\\
\gamma _{2}\left( t\right)& =&-\frac{i}{2t^{2}}+\circ\left( t\right) ,\nonumber\\
\gamma _{3}\left( t\right)& =&\frac{\beta }{t}+\circ\left(
t\right) .\nonumber\end{aligned}$$ Les diviseurs des pôles des fonctions $M$ et $\Gamma $ sont deux curbes algébriques $$\label{eqn:euler}
\mathcal{D}_{\varepsilon }:\beta ^{4}\left( \alpha ^{2}-1\right)
^{2}-\left( c_{1}\beta ^{2}-2\varepsilon c_{2}\beta -1\right)
\left( \alpha ^{2}-1\right) +c_{4}=0,\text{ }\varepsilon ^{2}=-1,$$ irréductibles isomorphes et chacune de genre $3.$ Ce sont deux revêtements $$\label{eqn:euler}
\mathcal{D}_{\varepsilon }\longrightarrow \mathcal{D}_{\varepsilon
}^{0},\quad \left( \alpha ,u,\beta \right) \longmapsto \left(
u,\beta \right) ,$$ doubles ramifiés en quatre points de courbes elliptiques : $$\label{eqn:euler}
\mathcal{D}_{\varepsilon }^{0}:u^{2}=\left( c_{1}\beta
^{2}-2\varepsilon c_{2}\beta -1\right) ^{2}-4c_{4}\beta ^{4}.$$ Soit $$\mathcal{L}\left( \mathcal{D}_{\varepsilon
=i}+\mathcal{D}_{\varepsilon =-i}\right)= \left\{ f\text{
m\'{e}romorphe sur }:\left( f\right) +\mathcal{D}_{\varepsilon
=i}+\mathcal{D}_{\varepsilon =-i}\geq 0\right\} ,$$ l’espace vectoriel des fonctions $f$ méromorphes telles que: $\left(
f\right) +\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon
=-i}\geq 0.$ , i.e., l’ensemble des fonctions holomorphes en dehors de $\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon
=-i}$ et ayant au plus des pôles le long de $\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon =-i}$. En utilisant les séries de Laurent obtenues précédemment, on montre que cet espace est engendré par les huit fonctions suivantes: $$\begin{aligned}
&&f_{0}=1,\text{ }f_{1}=m_{1},\text{ }f_{2}=m_{2},\text{
}f_{3}=m_{3},
\text{ }f_{4}=\gamma _{3},\text{ }f_{5}=f_{1}^{2}+f_{2}^{2},\nonumber\\
&&f_{6}=4f_{1}f_{4}-f_{3}f_{5},\text{ }f_{7}=\left( f_{2}\gamma
_{1}-f_{1}\gamma _{2}\right) f_{3}+2f_{4}\gamma _{2}.\text{ }\end{aligned}$$ En outre, l’application $$\left( \mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon =-i}\right) \longrightarrow
\mathbb{CP}^{7},p=\left( \alpha ,u,\beta \right) \longmapsto
\underset{t\rightarrow 0}{\lim }t\left(
1,f_{1}(p),...,f_{7}(p)\right)=$$$$\left( 0,\alpha,\pm i,\pm
i,\beta,\pm i\alpha\beta,\varepsilon(\alpha^2-1)\beta^2,\pm i (\mp
c_2+c_1\beta-2(\alpha^2-1)\beta^3)\right) ,$$ plonge $\left( \mathcal{D}_{\varepsilon
=i}+\mathcal{D}_{\varepsilon =-i}\right) $ dans $\mathbb{CP}^{7}$de telle façon que $\mathcal{D}_{\varepsilon =i}$ intersecte $\mathcal{D}_{\varepsilon =-i}$ transversalement en quatre points à l’infini $\left( \alpha =\pm 1,\text{ }u=\pm \beta
^{2}\sqrt{c_{1}^{2}-4c_{4}},\text{ }\beta =\infty \right)$ et que le genre géométrique de $\left( \mathcal{D}_{\varepsilon
=i}+\mathcal{D}_{\varepsilon =-i}\right) $ est $9.$ Les orbites du champ de vecteurs en question passant à travers $\left(
\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon =-i}\right)$ forment une surface lisse $\Sigma $ tout le long de $\left(\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon
=-i}\right) $ tel que : $\Sigma \backslash
\left(\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon
=-i}\right) \subset M_{c}.$ La variété $\widetilde{M_{c}}=M_{c}\cup \Sigma ,$ est lisse, compacte et connexe. En outre, le champ de vecteur est régulier le long du diviseur $\left( \mathcal{D}_{\varepsilon
=i}+\mathcal{D}_{\varepsilon =-i}\right) ,$ transversal en tout point $\beta \neq 0$ $\left( \text{resp. }\beta \neq \infty
\right) $ et doublement tangent en $\beta =0$ $\left( \text{resp.
}\beta =\infty \right)$. Les champs de vecteurs engendrés par $H_1$ et $H_4$ se prolongent de façon holomorphe et demeurent indépendants sur la variété $\widetilde{M_{c}}.$ La variété $\widetilde{M_{c}}$ est une surface abélienne sur laquelle le flot hamiltonien $\left(16\right)$ se linéarise. L’involution $\sigma $ $\left(17\right) $ sur la variété affine $M_{c}$ se transforme en une involution $$\sigma _{\varepsilon }:\mathcal{D}_{\varepsilon }\longrightarrow
\mathcal{D}_{\varepsilon },\text{ }\left( \alpha ,u,\beta \right)
\longmapsto \left( -\alpha ,u,\beta \right),$$ sur la surface de Riemann $\mathcal{D}_{\varepsilon }$ $\left(19\right) $ et admet huit points fixes donnés par les points de branchements de $\mathcal{D}_{\varepsilon }$ sur la courbe elliptique $\mathcal{D}_{\varepsilon }^{0}$ $\left(21\right) .$ Donc l’involution en question possède seize points fixes au total, confirmant ainsi le nombre de points fixes qu’une involution $z\longmapsto -z$ sur une variété abélienne doit en avoir. Il existe sur la surface abélienne $\widetilde{M_{c}}$ deux différentielles holomorphes $dt_{1}$ et $dt_{2}$ telles que : $$\begin{aligned}
dt_{1\mid _{\mathcal{D}_{\varepsilon }}}&=&\omega
_{1}=\frac{k_{1}\left( \alpha ^{2}-1\right) \beta ^{2}d\beta
}{\alpha u},\nonumber\\
dt_{2\mid _{\mathcal{D}_{\varepsilon }}}&=&\omega
_{2}=\frac{k_{2}d\beta }{\alpha u},\nonumber\end{aligned}$$ où $k_{1},k_{2}\in \mathbb{C}$ et $\omega _{1},\omega _{2}$ sont des différentielles holomorphes sur $\mathcal{D}_{\varepsilon }.$ En outre, l’espace des différentielles holomorphes sur le diviseur $\left(
\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon =-i}\right)
$ est $$\left\{ f_{1}^{\left( 0\right) }\omega _{2}\text{ }f_{2}^{\left( 0\right) }
\omega _{2},...,f_{7}^{\left( 0\right) }\omega _{2}\right\} \oplus \left\{ \omega _{1},\omega _{2}\right\}
,$$ où $f_{1}^{\left( 0\right) },$ $f_{2}^{\left( 0\right)
},...$ $f_{7}^{\left( 0\right) }$ sont les premiers coefficients des fonctions $f_{1},f_{2},...,f_{7}$ $\in \mathcal{L}\left(
\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon =-i}\right)
\left( 66\right) $ et le plongement de $\left(
\mathcal{D}_{\varepsilon =i}+\mathcal{D}_{\varepsilon =-i}\right)
$ dans $\mathbb{CP}^{7}$ est à deux différentielles holomorphes près le plongement canonique $$p=\left( \alpha ,u,\beta \right) \in \left( \mathcal{D}_{\varepsilon =i}
+\mathcal{D}_{\varepsilon =-i}\right) \longmapsto \left\{ \omega
_{2},f_{1}^{\left( 0\right) } \omega _{2},f_{2}^{\left( 0\right) }
\omega _{2},...,f_{7}^{\left( 0\right) }\omega _{2}\right\} \in
\mathbb{CP}^{7}.$$ La surface abélienne $\widetilde{M_{c}}$est caractérisée comme étant la duale de variété Prym $\left( \mathcal{D}_{\varepsilon
}/\mathcal{D}_{\varepsilon }^{0}\right) $ du revêtement double $\left(20\right)$. Les solutions sous forme de séries de Laurent (théorème 1), jouent un rôle crucial dans la preuve de ces résultats.
[99]{} ADLER M. AND VAN MOERBEKE P.: The complex geometry of the Kowalewski-Painlevé analysis. *Invent. Math.* **97** (1989) 3-51. ARNOLD V.I.: Ordinary differential equations. Springer-Textbook, 3nd edition, 1992. CARTAN H.: Théorie élémentaire des fonctions analytiques d’une ou plusieurs variables complexes. Hermann, 1961. FRANCOISE J.P.: Integrability of quasi-homogeneous vector fields (*preprint*). GRIFFITHS P.A. AND HARRIS, J.: Principles of algebraic geometry. Wiley-Interscience, New York, 1978. HAINE L.: Geodesic flow on $SO(4)$ and Abelian surfaces. *Math. Ann.* **263** (1983) 435-472. HILLE E.: Ordinary differential equations in the complex domain. Wiley-Interscience, New-York, 1976. LESFARI A.: Abelian surfaces and Kowalewski’s top. *Ann. Scient. École Norm. Sup.*, Paris, sér. 4 **21** (1988) 193-223. LESFARI A.: Eléments d’analyse. Sochepress-Université, Casablanca, 1991. LESFARI A.: Completely integrable systems : Jacobi’s heritage. *J. Geom. Phys* **31** (1999) 265-286. LESFARI A.: Le théorème d’Arnold-Liouville et ses conséquences. *Elem. Math.* (Issue 1) **58** (2003) 6-20. LESFARI A.: Analyse des singularités de quelques systèmes intégrables. *C. R. Acad. Sci.* Paris, Ser. **I 341** (2005) 85-88. LESFARI A.: Abelian varieties, surfaces of general type and integrable systems. *Beiträge Algebra Geom.*, Vol.48, 1 (2007) 95-114. LESFARI A.: Integrable systems and complex geometry. *arXiv : 0706.1579 \[math. CV\]* (2007) 1-45. LESFARI A.: Fonctions et Intégrales elliptiques. *arXiv : 0707.1137 \[math. DS\]* (2007) 1-46. PAINLEVE P.: Oeuvres: tomes 1,2,3. Edition du C.N.R.S. 1975.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a two-dimensional (2D) version of Thouless pumping which can be realized by using ultracold atoms in optical lattices. To be specific, we consider a 2D square lattice tight-binding model with obliquely introduced superlattice. We show that quantized particle transport occurs in this system, and that the transport is expressed as a solution of a Diophantine equation. This topological nature can be understood easily by mapping the Hamiltonian onto a 3D cubic lattice model with a homogeneous magnetic field. On the other hand, in the weak potential limit, we uncover that the pumping direction is restricted exactly to an $x$-axis or a $y$-axis direction. This difference in two limits causes a topological transition. Furthermore, we demonstrate by numerical simulation that the above phenomena are measurable in cold atom systems.'
author:
- 'Fuyuki Matsuda, Masaki Tezuka, and Norio Kawakami'
bibliography:
- 'ObliqueSuperlattice.bib'
title: 'Two-Dimensional Thouless Pumping of Ultracold Fermions in Obliquely Introduced Optical Superlattice'
---
Introduction
============
The studies of quantum Hall effect opened up a wide research area which is known as topological phases [@RevModPhys.82.3045; @RevModPhys.83.1057]. Topological phases are characterized by the associated invariants, which are prohibited to vary unless the energy gap closes. Moreover, topological phases have a characteristic property called bulk-edge correspondence, which guarantees the existence of the edge states, meaning that electrons can only move on the edge of a given material, as far as the bulk topological invariant has a non-trivial value. In quantum Hall systems, the topological invariant is defined by the Chern number, and it appears as the quantized Hall conductance. Thouless [*et al.*]{} showed that this Chern number satisfies a certain Diophantine equation[@PhysRevLett.49.405]. Similar relationships also appear in 3D quantum Hall systems[@PhysRevB.45.13488; @Koshino2001; @Koshino2002].
Topological phases can be realized in various experimental platforms, not only in materials, but also in optical waveguides[@Kraus2012], photonic systems[@PhysRevA.84.043804; @NaturePhys.7.907; @NatureMat.12.233], etc. In particular, great efforts have been made to realize topological insulators in cold atom systems by using synthetic gauge fields, which are induced by an optical lattice[@Dalibard2011]. Various topological systems have been realized in cold atom systems, for instance, Su-Schriefer-Heeger model [@Nature.9.795], Hofstadter model [@PhysRevLett.111.185301; @PhysRevLett.111.185302], Synthetic dimensions [@Mancini1510; @Stuhl1514], and Haldane model [@Nature.515.237]. Recently, topological charge pumping which is called Thouless pumping[@Review1983; @PhysRevB.34.5093] has been realized in cold atom systems [@NaturePhys.12.296; @NaturePhys.12.350], and also many theoretical proposals have been made in numerous studies[@PhysRevA.57.R2278; @PhysRevA.76.052304; @PhysRevA.84.013608; @PhysRevLett.111.026802; @PhysRevA.90.063638; @PhysRevA.92.013612; @PhysRevA.92.013609; @2018arXiv180108255Y; @JPSJ.83.083707].
Mapping Hamiltonian between different dimensions plays an important role in understanding physical properties. One of the significant examples of mapping can be seen in the problem of Hofstadter butterfly[@PhysRevB.14.2239]. Hofstadter mapped a two-dimensional (2D) system of square lattice under a uniform magnetic field into a 1D equation, which is called the Harper equation. Interestingly, a recent research of adiabatic pumping used this mapping backwards, and showed the topological equivalence over a certain group of Hamiltonians. This kind of mapping into higher dimensions can be applied not only to 1D systems, but also to two and higher dimensional systems[@Kraus2013].
In this paper, we consider a 2D tight-binding model of particles that hop on a square lattice in the presence of modulation in the on-site potential and the hopping amplitude, where the modulation amplitude depends only on the distance from an inclined line. We find that this model has topological nature. First, we show that edge-localized states appear when a strong binding potential which constrains particle around an inclined line exists. Also, if we pay our attention to the energy spectrum, these edge states cross the bulk energy band gap. These topological insulator-like behaviors can be understood clearly when we apply a mapping from higher dimensions to this 2D system. In a 3D mapped Hamiltonian, a set of weak Chern numbers of this system becomes non-zero when the chemical potential is controlled adequately, which implies that this system is a weak topological insulator. These weak Chern numbers determine the adiabatic pumping in the original 2D system.
This paper is organized as follows. In section II, we introduce a 2D square lattice tight-binding model with an obliquely introduced potential in order to investigate the effect of the superimposed potential in the context of topological phases. The existence of edge-localized states and band-crossing states, which are typical phenomena in topological phases, are shown. Also, we show that the Hofstadter butterfly structure appears in our model. In section III, we consider a continuum model in order to consider the above tight-binding model and treat it more precisely. We focus on two limits: weak superlattice and strong superlattice limit, which we call the Hofstadter regime and the sliding regime. In the Hofstadter regime, the amount of the charge pumping corresponds to the solution of the Diophantine equation. In the sliding limit, the pumping behavior completely changes to “rectified pumping”, in which a pumping direction is restricted to exactly $x$-direction or exactly $y$-direction. In section IV, we conclude our paper.
Topological aspects in 2D optical lattice with obliquely introduced potential
=============================================================================
Model
-----
Recently, the development of experimental techniques regarding cold atoms in optical lattice is remarkable, and the degree of freedom in designing the quantum simulation is high. Triangle lattice, square lattice, and hexagonal lattice have all been realized in cold atom systems. Moreover, it is possible to realize non-standard lattices in cold atom systems. One example of such lattices is the lattice made by superimposing a usual lattice potential and a superlattice potential. For example, in Ref., Taie [*et al.*]{} realized an optical Lieb lattice by adding a diagonal lattice to a usual square optical lattice.
Also, this paper is partly motivated by recent remarkable progress in solid state, i.e. twisted bilayer graphene. Twisted bilayer graphene, which is receiving a lot of attention lately because of the realization of unconventional superconductivity[@Nature26160]. Twisted bilayer graphene has been investigated in many contexts, such as its electronic structures[@PhysRevLett.99.256802; @ROZHKOV20161], flat bands[@PhysRevB.82.121407], topological band structures[@PhysRevB.84.045436], and Moiré butterflies[@PhysRevB.84.035440]. It provides also a prototypical example of superimposed lattice potential.
Stimulated by the above mentioned previous studies, it is natural to ask what happens if the additional diagonal lattice in the optical Lieb lattice is twisted. Now, we consider a fermion gas loaded in the 2D optical square lattice with a superlattice structure which is introduced obliquely. We study the ground-state properties of the system which is described by the following hamiltonian:
$$\begin{aligned}
\hat{\cal{H}}=&&\sum_{m,n} \left[ -t \left( \hat{c}_{m+1,n}^{\dag} \hat{c}_{m,n} + \hat{c}_{m,n+1}^{\dag} \hat{c}_{m,n}\right) + \mathrm{H.c.} \right] \nonumber \\
&&+ \sum_{m,n} \left[ V_{m,n} \hat{c}_{m,n}^{\dag} \hat{c}_{m,n} \right] \\
V_{m,n} &&= V(m+\alpha n -\delta )\end{aligned}$$
where $\hat{c}_{m,n}^{\dag} (\hat{c}_{m,n})$ is the fermion creation (annihilation) operator, $t$ is the hopping strength, $\delta$ is the parameter of the real space potential shift, and $\alpha$ is the tangent of the superlattice. The boundary conditions are taken both open and periodic.
Obliquely introduced single well
--------------------------------
![Plot of on-site potential in the case of $\alpha = (\sqrt{5}+1)/2$, $V(x)=-V e^{-x^2/w^2}$, $V=10$, $w=1$. []{data-label="onsite"}](figure1.pdf){width="8cm"}
![Energy spectrum plotted against the potential real-space shift $\delta$. The on-site potential is given in Fig. \[onsite\], with open boundary conditions. The bulk of the spectrum remains unchanged, with few of the states crossing across the band gap. []{data-label="spectrum"}](band50_50_d1.pdf){width="8cm"}
First, we consider the case of $\alpha = (\sqrt{5}+1)/2\simeq 1.618$ with open boundary conditions. We take $V(x)=-V e^{-x^2/w^2}$ as an oblique linear-shaped potential, which restricts particles to move around the potential in the low-energy region. On-site potential in this system is plotted in Fig. \[onsite\]. As a consequence of the overlap of the square lattice and obliquely introduced potential, this model can be effectively considered as a 1D tight binding model with a superlattice. This effective superlattice opens the energy gap. By the analogy to the previous study of topological aspects of 1D superlattice, the edge states seem to appear when we drive the phase of the superlattice, which corresponds to the parameter $\delta$ in this model. We can see the existence of band-crossing states in the energy spectrum as a function of $\delta$ in Fig. \[spectrum\]. Note that the bulk band is fixed because we choose an irrational number for tangent $\alpha$. Fig. \[edgelocalized\] (a) and (b) show the wave function of an in-band state and a band-crossing state. We confirm that the wave function of the band-crossing state is localized in the edge.
By changing $\alpha$, the distribution of energy band will also changes. Fig. \[butterfly\] shows the energy spectrum plotted against $\theta$, which is related to $\alpha$ by the formula $\alpha = \tan \theta$. As shown in the figure, a butterfly-like structure appears in the energy spectrum plot. This structure is similar to the “3D butterfly” which is introduced in Refs. . We will show the reason for this similarity later. The color in Fig. \[butterfly\] represents the inverse participation ratio (IPR) in each eigenfunction. Small IPR denotes that the eigenfunction is well localized. We can see that eigenstates in the band gap is well localized. These states correspond to edge states in topological systems.
[cc]{}
\(a) ![Wave functions of the energy eigenstate in xy plane: (a) 29th and (b) 30th excited state in real space with $\delta = 0.75$. While the wave functions are spread over the oblique potential in (a), they are localized at the edge in (b).[]{data-label="edgelocalized"}](50_50_vOFFhOFF_del0750000_29.pdf "fig:"){width="6.5cm"}
\(b) ![Wave functions of the energy eigenstate in xy plane: (a) 29th and (b) 30th excited state in real space with $\delta = 0.75$. While the wave functions are spread over the oblique potential in (a), they are localized at the edge in (b).[]{data-label="edgelocalized"}](50_50_vOFFhOFF_del0750000_30.pdf "fig:"){width="6.5cm"}
![Butterfly structure in plots of the energy spectrum against the tilting angle $\theta$. Color shows the inverse participation ratio (IPR). The IPR measures the spread of the eigenfunction.[]{data-label="butterfly"}](100butterfly-light.pdf){width="9cm"}
Obliquely introduced superlattice
---------------------------------
Now, we consider the case when $V(x)$ is periodic, in particular, we take $V_{m,n} = V \cos ( \phi_x m + \phi_y n - \delta)$. In this case, when we vary $\delta$ slowly, the hamiltonian will be periodic in time, and this setup corresponds to the adiabatic transport, which is first proposed by D. J. Thouless[@Review1983].
It is known that the mapping between a 1D superlattice system and a 2D Hofstadter model exists. Analogously, we can clearly understand the topological aspects of our model by mapping it onto a 3D hamiltonian. For simplicity, we start our discussion in the case of cosine type potential such as $V(x)=\cos (2\pi x/L_x)$, $\alpha = L_x/L_y$, where $L_x$ and $L_y$ determine the size of the system, and we use the periodic boundary condition. Also, we use $\phi \equiv \delta/L_x$ for the parameter of the potential shift. The Hamiltonian is written as follows: $$\begin{aligned}
{\cal H}(\phi)&&=t\sum_{m,n} \left[ \hat{c}^{\dag}_{m+1,n}(\phi)\hat{c}_{m,n}(\phi)+\hat{c}^{\dag}_{m,n+1} (\phi )\hat{c}_{m,n}( \phi) +\mathrm{H.c.} \right] \nonumber \\
&&+ V \sum_{m,n} \cos \left( 2\pi \left( \frac{m}{L_x}+\frac{n}{L_y} \right)+\phi \right)\hat{c}^{\dag}_{m,n}(\phi)\hat{c}_{m,n}( \phi ).
\label{hamiltonian2D}\end{aligned}$$ We can map this Hamiltonian to a 3D Hamiltonian by using Fourier transformation, $$\begin{aligned}
\hat{c}_{m,n}(\phi)=\cfrac{1}{\sqrt{2\pi L_z}}\sum_l e^{-il\phi} \hat{c}_{m,n,l} \\
\hat{c}_{m,n,l}=\cfrac{1}{\sqrt{2\pi L_z}}\sum_l e^{il\phi} \hat{c}_{m,n}(\phi).\end{aligned}$$ Hamiltonian (\[hamiltonian2D\]) will be $$\begin{aligned}
{\cal H}=t\sum_{m,n,l} \left[ \hat{c}^{\dag}_{m+1,n,l}\hat{c}_{m,n,l}+\hat{c}^{\dag}_{m,n+1,l}\hat{c}_{m,n,l} +\mathrm{H.c.} \right] \nonumber \\
+ \cfrac{V}{2} \sum_{m,n,l} \left[ e^{i\left( 2\pi \left( \frac{m}{L_x}+\frac{n}{L_y} \right)\right)}\hat{c}^{\dag}_{m,n,l+1}\hat{c}_{m,n,l}+\mathrm{H.c.} \right] .
\label{hamiltonian3D}\end{aligned}$$ This Hamiltonian implies the 3D tight-binding model on cubic lattice with a homogeneous magnetic field, whose direction is perpendicular to the $z$-axis but oblique in $xy$-plane. There are previous studies for this model, and it is known that the Hofstadter butterfly and the integer quantum Hall effect (IQHE) also appear in the 3D lattice. According to Refs. , each gap in the 3D Hofstadter butterfly is characterized by two Chern indices. In the 3D model, these indices are proportional to Hall conductivity. When we go back to our 2D model, these indices determine the quantum pump just like Thouless pump, and the bulk-edge correspondence shows that the quantum pump and the number of edge states match each other. In this model, quantum pump in $x$-direction and $y$-direction corresponds to the number of edge-localized states in edges of $x$-direction and $y$-direction, respectively. Also, by comparing (\[hamiltonian2D\]) and (\[hamiltonian3D\]), we can see that potential strength $V$ corresponds to the hopping strength in the $z$-direction. These correspondences explain why 3D butterfly appears in Fig. \[butterfly\].
Here, we have only considered a cosine-type potential, but a similar mapping is also possible for other types of potentials for which there are higher frequency components, and these components will be mapped into longer distance hopping in the $z$-direction.
Continuum model of 2D Thouless pumping
======================================
![Time evolution of the average of center of mass for the 10 lowest energy states in the case of (a) $p_x/q_x = 2/3, p_y/q_y = 2/5$ and (b) $p_x/q_x = 1/3, p_y/q_y = 1/4$. []{data-label="235_268"}](2_3_5.pdf "fig:"){width="6.5cm"} ![Time evolution of the average of center of mass for the 10 lowest energy states in the case of (a) $p_x/q_x = 2/3, p_y/q_y = 2/5$ and (b) $p_x/q_x = 1/3, p_y/q_y = 1/4$. []{data-label="235_268"}](2_6_8.pdf "fig:"){width="6cm"}
Model
-----
By considering the analogy of Thouless pumping and IQHE, the results in the previous section suggest that the amount of transport, as the Hamiltonian parameter $\phi$ changes its value from 0 to $2\pi$, is quantized and can be expressed by the Chern number. Since the discussion in the previous section is based on the tight-binding model, it is desirable to extend our discussion to continuum systems so as to apply the results to cold atom systems. To treat the system more precisely and to see the feasibility of the previous model in the experimental setups, let us consider next a 2D continuum model.
To treat the continuum model, we consider the following Hamlitonian, $$\begin{aligned}
H &&= \int dx dy \psi^\dagger (x,y) \biggl[ -\cfrac{\hbar^2}{2m} \cfrac{\partial^2}{\partial x^2} -\cfrac{\hbar^2}{2m} \cfrac{\partial^2}{\partial y^2} \\
&&+ V_x(x,y) + V_y(x,y) + V_{\mathrm{SL}}(x,y) \biggl] \psi(x,y), \\
&&V_j(x,y) = 2v_j \cos(q_j j) \ \ \ \ (j = x,y), \\
&&V_{\mathrm{SL}}(x,y) = 2v_{\mathrm{SL}} \cos(p_x x + p_y y - \varphi), \end{aligned}$$ where $\psi(x,y)$ is the field operator of a particle, $\hbar$ is Planck’s constant, and $m$ is the mass of a particle. There are three types of potential in this system: $V_x$, $V_y$, and $V_{\mathrm{SL}}$. $V_x$ and $V_y$ form the 2D square lattice, and $V_{\mathrm{SL}}$ form the superlattice which is introduced obliquely to the square lattice. Now we consider two limits: one is in the limit of weak superlattice (which is called the Hofstadter regime), and the other is in the limit of strong superlattice (which is called the sliding regime).
![Time evolution of the average of center of mass for the 10 lowest energy states in the case of (a)$v_{SL}=0.5$, (b)$v_{SL}=5.0$.[]{data-label="05to5"}](SLD05.pdf "fig:"){width="6cm"} ![Time evolution of the average of center of mass for the 10 lowest energy states in the case of (a)$v_{SL}=0.5$, (b)$v_{SL}=5.0$.[]{data-label="05to5"}](SLD5.pdf "fig:"){width="6cm"}
{width="13cm"} {width="13cm"}
Hofstadter Regime
-----------------
First, we consider the regime which is near the tight-binding limit, i.e. $V_x,V_y \ll V_{SL}^2/4E_L$. As we have mentioned above, in this limit, we can map our model onto a 3D cubic lattice model with an obliquely introduced homogeneous magnetic field in Eq. \[hamiltonian3D\], and the amount of adiabatic pumping in our model corresponds to the quantized Hall conductivity in the cubic lattice model. In Eq. \[hamiltonian3D\], $y$ and $z$ are cyclic coordinates, so that the wave function can be written as $\Psi_{lmn}=e^{ik_xm+ik_yn}F_l$, where $k_x$ and $k_y$ are the Bloch wave numbers along the $x$ and $y$ directions. Then the Schrödinger equation will be $$\begin{gathered}
t(F_{l-1}+F_{l+1}) \\
-V\{ \cos (2\pi l/L_x+k_x)+\cos (2\pi l/L_y+k_y) \} F_l = EF_l.\end{gathered}$$ This equation is equivalent to a 3D version of the Harper equation. By analyzing this equation, the following Diophantine equation can be shown[@PhysRevB.45.13488], $$\frac{r}{Q} = s + t_x\frac{p_x}{q_x} + t_y\frac{p_y}{q_y} .
\label{3Ddiophantos}$$ Here, $r$ is the number of filled bands and $Q$ is the least common multiple of $q_x$ and $q_y$. $t_x$ and $t_y$ correspond to the amount of the charge transport while $\varphi$ changes from $0$ to $\pi/Q$. The details of the proof of this equation is noted in Appendix.
Figure \[235\_268\] shows concrete examples of such quantized pumping. Figures \[235\_268\] (a) and (b) are the time evolution of the center of mass for the 10 lowest energy states in the case of (a) $p_x/q_x = 2/3, p_y/q_y = 2/5$ and (b) $p_x/q_x = 1/3, p_y/q_y = 1/4$, respectively. According to Fig. \[235\_268\], particles are pumped in the $y=x$ direction for (a), and the $y=-x$ direction for (b). Surprisingly, while the direction of the superlattice is changed only slightly, the pumping direction is completely changed. This pumping direction can be predicted by solving Eq. (\[3Ddiophantos\]) with the condition $r=1$ since the lowest band contribution is dominant. For example, in the case (a), the solution of Eq. (\[3Ddiophantos\]) is $(t_x, t_y) = (1+3m, 1+5n)$. $|t_x|$ and $|t_y|$ correspond to the order of perturbation. Therefore, although there are an infinite number of solutions, only the solution where $|t_x|$ and $|t_y|$ have the smallest values is dominant. In the case (a), that is $(t_x, t_y) = (1, 1)$. Similarly, in the case (b), the solution of the Diophantine equation is $(t_x, t_y) = (1+3m, -1+4n)$, however, the dominant solution is $(t_x, t_y) = (1, -1)$. These results are consistent with Fig. \[235\_268\].
Observing Fig. \[235\_268\] carefully, we find a small oscillations after the particle was pumped to the next lattice. This oscillation occurs at the moment when the ground state energies in two sites become close. At that moment, the potential shape can be approximated by a double well, and the particles oscillate between the double well. Since the angular frequency of this oscillation will be the same as the hopping amplitude $J$, the period will be $T = 2\pi / J$. To estimate $J$ in the optical lattice, we can use $J = \frac{4E_R}{\sqrt{\pi}} s^{3/4} \exp (-2\sqrt{s})$ where $s = V_L/E_R$ [@RevModPhys.80.885]. We have checked that this estimate is consistent with the numerical result in Fig. \[235\_268\]. We have also verified that the frequency of this small oscillation does not change while we change the time period $T$. This result implies that the small oscillation is caused by the pumping which is performed in finite time. Therefore, we expect that this small oscillation disappears in the adiabatic limit.
Sliding Regime
--------------
In the opposite limit, where the $V_\mathrm{SL}$ term is dominant, the pumping behavior changes completely. From our numerical simulation, we find that the pumping direction is restricted to the $x$-axis or the $y$-axis exactly in this limit. More precisely, the pumping direction will be whichever is closer to the $x$-axis or the $y$-axis.
When $V_\mathrm{SL}$ varies across the Hofstadter regime to the sliding regime, the topological transition occurs. Figures \[05to5\] and \[05to5density\] provide concrete examples of such topological transition. Figures \[05to5\] (a) and (b) show the time evolution of the center of mass for the 10 lowest energy states in the case of $p_x/q_x = 2/3, p_y/q_y = 1$, with (a) $v_\mathrm{SL} = 0.5$ and (b) $v_\mathrm{SL} = 5.0$, respectively. We note that only the difference of the superlattice potential depth causes the change of pumping direction. Figures \[05to5density\] (a) and (b) show particle density profiles computed in the conditions corresponding to Fig. \[05to5\] (a) and (b). We can clearly see that a particle moves in a different direction when $v_\mathrm{SL}$ is small and $v_\mathrm{SL}$ is large.
This kind of topological transition in 1D systems is mentioned in Ref. . A similar discussion is also possible in our 2D model. If the 2D square lattice is deep enough, eigenstates localized in each lattice can be approximated by eigenstates in the harmonic potential, and the eigenenergies are expressed as $E = \hbar \omega (n+1/2)$. The transition occurs when the energy difference between the ground state and the first excited state is close to $2v_\mathrm{SL}$. Since the transition point is $v_\mathrm{SL} \sim \sqrt{\frac{\pi^2 v_L}{md^2}}$, in Fig. \[05to5\] and Fig.\[05to5density\], we have chosen $v_L = 1.0, \frac{\pi^2}{md^2} = \frac{\pi^2}{4}$; therefore, the transition point is estimated to be $v_\mathrm{SL} = \pi/2 = 1.57\dots$ This is consistent with Fig. \[05to5\] and Fig. \[05to5density\].
Conclusion\[conclusion\]
========================
We have proposed a 2D version of Thouless pumping which can be realized by using ultracold atoms in an optical lattice. We considered a 2D square lattice tight-binding model with an obliquely introduced superlattice. We showed that quantized particle transport occurs in this system, and the transport is expressed as a solution of the Diophantine equation. This topological nature could be understood easily by mapping the hamiltonian onto a 3D cubic lattice model with a homogeneous magnetic field. On the other hand, in the weak potential limit, we have uncovered that the amount of the quantized transport is restricted to exactly $x$-axis or $y$-axis direction. This difference in two limits causes a topological transition. Furthermore, we have demonstrated by numerical simulations that the above phenomena are measurable in a cold atom system.
This work was supported by JSPS KAKENHI (Grant No. JP18H01140 and JP19H01838 (N. K.) and JP17K17822 (M. T.)) and a Grand-in-Aid for Scientic Research on Innovative Areas Topological Materials Science" (Grant No. JP15H05855).
Derivation of Diophantine equation in tight-binding limit
=========================================================
In this Appendix, we derive Eq. \[3Ddiophantos\] from the Hamiltonian, $$\begin{aligned}
{\cal H}(\phi)&=t_x\sum_{m,n} \left[ \hat{c}^{\dag}_{m+1,n}(\phi)\hat{c}_{m,n}(\phi)+\mathrm{h.c.} \right] \\ &+t_y\sum_{m,n} \left[ \hat{c}^{\dag}_{m,n+1} (\phi )\hat{c}_{m,n}( \phi) +\mathrm{h.c.} \right] \\
&+ V \sum_{m,n} \cos \left( 2\pi \left( \Phi_x m+\Phi_y n \right) +\phi \right)\hat{c}^{\dag}_{m,n}(\phi)\hat{c}_{m,n}.( \phi )
\end{aligned}$$ This Hamiltonian is periodic in $\phi$ with period $2\pi$. Here, we suppose that $\Phi_x, \Phi_y$ are rational numbers, such as $\Phi_x = p_x/q_x, \Phi_y = p_y/q_y$.
The Fourier-transformed Hamiltonian is $$\begin{aligned}
H &= - \int _ { - \pi } ^ { \pi } \frac { d k _ { x } } { 2 \pi } \int _ { - \pi } ^ { \pi } \frac { d k _ { y } } { 2 \pi } \\ & \biggl[ t \left( \cos \left( k _ { x } \right) + \cos \left( k _ { y } \right) \right) \hat{c} ^ { \dagger } ( k _ { x } , k _ { y } , k _ { z } ) \hat{c} ( k _ { x } , k _ { y }, k _ { z } ) \\ & + \frac{V}{2} \Bigl( e ^ { - i k _ { z } } \hat{c} ^ { \dagger } ( k _ { x } + 2 \pi \Phi_x , k _ { y } + 2 \pi \Phi_y, k_z ) \hat{c} ( k _ { x } , k _ { y }, k_z ) \\ &+ e ^ { i k _ { z } } \hat{c} ^ { \dagger } ( k _ { x } - 2 \pi \Phi_x , k _ { y } - 2 \pi \Phi_y, k_z ) \hat{c} ( k _ { x } , k _ { y }, k_z ) \Bigl) \biggr] .
\end{aligned}$$ Here, we define $k_z = \phi$ for convenience. However, there is a mixing between $(k_x, k_y, k_z) \to (k_x\pm 2 \pi \Phi_x, k_y \pm 2 \pi \Phi_y, k_z )$ and thus the Hamiltonian is not diagonal in $\bf{k}$. To diagonalize this, we need to separate $k_x, k_y$ into $q_x \times q_y$ regions as follows, $$\begin{aligned}
k_x & = k_x^0 + 2 \pi \Phi_x m \\
k_y & = k_y^0 + 2 \pi \Phi_y n .\end{aligned}$$ Then, the Hamiltonian can be written as: $$H = \frac { 1 } { ( 2 \pi ) ^ { 3 } } \int _ { - \pi / q_x } ^ { \pi / q_x } d k _ { x } ^ { 0 } \int _ { - \pi / q_y } ^ { \pi / q_y } d k _ { y } ^ { 0 } \int _ { - \pi } ^ { \pi } d k _ { z } \hat { H } ( k _ { x } ^ { 0 } , k _ { y } ^ { 0 } , k _ { z } )$$
$$\begin{aligned} \hat { H } ( k _ { x } ^ { 0 } , k _ { y } ) = \sum _ { m = 0 } ^ { q_x - 1 } \sum _ { n = 0 } ^ { q_x - 1 } & \left\{ - 2 t \left( \cos \left( k _ { x } ^ { 0 } + 2 \pi \Phi_x m \right) + \cos \left( k _ { y } ^ { 0 } + 2 \pi \Phi_y n \right) \right) \right. \\ & \left. \times \hat{c}^ { \dagger } \left( k _ { x } ^ { 0 } + 2 \pi \Phi_x m , k _ { y } ^ { 0 } + 2 \pi \Phi_y n , k _ { z } \right) \hat{c} \left( k _ { x } ^ { 0 } + 2 \pi \Phi_x m , k _ { y } ^ { 0 } + 2 \pi \Phi_y n , k _ { z } \right) \right. \\ & \left. - \frac{V}{2} \left( e ^ { - i k _ { z } } \hat{c} ^ { \dagger } \left( k _ { x } ^ { 0 } + 2 \pi \Phi_x ( m + 1 ) , k _ { y } ^ { 0 } + 2 \pi \Phi_y (n + 1) , k _ { z } \right) \hat{c} \left( k _ { x } ^ { 0 } + 2 \pi \Phi_x m , k _ { y } ^ { 0 } + 2 \pi \Phi_y n , k _ { z } \right) \right. \right. \\ & \left. \left. + e ^ { i k _ { z } } \hat{c}^ { \dagger } \left( k _ { x } ^ { 0 } + 2 \pi \Phi_x ( m - 1 ) , k _ { y } ^ { 0 } + 2 \pi \Phi_y ( n - 1 ) , k _ { z }\right) \hat{c} \left( k _ { x } ^ { 0 } + 2 \pi \Phi_x m , k _ { y } ^ { 0 } + 2 \pi \Phi_y n , k _ { z } \right) \right) \right\} . \end{aligned}$$
Note that the Brillouin zone is reduced to $[-\pi /q_x, \pi /q_x]$ for $k_x$ and $[-\pi /q_y, \pi /q_y]$ for $k_y$. Now, the Schrödinger equation $\hat { H } \left( k _ { x } ^ { 0 } , k _ { y } \right) | \psi \rangle = E _ { k _ { x } ^ { 0 } , k _ { y } } | \psi \rangle$ is reduced to that for a 1D tight-binding model. The single particle energy is obtained by expanding the state into single-particle states at each lattice point $m$, $$| \psi \rangle = \sum _ { m = 0 } ^ { q - 1 } a _ { m } \hat{c} ^ { \dagger } ( k _ { x } ^ { 0 } + 2 \pi \Phi_x m , k _ { y } ^ { 0 } + 2 \pi \Phi_y m , k _ { z } ) | 0 \rangle$$ where $| 0 \rangle$ is the vacuum. The eigenvalue equation is $$\begin{aligned}
\left( - 2 t \cos \left( k _ { x } ^ { 0 } + 2 \pi \Phi_x n \right) - 2 t \cos \left( k _ { y } ^ { 0 } + 2 \pi \Phi_y n \right) \right) a _ { n } \\ - \frac{V}{2} \left( e ^ { - i k _ { z } } a _ { n - 1 } + e ^ { i k _ { z } } a _ { n + 1 } \right) = E _ { k _ { x } ^ { 0 } ,k _ { y } ^ { 0 } , k _ { z } } a _ { n } .
\label{3Dharper}
\end{aligned}$$ This corresponds to the Harper equation for the Hofstadter model. Let $Q$ be the least common multiple of $q_x$ and $q_y$. For convenience, we perform the transformation $$a _ { j } = \sum _ { l = 0 } ^ { Q-1 } e ^ { i 2 \pi j l/Q } b _ { l }$$ and we obtain $$\begin{aligned}
&-t_x\left( e^{ik_x^0} b_{j+P_x} + e^{-ik_x^0} b_{j-P_x}\right) \\ &-t_y \left( e^{ik_y^0} b_{j+P_y} + e^{-ik_y^0} b_{j-P_y} \right) \\ &-V \cos \left( k_z + \frac{2\pi j}{Q}\right) b_j = E_{k_x^0, k_y^0, k_z} b_j .
\end{aligned}$$ To solve this equation, we apply the perturbation theory under the condition $t_x, t_y \ll V$. The solution at $t_x, t_y = 0$ is $$E _ { m } \left( k _ { x } ^ { 0 } , k _ { y } ^ { 0 } , k_z \right) = - 2 t _ { a } \cos \left( k _ { z } + \frac{2 \pi m}{Q} \right) , \quad \psi _ { j } = \delta _ { j , m } .$$ If we make $t_x, t_y$ finite, gaps open at the degeneracy points. To understand the details, we should make the condition for degeneracy clear. When the two bands $\psi_{m_1}$ and $\psi_{m_2}$ cross each other, the degeneracy condition is $$k _ z + \frac { 2 \pi } { Q } m _ { 1 } = - \left( k _ z+ \frac { 2 \pi } { Q } m _ { 2 } \right) \quad (\bmod 2 \pi) .$$ The degeneracy only occurs when $k_z = 0, \pm \pi / q$, so we can rewrite it as $$m_1+m_2+l \equiv 0 \quad (\bmod \ q) \ \ (l\in \{ 0, \pm 1 \}, l=q k_x^0/\pi) .
\label{banddege}$$ It is obvious that the lowest band is $m=0$. It is possible to determine all of the band indices by using Eq. \[banddege\]. There is a simple relation between the gap number $r$ and band indices $m_1, m_2$, $$r \equiv - |m_1-m_2| \quad (\bmod \ Q) .$$ Now we are ready to use the perturbation theory. Since the $t_x(t_y)$ term only mixes two sites which are $P_x(P_y)$ sites apart from each other, in order to hybridize $\phi_{m_1}$ and $\phi_{m_2}$, the following equation must be satisfied, $$|m_1-m_2| = P_x t_r + P_y u_r \quad (\bmod \ Q)$$ that is, $$r/Q = s_r + p_x/q_x t_r + p_y/q_y u_r .
\label{3Ddio}$$ The lowest order of perturbation which mixes $\phi_{m_1}$ and $\phi_{m_2}$ is the $|t_r|$th order for $t_x$ term, and $|u_r|$th order for the $t_y$ term, respectively.
The Hamiltonian around the $r$th gap is $$\left( \begin{array} { cc } \epsilon & { \Delta e ^ {- i k _ { x } t _ { r } } e ^ {- i k _ { y } u _ { r } } } \\ { \Delta e ^ { i k _ { x } t _ { r } }e ^ { i k _ { y } u _ { r } } } & { - \epsilon } \end{array} \right) \left( \begin{array} { l } { a } \\ { b } \end{array} \right) = E \left( \begin{array} { l } { a } \\ { b } \end{array} \right)$$ Now, it is possible to calculate the amount of pumping in one cycle from this Hamiltonian. The pumping amount will be $t_r$ for $x$-direction and $u_r$ for $y$-direction, and this is exactly the solution of Diophantine equation (\[3Ddio\]). However, we should be aware that “one cycle" here corresponds to the width of Brilliouin zone of $k_z$, while $k_z$ changes its value $2\pi/Q$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper is an elementary introduction to the theory of moduli spaces of curves and maps. As an application to enumerative geometry, we show how to count the number of bitangent lines to a projective plane curve of degree $d$ by doing intersection theory on moduli spaces.'
address: |
Department of Mathematics\
Stanford University\
450 Serra Mall, Bldg. 380\
Stanford, CA 94305-2125
bibliography:
- 'renzo.bib'
title: Counting Bitangents with Stable Maps
---
Introduction {#introduction .unnumbered}
============
Philosophy and Motivation {#philosophy-and-motivation .unnumbered}
-------------------------
The most apparent goal of this paper is to answer the following enumerative question:
*“What is the number $N_{\mathcal{B}}(d)$ of bitangent lines to a generic projective plane curve $Z$ of degree d?”*
This is a very classical question, that has been successfully solved with fairly elementary methods (see for example [@gh:ag], page 277). Here we propose to approach it from a very modern and “technological” angle: we think of lines in the projective plane as maps $\mu:\proj\rightarrow\Proj$ of degree $1$. We mark two points $p_1$ and $p_2$ on $\proj$ and keep track of their image via the map $\mu$. We then construct the space of all such marked maps, and ask ourselves: can we understand the locus $\mathcal{B}$ of all maps that are tangent to $Z$ at the images of both $p_1$ and $p_2$? The answer fortunately is yes. The description of $\mathcal{B}$ allows us to produce a closed formula for $N_{\mathcal{B}}$ in all degrees.
This brief description already reveals that there is something deep and interesting going on here, and that the journey is much more important than the destination itself. Our major goal is to introduce the reader to the rich and beautiful theory of Moduli Spaces in a hopefully “soft” way, with the final treat of seeing it concretely applied to solve our classical problem.
It is our intention for this paper to be a very readable expository work. We designed it to be accessible to a first year graduate student who is considering algebraic geometry as a specialty field. We emphasize geometric intuition and visualization above all, at the cost of silently glossing over some technical details here and there.
Outline of the Paper {#outline-of-the-paper .unnumbered}
--------------------
This paper is divided into three sections, getting progressively more advanced.
The first section introduces some basic ideas and techniques in modern algebraic geometry, necessary to develop and understand the later two sections and is intended for the unexperienced reader. We quickly tread through the most basic ideas in intersection theory; we introduce the concept of families of algebro-geometric objects; we discuss the specific example of vector bundles, and give a working sketch of the theory of Chern classes. Finally, we describe two interesting constructions: the blow-up and jet bundles.\
Entire books have been written on each of these topics, so we have no hope or pretense to be complete, or even accurate. Yet, we still think it valuable to present what lies in the back of a working mathematician’s mind, in the firm belief that a solid geometric intuition is the best stairway to understand and motivate the technicalities and abstract generalizations needed to make algebraic geometry “honest”.
The second section is the development of most of the theory. After a quick qualitative introduction to moduli spaces, we discuss our main characters: the moduli spaces of rational stable curves, and of rational stable maps. Intersection theory on the moduli spaces of stable maps, commonly referred to as Gromov-Witten Theory, is currently an extremely active area of research.
Finally, in the third section we apply all the theory developed so far to solve the bitangent problem. We explore in further detail the moduli spaces of rational stable maps of degree $1$ to $\Proj$, with one and two marked points. By intersecting appropriate cycle classes on these spaces we extract one of the classical *Plücker Formulas*, expressing the number of bitangents as a function of the degree $d$ of the curve.
References {#references .unnumbered}
----------
We suggest here some canonical references for the reader in search of more rigor and completeness. For intersection theory, [@f:it] is a fairly technical book, but definitely it has the last word on it. It also presents Chern classes from an algebraic point of view. A discussion of Chern classes from a geometric point of view can be found in [@bt:df].\
A good treatment of blow-ups can be found in any basic book in algebraic geometry, for example [@gh:ag] or [@h:ag].\
A very pleasant reference for jet bundles is [@v:jb]. An extensive treatment of jet bundles is found in [@s:jb].\
Our presentation of moduli spaces follows the spirit of [@k:kf]; for anybody interested in getting serious, [@hm:mc] is the way to go. Finally, a good introduction to $\psi$ classes is [@k:pc].
Acknowledgments {#acknowledgments .unnumbered}
---------------
We first of all owe the inspiration for this work to Joachim Kock, who outlines this strategy for counting bitangents in his talk [@k:bt], and is also responsible, with Israel Vainsencher, for the best elementary introduction to Gromov-Witten Theory we know of, [@k:kf] . We also thank Aaron Bertram, Herb Clemens, Tommaso de Fernex, Hugo Rossi and Ravi Vakil for their comments, suggestions and encouragement.
Preliminaries
=============
Intersection Theory
-------------------
It will be helpful, but not essential, that the reader be familiar with the Chow ring, $A^\ast(X)$, of an algebraic variety $X$. The ring[^1] $A^\ast(X)$ is, in some loose sense, the algebraic counterpart of the cohomology ring $H^\ast(X)$, and it allows us to make precise in the algebraic category the intuitive concepts of oriented intersection in topology.
We think of elements of the group $A^n(X)$ as formal finite sums of codimension $n$ closed subvarieties (cycles), modulo an equivalence relation called rational equivalence. $A^\ast(X)=\oplus_0^{dim X}A^n(X)$ is a graded ring with product given by intersection.\
Intersection is independent of the choice of representatives for the equivalence classes.
In topology, if we are interested in the cup product of two cohomology classes $\mathbf{a}$ and $\mathbf{b}$, we can choose representatives $a$ and $b$ that are transverse to each other. We can assume this since transversality is a generic condition: if $a$ and $b$ are not transverse then we can perturb them ever so slightly and make them transverse while not changing their classes. This being the case, then $a \cap b$ represents the cup product class $\mathbf{a} \cup \mathbf{b}$.
In algebraic geometry, even though this idea must remain the backbone of our intuition, things are a bit trickier. We will soon see examples of cycles that are rigid, in the sense that their representative is unique, and hence “unwigglable". Transversality then becomes an unattainable dream. Still, with the help of substantially sophisticated machinery (the interested reader can consult [@f:it]), we can define an algebraic version of intersection classes and a product that reduces to the “geometric" one when transversality can be achieved.
Throughout this paper, a bolded symbol will represent a class, the unbolded symbol a geometric representative. The intersection of two classes $\mathbf{a}$ and **b** will be denoted by **ab**.
**Example**: the Chow Ring of Projective Space. $$A^\ast(\mathbb{P}^n) = \frac{\mathbb{C}[\mathbf{H}]}{(\mathbf{H}^{n+1})},$$ where $\mathbf{H}\in A^1(\mathbb{P}^n)$ is the class of a hyperplane $H$.
Families and Bundles
--------------------
One of the major leaps in modern algebraic geometry comes from the insight that, to fully understand algebraic varieties, we should not study them one by one, but understand how they organize themselves in families.\
We are all familiar, maybe subconsciously, with the concept of a family. When, in high school, we dealt with “all parabolas of the form $y=ax^2$" or “all circles with center at the origin", we had in hand prime examples of families of algebraic varieties.\
The idea is quite simple: we have a parameter space, $B$, called the base of the family. For each point $b\in B$ we want an algebraic variety $X_b$ with certain properties. Further, we want all such varieties to be organized together to form an algebraic variety $\mathcal{E}$, called the total space of the family.
A little more formally we could define a family of objects of type $\mathcal{P}$ as a morphism of algebraic varieties $$\begin{array}{c}
\mathcal{E}\\
\pi\downarrow\\
B
\end{array}$$ where $\pi^{-1}(b)$ is an object of type $\mathcal{P}$.
A **section** of a family $\pi:\mathcal{E}\rightarrow B$ is a map $s:B\rightarrow \mathcal{E}$ such that $\pi\circ s:B\rightarrow B$ is the identity map. Often, the section $s$ is written $$\begin{array}{c}
\mathcal{E}\\
\pi\downarrow\uparrow s\\
B.
\end{array}$$ Notice that $s(b)\in\pi^{-1}(b)$.
Given a family $\pi:
\mathcal{E}\longrightarrow B$ and a map $f:M\rightarrow B$ we can construct a new family $$\begin{array}{c}
f^\ast \mathcal{E}\\
\downarrow\pi_f\\
M,
\end{array}$$ called the **pull-back of $\pi$ via $f$**: $$f^{\ast}\mathcal{E}=\{(m,e)\in M\times \mathcal{E} \mid f(m)=\pi(e)\}.$$ Intuitively, the fibre of $\pi_f$ over a point $m\in M$ will be the fibre of $\pi$ over $f(m)$. An essential property of this construction is that it is natural, up to isomorphism.
### Vector Bundles
A **vector bundle** of rank $n$ is a family $\pi:\mathcal{E}\rightarrow B$ of vector spaces over $\mathbb{C}$ of dimension $n$ which is **locally trivial**[^2]. By locally trivial we mean that there is on open cover $\{U_\alpha\}$ of $B$ such that $\pi^{-1}(U_\alpha)\cong U_\alpha\times\mathbb{C}^n$. Our vector bundle is uniquely determined by how these trivial pieces glue together.
A vector bundle of rank one is called a **line bundle**, as its fibers are (complex) lines.
Given two vector bundles $$
[ccc]{} \_1 & & \_2\
\_1& $and$ & \_2\
B & & B
$$over the same base space, one can define their Whitney sum$$
[c]{} \_1\_2\
\
B,
$$ where a fibre $\pi^{-1}(b)$ is the direct sum of the vector spaces $\pi_1^{-1}(b)\oplus\pi_2^{-1}(b)$. It can be easily verified that this family satisfies the local triviality condition.
Similarly, one can define the tensor product $\mathcal{E}_1\otimes \mathcal{E}_2$, the dual bundle $\mathcal{E}^\ast$, the wedge product $\bigwedge^p(\mathcal{E})$ and the bundle $Hom(\mathcal{E}_1, \mathcal{E}_2)=(\mathcal{E}_2\otimes \mathcal{E}_1^\ast)$.
### Characteristic Classes of Bundles {#Chern}
For every vector bundle there is a natural section $s_0:B\rightarrow
\mathcal{E}$ defined by $$s_0(b)=(b,0)\in\{b\}\times\mathbb{C}^n.$$ It is called the zero section, and it gives an embedding of $B$ into $\mathcal{E}$.
A natural question to ask is if there exists another section $s:B\rightarrow \mathcal{E}$ which is disjoint form the zero section, i.e. $s(b)\not=s_0(b)$ for all $b\in B$. The **Euler class** of this vector bundle ($\mathbf{e}(\mathcal{E})\in A^n(B)$) is defined to be the class of the self-intersection of the zero section: it measures obstructions for the above question to be answered affirmatively. This means that $\mathbf{e}(\mathcal{E})=0$ if and only if a never vanishing section exists. It easily follows from the Poincaré-Hopf theorem that for a manifold $M$, the following formula holds: $$\mathbf{e}(TM)\cap[M]=\chi(M).$$ That is, the degree of the Euler class of the tangent bundle is the Euler characteristic.
The Euler class of a vector bundle is the first and most important example of a whole family of “special" cohomology classes associated to a bundle, called the **Chern classes** of $\mathcal{E}$. The $k$-th Chern class of $\mathcal{E}$, denoted $\mathbf{c}_k(\mathcal{E})$, lives in $A^k(B)$. In the literature you can find a wealth of definitions for Chern classes, some more geometric, dealing with obstructions to finding a certain number of linearly independent sections of the bundle, some purely algebraic. Such formal definitions, as important as they are (because they assure us that we are talking about something that actually exists!), are not particularly illuminating. In concrete terms, what you really need to know is that Chern classes are cohomology classes associated to a vector bundle that satisfy a series of really nice properties, which we are about to recall.
Let $\mathcal{E}$ be a vector bundle of rank $n$:
identity:
: by definition, $\mathbf{c}_0(\mathcal{E}) = 1$.
normalization:
: the $n$-th Chern class of $\mathcal{E}$ is the Euler class: $$\mathbf{c}_n(\mathcal{E})=\mathbf{e}(\mathcal{E}).$$
vanishing:
: for all $k>n$, $\mathbf{c}_k(\mathcal{E})=0$.
pull-back:
: Chern classes commute with pull-backs: $$f^\ast\mathbf{c}_k(\mathcal{E})=\mathbf{c}_k(f^\ast \mathcal{E}).$$
tensor products:
: if $L_1$ and $L_2$ are line bundles, $$c_1(L_1\otimes L_2)= c_1(L_1) + c_1(L_2).$$
Whitney formula:
: for every extension of bundles\[whit\] $$0\rightarrow \mathcal{E}' \rightarrow \mathcal{E} \rightarrow \mathcal{E}'' \rightarrow 0 ,$$ the $k$-th Chern class of $\mathcal{E}$ can be computed in terms of the Chern classes of $\mathcal{E}'$ and $\mathcal{E}''$, by the following formula: $$\mathbf{c}_k(\mathcal{E})=\sum_{i+j=k} \mathbf{c}_i(\mathcal{E}')\mathbf{c}_j(\mathcal{E}'').$$
Using the above properties it is immediate to see:
1. all the Chern classes of a trivial bundle vanish (except the $0$-th, of course);
2. for a line bundle $L$, $\mathbf{c}_1(L^\ast)= -\mathbf{c}_1(L)$.
To show how to use these properties to work with Chern classes, we will now calculate the first Chern class of the tautological line bundle over $\proj$. The tautological line bundle is $$\begin{array}{c}
\mathcal{S}\\
\pi\downarrow \ \ \\
\ \ \proj,
\end{array}$$ where $\mathcal{S}=\{(p,l)\in\mathbb{C}^2\times\proj\mid p\in l\}$. It is called tautological because the fiber over a point in $\proj$ is the line that point represents.
Our tautological family fits into the short exact sequence of vector bundles over $\proj$ $$\begin{array}{ccccccccc}
0 &\rightarrow & \mathcal{S} & \rightarrow & \mathbb{C}^2\times\proj & \rightarrow
& \mathcal{Q} & \rightarrow & 0 \\
& & & \searrow & \downarrow & \swarrow &
& & \\
& & & & \proj & & & &
\end{array}$$ where $\mathcal{Q}$ is the bundle whose fibre over a line $l\in\proj$ is the quotient vector space $\mathbb{C}^2/l$. Notice that $\mathcal{Q}$ is also a line bundle. From the above sequence, we have that $$\begin{aligned}
0=\mathbf{c}_1(\mathbb{C}^2\times\proj)=\mathbf{c}_1(\mathcal{S})+\mathbf{c}_1(\mathcal{Q}).\label{triv}\end{aligned}$$
Since $\proj$ is topologically a sphere, which has Euler characteristic 2, then $$\begin{aligned}
2=\mathbf{c}_1(T\proj)=\mathbf{c}_1(\mathcal{S}^\ast)+\mathbf{c}_1(\mathcal{Q})
=-\mathbf{c}_1(\mathcal{S})+\mathbf{c}_1(\mathcal{Q}).\label{tang}\end{aligned}$$ The second equality in \[tang\] holds because $T\proj$ is the line bundle $Hom(\mathcal{S},\mathcal{Q})=
\mathcal{S}^\ast\otimes\mathcal{Q}$.
It now follows from (\[triv\]) and (\[tang\]) that $\mathbf{c}_1(\mathcal{S})=-1$.
Blow-up
-------
Let us begin by discussing the prototypical example of the blow up of a point on a surface: first off, the blow up is a local construction and so we need only understand the local picture.\
Consider the map $$\begin{array}{rccc}
\phi: & \mathbb{C}^2 & \rightarrow & \mathbb{C}\\
& (x,y) & \mapsto & y/x.\\
\end{array}$$ This is a rational map and is not defined on the line $\{x=0\}$. We may try to fix this by modifying our target space to $\proj$. Still, $\phi$ cannot be defined at $\mathbf{0}:=(0,0)$. In fact, along any line $l$ through the origin, the limit of $\phi$ at $\mathbf{0}$ is the slope of $l$. We would like to modify $\mathbb{C}^2$ to a smooth surface birational to it, where the map $\phi$ can be defined everywhere. We would like points outside $\mathbf{0}$ to remain “untouched" and $\mathbf{0}$ to be replaced by a $\proj$ whose points represent all tangent directions at $\mathbf{0}$.
Here is how to do it: consider the graph of $\phi$, $\Gamma_\phi\subset\mathbb{C}^2\times\proj$. We have the commutative diagram $$\begin{array}{rrc}
& \Gamma_\phi & \subset\mathbb{C}^2\times\proj \\
& \stackrel{(id,\phi)}{}{\nearrow}\swarrow \pi_1 &\downarrow \pi_2 \\
&\mathbb{C}^2\setminus\{0\} \stackrel{\phi}{\longrightarrow} & \proj
\end{array}$$ The closure $\overline{\Gamma}_\phi$ is what we are looking for. It is birational to $\mathbb{C}^2$; the left projection ${\pi_1}_{\mid\Gamma_\phi}$ is an isomorphism onto $\mathbb{C}^2-\{\mathbf{0}\}$ and $\pi_1^{-1}(\mathbf{0})=\proj$. We define the blow-up of $\mathbb{C}^2$ at $\mathbf{0}$ as $Bl_\mathbf{0}(\mathbb{C}^2):=\overline{\Gamma}_\phi=\Gamma_\phi\cup\proj$; the projective line $\pi_1^{-1}(\mathbf{0})$ is called the exceptional divisor of the blow-up and denoted $E$. We have obtained a new (smooth!) space by replacing, in a particularly favorable way, a point with the projectivization of its normal bundle.
{height="6cm"}
In general, for $Y\subset X$ a closed subvariety of codimension $k\geq0$, one can construct a new space $Bl_Y(X)$ such that:
1. $Bl_Y(X)$ is birational to $X$;
2. points outside $Y$ are untouched;
3. a point in $Y$ is replaced by $\mathbb{P}^{k-1}$, representing the “normal” directions to $Y$ at that point.
The total space of the blow-up of $\mathbb{C}^2$ at $\mathbf{0}$ admits a natural map to the exceptional divisor, consisting of projecting points along lines through the origin. This realizes $Bl_0 (\mathbb{C}^2)\rightarrow E$ as a line bundle over $\proj$. This is the tautological bundle, which does not have any global sections besides the $0$-section. It follows that the class $\mathbf{E}$ of the exceptional divisor admits only one representative, namely $E$ itself. It is therefore impossible to compute the self-intersection $\mathbf{EE}$ by means of intersecting two transverse representatives of the class.
Jet Bundles {#jets}
-----------
Let $L$ be a line bundle over a variety, $X$. Then the local sections of this line bundle form a vector space. In fact, locally, such a section is just a complex valued function on some open set in X. We will now describe a new vector bundle over $X$ whose fibre over $x\in X$ consists of all Taylor expansions of these sections centered about $x$ and truncated after degree $k$. To see how the locally trivial charts of this bundle glue together is simply a matter of shifting the center of a Taylor expansion. We call this bundle the $k$th jet bundle of $L$ and denote it by $J^kL$. In particular $J^0L=L$.\
Notice that the first jet bundle keeps track of all locally defined functions and differential forms and so there is an obvious surjection $J^1L\rightarrow L$. This gives us the short exact sequence $$0\rightarrow L\otimes\Omega\rightarrow J^1L\rightarrow
L\rightarrow0$$ which will be an essential tool later on.
The previous statement is a particular case of what can be considered the “fundamental theorem of jet bundles".
For all $n\geq 0$, the sequence $$\begin{aligned}
0\rightarrow L\otimes Sym^n\Omega\rightarrow J^nL\rightarrow
J^{n-1}L\rightarrow0
\label{jet}\end{aligned}$$ is exact.
For a slightly more rigorous and still enjoyable account of jet bundles, refer to [@v:jb].
Moduli Spaces
=============
A “High School" Example
------------------------
What is the idea of a moduli space? A moduli space of geometric objects of a certain type is a space which “encodes” information about collections of geometric objects of a given type, in the sense that:
1. points in the moduli space correspond bijectively to the desired geometric objects;
2. the moduli space itself has an algebraic structure that respects how the objects can organize themselves in families.
For example, suppose that we would like to consider the space of all circles in the plane. Since a circle is uniquely the zero locus of a second degree polynomial of the form $(x-x_0)^2+(y-y_0)^2-r^2$, upon specifying the coordinates of the center and its radius, we have completely identified the circle. Thus, the space of all circles in the plane can be represented by $\mathcal{M}:=\mathbb{R}^2\times\mathbb{R}_+$. This is indeed much more than just a set-theoretic correspondence.
Consider the tautological family
$$\begin{array}{c}
\mathcal{U} \ \\
\downarrow\pi\\
\mathcal{M}
\end{array}$$ where $\mathcal{U}:=\{((x_0,y_0),r,(x,y))\mid (x-x_0)^2+(y-y_0)^2=r^2\}\subset\mathcal{M}\times\mathbb{R}^2$ and $\pi$ is the projection onto the first factor. This family enjoys the following properties, that clarify the vague point $2$ above:
1. for any family of circles in the plane $p:E\rightarrow B$, there is a map $m:B\rightarrow\mathcal{M}$ defined by $m(b)=p^{-1}(b)$;
2. to every map $m:B\rightarrow\mathcal{M}$ there uniquely corresponds a family of circles parametrized by $B$, i.e. $$\begin{array}{c}
m^\ast \mathcal{U}\\
p\downarrow\\
B
\end{array}$$ such that the fibre $p^{-1}(b)$ is the circle $m(b)$.
This is the best that we could have hoped for. In this case we say that $\mathcal{M}$ is a **fine moduli space** with $\mathcal{U}$ as its **universal family**.
Often, due to the presence of automorphisms of the parametrized objects, it is impossible to achieve this perfect bijection between families of objects and morphisms to the moduli space. If only property 1 holds we call the moduli space **coarse**.
Moduli of $n$ Points on $\proj$
-------------------------------
Let us now consider the moduli space $M_{0,n}$ of all isomorphism classes of $n$ ordered distinct marked points $p_i\in\proj$. The subscript 0 is to denote the genus of our curve $\proj$. Since the automorphism group $Aut(\proj)=PSL_2(\mathbb{C})$ allows us to move any three points on $\proj$ to the ordered triple $(0,1,\infty)$, the space $M_{0,n}$ reduces to a single point for $n\leq3$.
Going one step up, $M_{0,4}=\proj-\{0,1,\infty\}$ : given a quadruple $(p_1,p_2,p_3,p_4)$, we can always perform the unique automorphism of $\proj$ sending $(p_1,p_2,p_3)$ to $(0,1,\infty)$; the isomorphism class of the quadruple is then determined by the image of the fourth point.
The general case is similar. Any $n$-tuple $\underline{p}=(p_1,\ldots, p_n)$ is equivalent to a $n$-tuple of the form $(0,1,\infty,\phi(p_4),\ldots,\phi(p_n))$, where $\phi$ is the unique automorphism of $\proj$ sending $(p_1,p_2,p_3)$ to $(0,1,\infty)$. This shows $$M_{0,n}=\overbrace{M_{0,4}\times...\times M_{0,4}}^{n-3\ \mbox{times}} \setminus \{\mbox{all
diagonals}\}.$$ If we define $U_n:=M_{0,n}\times\proj$, then the projection of $U_n$ onto the first factor gives rise to a universal family $$\begin{array}{c}
U_n\\
\pi\downarrow\uparrow\sigma_i\\
M_{0,n}
\end{array}$$ where the $\sigma_i$’s are the universal sections:
- $\sigma_i(\underline{p})= (\underline{p},\phi(p_i))\in U_n$.
This family is tautological since the fibre over a moduli point, which is the class of a marked curve, is the marked curve itself.\
With $U_n$ as its universal family, $M_{0,n}$ becomes a fine moduli space for isomorphism classes of $n$ ordered distinct marked points on $\proj$.
This is all fine except $M_{0,n}$ is not compact for $n\geq 4$. There are many reasons why compactness is an extremely desirable property for moduli spaces. As an extremely practical reason, proper (and if possible projective) varieties are much better behaved and understood than non compact ones. Also, a compact moduli space encodes information on how our objects can degenerate in families. For example, what happens when $p_1\rightarrow p_2$ in $M_{0,4}$?
In general there are many ways to compactify a space. A “good” compactification $\overline{\mathcal{M}}$ of a moduli space $\mathcal{M}$ should have the following properties:
1. $\overline{\mathcal{M}}$ should be itself a moduli space, parametrizing some natural generalization of the objects of $\mathcal{M}$.
2. $\overline{\mathcal{M}}$ should not be a horribly singular space.
3. the boundary $\overline{\mathcal{M}}\setminus\mathcal{M}$ should be a normal crossing divisor.
4. it should be possible to describe boundary strata combinatorially in terms of simpler objects. This point may appear mysterious, but it will be clarified by the examples of stable curves and stable maps.
In the case of rational $n$-pointed curves there is a definite winner among compactifications.
Moduli of Rational Stable Curves
--------------------------------
We will discuss the simple example of $M_{0,4}$; this hopefully will, without submerging us in combinatorial technicalities, provide intuition on the ideas and techniques used to compactify the moduli spaces of $n$-pointed rational curves.
A natural first attempt would be to just allow the points to come together, i.e. enlarge the collection of objects that we are considering from $\proj$ with $n$ ordered distinct marked points to $\proj$ with $n$ ordered, not necessarily distinct, marked points.
However, this will not quite work. For instance, consider the families $$C_t=(0,1,\infty,t) \mbox{\ \ and\ \ }
D_t=(0,t^{-1},\infty,1).$$ For each $t\not=0$, up to an automorphism of $\proj$, $C_t=D_t$, thus corresponding to the same point in $M_{0,4}$. But for $t=0$, $C_0$ has $p_1=p_4$ whereas $D_0$ has $p_2=p_3$. These configurations are certainly not equivalent up to an automorphism of $\proj$ and so should be considered as distinct points in our compactification of $M_{0,4}$. Thus, we have a family with two distinct limit points (in technical terms we say that the space is **nonseparated**). This is not good.
Our failed attempt was not completely worthless though since it allowed us to understand that we want the condition $p_1=p_4$ to coincide with $p_2=p_3$, and likewise for the other two possible disjoint pairs. On the one hand this is very promising: $3$ is the number of points needed to compactify $\proj\setminus\{0,1,\infty\}$ to $\proj$. On the other hand, it is now mysterious what modular interpretation to give to this compactification.
To do so, let us turn carefully to our universal family, illustrated in Figure \[uf\]. The natural first step is to fill in the three points on the base, to complete $U_4$ to $\proj\times\proj$ and extend the sections by continuity.
![first attempt at compactifying $U_{4}$[]{data-label="uf"}](pic4.eps){height="6cm"}
We immediately notice a bothersome asymmetry in this picture: the point $p_4$ is the only one allowed to come together with all the other points: yet common sense, backed up by the explicit example just presented, suggests that there should be democracy among the four points. This fails where the diagonal section $\sigma_4$ intersects the three constant ones, i.e. at the three points $(0,0)$, $(1,1)$, $(\infty,\infty)$. Let us blow-up $\proj\times\proj$ at these three points. This will make all the sections disjoint, and still preserve the smoothness and projectivity of our universal family.\
The fibres over the three exceptional points are $\proj \cup E_i$: nodal rational curves. These are the new objects that we have to allow in order to obtain a good compactification of $M_{0,4}$.\
Let us finally put everything together, and state things carefully.
A **tree of projective lines** is a connected curve with the following properties:
1. Each irreducible component is isomorphic to $\proj$.
2. The points of intersection of the components are ordinary double points.
3. There are no closed circuits, i.e., if a node is removed then the curve becomes disconnected.
These three properties are equivalent to saying that the curve has arithmetic genus zero. Each irreducible component will be called a twig. We will often draw a marked tree as in fig \[mt\], where each line represents a twig.
![stable marked trees.[]{data-label="mt"}](pic0.eps){height="4cm"}
A marked tree is **stable** if every twig has at least three special points (marks or nodes).
This stability condition is equivalent to the existence of no nontrivial automorphisms of the tree that fix all of the marks.
$\overline{M}_{0,4}\cong \proj$ is the moduli space of isomorphism classes of four pointed stable trees. It is a fine moduli space, with universal family $U_4 = Bl(\proj\times \proj)$.
These results generalize to larger $n$.\
**Fact:** The space $\overline{M}_{0,n}$ of $n$-pointed rational stable curves is a fine moduli space compactifying $M_{0,n}$. It is projective, and the universal family $\overline{U}_n$ is obtained from $U_n$ via a finite sequence of blow-ups. (In particular all the diagonals need to be blown up in an appropriate order) . For further details see [@k:kf] or [@k:2], [@k:1].
One of the exciting features of this theory is that all these spaces are related to one another by natural morphisms. Consider the map $$\pi_i:\overline{M}_{0,n+1}\rightarrow\overline{M}_{0,n},$$ defined by forgetting the $i$th mark. It is obviously defined if the $i$th mark does not belong to a twig with only three special points. If it does belong to such a twig, then our resulting tree will no longer be stable. In this case, we must perform what is called contraction.
Contraction:
: We need to consider two cases:
1. The remaining two special points are both nodes. We make the tree again stable by contracting this twig so that the two nodes are now one (see Figure \[c1\]).
![contracting a twig with only two nodes.[]{data-label="c1"}](pic5.eps){width="13.5cm"}
2. There is one other mark and one node on the twig in question. We make the tree stable by forgetting the twig and placing the mark where the node used to be (Figure\[c2\]).
![contracting a twig with one node and one mark.[]{data-label="c2"}](pic6.eps){width="14cm"}
We would like to construct a section $\sigma_i$ of the family $$\begin{array}{c}
\overline{M}_{0,n+1}\\
\pi_k\downarrow\uparrow\sigma_i\\
\overline{M}_{0,n}
\end{array}$$ by defining the $k$th mark to coincide with $i$th one. It should trouble you that in doing so we are not considering curves with distinct marked points, but we can get around this problem by “sprouting” a new twig so that the node is now where the $i$th mark was. The $k$th and the $i$th points now belong to this new twig.\
This process of making stable a tree with two coinciding points is called **stabilization**.
{width="13cm"}
Finally, we may now identify our universal family $$
[ccc]{} \_n & & \_[0,n+1]{}\
& $with the family$ & \_i\
\_[0,n]{} & & \_[0,n]{}
$$ as follows.\
The fibre $\pi^{-1}([(C,p_1, \cdots, p_n)])\subset\overline{U}_n$ is the marked curve itself. So any point $p\in\overline{U}_n$ belonging to the fibre over $C$ is actually a point on the stable $n$-pointed tree $C$, and may therefore be considered as an additional mark; stabilization may be necessary to ensure that our new $(n+1)$-marked tree is stable. Vice-versa, given an $(n+1)$ pointed curve $C'$, we can think of the $(n+1)$st point as being a point on the $n$-marked curve obtained by forgetting the last marked point (eventually contracting, if needed); this way $C'$ corresponds to a point on the universal family $\overline{U_n}$. These constructions realize an isomorphism between $\overline{U}_n$ and $\overline{M}_{0,n+1}$.
### The boundary
We define the boundary to be the complement of $M_{0,n}$ in $\overline{M}_{0,n}$. It consists of all nodal stable curves.
**Fact:** the boundary is a union of irreducible components, corresponding to the different possible ways of arranging the marks on the various twigs; the codimension of a boundary component equals the number of nodes in the curves in that component. See [@k:kf] for more details.
The codimension $1$ boundary strata of $\overline{M}_{0,n}$, called the **boundary divisors**, are in one-to-one correspondence with all ways of partitioning $[n]=A\cup B$ with the cardinality of both $ A$ and $B$ strictly greater than 1.\
A somewhat special class of boundary divisors consists of those with only two marked points on a twig. Together, these components are sometimes called the **soft boundary** and denoted by $D_{i,j}$. We can think of $D_{i,j}$ as the image of the $i$th section, $\sigma_i$, of the $j$th forgetful map, $\pi_j$ (or vice-versa).
![irreducible components of the boundary of $\overline{M}_{0,4}$[]{data-label="m04"}](pic8.eps){width="13cm"}
![boundary cycles of $\overline{M}_{0,5}$[]{data-label="m05"}](m05.eps){width="16cm"}
There is plenty more to be said about the spaces $\overline{M}_{0,n}$, their relationships, and their boundaries, but we will leave our treatment of $\overline{M}_{0,n}$ here, suggesting [@k:kf] as an excellent reference for beginners. In Figures \[m04\] and \[m05\] we draw all boundary strata for $\overline{M}_{0,4}$ and $\overline{M}_{0,5}$.
Moduli of Rational Stable Maps
------------------------------
Let us now move on to the moduli spaces of greatest interest for solving the bitangent problem. We would like to study, in general, rational curves in projective space. The characteristic property of an irreducible rational curve is that it can be parametrized by the projective line, $\proj$. For this reason, it is natural to study maps $\mu:\proj\rightarrow\mathbb{P}^r$.
When we talk about the **degree** of such a map we mean the degree of $\mu_\ast[\proj]$ as in homology. Be careful, the degree of the map may be different from the degree of the image curve! For example the map $$\begin{array}{cccc}
\mu: &\proj & \rightarrow & \Proj \\
& (x_0:x_1) & \mapsto & (x_0^2:x_1^2 :0) \\
\end{array}$$ has degree two, but its image is a line.
Define $W(r,d)$ as the space of all maps from $\proj$ to $\mathbb{P}^r$ of degree $d$. A map in $W(r,d)$ is specified, up to a constant, by $r+1$ binary forms of degree $d$ that do not all vanish at any point. It can then be seen that $dim W(r,d)=(r+1)(d+1)-1$.
We also have the family $$\begin{array}{ccc}
W(r,d)\times\proj & \stackrel{\rho}{\rightarrow} & \mathbb{P}^r\\
\downarrow & &\\
W(r,d) & &
\end{array}$$ where $\rho(\mu,x)=\mu(x)$. This family is tautological in the sense that the fibre over the map $\mu$ is the map $\rho\mid_{\{\mu\}\times\proj}=\mu$. In fact, this is a universal family. Thus, $W(r,d)$ is a fine moduli space for maps $\proj\rightarrow\mathbb{P}^r$ of degree $d$.
However, $W(r,d)$ is not the moduli space that we would like to study. For one, it is not compact. For another, reparametrizations of the source curve are considered as different points in $W(r,d)$. To fix the latter problem, let us simply consider the space $M_{0,0}(\mathbb{P}^r,d):=W(r,d)/Aut(\proj)$. For a detailed account on why this quotient is indeed a space, see chapter five of [@hm:mc]
Another way to eliminate automorphisms is to consider $n$-pointed maps (maps $\mu:C\rightarrow\mathbb{P}^r$ with an $n$-marked source $C\simeq\proj)$. It should be no surprise that there is a fine moduli space $M_{0,n}(\mathbb{P}^r,d)$ for isomorphism classes of $n$-pointed maps $\proj\rightarrow\mathbb{P}^r$ of degree $d$, namely $M_{0,n}\times
W(r,d)$. But we still have not dealt with the non-compactness of this moduli space. The idea is to parallel the construction that led us to stable curves.
An **n-pointed stable map** is a map $\mu:C\rightarrow\mathbb{P}^r$, where:
1. $C$ is a $n$-marked tree.
2. Every twig in $C$ mapped to a point must have at least three special points on it.
**Fact:** Moduli spaces of $n$-pointed rational stable maps to $\mathbb{P}^r$ of degree $d$ (denoted $\overline{M}_{0,n}(\mathbb{P}^r,d))$ can be constructed; they compactify the moduli spaces of smooth maps. It is straightforward to verify that an $n$-pointed map is stable if and only if it has only a finite number of automorphisms. Unfortunately, there is no way to eliminate all nontrivial automorphism. Details can be found in [@k:kf].
**Example:** An element $\mu\in\overline{M}_{0,2}(\mathbb{P}^2,2)$ that is the double cover of a line, marking the ramification points, admits a nontrivial automorphism exchanging the two covers. This allows us to construct a nontrivial family of maps $\mu_t$ that maps constantly to one point in the moduli space. Consider: $$\begin{array}{cccc}
\mu_t: &[0,1]/\{0=1\} \times \proj & \longrightarrow & \Proj \\
& & & \\
& (t,(x_0:x_1)) & \mapsto & (0:x_0^2: e^{2\pi it}x_1^2).
\end{array}$$
Because of this phenomenon there is no universal family associated to the spaces $\overline{M}_{0,n}(\mathbb{P}^r,d)$, and the corresponding moduli spaces are only coarse.
Since $M_{0,n}(\mathbb{P}^r,d)$ is dense in $\overline{M}_{0,n}(\mathbb{P}^r,d)$, the latter has dimension $$(r+1)(d+1)-1+(n-3)=rd+r+d+n-3.$$
**Example:** in particular, $\overline{M}_{0,n}(\mathbb{P}^r,0) = \overline{M}_{0,n} \times \mathbb{P}^r.$
There are some useful maps among the spaces $\overline{M}_{0,n}(\mathbb{P}^r,d)$. For instance, as with the spaces $\overline{M}_{0,n}$, we have the forgetful maps $\pi_i$ defined by simply forgetting the $i$th mark and the sections, $\sigma_j$, of the family $$\begin{array}{c}
\overline{M}_{0,n+1}(\mathbb{P}^r,d)\\
\pi_i\downarrow\uparrow\sigma_j\\
\overline{M}_{0,n}(\mathbb{P}^r,d)
\end{array}$$ defined by declaring the $j$th and the $i$th mark to coincide. Contraction and stabilization are performed to make these maps defined everywhere.
In addition, there are **evaluation maps** $$\nu_i:\overline{M}_{0,n}(\mathbb{P}^r,d)\rightarrow\mathbb{P}^r$$ defined by $\nu_i(\mu)=\mu(p_i)$ where $p_i$ is the $i$th mark on the source curve $C$.
The forgetful and evaluation morphisms allow us to identify $\overline{M}_{0,n+1}(\mathbb{P}^r,d)$ as a tautological family for $\overline{M}_{0,n}(\mathbb{P}^r,d) $:
$$\begin{array}{ccc}
\overline{M}_{0,n+1}(\mathbb{P}^r,d) & \stackrel{\nu_{n+1}}{\longrightarrow} &
\mathbb{P}^r\\
\pi_{n+1}\downarrow & & \\
\overline{M}_{0,n}(\mathbb{P}^r,d). & &
\end{array}$$
This way we can think of points of $\overline{M}_{0,n+1}(\mathbb{P}^r,d)$ either as $n+1$ pointed maps or as points on an $n$-marked curve mapped to $\mathbb{P}^r$. Being comfortable with this identification comes in very handy when making computations.
### The Boundary
The boundary consists of maps whose domains are reducible curves. In fact, its description is very similar to that of $\overline{M}_{0,n}$. Boundary strata are determined now not only by the combinatorial data of the arrangement of the marks, but also by the degree the maps restrict to on each twig.
Boundary divisors are in one to one correspondence with all ways of partitioning $[n]=A\cup B$ and $d=d_A+d_B$ such that:
- $\#A\geq 2$ if $d_A=0$;
- $\#B\geq 2$ if $d_B=0$.
Psi classes
-----------
Consider a family of curves admitting a section. $$\begin{array}{c}
\mathcal{C}\\
\pi\downarrow\uparrow \sigma\\
B
\end{array}$$ We can define the **cotangent line bundle**, $\mathbb{L}_\sigma$, as the line bundle on $B$ whose fibre at a point $b\in B$ is the cotangent space of $\mathcal{C}_b=\pi^{-1}(b)$ at the point $\sigma(b)$. We call the $\mathbf{\psi}$ **class** of the family the first Chern class of this line bundle. Observe that, for a trivial family of curves with a constant section, the $\psi$ class vanishes.
This construction can be extended in a natural way to the moduli space $\overline{M}_{0,n}(\mathbb{P}^r,d)$.\
Informally, we have a sheaf on the tautological family $\overline{M}_{0,n+1}(\mathbb{P}^r,d)$ whose local sections away from nodes are differential forms on the curves. We obtain it by considering $1$-forms on the tautological family and quotienting by forms that are pulled back from the moduli space. This sheaf is called the **relative dualizing sheaf**[^3] , and denoted by $\omega_{\pi_{n+1}}$.
Consider now the $i$th tautological section $\sigma_i$. If we restrict $\omega_{\pi_{n+1}}$ to this section, we obtain a line bundle on the moduli space whose fibres are naturally identified with the cotangent spaces of the curves at the $i$th marked point. Then we can define the class: $$\mathbf{\psi}_i:=\mathbf{c}_1(\sigma_i^\ast\omega_{\pi_{n+1}})\in
A^1(\overline{M}_{0,n}(\mathbb{P}^r,d)).$$
The construction of $\psi_i$ is natural in the sense that if we have a family of pointed stable maps, inducing a morphism to the moduli space, the $\psi_i$ class of the family is the pull-back of the $\psi_i$ class on the moduli space.
It may seem that there is no difference between the information carried by $\mathbf{\psi}_i\in
A^1(\overline{M}_{0,n+1}(\mathbb{P}^r,d))$ and $\mathbf{\psi}_i\in
A^1(\overline{M}_{0,n}(\mathbb{P}^r,d))$. We have a natural map between these two spaces, the tautological family $\pi_{n+1}:\overline{M}_{0,n+1}(\mathbb{P}^r,d)\rightarrow\overline{M}_{0,n}(\mathbb{P}^r,d)$. It may seem that $\mathbb{L}_{i,n+1}:=\sigma_i^\ast\omega_{\pi_{n+1}}$ and $\pi_{n+1}^\ast \mathbb{L}_{i,n}$ are the same line bundle, thus yielding the relation $\mathbf{\psi}_i=\pi_{n+1}^\ast\mathbf{\psi}_i$. In reality, this is almost true. Surely these line bundles agree off the component $D_{i,n+1}$ of the soft boundary. From this consideration, we conclude $$\begin{aligned}
\mathbb{L}_{i,n+1}=\pi_{n+1}^\ast \mathbb{L}_{i,n}\otimes\mathcal{O}(mD_{i,n+1})\end{aligned}$$ for some integer $m$.
Next, observe that $\mathbb{L}_{i,n+1}$ restricted to $D_{i,n+1}$ is a trivial line bundle: we are looking at curves with a twig having only three special points; the node, the $i$th and the $(n+1)$st mark. By an automorphism of the twig, we can assume that the node is at $0$ and the two marks are at $1$ and $\infty$. Therefore, this line bundle restricted to $D_{i,n+1}$ is the cotangent space at a single unchanging point of $\proj$. This implies $$\mathcal{O}_{D_{i,n+1}} = {\mathbb{L}_{i,n+1}}_{\mid D_{i,n+1}} =$$ $$= (\pi_{n+1}^\ast
\mathbb{L}_{i,n}\otimes\mathcal{O}(mD_{i,n+1}))_{\mid D_{i,n+1}} =
\mathbb{L}_{i,n} \otimes{\mathcal{O}(mD_{i,n+1})}_{\mid D_{i,n+1}}.$$
By the adjunction formula ([@gh:ag], page 146), the line bundle $\mathcal{O}{(D_{i,n+1})}_{\mid D_{i,n+1}}$ is the normal bundle of the divisor $D_{i,n+1}$.
But the normal directions to a section in the moduli space are precisely the tangent directions to the fibres. Hence $\mathcal{O}{(D_{i,n+1})}_{\mid D_{i,n+1}}$ is the dual to the relative cotangent bundle $\mathbb{L}_{i,n}$. It follows that $m$ must be $1$.
Finally, by taking Chern classes in $(3)$, we can deduce the fundamental relation: $$\begin{aligned}
\psi_i=\pi_{n+1}^\ast \psi_i + D_{i,n+1}.\end{aligned}$$ The above pull-back relation can be used to describe explicitly $\psi$ classes for moduli spaces of rational stable curves in terms of boundary strata. A closed formula can be found in [@k:pc].
Counting Bitangents
===================
The Strategy
------------
We now have all the necessary machinery to tackle our problem of counting bitangents. Before we start digging deep into details and computations, let us outline our strategy.\
Let $Z:=\{f=0\}$ be a projective plane curve of degree $d$:
- we consider the moduli space $\maps$, of two-pointed, rational stable maps of degree one;
- we construct a jet bundle on this space with the property that the zero set of a section of this bundle consists of stable maps having at least second order contact with $Z$ at the image of the $i$th marked point; we name this cycle $\Phi_i(Z)$;
- we represent $\Phi_i(Z)$ in the Chow ring in terms of $\psi$ classes and other natural classes;
- we step by step compute the intersection $\Phi_1(Z)\Phi_2(Z)$;
- we identify and clean up some garbage that lives in that intersection and corresponds to maps that are not bitangents;
- finally, we have counted two pointed maps that are tangent to $Z$ at each mark; we just need to divide by $2$, since we are not interested in the ordering of the marks.
Easy enough? Now let us start over slowly and do everything carefully.
$\map$
------
This space has dimension 3, and it is explicitly realized by the following incidence relation: $$\map
=
\{(p,l)
\in
\Proj\times\check{\Proj}
\mid
p\in
l\}
=:\I
\subset
\Proj\times\check{\Proj}.$$ There are two projections of $\mathcal{I}$ onto $\Proj$ and $\check{\Proj}$, that we will denote $q$ and $\check{q}$. The latter makes $\I$ into a tautological family of lines in $\Proj$: $$\begin{array}{ccc}
\proj &
\longrightarrow &
\I \\
& & \downarrow\check{q} \\
& & \check{\Proj}.
\end{array}$$ This family is tautological in the sense that the fibre over $l\in\check{\Proj}$ is $l$ itself.
A fibre over $p\in\Proj$ under $q$ is the pencil of lines in $\Proj$ passing through $p$ and so this projection also gives rise to a $\proj$ bundle over $\Proj$. Observe that $q$ is precisely the evaluation map $\nu_x$.
**Notation:** We denote by $x$ the unique mark in the space of one-pointed maps, and add the subscript $x$ to any entity (class, map ...) related to it. We do so to keep track of the conceptual difference from the marked points on the two-pointed maps, which will be numbered $1$ and $2$.
We identify and name two natural divisors on $\map$.
$\mathbf{\iota}(p):$
: Let us look at the hyperplane divisor $\check{p}\subset\check{\Proj}$ of all lines passing through a point $p$, and consider the cycle of its pull-back $\check{q}^\ast(\check{p}):=\mathbf{\iota}(p)$.
$\mathbf{\eta}_x(l):$
: Similarly, we look at the hyperplane divisor $l\subset\Proj$ and define $\mathbf{\eta}_x(l):=q^\ast (l)$ as its pull-back under $q$.
In general, define $\mathbf{\eta}_x(Z):=q^\ast(Z)=\nu^\ast_x(Z)$ as the cycle of maps whose mark is sent into $Z$.
There is only one class of a line and only one of a point in $A^\ast(\Proj)$, hence $\mathbf{\eta}_x:=[\mathbf{\eta}_x(l)]$ and $\mathbf{\iota}:=[\mathbf{\iota}(p)]$ are independent of $l$ and $p$ respectively.\
Since $\map$ is a $\proj$ bundle over $\Proj$, it follows that $\mathbf{\iota}$ and $\mathbf{\eta}_x$, i.e. the pull-backs of hyperplane divisors in the base and in the fiber, generate $A^1(\map)$. It is therefore useful to know all intersections of the two classes.
It is a good exercise to construct a picture verifying each of the following statements about classes.
- $\mathbf{\eta}_x=[\{(p^\prime,l^\prime)\mid p^\prime\in l,
l^\prime\in\check{p^\prime}$ with $l$ fixed$\}]$
- $\mathbf{\iota}=[\{(p^\prime,l^\prime)\mid l^\prime\in\check{p},
p^\prime\in l^\prime$ with $p$ fixed$\}]$
- $\mathbf{\eta}_x^2=[\{(p,l^\prime)\mid l^\prime\in\check{p}$ with $p$ fixed$\}]$
- $\mathbf{\iota\eta}_x=[\{(p^\prime,l^\prime)\mid p^\prime\in l,
l\in\check{p}$ with $p$ and $l$ fixed$\}]$
- $\mathbf{\iota}^2=[\{(p^\prime,l)\mid p^\prime\in\l$ with $l$ fixed$\}]$
- $\mathbf{\iota\eta}_x^2=[\{(p,l^\prime)\mid l\in\check{p}\cap\check{p}_0$ with $p,p_0$ fixed$\}]=\mathbf{pt}$ since there is exactly one line passing through two distinct points.
- $\mathbf{\eta}_x\iota^2=[\{(p^\prime,l)\mid p^\prime\in l\cap l_0$ with $l,l_0$ fixed$\}]=\mathbf{pt}$ since there is exactly one point in the intersections of two distinct lines.
- $\mathbf{\eta}_x^3=[\{(p^\prime,l^\prime)\mid
l^\prime\in\check{p}_1\cap\check{p}_2\cap\check{p}_3, p^\prime\in
l^\prime$ with $p_1,p_2,p_3$ fixed$\}]=0$ since in general three points do not lie on a common line.
- $\mathbf{\iota}^3=[\{(p^\prime,l^\prime)\mid p^\prime\in l_1\cap
l_2\cap l_3, p^\prime\in l^\prime$ with $l_1,l_2,l_3$ fixed$\}]=0$ since in general three lines do not share a common point.
The following two lemmas prove identities that will be crucial for our later computations.
$$\begin{aligned}
\mathbf{\eta}_x\iota=\mathbf{\eta}_x^2+\mathbf{\iota}^2. \label{etaiota}\end{aligned}$$
<span style="font-variant:small-caps;">Proof:</span> We construct a one-parameter family of cycles, parametrized by $[0,1]$, with the left-hand side of our identity as one endpoint of this family and the right-hand side as the other.\
To choose a representative of the class $\mathbf{\eta}_x\iota$, one must specify a fixed point $p$ and a fixed line $l$. Let us fix $l$ once and for all and let $p_t$ be a path in $\Proj$ such that $p_t\in l$ if and only if $t=0$. Our one parameter family $\mathbf{\alpha}_t$ is defined as follows: $$\mathbf{\alpha}_t=\{(p^\prime,l^\prime)\mid p^\prime\in l, l\in\check{p_t}\}.$$ Notice that $[\mathbf{\alpha}_t]=\mathbf{\eta}_x\iota$ for $t\not=0$.
{width="12cm"}
We now examine what happens as $t\rightarrow0$.
$p^\prime\not=p_0$:
: necessarily, $l^\prime=l$ and our resulting one dimensional class is parametrized solely by $p^\prime\in l$ and is therefore $\mathbf{\iota}^2$.
$p^\prime=p_0$:
: Then $l^\prime$ is only required to be in $\check{p_0}$ and so such $l^\prime$’s in $\check{p_0}$ parametrize our resulting one dimensional class. We have arrived at $\mathbf{\eta}_x^2$.
We now have that $\alpha_0=\eta_x^2+\iota^2$, which allows us to conclude (\[etaiota\]).
$$\begin{aligned}
\mathbf{\psi}_x=\mathbf{\iota}-2\mathbf{\eta}_x. \label{psi}\end{aligned}$$
<span style="font-variant:small-caps;">Proof:</span> As $\mathbf{\psi}_x\in A^1(\map)$, it is possible to express $\mathbf{\psi}_x=a\mathbf{\iota}+b\mathbf{\eta}_x$ for some integers $a$ and $b$. Let us determine these two integers.
a:
: intersecting $\psi_x$ with $\eta_x^2$ we obtain $$\psi_x\eta_x^2=a\mathbf{\iota}\mathbf{\eta}_x^2+b\mathbf{\eta}_x^3=a.$$
Consider $\sigma_x^\ast\omega_{\pi_2}$ restricted to $\eta_x^2=\{(p,l^\prime)\mid p$ is fixed and $l^\prime\in\check{p}\}$. It is the line bundle over $\eta_x^2$ whose fibre over a point $(p,l^\prime)\in\eta_x^2$ is the cotangent space of $l^\prime$ at the fixed point $p$.
It is worth convincing yourself that this is the dual of the tautological line bundle $$\begin{array}{cl}
\mathcal{S} & \\
\downarrow\pi & \\
\proj & = \{l\subset\Proj\mid p\in l\} = \eta_x^2.
\end{array}$$
We computed (section \[Chern\]) that $\mathbf{c}_1(\mathcal{S})=-1$ and so $a=1$.
b:
: similarly, $b$ is the product $\mathbf{\psi}_x{\iota}^2$. To find this intersection we must consider the line bundle $\sigma_x^\ast\omega_{\pi}$ restricted to $\iota^2=\{p\in l\mid l$ is fixed$\}$. A fibre of this line bundle over a point $p\in l$ is the cotangent space of our fixed $l$ at $p$. This is simply the cotangent bundle of $l$.
Since $l=\proj=S^2$ has Euler characteristic $2$, then the degree of the first Chern class of the cotangent bundle is $-2$. We thus have that $b=-2$.
$\maps$
-------
First off note that $dim\maps=4$. The description of $\maps$ is slightly more complicated largely due to the existence of its only boundary divisor which we call $\beta$.
{height="4cm"}
For a two-pointed map $\mu$ not in the boundary, $\mu(p_1)\not=\mu(p_2)$. Since we are considering maps of degree 1, i.e. isomorphisms of lines, $\mu(p_1)$ and $\mu(p_2)$ completely determine (up to reparametrization)j our map $\mu$. It follows that $$\maps\setminus\beta=\Proj\times\Proj\setminus\Delta.$$ On $\beta$, $\mu(p_1)=\mu(p_2)$: for any line $l$ through $p$, there is a map in $\beta$ that contracts the twig with the marks to $p$ and maps the other twig isomorphically to $l$.\
So for our description to be complete, we need to replace $(p,p)\in\Delta\subset\Proj\times\Proj$ with a $\proj$ worth of maps. We arrive at $$\maps=Bl_\Delta(\Proj\times\Proj).$$
Consider the tautological families $$\begin{array}{c}
\maps\\
\pi_1\downarrow\downarrow\pi_2\\
\map.
\end{array}$$ Both families have a natural common section $\sigma:=\sigma_1=\sigma_2$. The image of $\sigma$ is the unique boundary divisor $\beta$ in $\maps$.
Define $\eta_i(Z):= \nu_i^\ast(Z)$ It’s easy to check the following identities:
1. $\sigma_\ast(\mathbf{\eta}_x(Z))=\mathbf{\beta\eta}_1(Z)=\mathbf{\beta\eta}_2(Z)$.
2. $\sigma^\ast(\eta_i(Z))=\eta$.
3. $\pi_i^\ast\mathbf{\eta}_x(Z)=\mathbf{\eta}_i(Z)$.
Tangency Conditions
-------------------
Let us define $\mathbf{\Phi}_i(Z)\in
A^\ast(\overline{M}_{0,n}(\mathbb{P}^2,1))$ as the cycle of maps tangent to a plane curve $Z=\{f=0\}$ at the image of the $i$th marked point. Formally, $$\mathbf{\Phi}_i(Z)=\{\mu\in\overline{M}_{0,n}(\mathbb{P}^2,1)|\mu^\ast
f \mbox{ vanishes at }p_i\mbox{ with
multiplicity }\geq2\}.$$
We now want to obtain an expression for $\mathbf{\Phi}_x(Z)\in
A^\ast(\map)$ in terms of $\mathbf{\eta}_x$, $\mathbf{\beta}$, and $\mathbf{\psi}_x$.\
Consider the tautological family $$\begin{array}{ccc}
\maps & \stackrel{\nu_2}{\longrightarrow} & \Proj \\
& & \\
\sigma\uparrow \downarrow \pi_2& \nearrow \nu_x & \\
& & \\
\map. & &
\end{array}$$ Let us pull-back the line bundle $\mathcal{O}(Z)$ via $\nu_2$, and consider the first jet bundle $J_{\pi_2}^1\nu_2^\ast\mathcal{O}(Z)$ relative to $\pi_2$. Relative here means that we quotient out by everything that can be pulled back from $\map$. Let $Z$ be defined by the vanishing of the polynomial $f$, and let us consider the zero locus of the section $\tau:=\nu_2^\ast f + (\partial_{\pi_2}^1\nu_2^\ast f) dt\in \Gamma(J_{\pi_2}^1\nu_2^\ast\mathcal{O}(Z)$); what we obtain is the locus of maps in $\maps$ such that the pull-back of $f$ at the second marked point vanishes to second order. If we interpret $\maps$ as the universal family for $\map$ [^4], it follows that to obtain $\mathbf{\Phi}_x(Z)$, the locus in $\map$ of lines tangent to $Z$ at the unique marked point, we need to pull-back via the section $\sigma$.
In formulas, this translates to $$\mathbf{\Phi}_x(Z) =\mathbf{e}(\sigma^\ast
J_{\pi_2}^1\nu_2^\ast\mathcal{O}(Z)).$$ Since the rank of the bundle in question is $2$, the Euler class is the second Chern class.
To calculate $\mathbf{c}_2(\sigma^\ast
J_{\pi_2}^1\nu_2^\ast\mathcal{O}(Z))$ we use the following short exact sequence discussed in section \[jets\]. $$0\rightarrow\nu_2^\ast\mathcal{O}(Z)\otimes\omega_{\pi_2}\rightarrow
J_{\pi_2}^1\nu_2^\ast\mathcal{O}(Z)\rightarrow\nu_2^\ast\mathcal{O}(Z)\rightarrow0.$$ Notice that the first and last terms of this sequence are line bundles. We then want to consider the pull-back along $\sigma$ of this exact sequence. Using the Whitney formula, we now have that $$\mathbf{\Phi}_x(Z)=\mathbf{c}_1(\sigma^\ast\nu_2^\ast\mathcal{O}(Z))\mathbf{c}_1(\sigma^\ast\nu_2^\ast\mathcal{O}(Z)\otimes\sigma^\ast\omega_{\pi_2})=d\mathbf{\eta}_x(d\mathbf{\eta}_x+\mathbf{\psi}_x)$$ in $\map$.
The last equality follows from the two facts:
- $[Z]= d\mathbf{H} \in A^1(\Proj)$, where $\mathbf{H}$ is the hyperplane class generating $A^\ast(\Proj)$ and $d=deg f$;
- $\nu_x = \nu_2\sigma$, and $\eta_x$ is by definition $\nu_x^\ast(\mathbf{H})$.
Now we want to consider the case when we have more than one marked point: let us say we want to compute $\Phi_1(Z)$ in $\maps$. The obvious guess is $\Phi_1(Z)=d\mathbf{\eta}_1(d\mathbf{\eta}_1+\mathbf{\psi}_1)$. We need to be careful, though: for maps in $\beta\mathbf{\eta}_1(Z)\subset d\mathbf{\eta}_1(d\mathbf{\eta}_1+\mathbf{\psi}_1)$, the whole twig containing the two marks is mapped to $Z$. So, for $\mu$ such a map, $\mu^\ast f$ vanishes identically along the contracting twig and thus to all orders. We do not want to consider these maps as tangents to $Z$.
**Fact:** This simple correction works. The formula in $\maps$ is $$\begin{aligned}
\mathbf{\Phi}_1(Z)=d\mathbf{\eta}_1(d\mathbf{\eta}_1+\mathbf{\psi}_1-\mathbf{\beta}).\end{aligned}$$ Lastly, note that $\mathbf{\beta\Phi}_1(Z)=\sigma_\ast\mathbf{\Phi}_x(Z)$.
**Remark:** It would be nice to use higher order jet bundles to describe cycles of maps having higher order contact with our curve $Z$. Unfortunately in general it is quite difficult, as fairly complicated problems of excess intersection arise.
For our application, we only need to push our luck a little further: we need to describe the cycle $\mathbf{\Phi}_x^{(3)}(Z)$ of inflection tangents to $Z$. Luckily, thanks to the fact that $\map$ has no boundary, the argument carries through: $\mathbf{\Phi}_x^{(3)}(Z)$ can be computed as the Euler class of the bundle $E^\prime:=\sigma^\ast
J_{\pi_2}^2\nu_2^\ast\mathcal{O}(Z)$. Here, rank$ E^\prime=3$. By the exact sequence (\[jet\]) and the Whitney formula: $$\mathbf{\Phi}_x^{(3)}(Z)=\mathbf{c}_3(E^\prime)=\mathbf{\eta}_x(Z)(\mathbf{\eta}_x(Z)+\mathbf{\psi}_x)(\mathbf{\eta}_x(Z)+2\mathbf{\psi}_x).$$ Using our previous calculations, $$\begin{aligned}
\mathbf{\Phi}_x^{(3)}(Z)=(3d^2-6d)\mathbf{pt}.\end{aligned}$$ We will abuse notation from now on and leave off the class of a point in our calculations. When writing a dimension zero class we will simply write its integral over the fundamental class.
The Computation
---------------
Let us finally get down to business. We define in general $$\Lambda_Z(m_1p_1+m_2p_2)\subset\maps$$ as the cycle of maps $\mu$, such that $\mu^\ast\mathbf{Z}\geq m_1p_1+m_2p_2$. Note that $\mathbf{\Lambda}_Z(2p_1)=\mathbf{\Phi}_1(Z)$.\
Our ultimate goal is to compute $\mathbf{\Lambda}_Z(2p_1+2p_2)$, i.e. the class of maps in $\maps$ which are tangent to $Z$ at both $p_1$ and $p_2$.\
Note that $dim\mathbf{\Lambda}_Z(2p_1+2p_2)=0$ since we are imposing 4 independent conditions in a space of dimension 4. This tells us that our enumerative problem makes sense.
On first thought, one might suggest $\mathbf{\Lambda}_Z(2p_1+2p_2)=\mathbf{\Phi}_1(Z)\mathbf{\Phi}_2(Z)$. After all, $\mathbf{\Phi}_1(Z)$ is the class of all maps tangent to $Z$ at $p_1$ and similarly for $\mathbf{\Phi}_2(Z)$ at $p_2$ so their intersection seems to be what we are after. However, one must be careful: for example, $\mathbf{\beta\Phi}_1(Z)$ is in this intersection and we don’t want to consider such maps.
We will proceed in two steps and obtain $\mathbf{\Lambda}_Z(2p_1+2p_2)$ from $\mathbf{\Phi}_1(Z)\mathbf{\Phi}_2(Z)$ by “throwing away the spurious maps”.
**Step 1:** We consider the intersection $\mathbf{\Phi}_1(Z)\mathbf{\eta}_2(Z)$, consisting of all maps tangent to $Z$ at $p_1$ that intersect $Z$ at $p_2$. What we get are two parts:
- $\mathbf{\Lambda}_Z(2p_1+p_2)$, in which $p_1$ and $p_2$ do not lie on the same twig of degree zero;
- $\mathbf{\beta\Phi}_1(Z)=\sigma_\ast\mathbf{\Phi}_x(Z)$, corresponding to maps with both marked points on a degree $0$ twig.
Set theoretically, $$\eta_2(Z)\Phi_1(Z)=\Lambda_Z(2p_1+p_2)\cup\sigma_\ast\Phi_x(Z).$$ The correct multiplicities are 1 and 2 respectively, giving: $$\mathbf{\Lambda}_Z(2p_1+p_2)=\mathbf{\eta}_2(Z)\mathbf{\Phi}_1(Z)-2\sigma_\ast\mathbf{\Phi}_x(Z).$$
**Step 2:**We now intersect $
\mathbf{\Lambda}_Z(2p_1+p_2)$ with $\mathbf{\eta}_2(Z)+\mathbf{\psi}_2-\mathbf{\beta}$. This imposes second order vanishing at the second marked point.
Again, this intersection gives one part with multiplicity 1, which is not on the boundary $\mathbf{\beta}$, and another part with multiplicity 2 in $\mathbf{\beta}$. The first is what we are looking for: $\mathbf{\Lambda}_Z(2p_1+2p_2)$. To find the other part, remember we are already working inside $\mathbf{\Lambda}_Z(2p_1+p_2)$, so to lie in $\mathbf{\beta}$ means $p_2\rightarrow p_1$. We thus obtain $\mathbf{\beta\Lambda}_Z(3p_1)=\mathbf{\beta\Phi}_1^{(3)}$ with multiplicity 2.
**Note:** To explain these multiplicities rigorously is a subtle business which we will not go into here. The intuition behind these $2$’s is that the boundary contribution can be carried by either of the classes we are intersecting, and hence shows up twice in the intersection. The reader interested in how to carry out these computation can consult [@g:tnopc].
Finally, we have identified
$$\begin{aligned}
%\hspace{-1cm}
\mathbf{\Lambda}_Z(2p_1+2p_2) &=& -2\sigma_\ast\mathbf{\Phi}_x^{(3)}(Z)+(z\mathbf{\eta}_2+\mathbf{\psi}_2-\mathbf{\beta})\mathbf{\Lambda}_Z(2p_1+p_2)\nonumber\\
&=& -2\sigma_\ast\mathbf{\Phi}_x^{(3)}+\mathbf{\Phi}_1(Z)\mathbf{\Phi}_2(Z)-2\sigma_\ast\mathbf{\Phi}_x(Z)(z\mathbf{\eta}_2+\mathbf{\psi}_2-\mathbf{\beta}).\nonumber \\\end{aligned}$$
Now all that stands in our way of calculating bitangents are substitutions and computations.
Before we move on, take a second to remember or derive the following easy facts, that we will use in the forthcoming computations:
(a)
: $\mathbf{\iota\eta}_x^2=\mathbf{\iota}^2\mathbf{\eta}_x=1$
(b)
: $\mathbf{\iota\eta}_x=\mathbf{\eta}_x^2+\mathbf{\iota}^2$
(c)
: $\mathbf{\psi}_x=\mathbf{\iota}-2\mathbf{\eta}_x$
(d)
: $\mathbf{\eta}_i^3=0$
(e)
: $\mathbf{\psi}_1=\pi_2^\ast\mathbf{\psi}_x+\mathbf{\beta}$
(f)
: $\mathbf{\beta\psi}_i=0$
(g)
: $\mathbf{\beta}^2=-\mathbf{\beta}\pi_i^\ast\mathbf{\psi}_x$
(h)
: $\mathbf{\beta\eta}_1=\mathbf{\beta\eta}_2$
(i)
: $\sigma_\ast\mathbf{\Phi}_x(Z)=\mathbf{\beta\Phi}_1(Z)=\mathbf{\beta\Phi}_2(Z)$
(j)
: $\mathbf{\eta}_1\pi_2^\ast\mathbf{\alpha}=\pi_2^\ast(\mathbf{\alpha\eta}_x)$ for any class $\mathbf{\alpha}\in A^\ast(\map)$.
Let us expand the first term in (9), $\sigma_\ast\mathbf{\Phi}_x^{(3)}$. We have already found $\mathbf{\Phi}_x^{(3)}(Z)=(3d^2-6d)$ in $\map$. Since pushing forward preserves dimension and $\sigma_i$ is injective, thus sending the class of a point to the class of a point, then our first term becomes $$\sigma_\ast\mathbf{\Phi}_x^{(3)}=3d^2-6d.$$
Now for the second and third terms. $$\begin{aligned}
\mathbf{\Phi}_1(Z)\mathbf{\Phi}_2(Z) & =
& d\mathbf{\eta}_1(d\mathbf{\eta}_1+\mathbf{\psi}_1-\mathbf{\beta})d\mathbf{\eta}_2(d\mathbf{\eta}_2+\mathbf{\psi}_2-\mathbf{\beta})\nonumber\\
& = &
d^4(\mathbf{\eta}_1^2\mathbf{\eta}_2^2)+d^3(\mathbf{\eta}_1^2\mathbf{\eta}_2\mathbf{\psi}_2+\mathbf{\eta}_1\mathbf{\eta}_2^2\mathbf{\psi}_1)+d^2(\mathbf{\eta}_1\mathbf{\eta}_2\mathbf{\psi}_1\mathbf{\psi}_2+\mathbf{\eta}_1^2\mathbf{\beta}^2)\nonumber.\end{aligned}$$ $$\begin{aligned}
\sigma_\ast\mathbf{\Phi}_x(Z)(d\mathbf{\eta}_2+\mathbf{\psi}_2-\mathbf{\beta})
& = &
\mathbf{\beta}((d\mathbf{\eta}_x)(d\mathbf{\eta}_1+\mathbf{\psi}_1-\mathbf{\beta}))(d\mathbf{\eta}_2+\mathbf{\psi}_2-\mathbf{\beta})\nonumber\\
& = & -2d^2(\mathbf{\eta}_1^2\mathbf{\beta}^2)+d(\mathbf{\eta}_1\mathbf{\beta}^3)\nonumber.\end{aligned}$$ We must now compute each intersection in the above expressions[^5].
- $\mathbf{\eta}_1^2\mathbf{\eta}_2^2=1$, as seen from the figure \[eta\] illustrating the fact that there is exactly one line passing through two prescribed points.
![the intersection $\mathbf{\eta}_1^2\mathbf{\eta}_2^2$. []{data-label="eta"}](pic12.eps){height="5cm"}
- $\mathbf{\eta}_1^2\mathbf{\beta}^2
\ \stackrel{(f)}{=}\ \mathbf{\eta}_1^2\mathbf{\beta}(-\pi_2^\ast\mathbf{\psi}_x)
\ \stackrel{(i)}{=}\ -\mathbf{\beta}\pi_2^\ast(\mathbf{\eta}_x^2\mathbf{\psi}_x)
\ \stackrel{(b+c)}{=}\ -\mathbf{\beta}\pi_2^\ast(a)=-\mathbf{\beta}[fibre]$.\
But $\beta$ is the image of a section, hence it intersects all fibres transversely. Thus, $\mathbf{\eta}_1^2\mathbf{\beta}^2=-1$.
- $\mathbf{\eta}_1\mathbf{\beta}^3
\ \stackrel{(f)}{=}\ -\mathbf{\eta}_1\mathbf{\beta}^2\pi_2^\ast\mathbf{\psi}_x
\ \stackrel{(f)}{=}\ \mathbf{\eta}_1\mathbf{\beta}(\pi_2^\ast\mathbf{\psi}_x)^2
=\mathbf{\eta}_1\mathbf{\beta}(\pi_2^\ast\mathbf{\psi}_x^2)
\ \stackrel{(i)}{=}\ \mathbf{\beta}\pi_2^\ast(\mathbf{\psi}_x^2\mathbf{\eta}_x)
\ \stackrel{(a+b+c)}{=}\ \mathbf{\beta}\pi_2^\ast(-3)=-3$.
- $\mathbf{\eta}_1^2\mathbf{\eta}_2\mathbf{\psi}_2
\ \stackrel{(d)}{=}\mathbf{\eta}_1^2\mathbf{\eta}_2(\pi_1^\ast\mathbf{\psi}_x+\mathbf{\beta})
\ \stackrel{(i)}{=}\mathbf{\eta}_1^2\pi_1^\ast(\mathbf{\eta}_x\mathbf{\psi}_x)
\ \stackrel{(a+b)}{=}\mathbf{\eta}_1^2\pi_1^\ast(\mathbf{\iota}^2-\mathbf{\eta}_x^2)$.\
Notice first of all that $\mathbf{\eta}_1^2\pi_1^\ast\mathbf{\eta}_x^2= \mathbf{\eta}_1^2\mathbf{\eta}_2^2=1$, as shown in figure \[eta\].
Next, we claim that $\mathbf{\eta}_1^2\pi_1^\ast\mathbf{\iota}^2=0$. In fact $\mathbf{\eta}_1^2$ is the class of all two-pointed lines passing through a fixed point, where the first mark is at the fixed point while the second mark is free to move.
![the intersection $\mathbf{\eta}_1^2\pi_1^\ast(\mathbf{\iota}^2)$. []{data-label="neta"}](pic13.eps){height="5cm"}
Intersecting with $\pi_1^\ast(\mathbf{\iota}^2)$ means to require that our fixed point intersects our fixed line transversely, which is to say that they don’t intersect at all. This is illustrated in figure \[neta\].
Thus, $\mathbf{\eta}_1^2\mathbf{\eta}_2\mathbf{\psi}_2=-1$; by symmetry, we also have $\mathbf{\eta}_2^2\mathbf{\eta}_1\mathbf{\psi}_1=-1$.
- Finally, $$\begin{aligned}
\hspace{-1.5cm}\mathbf{\eta}_1\mathbf{\eta}_2\mathbf{\psi}_1\mathbf{\psi}_2
& \stackrel{(d)}{=} &
\mathbf{\eta}_1(\pi_2^\ast\mathbf{\psi}_x+\mathbf{\beta})\mathbf{\eta}_2(\pi_1^\ast\mathbf{\psi}_x+\mathbf{\beta})\nonumber\\
& \stackrel{(i)}{=} &
(\pi_2^\ast(\mathbf{\psi}_x\mathbf{\eta}_x)+\mathbf{\beta\eta}_1)(\pi_1^\ast(\mathbf{\psi}_x\mathbf{\eta}_x+\mathbf{\beta\eta}_2)\nonumber\\
& \stackrel{(a+b)}{=} &
\pi_1^\ast(\mathbf{\iota}^2-\mathbf{\eta}_x^2)\pi_2^\ast(\mathbf{\iota}^2-\mathbf{\eta}_x^2)+\mathbf{\beta\eta}_2\pi_1^\ast(\mathbf{\iota}^2-\mathbf{\eta}_x^2)+\mathbf{\beta\eta}_1\pi_2^\ast(\mathbf{\iota}^2-\mathbf{\eta}_x^2)+\mathbf{\beta}^2\mathbf{\eta}_1^2\nonumber\\
& = &
\pi_1^\ast(\mathbf{\iota}^2)\pi_2^\ast(\mathbf{\iota}^2)-\pi_1^\ast(\mathbf{\iota}^2)\pi_2^\ast(\mathbf{\eta}_x^2)-\pi_1^\ast(\mathbf{\eta}_x^2)\pi_2^\ast(\mathbf{\iota}^2)+\pi_1^\ast(\mathbf{\eta}_x^2)\pi_2^\ast(\mathbf{\eta}_x^2)\nonumber\\
& &+\mathbf{\beta}\pi_1^\ast(\mathbf{\iota}^2\mathbf{\eta}_x)+\mathbf{\beta}\pi_2^\ast(\mathbf{\iota}^2\mathbf{\eta}_x)+\mathbf{\eta}_x^2\mathbf{\beta}^2.\nonumber\end{aligned}$$
Via our previous calculations, the last three terms can be easily seen as $1,1,$ and $-1$ respectively. So we must now find the first four terms above. We do so by recalling pictures.
$\pi_1^\ast(\mathbf{\iota}^2)$ is the class of a fixed line and all ordered pairs of points on it.
To intersect two such classes is to require that our line is fixed as *two* transverse lines, which is impossible. Thus $\pi_1^\ast(\mathbf{\iota}^2)\pi_2^\ast(\mathbf{\iota}^2)=0$.
$\pi_1^\ast(\mathbf{\eta}_x^2)= \eta_2^2$, and notice that now symmetry implies that $\eta_2^2\pi_2^\ast(\mathbf{\iota}^2)=\eta_1^2\pi_1^\ast(\mathbf{\iota}^2)= 0$, as shown in figure \[neta\]. Similarly, $\pi_2^\ast(\mathbf{\eta}_x^2)\pi_1^\ast(\mathbf{\iota}^2)=0$.
Lastly, let us intersect $\pi_2^\ast(\mathbf{\eta}_x^2)$ with $\pi_1^\ast(\mathbf{\eta}_x^2)$. But it’s clear that $\pi_2^\ast(\mathbf{\eta}_x^2)\pi_1^\ast(\mathbf{\eta}_x^2)=\eta_1^2\eta_2^2$, and we have shown in figure \[eta\] that this intersection is $1$.
We have then found $$\mathbf{\eta}_1\mathbf{\eta}_2\mathbf{\psi}_1\mathbf{\psi}_2=0+0+0+1+1+1-1=2.$$
Putting this all together, we have now calculated $$\begin{aligned}
\mathbf{\Lambda}_Z(2p_1+2p_2) & = &
-2(3d^2-6d)+(d^4-2d^3+d^2)-2(-2d^2-3d)\nonumber\\
& = & d^4-2d^3-9d^2+18d\nonumber\\
& = & d(d-2)(d-3)(d+3).\nonumber\end{aligned}$$
After the dust has settled, we now know that a generic plane curve $Z$ of degree $d$ has $$N_{\mathcal{B}}(d)=\frac{1}{2}d^4-d^3-\frac{9}{2}d^2+9d$$ bitangents. Remember we are dividing by $2$ because we do not care about the order of the marked points.
Notice that for $d=2$ and $d=3$ we get that there are no bitangents as should be the case. For $d=4$ we find $28$ bitangents, the first interesting result.
[^1]: This is probably our greatest sloppiness. In order for $A^\ast(X)$ to be a ring we need $X$ to be smooth. Since the spaces we will actually work with satisfy these hypotheses, we do not feel too guilty.
[^2]: To be precise, more structure is needed: the clutching functions must take values in $GL(n,\mathbb{C})$.
[^3]: The word relative refers to the fact that we are quotienting by everything coming from downstairs. In other words, we are constructing a sheaf on the universal family of the moduli space by “gluing” together sheaves defined on the curves.
[^4]: this is true because one pointed maps of degree one have no nontrivial automorphisms!
[^5]: The little numbers over the equal signs refer to the identities from page that are used at each step.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We construct a new representation formula for indefinite improper affine spheres in terms of two para-holomorphic functions and study singularities which appear in this representation formula. As a result, it follows that cuspidal cross caps never appear as the singularities on indefinite improper affine spheres and so on. Comparison with other representation formulae are also studied.'
address: ' Graduate School of Mathematics, Kyushu University, Higashi-ku, Fukuoka 812-8581, Japan'
author:
- Daisuke Nakajo
date: 'December 7, 2007'
title: A representation formula for indefinite improper affine spheres
---
Introduction {#introduction .unnumbered}
============
Affine spheres are important objects in the study of affine differential geometry. They have close relations to the theory of minimal surfaces, Monge-Ampère equations, Liouville equation, Tzitzeica equation and so on. In this paper, we construct a representation formula for indefinite improper affine spheres in the affine 3-space (Theorem \[thm:rep\]), in terms of two para-holomorphic functions. This may be regarded as an indefinite version of the representation formula for locally strongly convex improper affine spheres obtained by A. Martínez [@Martinez]. Here, indefinite improper affine spheres mean improper affine spheres with indefinite affine metric. It is found that affine spheres which are represented by these formulae may have singularities and we investigate them. As a result, we have peculiar examples of indefinite improper affine spheres with singularities (Section \[sec:ex\]).
So far, various kinds of representation formulae for (improper) affine spheres are studied by several authors. Perhaps the earliest version, which is discovered in early twentieth century, is due to Blaschke [@Blaschke]. In this representation formula, improper affine spheres are represented in terms of two smooth functions. Another representation formula is due to Cortés [@Cortes]. His formula represents special class of improper affine spheres which are related to special Kähler structure, in terms of one holomorphic function. The feature of his representation formula is that it even covers higher (even-) dimensional improper affine spheres.
Recently, A. Martínez made a representation formula for locally strongly convex improper affine spheres in terms of two holomorphic functions ([@Martinez]). The feature of Martínez’ representation formula is that it includes even improper affine spheres with singularities. In the same paper, he also introduced the notion of IA-map, a class of (locally strongly convex) improper affine spheres with singularities which have close relations with his representation formula. He also studied a correspondence between improper affine spheres and flat fronts in hyperbolic 3-space (see [@KRSUY], [@flat; @fronts]) in [@IGV].
Following Martínez’ manner, we introduce *generalized IA-maps* in Section \[sec: rep\]. (Definitions \[def: Martinez\], \[def: indefinite\]). The singularities on indefinite generalized IA-maps have different properties from those on locally strongly convex ones. The singularities on locally strongly convex generalized IA-maps are either singularity of fronts, or branch points. Contrarily, the singularities on indefinite generalized IA-maps, however, are not necessarily fronts nor branch points. Moreover, the cuspidal cross caps never appear as singularities of indefinite generalized IA-maps (Theorem \[thm: not CCR\]), while they appear often in the case of spacelike maximal surfaces in Lorentzian 3-space [@FSUY]. Furthermore, we found a strange example of singularity on indefinite generalized IA-map (Section \[sec:ex\]). This is not $\mathcal{A}$-equivalent to any singularities on improper affine spheres obtained by the projection of a generalized geometric solution of certain Monge-Ampère system which are considered in [@Ishikawa-Machida] (Section \[sec:comparison\]).
Preliminaries {#sec: prel}
=============
Before explaining the main part of results, we would like to review some definitions and basic known facts about affine differential geometry, para-complex number and singularity theory.
Improper affine spheres
-----------------------
At first, we would like to introduce briefly the affine differential geometry. For the detailed exposition, see [@Li-Simon-Zhao] and [@Nomizu-Sasaki].
Let $(f,\xi)$ be a pair of an immersion $f: M^n \rightarrow {{\boldsymbol{R}}}^{n+1}$ into the affine space ${{\boldsymbol{R}}}^{n+1}$ and the vector field $\xi$ on $M^n$ along $f$ which is transversal to $f_{*}(TM)$. Then, the Gauss-Weingarten formula become $$\begin{cases}
\label{eq: Gauss-Weingarten}
D_{X}{f_*Y} &= f_*({\nabla}_XY)+g(X,Y)\xi, \\
D_{X}\xi &= -f_*(SX) +\tau(X)\xi, \\
\end{cases}$$ where $D$ is the standard connection on ${{\boldsymbol{R}}}^{n+1}$. Here, $g$ is called the affine metric of a pair $(f,\xi)$. It can be easily shown that the rank of the affine metric $g$ is invariant under the change of transversal vector field $\xi$. So we call $f$ a *locally strongly convex immersion* (respectively *indefinite immersion*) if $g$ is positive definite (resp. indefinite). Given an immersion $f: M^n \rightarrow {{\boldsymbol{R}}}^{n+1}$, we can choose uniquely the transversal vector field $\xi$ which satisfy the following conditions,
1. ${\tau}\equiv 0$, (or equivalently $D_X\xi \in f_*(TM)$ for all $X \in \mathfrak{X}(M)$),
2. ${{\operatorname{vol}}}_g(X_1, \cdots, X_n) = {\det}(f_*X_1, \cdots, f_*X_n, \xi)$ for all $X_1, \cdots, X_n \in \mathfrak{X}(M)$,
where ${{\operatorname{vol}}}_{g}$ is the volume form of the (pseudo-)Riemannian metric $g$ and $\det$ is the standard volume element of ${{\boldsymbol{R}}}^{n+1}$. The transversal vector field $\xi$ which satisfies above two conditions is called a *Blaschke normal* (or *affine normal*) and a pair $(f,\xi)$ of a immersion and its Blaschke normal is called a *Blaschke immersion*.
A Blaschke immersion $(f,\xi)$ with $S=0$ in is called an *improper affine sphere*. In this case $\xi$ becomes constant vector because $\tau=0$. Hence hereafter, we can think of a transversal vector field $\xi$ of a improper affine sphere as $\xi=(0,\cdots, 0,1)$ after certain affine transformation of ${{\boldsymbol{R}}}^{n+1}$.
The *conormal map* ${\nu}: M^n \rightarrow ({{\boldsymbol{R}}}^{n+1})^{*}$ for a given Blaschke immersion $(f, \xi)$ is defined as the immersion which satisfy the following conditions,
1. ${\nu}(f_{*}X) =0$ for all $X \in \mathfrak{X}(M)$,
2. ${\nu}(\xi)=1$.
For an improper affine sphere with the Blaschke normal $(0, \cdots, 1)$, we can write $\nu=(n,1)$ with a smooth map $n: M^2 \rightarrow {{\boldsymbol{R}}}^n$.
Using the notations defined as above, we can now state the representation formula for locally strongly convex improper affine spheres (possibly with singularities) by Martinez in [@Martinez]. In that paper, he define the notion of improper affine maps, a generalization of improper affine spheres which possibly have singularities (Definition \[def: Martinez\]) and give its representation formula as follows:
\[thm:lscIA-map\] Let $\psi=(x,{\varphi}): {\Sigma}^2\rightarrow {{\boldsymbol{R}}}^2\times {{\boldsymbol{R}}}$ be an improper affine map. Then there exists a regular complex curve ${\alpha}:=(F,G):{\Sigma}^2 \rightarrow {{\boldsymbol{C}}}^2$ such that, $$\label{eq:IA-map}
\psi=\left( G+\bar{F}, \dfrac{1}{2}(|G|^2-|F|^2)+{{\operatorname{Re}}}\left( GF-2\int{FdG} \right) \right).$$ Here, the conormal map of $\psi$ becomes $$\nu = (\bar{F}-G ,1).$$ Conversely, given a Riemann surface ${\Sigma}$ and a complex curve $\alpha:=(F,G):{\Sigma}\rightarrow
{{\boldsymbol{C}}}^2$, then gives an improper affine map which is well defined if and only if $\int{FdG}$ does not have real periods.
Para-complex numbers
--------------------
The set of para-complex numbers $\widetilde{{\boldsymbol{C}}}$, is an algebra over ${\boldsymbol{R}}$ which is defined as $${\widetilde{{\boldsymbol{C}}}}:=\left\{a+jb| a,b \in {{\boldsymbol{R}}}, j^2=1 \right\}.$$ For $z = u+jv \in {\widetilde{{\boldsymbol{C}}}}$ the conjugate $\bar{z}$ is defined as $\bar{z}:= u-jv$ and absolute value $|z|$ is defined as $|z|:=z{\bar{z}}=u^2-v^2$. As an analogy to the holomorphic function, we can define the so-called para-holomorphic function: A map $F: {\widetilde{{\boldsymbol{C}}}}\rightarrow {\widetilde{{\boldsymbol{C}}}}$ is called a *para-holomorphic function* if $F$ satisfies the para-Cauchy-Riemann equation, that is, $$\label{eq: para-CR}
\begin{cases}
&\dfrac{\partial{f^1}}{\partial{u}} = \dfrac{\partial{f^2}}{\partial{v}}, \\[6pt]
&\dfrac{\partial{f^1}}{\partial{v}} = \dfrac{\partial{f^2}}{\partial{u}}. \\
\end{cases}$$ where $F(u+jv)=f^1(u,v)+jf^2(u,v)$. The set of para-holomorphic functions forms an algebra. For a para-holomorphic function $F: {\widetilde{{\boldsymbol{C}}}}\rightarrow {\widetilde{{\boldsymbol{C}}}}$, we define its differential $F'$ as $$F'(u+iv):=\dfrac{\partial{f^1}}{\partial{u}}+j\dfrac{\partial{f^2}}{\partial{u}} \ \ \ \
( \textrm{or equivalently}, F'(u+iv):=\dfrac{\partial{f^2}}{\partial{v}}+j\dfrac{\partial{f^1}}{\partial{v}})$$ where $F(u+jv)=f^1(u,v)+jf^2(u,v)$.
Since is reduced to the wave equation $$\frac{{\partial}^2{f^j}}{\partial{u^2}}
-\frac{{\partial}^2{f^j}}{\partial{v^2}}=0 \ \ \ \ \ (j=1,2)$$ and the general solution of the wave equation, which is known as d’Alembert solution, is given by arbitrary two smooth functions, a para-holomorphic function is described in terms of two smooth functions. Concretely, for a para-holomorphic function $F$, there exist two smooth functions $\rho$ and $\sigma$ on ${{\boldsymbol{R}}}^2$, such that $$F(u+jv) = \rho (u+v)+\sigma(u-v) + j \left\{ \rho(u+v)-\sigma(u-v) \right\}$$ holds. For more detailed exposition on para-complex numbers, see [@Inoguchi-Toda].
Criteria for singularities
--------------------------
Here, we explain criteria for singularities of smooth maps (see [@FSUY], [@KRSUY] and [@flat; @fronts]). The explanation is restricted to the case for smooth maps from $2$-dimensional manifolds $M^2$ to the affine $3$-space ${{\boldsymbol{R}}}^3$ because all the affine spheres are considered to be $2$-dimensional in this paper.
Consider a smooth map $f:U\rightarrow {{\boldsymbol{R}}}^3$ from a open subset $U$ of ${{\boldsymbol{R}}}^2$ to the affine $3$-space ${{\boldsymbol{R}}}^3$. A point $p$ on $U$ is called a *singular point of $f$*, if $f$ is not immersive at $p \in U$. On studying the local properties of singularity on smooth maps, one usually consider its map-germ. Let $f:U\rightarrow {{\boldsymbol{R}}}^3$ and $g: V \rightarrow {{\boldsymbol{R}}}^3$ are smooth maps from open sets $U$, $V$ of ${{\boldsymbol{R}}}^2$ to ${{\boldsymbol{R}}}^3$ respectively, with $p \in U\cap V$. Then we call *$f$ and $g$ defines the same map-germ at $p$* if there exists an open set $W \subset {{\boldsymbol{R}}}^2$ with $p \in W$ and $W \subset {U\cap V}$ such that $f=g$ holds on $W$. Next, we introduce $\mathcal{A}$-equivalence, an equivalence relation between two singularities. Given two smooth maps $f:U\rightarrow {{\boldsymbol{R}}}^3$ and $g:\widetilde{U}\rightarrow {{\boldsymbol{R}}}^3$, and assume $p \in U$ and $\tilde{p} \in {\widetilde{U}}$ are singular points. Consider two map-germ $(f,p)$ around $p \in U$ and $(g, \widetilde{p})$ around $\tilde{p} \in \widetilde{U}$. Then $(f,p)$ and $(g, \tilde{p})$ is called *$\mathcal{A}$-equivalent* if there exists two diffeomorphism-germ ${\varphi}: U\rightarrow {\widetilde{U}}$ and $\psi:{{\boldsymbol{R}}}^3\rightarrow {{\boldsymbol{R}}}^3$ such that $\psi \circ f = \widetilde{f} \circ {\varphi}$ as a map-germ around $p \in U$ and $\tilde{p}= {\varphi}(p)$ hold.
In this paper, we mainly consider the special class of singularities, that is, frontals and fronts. A smooth map (possibly with singularities) $f:U\rightarrow {{\boldsymbol{R}}}^3$ from open subset $U \subset {{\boldsymbol{R}}}^2$ to the affine $3$-space ${{\boldsymbol{R}}}^3$ is called a *frontal* (or a *frontal map*) if there exists a unit normal vector field $\nu$ to $f$ (even on singular points). This is equivalent to the existence of a map to the 2-sphere $\nu:U \rightarrow S^2\subset{{\boldsymbol{R}}}^3$ which satisfies $$\label{eq:frontal map}
\langle df, \nu \rangle = 0,$$ where $\langle , \rangle$ is the Euclidean metric of ${{\boldsymbol{R}}}^3$. By the identification of the unit tangent bundle $T_{1}{{\boldsymbol{R}}}^3$ and the unit cotangent bundle $T_{1}^{*}{{\boldsymbol{R}}}^3$, frontals can be considered as the projection of a Legendrian map into $T_{1}^{*}{{\boldsymbol{R}}}^3$ with respect to the canonical contact structure as follows. Let $f:U\rightarrow {{\boldsymbol{R}}}^3$ be a frontal and $\nu$ be its unit normal vector field. Then $f$ and $\nu$ defines a map into the unit tangent bundle, $(f,\nu):U \rightarrow T_1{{\boldsymbol{R}}}^3$. We can consider this map as $(f,\nu):U \rightarrow T_1^{*}{{\boldsymbol{R}}}^3$ by the above identification. Then the condition is equivalent to that $(f,\nu)$ is Legendrian map into $T_{1}^{*}{{\boldsymbol{R}}}^3$. Conversely, from a Legendrian map into $T_{1}^{*}{{\boldsymbol{R}}}^3$, we can make a frontal into ${{\boldsymbol{R}}}^3$ by the projections into ${{\boldsymbol{R}}}^3$ and $S^2$ respectively. A frontal $f:U\rightarrow {{\boldsymbol{R}}}^3$ is called a *front* if $(f,\nu)$ defined as above is an immersion. From above arguments, it can be seen that the notion of fronts and frontals do not depend on the choice of the Riemannian metric $g$ of ${{\boldsymbol{R}}}^3$.
Let $\det$ be the standard volume element of ${{\boldsymbol{R}}}^3$. Then the function $\lambda(u,v):=\det(f_u, f_v, \nu)$ is called the *signed area density function* where $\tilde{nu}$ is the unit normal of $f$. It is obvious from the definition that $p \in U$ is a singular point if and only if $\lambda(p)=0$ holds. A singular point $p \in M^2$ is called *non-degenerate* if the derivative $d{\lambda}$ of the signed area density function does not vanish at $p$.
The typical examples of fronts are *cuspidal edges* and *swallowtails*, while those of frontals are a *cuspidal cross caps*. Here, cuspidal edges, swallowtails and cuspidal cross caps are defined as a smooth map $f:M^2\rightarrow N^3$ which are $\mathcal{A}$-equivalent to the following maps, $f_C$, $f_S$ and $f_{CCR}$ respectively: $$f_C(u,v):=(u^2,u^3,v), f_S(u,v):=(3u^4+u^2v,4u^3+2uv,v),$$ $$f_{CCR}(u,v):=(u,v^2,uv^3).$$
In [@FSUY] and [@KRSUY], a geometric criteria for singular points to be cuspidal edges, swallowtails and cuspidal cross caps are given. To make the criterion, they define *singular curve*, *null direction* and so on. If $p \in U$ is a non-degenerate singular point, then the implicit function theorem implies that the singular set $\Sigma_f$ around $p$ becomes locally a regular curve because $\Sigma_f$ coincides with the zero sets of signed area density function $\lambda$. This curve is called a singular curve and denoted by $\gamma: (-{\varepsilon}, {\varepsilon}) \rightarrow U $. Usually, we fix the parametrization of $\gamma$ as to $\gamma(0)=p$. The *singular direction* of $f$ at $\gamma(t)$ is the 1-dimensional subspace of $T_pU$ which is spanned by $\gamma'(0) \in T_pU$. The null direction is defined as the kernel of the differential map $f_*$. This is defined uniquely only in the case that image of the differential map $f_*$ become 1-dimensional. A vector field $\eta(t)$ along the singular curve $\gamma(t)$ is called a *null vector field* if it associates the non-zero vector belonging to the null direction in $T_{\gamma(t)}U$ to each $t$.
Using the above definitions, criteria for cuspidal edges, swallow tails and cuspidal cross caps are given as follows.
\[fact: CS\] Let $p$ be a non-degenerate singular point of a front $f$, $\gamma$ the singular curve passing through $p$, and $\eta$ a null vector field along $\gamma$. Then
1. $p=\gamma(0)$ is a cuspidal edge if and only if the null direction and the singular direction are transversal, that is, $\det(\gamma'(0), \eta(0)) \not= 0$ holds, where $\det$ denotes the determinant of $2\times 2$ matrices and where we identify the tangent space in $T_pU$ with ${{\boldsymbol{R}}}^2$.
2. $p=\gamma(0)$ is a swallowtail if and only if $$\det(\gamma'(0), \eta(0)) = 0 \ \ \ \ and\ \ \ \ \dfrac{d}{dt}{\Big|}_{t=0}\det(\gamma'(t), \eta(t)) \not= 0$$hold.
\[fact: CCR\] Let $f:U\rightarrow {{\boldsymbol{R}}}^3$ be a frontal and $\gamma(t)$ a singular curve on $U$ passing through a non-degenerate singular point $p=\gamma(0)$. We set $$\Psi(t):=\det(\tilde{\gamma}', D_{\eta}^{f}\nu, \nu),$$ where $\tilde{\gamma}=f\circ{\gamma}$, $D_{\eta}^{f}\nu$ is the canonical covariant derivative along a map $f$ induced from the standard connection on ${{\boldsymbol{R}}}^3$, and $'=d/dt$. Then the germ of $f$ at $p=\gamma(0)$ is $\mathcal{A}$-equivalent to a cuspidal cross cap if and only if
1. $\eta(0)$ is transversal to $\gamma'(0)$,
2. $\Psi(0)=$ and $\Psi'(0)\not=0$.
Representation formula for indefinite improper affine spheres {#sec: rep}
=============================================================
In this section, we introduce a representation formula for indefinite improper affine spheres by modifying Martínez’ method in [@Martinez]. To do this, we should review a duality relation for improper affine spheres with indefinite affine metric. Throughout this paper, we fix its affine normal to $\xi=(0,0,1)$ by the appropriate affine transformation of ${{\boldsymbol{R}}}^3$.
\[prop: duality\] Let $f:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3$ be an indefinite improper affine sphere, $\xi$ its affine normal and ${\nu}:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3$ the conormal map of $f$, where $\Sigma^2$ is an open subset of ${{\boldsymbol{R}}}^2$. Take a coordinate system $(u,v)$ of ${\Sigma}^2$ so that the affine metric $g$ of $f$ (which is indefinite in this case) is represented as $g=E(du^2-dv^2)$. Then, the following identities hold; $$\hspace{0.1cm}
\left\{
\begin{array}{cll}
f_u&= {\nu}\times {\nu}_v \ \ \ {\nu}_u= f_v\times {\xi}\\
f_v&= {\nu}\times {\nu}_u \ \ \ {\nu}_v= f_u\times {\xi}\\
\end{array}
\right.$$ where $\times$ is the outer product on ${{\boldsymbol{R}}}^3$, $f_u:={\dfrac{\partial{f}}{\partial{u}}}$ and $f_v:={\dfrac{\partial{f}}{\partial{v}}}$
The duality relations for affine spheres are well-known for the other kinds of affine spheres. The proof is almost parallel to them. In fact, we can obtain the duality relations for indefinite improper affine spheres by taking care of the definition of conormal maps and the Gauss-Codazzi equations. The case of definite improper affine spheres is presented in [@Matsuura-Urakawa].
Let $\psi=(x,{\varphi}):{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3(={{\boldsymbol{R}}}^2\times{{\boldsymbol{R}}})$ be an indefinite improper affine sphere with the Blaschke normal ${\xi}=(0,0,1)$, where $x:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^2$ and ${\varphi}:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}$ are smooth maps. Then the conormal vector field ${\nu}:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3(={{\boldsymbol{R}}}^2\times{{\boldsymbol{R}}})$ can be represented as ${\nu}=(n,1)$ and the affine metric is written as $g=-\langle dx, dn \rangle $, where $\langle\ ,\ \rangle$ is the standard Euclidean inner product of ${{\boldsymbol{R}}}^2$.
Using Proposition \[prop: duality\], we obtain the following characterization of indefinite improper affine spheres.
\[prop:duality\] Define ${\tilde{L}}_{\psi}:{\Sigma}^2 \rightarrow {\widetilde{{\boldsymbol{C}}}}^2$ as $${\tilde{L}}_f:=x+jn,$$ where $x$ and $n$ are defined as above, that is, $x:=\pi \circ {\psi}$ and $n:=\pi \circ \nu$ with the projection $\pi:{{\boldsymbol{R}}}^3 \rightarrow {{\boldsymbol{R}}}^2$. Identify ${\widetilde{{\boldsymbol{C}}}}^2$ with ${{\boldsymbol{R}}}^4$ as, $${\widetilde{{\boldsymbol{C}}}}^2 \ni z_1=y_0+jy_1, z_2=y_2+jy_3 \longleftrightarrow (y_0,y_1,y_2,y_3) \in {{\boldsymbol{R}}}^4$$ Then $\tilde{L}_{\psi}$ annihilates $2$-forms $dy_0{\wedge}dy_1+dy_2{\wedge}dy_3$ and $dy_0{\wedge}dy_2+dy_1{\wedge}dy_3$, in other words, $$\label{eq:2-forms}
\begin{cases}
dy_0{\wedge}dy_1+dy_2{\wedge}dy_3|_{{\tilde{L}}_{\psi}({\Sigma}^2)} &= 0, \\
dy_0{\wedge}dy_2+dy_1{\wedge}dy_3|_{{\tilde{L}}_{\psi}({\Sigma}^2)} &= 0. \\
\end{cases}$$ hold.
Applying Proposition \[prop:duality\], we have $$\begin{split}
{{\tilde{L}}_{\psi}}^{*}({dy_0{\wedge}dy_1+dy_2{\wedge}dy_3}) &=dx^1{\wedge}dx^2+dn^1{\wedge}dn^2 \\
&=\left\{(x^1_ux^2_v-x^1_vx^2_u)+(n^1_un^2_v-n^1_vn^2_u)\right\}du{\wedge}dv \\
&= 0
\end{split}$$ and $$\begin{split}
{{\tilde{L}}_{\psi}}^{*}({dy_0{\wedge}dy_2+dy_1{\wedge}dy_3}) &=dx^1{\wedge}dn^1+dx^2{\wedge}dn^2 \\
&=\left\{(x^1_un^1_v-x^1_vn^1_u)+(x^2_un^2_v-x^2_vn^2_u)\right\}du{\wedge}dv \\
&= 0
\end{split}$$ where $x=(x^1,x^2)$ and $n=(n^1,n^2)$.
Proposition \[prop:duality\] immediately leads us to the following representation formula.
\[thm:rep\]
1. Let ${\psi}=(x,{\varphi}):{\Sigma}^2 \rightarrow {{\boldsymbol{R}}}^2 \times {{\boldsymbol{R}}}$ be an indefinite improper affine sphere whose affine normal is fixed to ${\xi}=(0,0,1)$. Let ${\nu}=(n,1):{\Sigma}^2 \rightarrow {{\boldsymbol{R}}}^2 \times {{\boldsymbol{R}}}$ be the conormal map of $\psi$. Then there exists a para-holomorphic curve ${\alpha}:=(F,G):{\Sigma}^2 \rightarrow {\widetilde{{\boldsymbol{C}}}}^2$ such that $|dF|\not=|dG|$ and $$\label{eq:F and G}
\begin{cases}
x&= {F}-{\bar{G}}, \\
n&= {\bar{F}}+{G}. \\
\end{cases}$$ hold everywhere.
2. Conversely, given a para-holomorphic curve $(F,G):
{\Sigma}\rightarrow {\widetilde{{\boldsymbol{C}}}}^2$, then $$\psi=\left(x, -\int{{\langle\ n,dx \rangle}}\right)$$ gives an indefinite improper affine sphere (possibly with singularities), where $x$ and $n$ are defined by . Moreover, if we write $F$ and $G$ as $F=f^1+jf^2, G=g^1+jg^2$, then $\psi$ can be expressed as $$\begin{gathered}
\label{eq:explicit}
\psi=\left(
f^1-g^1,
f^2+g^2,
\vphantom{\int}\right.\\
-\int{\left\{(f^1+g^1)(f^1_u-g^1_u)+(-f^2+g^2)(f^2_u+g^2_u)\right\}du-}
\\
\left.
\int{\left\{(f^1+g^1)(f^2_u-g^2_u)+(-f^2+g^2)(f^1_u+g^1_u)\right\}dv}
\right)
\end{gathered}$$
<!-- -->
1. First, define $F:{\Sigma} \rightarrow {\widetilde{{\boldsymbol{C}}}}$ as $$F:=f^1+jf^2:={\frac{x^1+n^1}{2}}+j{\frac{x^2-n^2}{2}}.$$ Then, by Proposition \[prop:duality\], we have $$f^1_u=f^2_v, \ \ \ f^1_v=f^2_u$$ and $F$ is a para-holomorphic function on ${\Sigma}^2$ with respect to the para-complex structure induced from $g$.\
Take a new coordinate system $\left\{w_1, w_2 \right\}$ on ${\widetilde{{\boldsymbol{C}}}}^2$ as, $$\begin{cases}
w_1 &={\dfrac{y_0+y_1}{2}}+j{\dfrac{y_2-y_3}{2}}, \\
w_2 &={\dfrac{-y_0+y_1}{2}}+j{\dfrac{y_2+y_3}{2}}. \\
\end{cases}$$ Then, is equivalent to $$\label{eq:2-form}
dw_1{\wedge}dw_2|_{{\tilde{L}}_{\psi}({\Sigma})}=0.$$ Here $w_1=F$ and $w_1$ induce the same para-complex structure on ${\Sigma}$ as the one induced from $g$. Thus implies that $w_2$ also defines a para-holomorphic function on ${\Sigma}$. Define a para-holomorphic function $G:{\Sigma} \rightarrow {\tilde{{\boldsymbol{C}}}}$ as $G:=w_2$, then we have . The condition that $|dF|\not=|dG|$ is equivalent to the absence of singularity on $\psi$.
2. We can check that $\psi$ defines an indefinite improper affine sphere by direct calculation. However, it is almost obvious that $\psi$ defines an indefinite improper affine sphere by the following reason. For $\psi:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3$, a duality relation is equivalent to the condition to be an affine sphere. On the other hand, a duality relation is equivalent to para-Cauchy-Riemann equation in this case.
Next, we introduce the notion of IA-maps and generalized IA-maps, which is a generalization of improper affine sphere. IA-maps are firstly defined by A. Martínez in [@Martinez], as a generalization of improper affine sphere which permits a singularity. Modifying the Martínez’ definition of IA-maps, we define generalized IA-maps as the following.
\[def: Martinez\] A map $\psi=(x,{\varphi}):{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3(={{\boldsymbol{R}}}^2\times{{\boldsymbol{R}}})$, where $x:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^2$ and ${\varphi}:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}$ are smooth maps, is called an *IA-map* (respectively *a generalized IA-map*) if ${\tilde{L}}_{\psi}:{\Sigma}^2 \rightarrow {{\boldsymbol{C}}}^2$, which is defined by $${\tilde{L}}_{\psi}:=x+\sqrt{-1}n,$$ becomes a SL-immersion (resp. SL-map) with respect to the symplectic structure ${\omega}'$ and the calibration ${\operatorname{Re}}(\sqrt{-1}{\Omega}')$, where ${\omega}'=\dfrac{\sqrt{-1}}{2}(d{\zeta}_1 {\wedge} d{\bar{{\zeta}_1}}+d{\zeta}_2 {\wedge} d{\bar{{\zeta}_2}})$ and ${\Omega}'=d{\zeta}_1 {\wedge} d{\zeta}_2$.
As an analogy to this definition, it is appropriate to define the class of indefinite improper affine spheres with singularities as follows.
\[def: indefinite\] A map $\psi=(x,{\varphi}):{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3(={{\boldsymbol{R}}}^2\times{{\boldsymbol{R}}})$, where $x:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^2$ and ${\varphi}:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}$ are smooth maps, is called an *indefinite IA-map* (respectively, an *indefinite generalized IA-map*) if ${\tilde{L}}_{\psi}:{\Sigma}^2 \rightarrow {\widetilde{{\boldsymbol{C}}}}^2$, which is defined by $${\tilde{L}}_{\psi}:=x+jn,$$ becomes an immersion (resp. a map which is not necessarily an immersion) and ${\tilde{L}}_{\psi}^{*}(dy_0{\wedge}dy_1+dy_2{\wedge}dy_3)
={\tilde{L}}_{\psi}^{*}(dy_0{\wedge}dy_2+dy_1{\wedge}dy_3)=0$ holds.
From now on, to avoid the confusion, we call an IA-map (respectively, a generalized IA-map) with positive definite affine metric an *locally strongly convex IA-maps* (resp. an *locally strongly convex generalized IA-map*).
Singularities of IA-maps
========================
The representation formula for locally strongly convex improper affine spheres and that for indefinite ones look very similar. However, singular points which appear in two representation formulae have quite different properties each other. Concretely, in the locally strongly convex case, except for on branch points, all the singular points are fronts. On the other hand, in the indefinite case, a singular point is not necessarily a front even if it is not a branch point. Indeed, we can easily find examples of frontal maps which are not fronts in indefinite case by direct calculations.
Locally strongly convex case
----------------------------
To study the properties of singularities on a locally strongly convex IA-map ${\psi}:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3$, we will first check the condition for $p \in \Sigma^2$ to be a singular point of $\psi$, which is also given in [@Martinez].
\[prop: sing. pt. on l.s.c.\] Let $\psi:{\Sigma} \rightarrow {{\boldsymbol{R}}}^{3}$ be an locally strongly convex generalized IA-map¡¥ From the representation formula for locally strongly convex IA-maps in Theorem \[thm:lscIA-map\], $p\in{\Sigma}^2$ is a singular point of $\psi$ if and only if $|dF|=|dG|$ holds on $p\in{\Sigma}^2$.
From the explicit form of the representation formula , the differentiation of $\psi$ becomes $$\begin{split}
{\psi}_u &= (f^1_u+g^1_u, g^2_u-f^2_u, (g^1-f^1)(g^1_u+f^1_u)+(g^2+f^2)(g^2_u-f^2_u)), \\
{\psi}_v &= (-f^2_u-g^2_u, g^1_u-f^1_u, (-g^1+f^1)(g^2_u+f^2_u)+(g^2+f^2)(g^1_u-f^1_u)). \\
\end{split}$$ This can be rewritten as $$\label{eq: indefinite}
\begin{split}
{\psi}_u &= (f^1_u+g^1_u)(1,0,g^1-f^1)+(g^2_u-f^2_u)(0,1,g^2+f^2) \\
{\psi}_v &= -(g^2_u+f^2_u)(1,0,g^1-f^1)+(g^1_u-f^1_u)(0,1,g^2+f^2). \\
\end{split}$$ The linearly independence of the above two vectors implies that the condition to be a singular points is $$(f^1_u+g^1_u)(g^1_u-f^1_u)-(g^2_u-f^2_u)\left\{-(g^2_u+f^2_u)\right\}=0 .$$ This is equivalent to $|dF|=|dG|$.
Next we show that the singularity of locally strongly convex improper affine spheres is always locally a front unless it is a branch point, that is, $\psi_u=\psi_v=0$ holds on that point. In other words, Legendrian lift $L_{\psi}=(\psi,{\tilde{\nu}}): {\Sigma}^2
\rightarrow {T_1{{\boldsymbol{R}}}^3}$ is immersive on ${\Sigma}^2$ (even on the singular point of $\psi$), where $\tilde{\nu}$ is the unit normal of $\psi$.
\[prop:front\] Singularity of a locally strongly convex generalized IA-map is always locally a front if $|dF|=|dG|=0$ does not hold.
The proof follows from the direct calculation. We check that ${{\operatorname{Ker}}}d{\psi} \not= {{\operatorname{Ker}}}{d{\tilde{\nu}}}$ holds at singular points of $\psi$.
From Theorem \[thm:lscIA-map\], the explicit representation of ${\tilde{\nu}}$ is $$\tilde{\nu} = \frac{1}{(1+(f^1-g^1)^2+(f^2+g^2)^2)^{\frac{1}{2}}}(f^1-g^1, -(f^2+g^2), 1),$$and its differentiations are $$\begin{aligned}
{2}
{\tilde{\nu}}_u &=
\frac{1}{{\sqrt{1+(f^1-g^1)^2+(f^2+g^2)^2}}^3}\cdot\\
&
\hspace{1in}
\begin{aligned}
\biggl(
&(f^1_u-g^1_u)(1+(f^2+g^2)^2)-(f^2+g^2)(f^1-g^1)(f^2_u+g^2_u), \\
&-(f^2_u+g^2_u)(1+(f^1-g^1)^2)+(f^2+g^2)(f^1-g^1)(f^1_u-g^1_u), \\
&-(f^1-g^1)(f^1_u-g^1_u)-(f^2+g^2)(f^2_u+g^2_u)
\biggr),
\end{aligned} \\
{\tilde{\nu}}_v &=
\frac{1}{{\sqrt{1+(f^1-g^1)^2+(f^2+g^2)^2}}^3}\cdot\\
&
\hspace{1in}
\begin{aligned}
\biggl(
&-(f^2_u-g^2_u)(1+(f^2+g^2)^2)-(f^2+g^2)(f^1-g^1)(f^1_u+g^1_u), \\
&-(f^1_u+g^1_u)(1+(f^1-g^1)^2)-(f^2+g^2)(f^1-g^1)(f^2_u-g^2_u), \\
&(f^1-g^1)(f^2_u-g^2_u)-(f^2+g^2)(f^1_u+g^1_u)
\biggr).
\end{aligned} \end{aligned}$$ The proof of Proposition \[prop: sing. pt. on l.s.c.\] implies that $$(g^2_u+f^2_u){\psi}_u+(g^1_u+f^1_u){\psi}_v = 0$$ holds at a singular point of $f$. Then, to prove ${L_{\psi}}$ is an immersion, we have only to show that $$(g^2_u+f^2_u){\tilde{\nu}}_u+(g^1_u+f^1_u){\tilde{\nu}}_v \not= 0$$ holds at any singular point.\
Here, $$\begin{split}
&{{\sqrt{1+(f^1-g^1)^2+(f^2-g^2)^2}}^3}\left\{(g^2_u+f^2_u){\tilde{nu}}_u+(g^1_u+f^1_u){\tilde{\nu}}_v\right\} \\
=(&-(g^1-f^1)(g^2+f^2)((g^1_u+f^1_u)^2+(g^2_u+f^2_u)^2)+2(1+(f^2+g^2)^2)(f^1_ug^2_u-f^2_ug^1_u), \\
& -(1+(f^1-g^1)^2)((g^1_u+f^1_u)^2+(g^2_u+f^2_u)^2)+2(f^1-g^1)(f^2+g^2)(f^1_ug^2_u-f^2_ug^1_u), \\
& -(f^2+g^2)((f^1_u+g^1_u)^2+(g^2_u+f^2_u)^2)-2(f^1-g^1)(f^1_ug^2_u-f^2_ug^1_u) ) \\
=(&-(g^1-f^1)(g^2+f^2)A+2(1+(f^2+g^2)^2)B, -(1+(f^1-g^1)^2)A+2(f^1-g^1)(f^2+g^2)B, \\
& -(f^2+g^2)A-2(f^1-g^1)B ). \\
\end{split}$$ where, $A=(f^1_u+g^1_u)^2+(g^2_u+f^2_u)^2$ and $B=(f^1_ug^2_u-f^2_ug^1_u)$. Here $A>0$ because $|dF|=|dG|=0$ does not hold. If$${{\sqrt{1+(f^1-g^1)^2+(f^2-g^2)^2}}^3}\left\{(g^2_u+f^2_u){\tilde{\nu}}_u+(g^1_u+f^1_u){\tilde{\nu}}_v\right\}=0$$ holds, then the computation of $B/A$ by using the fact that each element becomes 0 gives $1+(f^1-g^1)^2+(f^2+g^2)^2=0$. However, this is contradiction. So $$(g^2_u+f^2_u){\nu}_u+(g^1_u+f^1_u){\nu}_v \not= 0$$ holds at any point (even at a singular point)¡¥\
This completes the proof of Proposition \[prop:front\].
\[rmk: branch\] We can state the relation between $\psi$, $\tilde{L}_{\psi}$ and $L_{\psi}$ as follows. $L_{\psi}$ is an immersion if and only if ${\tilde{L}}_{\psi}$ is an immersion. So a locally strongly convex generalized IA-map $\psi$ is a front if and only if ${\tilde{L}}_{\psi}$ is an immersion.
On the other hand, if ${\tilde{L}}_{\psi}$ is not immersive, then ${\psi}_u={\psi}_v=0$ holds at that point. On such a point $|dF|=|dG|=0$ holds. We call such a point a branch point. Note that branch points are isolated because $F$ and $G$ are both holomorphic functions. Above fact can be considered as another proof of Proposition\[prop:front\].
Indefinite case
---------------
So far in this section, we consider the singularity of locally strongly convex IA-maps. Unlike the locally strongly convex case, any indefinite improper affine sphere (with singularities) is not necessarily a front. Indeed, we can construct several concrete examples which are frontal maps, but not fronts. So we shall check the property of singularity of indefinite improper affine spheres in the rest of this section.
First, similar to the locally strongly convex case, we check the condition for a pair of para-holomorphic functions so that the corresponding indefinite improper affine sphere has a singular point.
\[prop: sing. pt. on indefinite\] Let $\psi:{\Sigma}^2 \rightarrow {{\boldsymbol{R}}}^{3}$ be an indefinite generalized IA-map¡¥ Then, $p\in{\Sigma}^2$ is a singular point of $\psi$ if and only if $|dF|=|dG|$ holds on $p\in{\Sigma}^2$.
The proof is almost the same as that of Theorem\[prop: sing. pt. on l.s.c.\] except for the detailed calculation.
From the explicit form of the representation formula , the differentiation of $\psi$ becomes $$\begin{split}
{\psi}_u &= (f^1_u-g^1_u, f^2_u+g^2_u, -(f^1+g^1)(f^1_u-g^1_u)-(-f^2+g^2)(f^2_u+g^2_u)), \\
{\psi}_v &= (f^2_u-g^2_u, f^1_u+g^1_u, -(f^1+g^1)(f^2_u-g^2_u)-(-f^2+g^2)(f^1_u+g^1_u)). \\
\end{split}$$ This can be rewritten as $$\label{eq:frontal}
\begin{split}
{\psi}_u &= (f^1_u-g^1_u)(1,0,-(f^1+g^1))+(f^2_u+g^2_u)(0,1,f^2-g^2) \\
{\psi}_v &= (f^2_u-g^2_u)(1,0,-(f^1+g^1))+(f^1_u+g^1_u)(0,1,f^2-g^2). \\
\end{split}$$ The linearly independence of the above two vectors implies that the condition to be a singular points is $$(f^1_u-g^1_u)(f^1_u+g^1_u)-(f^2_u+g^2_u)(f^2_u-g^2_u)=0 .$$ This is equivalent to $|dF|=|dG|$.
As mentioned above, there exist frontal maps which are not fronts in the representation formula . Next, we shall check the condition for an indefinite improper affine sphere to have this type of singularity.
\[prop:not front\] For an indefinite generalized IA-map $\psi:{\Sigma}^2 \rightarrow {{\boldsymbol{R}}}^{3}$, the following two conditions are equivalent:
1. $\psi$ is a frontal map which is not a front at $p \in {\Sigma}^2$ and $p$ is not a branch point.
2. One of the following two conditions are satisfied at $p\in{\Sigma}^2$.
1. $f^1_u=f^2_u$ and $g^1_u=g^2_u$,
2. $f^1_u=-f^2_u$ and $g^1_u=-g^2_u$.
From , we already know that at any point, the tangent space of any indefinite generalized IA-map is spanned by two vectors $(1,0,-(f^1+g^1))$ and $(0,1,f^2-g^2)$. This means every infinite generalized IA-map has an unit normal, which is parallel to $(f^1+g^1, -f^2+g^2, 1)$. So every indefinite generalized IA-map is a frontal map. Therefore the condition (1) is satisfied if and only if $\psi$ is not a front, i.e., $L_{\psi}$ is not a immersion. Thus we have only to check the condition for $L_{\psi}$ not to be an immersion.
From Theorem\[thm:rep\], the unit normal ${\tilde{\nu}}$ of $\psi$ is $${\tilde{\nu}} = \frac{1}{\sqrt{\Delta}}(f^1+g^1, -f^2+g^2, 1)$$where $\Delta=(f^1)^2+(f^2)^2+(g^1)^2+(g^2)^2+2f^1g^1-2f^2g^2+1$.\
Then, we have $$\label{eq:unit normal}
\begin{split}
{\tilde{\nu}}_u &= -{\frac{{\Delta}_u}{2{\sqrt{\Delta}}^3}}(f^1+g^2, -f^2+g^2, 1) + {\frac{1}{\sqrt{\Delta}}}(f^1_u+g^1_u, -f^2_u+g^2_u, 0) \\
&= \frac{1}{2{\sqrt{\Delta}}^3} \left\{-\Delta_u(f^1+g^1, -f^2+g^2,1) + 2\Delta(f^1_u+g^1_u, -f^2_u+g^2_u, 0) \right\} \\
&= \frac{1}{2{\sqrt{\Delta}}^3} (-2(f^2_u-g^2_u)(f^2-g^2)(f^1+g^1)+2(-f^1_u+g^1_u)(f^2-g^2)^2+2(f^1_u+g^1_u), \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ 2(f^1_u+g^1_u)(f^1+g^1)(f^2-g^2)+2(-f^2_u+g^2_u)(f^1+g^1)^2+2(-f^2_u+g^2_u), \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ -2(f^1_u+g^1_u)(f^1+g^1)-2(f^2_u-g^2_u)(f^2-g^2) ), \\
{\tilde{\nu}}_v &= -{\frac{\Delta_v}{2{\sqrt{\Delta}}^3}}(f^1+g^2, -f^2+g^2, 1) + {\frac{1}{\sqrt{\Delta}}}(f^2_u+g^2_u, -f^1_u+g^1_u, 0) \\
&= \frac{1}{2{\sqrt{\Delta}}^3} \left\{ -\Delta_v(f^1+g^1, -f^2+g^2,1) + 2\Delta(f^2_u+g^2_u, -f^1_u+g^1_u, 0) \right\} \\
&= \frac{1}{2{\sqrt{\Delta}}^3} (-2(f^1_u-g^1_u)(f^1+g^1)(f^2-g^2)+2(f^2_u+g^2_u)(f^2-g^2)^2+2(f^2_u+g^2_u), \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ 2(f^2_u+g^2_u)(f^1+g^1)(f^2-g^2)+2(-f^1_u+g^1_u)(f^1+g^1)^2+2(-f^1_u+g^1_u), \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ -2(f^2_u+g^2_u)(f^1+g^1)-2(f^1_u-g^1_u)(f^2-g^2) ). \\
\end{split}$$ First, we consider the case that neither $f^1_u+g^1_u$, $f^1_u-g^1_u$, $f^2_u+g^2_u$ nor $f^2_u-g^2_u$ does not equal to $0$. Since $|dF|=|dG|$ holds on a singular point of $\psi$, implies that $$\label{eq:f}
{\psi}_u=\frac{f^1_u-g^1_u}{f^2_u-g^2_u}{\psi}_v= \frac{f^2_u+g^2_u}{f^1_u+g^1_u}{\psi}_v$$also holds on a singular point of $\psi$. On the other hand, by the same reason, implies that $$\label{eq:nu}
{\tilde{\nu}}_u = \frac{f^2_u-g^2_u}{f^1_u-g^1_u}{\tilde{\nu}}_v= \frac{f^1_u+g^1_u}{f^2_u+g^2_u}{\tilde{\nu}}_v$$ holds on a singular point of $\psi$.
By comparing and , and taking care of the fact that $\psi$ is not a front if and only if $(L_{\psi})_u$ and $(L_{\psi})_v$ are parallel, we have condition (2) after elementary calculation.
In the case that either $f^1_u+g^1_u$, $f^1_u-g^1_u$, $f^2_u+g^2_u$ or $f^2_u-g^2_u$ equals $0$, we have condition (2) by taking care of the fact that $\psi$ is not a front at $p$ and that $p$ is not a branch point.
In the same manner as Remark \[rmk: branch\], we have the following relation between $\psi$, $\tilde{L_{\psi}}$ and $L_{\psi}$ for an indefinite generalized IA-map.
$L_{\psi}$ is an immersion if and only if ${\tilde{L}}_{\psi}$ is an immersion. So an indefinite generalized IA-map $\psi$ is a front if and only if ${\tilde{L}}_{\psi}$ is an immersion.
On the other hand, if $\tilde{L_{\psi}}$ is not an immersion, then $|dF|=|dG|=0$ hold and it follows that $\psi$ is a frontal which is not a front or $p$ is a branch point.
Now, we end this section by remarking that cuspidal cross caps never appear as the singularities of indefinite generalized IA-maps. This fact is proved by using the criterion for cuspidal cross caps (Fact \[fact: CCR\]) and the condition for indefinite generalized IA-map to be a frontal but not a front (Proposition \[prop:not front\]).
\[thm: not CCR\] Let $\psi:{\Sigma}^2\rightarrow {{\boldsymbol{R}}}^3$ be an indefinite generalized IA-map and $p\in{\Sigma}^2$ be a point. Then the germ of $\psi$ at $p$ is not $\mathcal{A}$-equivalent to the cuspidal cross cap.
From Fact \[fact: CCR\], it is sufficient to check that ${\Psi}'(0) = 0$. Here, we have only to consider the case of $f^1_u=f^2_u \ and\ g^1_u=g^2_u$ because the proof of another case is parallel to it. Since ${\gamma}'(0)=(1,1)$, $\eta(0)=(1,-1)$ and ${\nu}_{uu}(p)={\nu}_{vv}(p)$, we have ${\Psi}'(0) = 0$.
\[rmk: IM\] Theorem\[thm: not CCR\] is proved in [@Ishikawa-Machida] for another setting. In [@Ishikawa-Machida], Ishikawa and Machida considered singularities on improper affine spheres given by projection derived from the framework of differential systems. They also have the same results as Theorem \[thm: not CCR\]. In Section \[sec:comparison\], we will discuss the relationship between the singularities on generalized indefinite IA-maps and those which appear in [@Ishikawa-Machida].
Comparison with other representation formulae
=============================================
In addition to Martínez type representation formulae, there are several representation formulae for improper affine spheres. In this section, we compare the representation formulae for (mainly indefinite) improper affine spheres which are obtained in the previous sections, to the other representation formulae for improper affine spheres which are studied in [@Blaschke], [@Cortes], [@Cortes-Lawn-Schafer] and [@Ishikawa-Machida]. The idea that relates the Martínez’ representation formula to the Cortés’ representation formula is originally due to T. Kurose ([@Kurose; @lecture]).
At first, the relation between Martínez type representation formula and that obtained by Cortés-Lawn-Schäfer is as follows.
For 2-dimensional case, we can derive Cortés-Lawn-Schäfer representation formula in [@Cortes-Lawn-Schafer] from Theorem \[thm:rep\] by taking $ F=\frac{1}{2}(z-jf'(z)), G=\frac{1}{2}(z+jf'(z)) $.
This fact is the analogy for the case of locally strongly convex ones as below (Remark \[rmk: lsc\]).
The following two remarks are pointed out by T. Kurose.
\[rmk: lsc\] For 2-dimensional case, we can derive Cortés’ representation formula in [@Cortes] from Martínez’ representation formula in [@Martinez] by taking $ F=\frac{1}{2}(z-\sqrt{-1}f'(z)), G=\frac{1}{2}(z+\sqrt{-1}f'(z)) $.
By the identification of a para-holomorphic function to a pair of smooth functions explained in Section \[sec: prel\], we can derive Blaschke’s representation formula from the Martínez type representation formula in Theorem \[thm:rep\].
Here, two para-holomorphic functions $F,G$ can be written as $$\begin{cases}
F(u,v)&={\rho}_1(u+v)+{\sigma}_1(u-v)+j\left\{{\rho}_1(u+v)-{\sigma}_1(u-v)\right\}, \\
G(u,v)&={\rho}_2(u+v)+{\sigma}_2(u-v)+j\left\{{\rho}_2(u+v)-{\sigma}_2(u-v)\right\} \\
\end{cases}$$ by four smooth function ${\rho}_1, {\rho}_2, {\sigma}_1, {\sigma}_2$. Therefore if we take $$\begin{cases}
U_1&= {\rho}_1+{\rho}_2, \ \ V_1 = {\sigma}_1+{\sigma}_2, \\
U_2&= -{\rho}_1+{\rho}_2, \ \ V_2={\sigma}_1-{\sigma}_2, \\
\end{cases}$$ then we have the Blaschke’s representation formula in [@Blaschke].
Examples {#sec:ex}
========
In [@Martinez], Martínez gave several examples of locally strongly convex improper affine spheres by substituting holomorphic functions to his representation formula. Here, we construct indefinite improper affine spheres in the same manner as [@Martinez].
\[ex: 23\] Let $(F,G)=( z^2, z^3)$, then Theorem \[thm:rep\] gives a concrete example of indefinite generalized IA-maps, which is a frontal map but not a front.
For this example, we have $$\begin{cases}
f^1&=u^2+v^2, \ \ f^2=2uv, \\
g^1&=u^3+3uv^2, \ \ g^2=3u^2v+v^3, \\
\end{cases}$$and this implies $${\lambda}=(f^1_u)^2-(f^2_u)^2-(g^1_u)^2+(g^2_u)^2 =(u^2-v^2)(4-9(u^2-v^2))$$and $$\begin{cases}
{\lambda}_u&=4u(2-9u(u^2-v^2)), \\
{\lambda}_v&=-4v(2-9v(u^2-v^2)). \\
\end{cases}$$ Therefore, singular locus is $\left\{(u,v)\in{{\boldsymbol{R}}}^2|(u^2-v^2)(4-9(u^2-v^2))=0\right\}$. Among them, only the origin is degenerate singular point, and $\psi$ is a frontal but not a front on $\left\{ u^2-v^2=0 \right\} $ (they are all corank 1 map-germs except for on the origin.) while $\psi$ is a front on $C:=\left\{ 4-9(u^2-v^2)=0 \right\}$ (Figure 1).
By using the criteria for cuspidal edges and swallowtails, we can completely classify the singularities on $C$. In fact, at any point on $C$ except for the point $(-\frac{2}{3},0)$, singularities are $\mathcal{A}$-equivalent to the cuspidal edges while swallowtail appears at $(-\frac{2}{3},0)$. This assertion is verified as below.
Let $C_1:=C \cap \left\{u \geq 0 \right\}$ and $C_2:=C \cap \left\{u \leq 0 \right\}$. First, we show that the singularity is $\mathcal{A}$-equivalent to the cuspidal edge at any point on $C_1$. $C_1$ is parametrized as $\gamma(t):= \frac{2}{3}(\cosh t, \sinh t)$. Since $$\begin{cases}
(f^2+g^2)_u(\gamma(t)) &={\frac{4}{3}}{\sinh t}(1+2{\cosh t}), \\
(f^2+g^2)_v(\gamma(t)) &={\frac{4}{3}}({\cosh t}+1)(2{\cosh t}-1), \\
\end{cases}$$ the null vector field at $\gamma(t)$ is parallel to $(-({\cosh t}+1)(2{\cosh t}-1), {\sinh t}(1+2{\cosh t}))$. Therefore $$\det(\gamma'(t), \eta(t)) =- ({\cosh t}+1)$$ and it follows that the null vector field and the singular direction are transversal on $C_1$. $C_2$ is parametrized as $\gamma(t):= -\frac{2}{3}(\cosh t, \sinh t)$ and $$\begin{cases}
(f^1-g^1)_u(\gamma(t)) &= -{\frac{4}{3}}({\cosh t}+1)(2{\cosh t}-1), \\
(f^1-g^1)_v(\gamma(t)) &= -{\frac{4}{3}}{\sinh t}(1+2{\cosh t}), \\
\end{cases}$$ hold. Thus the null vector field at $\gamma(t)$ is parallel to $(-{\sinh t}(1+2{\cosh t}), ({\cosh t}+1)(2{\cosh t}-1))$. Therefore $$\det(\gamma'(t), \eta(t)) = -{\sinh t}$$ and it follows that the null vector field and the singular direction are transversal on $C_2$ except for on the point $(-\frac{2}{3}, 0)$ and that $\det{\big |}_{t=0}(\gamma'(t), \eta(t)) \not= 0$ on $(-\frac{2}{3}, 0)$.
\[ex: 34\] Let $(F,G)=( z^3, z^4)$, then Theorem \[thm:rep\] also gives a concrete example of indefinite generalized IA-maps, which is a frontal map but not a front.
For this example, we have $$\begin{cases}
f^1&=u^3+3uv^2, \ \ f^2=3u^2v+v^3, \\
g^1&=u^4+6u^2v^2+v^4, \ \ g^2=4u^3v+4uv^3, \\
\end{cases}$$ and this implies $${\lambda}=(f^1_u)^2-(f^2_u)^2-(g^1_u)^2+(g^2_u)^2 =(u^2-v^2)^2(9-16(u^2-v^2))$$and $$\begin{cases}
{\lambda}_u&=12u(u^2-v^2)(3-8(u^2-v^2)), \\
{\lambda}_v&=-12v(u^2-v^2)(3-8v(u^2-v^2)). \\
\end{cases}$$ Therefore, singular locus is $\left\{(u,v)\in{{\boldsymbol{R}}}^2|(u^2-v^2)^2(9-16(u^2-v^2))=0 \right\}$. Among them, $\left\{u^2-v^2=0 \right\}$ is the set of degenerate singular point. Moreover, $\psi$ is a frontal but not front on $\left\{ u^2-v^2=0 \right\}$ (they are all corank 1 map-germs except for on the origin.) while $\psi$ is a front on $\widetilde{C}:=\left\{ 9-16(u^2-v^2)=0 \right\} $. Above calculation shows that (1,1) is the point where the following three conditions are satisfied, that is,
1. \[item:1\] $(1,1)$ is a degenerate singular point.
2. \[item:2\] $\psi$ is a frontal but not a front of corank 1 on $(1,1)$.
3. \[item:3\] the set of degenerate singular point around $(1,1)$ is locally a smooth curve.
It is worth mentioning that any map-germs satisfying above conditions which appear in Ishikawa-Machida’s formulation is not $\mathcal{A}$-equivalent to this example (Figure 2). We prove this fact in Section \[sec:comparison\].
As similar to the Example \[ex: 23\], the singularities on $\widetilde{C}$ is completely classified by the same way. At any point on $\widetilde{C}$ except for the point $(-\frac{2}{3},0)$, singularity is $\mathcal{A}$-equivalent to the cuspidal edge while swallowtail appears at $(-\frac{2}{3},0)$.
------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------
![Example \[ex: 23\] and Example \[ex: 34\].[]{data-label="fig:lip-and-nondeg"}](z2z3.eps "fig:"){width="5.5cm"} ![Example \[ex: 23\] and Example \[ex: 34\].[]{data-label="fig:lip-and-nondeg"}](z3z4.eps "fig:"){width="5.5cm"}
Example \[ex: 23\] [Example \[ex: 34\]]{}
------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------
Comparison with formulation by Ishikawa-Machida {#sec:comparison}
===============================================
As mentioned in Remark \[rmk: IM\], Ishikawa and Machida also studied improper affine spheres with singularities in another setting. They considered (generalized) geometric solutions of a certain Monge-Ampère system $\mathcal{M}$ on ${{\boldsymbol{R}}}^5$ and studied the singularities of the projections of such (generalized) geometric solutions. Here, the projection of a (generalized) geometric solution of $\mathcal{M}$ is nothing but an improper affine sphere outside singular points. The singularities of indefinite generalized IA-maps and those of the projection of generalized geometric solutions of $\mathcal{M}$ share some same properties as mentioned in Remark \[rmk: IM\]. However, there are also different properties between singularities which appear in both formulations. Indeed, we can find the singularity on indefinite generalized IA-maps which does not appear on the generalized geometric solution of $\mathcal{M}$.
Before proving this, we should review the Ishikawa-Machida’s formulation briefly.
In [@Ishikawa-Machida], they studied the singularities of graphs $z=f(x,y)$ in $xyz$-space ${{\boldsymbol{R}}}^3$ where $f$ is a solutions of the Monge-Ampère type equation $f_{xx}f_{yy}-f_{xy}^2=c$. Here, the equation $f_{xx}f_{yy}-f_{xy}^2=c$ can be considered geometrically in terms of the differential system $\mathcal{M}$ on $xyzpq$-space ${{\boldsymbol{R}}}^5$, which is generated by $$\label{eq: differential system}
\omega= cdx{\wedge}dy-dp{\wedge}dq, \ \ \theta=dz-pdx-qdy,$$ where $p,q$ represent $z_x=f_x, z_y=f_y$ respectively. Here, $D=\left\{ \theta=0 \right\}$ is the contact structure on $T{{\boldsymbol{R}}}^5$. For the graph $z=f(x,y)$ of a solution of the Monge-Ampère type equation $f_{xx}f_{yy}-f_{xy}^2=c$ in $xyz$-space ${{\boldsymbol{R}}}^3$, we define its lift $L_f:{{\boldsymbol{R}}}^2 \rightarrow {{\boldsymbol{R}}}^5$ to ${{\boldsymbol{R}}}^5$ as $L_f(x,y,z):=(x,y,f(x,y),f_x(x,y),f_y(x,y))$. Obviously, the lift of the graph of $f_{xx}f_{yy}-f_{xy}^2=c$ annihilates both $\omega$ and $\theta$, that is, it is a Legendrian immersion into ${{\boldsymbol{R}}}^5$ which annihilates $\omega$. Taking this into consideration, a *geometric solution* (respectively, *generalized geometric solution*) of $\mathcal{M}$ is defined as a Legendrian immersion with respect to the contact structure $D$ (respectively, a map annihilating $\omega$ which is not necessarily an immersion) of ${{\boldsymbol{R}}}^2$ into ${{\boldsymbol{R}}}^5$, which also annihilates $\theta$.
Among the singularities of improper affine spheres $\psi=(x,y,z):{{\boldsymbol{R}}}^2 \rightarrow {{\boldsymbol{R}}}^3$ which appear as the projection of (generalized) geometric solution $f=(x,y,z,p,q):{{\boldsymbol{R}}}^2 \rightarrow {{\boldsymbol{R}}}^5$ of the Monge-Ampère system $\mathcal{M}$, we are especially interested in singularities of corank 1. In this case, by implicit function theorem, we can take a coordinate system $(u,v)$ of ${{\boldsymbol{R}}}^2$ around $(0.0)$ such that $\psi(u,v)=(u,y(u,v),z(u,v))$. Since $f^{*}\theta =0$, we have $$\label{eq:theta}
z_u=p+qy_u \ \ \ \textrm{and} \ \ \ z_v=qy_v.$$ So the unit normal $\tilde{\nu}$ of $f$ becomes $\tilde{\nu}=\dfrac{1}{\sqrt{p^2+q^2+1}}(-p,-q,1)$ because $f_u=(1,y_u,z_u)$ and $f_v=(0,y_v,z_v)$. Hence the signed area density of $f$ is $\lambda=(1+q^2)y_v$ and we can conclude that a point $(u_0,v_0)$ is a singular point if $y_v(u_0,v_0)=0$ and that a singular point $(u_0,v_0)$ is degenerate (respectively, non-degenerate) if $y_{uv}(u_0,v_0)=y_{vv}(u_0,v_0)=0$ (resp. neither $y_{uv}(u_0,v_0)\not=0$ nor $y_{vv}(u_0,v_0)\not=0$ holds).
Now, based on the above review, we can prove the following proposition on singularities on improper affine spheres.
\[prop:comparison\] The germ of the map of Example \[ex: 34\] at $(1,1)\in{{\boldsymbol{R}}}^2$ is not $\mathcal{A}$-equivalent to any germ of generalized geometric solution of $\mathcal{M}$.
The proof is based on the following well-known facts about singularities.
\[fact:Mather\] Let $F$ be a smooth real-valued function defined on a neighborhood of $0$ in ${{\boldsymbol{R}}}\times {{\boldsymbol{R}}}$ such that $F(0,t)=g(t)t^k$ where $g(0)\not=0$ and $g$ is smooth on some neighborhood of $0$ in ${{\boldsymbol{R}}}$. Then given any smooth real-valued function $G$ defined on a neighborhood of $0$ in ${{\boldsymbol{R}}}\times{{\boldsymbol{R}}}$, there exist smooth functions $q$ and $r$ such that
1. $G=qF+r$ on a neighborhood of $0$ in ${{\boldsymbol{R}}}\times{{\boldsymbol{R}}}$, and
2. $r(x,t)= \displaystyle \sum_{i=0}^{k-1} r_i(x)t^i$ for $(t,x)\in{{\boldsymbol{R}}}\times{{\boldsymbol{R}}}$ near $0$.
Next, we denote by $\mathcal{E}_2$ the set of smooth functions on ${{\boldsymbol{R}}}^2$. For a smooth map-germ $f=(f^1, f^2, f^3):({{\boldsymbol{R}}}^2,0) \rightarrow ({{\boldsymbol{R}}}^3,0)$, we denote by $I(f)$ the ideal of $\mathcal{E}_2$ generated by $f^1, f^2$ and $f^3$ and define $Q(f):=\mathcal{E}_2/I(f)$.
Let $f,g:({{\boldsymbol{R}}}^2,0) \rightarrow ({{\boldsymbol{R}}}^3,0)$ be two map-germs from ${{\boldsymbol{R}}}^2$ to ${{\boldsymbol{R}}}^3$. If $f$ and $g$ are $\mathcal{A}$-equivalent each other, then $Q(f)$ and $Q(g)$ are isomorphic as ${\boldsymbol{R}}$-algebras.
The map of Example \[ex: 34\] is concretely expressed as $$\psi(u,v)=(u^3 + 3uv^2-u^4-6u^2v^2-v^4, v^3+3u^2v+4u^3v+4uv^3, \Phi)$$ where $$\begin{aligned}
{1}
{\Phi} &=
\frac{1}{2}u^6-\frac{3}{2}u^4v^2+\frac{3}{2}u^2v^4-\frac{1}{2}v^6-\frac{1}{7}u^7-3u^5v^2-5u^3v^4-uv^6 \\
&
\hspace{0.3in}
\begin{aligned}
-\frac{1}{2}u^8+2u^6v^2-3u^4v^4 + 2u^2v^6 - \frac{1}{2}v^8.
\end{aligned} \end{aligned}$$
Define $\tilde{\psi}:= (\tilde{\psi}^1, \tilde{\psi}^2, \tilde{\psi}^3):=\psi(u-1, v-1)$, then $Q(\tilde{\psi}):={\mathcal{E}}_2/{\langle\tilde{\psi}^1, \tilde{\psi}^2, \tilde{\psi}^3 \rangle}_{\mathcal{E}_2}
\cong {\langle 1, v, v^2 \rangle}_{{\boldsymbol{R}}}$. Recall that the map-germ $(\psi,(1,1))$ (and so $(\tilde{\psi},(0,0))$) possesses the following property at a singular point $p\in{\Sigma}^2$. (Example \[ex: 34\])
1. \[item:1\] $p$ is a degenerate singular point.
2. \[item:2\] $\psi$ is frontal but not front of corank 1 on $p$.
3. \[item:3\] the set of degenerate singular point around $p$ is locally a smooth curve.
In the following, we will show that for $f=(x,y,z):{{\boldsymbol{R}}}^2\rightarrow {{\boldsymbol{R}}}^3$, the projection of a (generalized) geometric solution of $\mathcal{M}$ and for $(u_0,v_0) \in {{\boldsymbol{R}}}^2$, if the map-germ $(f,(u_0,v_0))$ is $\mathcal{A}$-equivalent to $(\tilde{\psi},(0,0))$ then contradiction occurs. From assumption, $(f,(u_0,v_0))$ also have the above properties (i), (ii) and (iii) because these three conditions are preserved under the same $\mathcal{A}$-equivalent class. By the change of coordinates of ${{\boldsymbol{R}}}^2$ and ${{\boldsymbol{R}}}^3$, we can assume that $(u_0,v_0)=(0,0)$ and $f(0,0)=(0,0,0)$.
First, the condition (i) and (ii) implies that $x(u,v)=u$ and $y_v(0,0)=y_{uv}(0,0)=y_{vv}(0,0)=0$. Therefore the Taylor expansion of $y(u,v)$ around $(0,0)$ becomes as the following form: $$\label{eq:y}
y(u,v)=a_kv^kq(v)+{ \sum_{i\not=0} a_{ij}u^iv^j}$$ where $k \geq 3$, $a_k \not=0 $ and $q(v)$ is a smooth function with $q(0)\not=0$ because otherwise, $y(u,v),z(u,v) \in I(u)$ from . It follows from and that $z(u,v) \in I(y(u,v))$. Hence $Q(f) \cong {\langle 1, v, \cdots, v^{k-1} \rangle}_{{\boldsymbol{R}}}$ and $Q(\tilde{\psi}) \cong Q(f)$ implies that $k=3$.
By the way, the condition (iii) implies that
> \(iv) $\left\{y_v=0 \right\} \cap \left\{y_{vv}=0\right\}$ is locally a smooth curve.
From , Fact \[fact:Mather\] implies that $$y_v(u,v)=P(u,v)y_{vv}(u,v)+R(u)$$ holds. The condition (iv) implies that $y_v(u,v)=0$ if $y_{vv}(u,v)=0$ because $y_{vv}(u,v)=0$ is locally a smooth curve and that there exists some $v$ for any $u$ such that $y_v(u,v)$ and $y_{vv}(u,v)$ holds around $(0,0)$ unless the singular set is $\left\{u=0 \right\}$. Thus $R(u)=0$ holds for any $u$ around 0. Therefore, locally $$\label{eq:MA}
y_v(u,v)=P(u.v)y_{vv}(u,v)$$ holds. The general solution of is $$y(u,v)=\int{\exp{\left(\int{\frac{dv}{P(u,v)}}+\alpha(u)\right)}dv}+\beta(u),$$ but this contradicts to the condition (iv) because $\left\{y_v=0 \right\} = \emptyset$.
The author would like to thank Kotaro Yamada for reading carefully the manuscript and giving him some suggestive comments. He also would like to thank Hitoshi Furuhata, Jun-ichi Inoguchi, Go-o Ishikawa and Takashi Kurose for valuable suggestions.
[1]{} W. Blaschke: [*Vorlesungen über Differentialgeometrie II, Affine Differentialgeometrie*]{}. Springer, Berlin, (1923).
V. Cortés: [*A holomorphic representation formula for parabolic hypherspheres*]{}. Banach Center Publications. [**57**]{}, 11-16 (2002).
V. Cortés, M.-A. Lawn, L. Schäfer: [*Affine hyperspheres associated to special para-Kähler manifolds*]{}. preprint
S. Fujimori, K. Saji, M. Umehara, K. Yamada: [*Singularities of maximal surfaces*]{}. to appear in Math. Z.
M. Golubitsky, V. Guillemin: [*Stable mappings and their singularities*]{}. Springer-Verlag, New York (1973)
M. Kokubu, W. Rossman, K. Saji, M. Umehara, K. Yamada: [*Singularities of flat fronts in hyperbolic space*]{}. Pacific J. Math. [**221**]{} (2005), 303-351.
M. Kokubu, M. Umehara, K. Yamada: [*Flat fronts in hyperbolic 3-space*]{}. Pacific J. Math. [**216**]{} (2004), 149-175.
J. Inoguchi, M. Toda: [*Timelike minimal surfaces via loop groups*]{}. Acta Appl. Math. [**83**]{} (2004), 313-355.
G. Ishikawa: A private communication. (2007).
G. Ishikawa, Y. Machida: [*Singularities of improper affine spheres and surfaces of constant Gaussian curvature*]{}. Internat. J. Math. [**17**]{}, (2007), 269-293.
T. Kurose: Lecture given at Kyushu University. (2006).
T. Kurose: A private communication. (2007).
A.-M. Li, U. Simon, Z. Zhao: [*Global affine differential geometry of hypersurfaces*]{}. Walter de Gruyter, Berlin-New York (1993)
A. Martínez: [*Relatives of flat surfaces in $H^3$*]{}. Proceedings of the International Workshop on integrable systems, geometry and visualization, Kyushu Univ. Fukuoka, Japan. (2005) 115-132. A. Martínez: [*Improper affine maps*]{}. Math. Z. [**249**]{} (2005), 755-766.
J. Mather: [*Stability of $C^{\infty}$ mappings. IV: Classification of stable germs by R-algebras*]{}. Publ. Math. I.H.E.S. [**37**]{} (1969), 223-248.
N. Matsuura, H. Urakawa: [*Discrete improper affine spheres*]{}. J. Geom. Phys. [**45**]{} (2003), 164-183.
K. Nomizu, T. Sasaki: [*Affine differential geometry*]{}. Cambridge University Press. (1994)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The nonlinear regime of laser-plasma interaction including both two-plasmon–decay (TPD) and stimulated Raman scattering (SRS) instabilities has been studied in three-dimensional (3-D) particle-in-cell simulations with parameters relevant to the inertial confinement fusion (ICF) experiments. SRS and TPD develop in the same region in plasmas, and the generation of fast electrons can be described accurately with only the full model including both SRS and TPD. The growth of instabilities in the linear stage is found to be in good agreement with analytical theories. In the saturation stage the enhanced low-frequency density perturbations driven by the daughter waves of the SRS sidescattering can saturate the TPD and consequently inhibit the fast-electron generation. The fast-electron flux in 3-D modeling is up to an order of magnitude smaller than previously reported in 2-D TPD simulations, bringing it close to the results of ICF experiments.'
author:
- 'H. Wen$^{1,2}$'
- 'A. V. Maximov$^{1,2}$'
- 'R. Yan$^{1,2}$'
- 'J. Li$^{1,2}$'
- 'C. Ren$^{1,2,3}$'
- 'F. S. Tsung$^4$'
title: 'Three-dimensional particle-in-cell modeling of parametric instabilities near the quarter-critical density in plasmas'
---
Since the 1960s, the pursuit of inertial confinement fusion (ICF) driven by lasers has led to large-scale research on laser interaction with the plasmas of ICF targets [@Craxton2015]. Decades of laser–plasma interaction (LPI) research [@myatt14] have concentrated on several processes in laser-produced plasmas that can grow as parametric instabilities at high-enough laser intensities, namely stimulated Raman scattering (SRS), stimulated Brillouin scattering (SBS), and two-plasmon decay (TPD).
Laser light can propagate in a plasma up to the critical density ($n_\text{c}$) determined by the laser frequency. The region near quarter-critical density ($\sfrac{1}{4}$ $n_\text{c}$) is a possible place for the interplay between SRS, SBS, and TPD as all three instabilities can develop at that region. plasmons produced by SRS and TPD generate fast electrons that can preheat the fusion fuel and degrade the performance of the ICF targets [@Craxton2015], making LPI a concern in ICF experiments. Several mechanisms of fast–electron acceleration have been studied before, namely staged acceleration [@yan09; @yan12], Langmuir cavitation [@vu12; @vu12a], and wave breaking [@coffey71].
In this Letter, LPI is studied using particle-in-cell (PIC) modeling [@Dawson83], which can describe the interplay between different instabilities and the particle distributions including fast-electron generation. Usually, few hot electrons are found in the simulations at the linear stage of the TPD and SRS instabilities. The electron acceleration becomes effective after the instabilities saturate [@yan12].
The TPD-related waves are mostly localized in the plane of polarization [@simon83], which is defined by the incident laser wave vector (in the $x$ direction) and the laser electric field vector (in the $y$ direction). The SRS sidescattering develops mostly outside of the polarization plane, and its scattered-light wave vector is almost perpendicular to the incident laser wave vector [@liu74; @afeyan97c]. Scattered light waves can also propagate in the direction parallel or anti-parallel to the laser wave vector (forward- and backscattering, respectively) [@drake73]. A 2-D simulation in the polarization plane ($x$–$y$) or in the perpendicular plane ($x$–$z$) will be referred to as $p$ polarized (PP) or $s$ polarized (SP), respectively. Two-dimensional simulations can model only the interaction where either (in PP simulations) TPD or (in SP simulations) SRS dominates except for the high-frequency hybrid instability (HFHI) [@afeyan97b] case when the SRS scattered light propagates in the backward direction and the SRS-related and TPD-related waves are in the same ($x$–$y$) plane. The 3-D simulations are required to study the interaction including both TPD and SRS. In this Letter, the results of several 3-D simulations for different plasma parameters and incident laser profiles are presented and compared with the respective 2-D simulations to illustrate that both TPD and SRS strongly influence the LPI near $\sfrac{1}{4}$ $n_\text{c}$. In the 3-D modeling including both TPD and SRS the fast-electron flux is reduced by up to an order of magnitude compared to 2-D TPD simulation results published before [@yan12].
Here we describe in detail a 3-D simulation for the parameters relevant to ICF experiments [@seka09; @Michel13]. A CH plasma is initialized with the electron temperature $T_\text{e}=2$ keV, and the temperatures for both ion species $T_\text{i}=1$ keV. The incident laser beam with intensity $I = 9\times10^{14}$ $\text{W/cm}^2$ propagates in the direction of density inhomogeneity ($x$). A linear density profile with the scale length $L= 100$ $\mu\text{m}$ is assumed at the initial time. The size of the simulation box is $21~\mu\text{m}\times8.4~\mu\text{m}\times6.7~\mu\text{m}$ modeling the density range from 0.21 $n_\text{c}$ to 0.26 $n_\text{c}$.
Two 2-D simulations (PP and SP) with the same physical parameters were also performed. The TPD threshold parameter $\eta$ [@simon83] is 1.9 ($\eta=1$ at threshold), and the SRS backscattering threshold parameter $N$ [@drake73] is 0.5 ($N=0.26$ at SRS threshold) for these simulations. The SRS sidescattering threshold [@liu74; @afeyan97c] is close to the backscattering threshold for these parameters, and both absolute TPD and absolute SRS instabilities are expected to grow. The threshold of the convective SRS [@liu74] is not exceeded for the parameters described above. The time evolution of the energy of the field components in the simulation region is shown in Fig. \[fig:figure1\]. The field energy is defined as the square of the electric- or magnetic-field amplitudes integrated over the simulation region normalized to the respective laser field energy at early time (when there are no instabilities). In the 2-D PP simulation \[Fig. \[fig:figure1\](a)\], the $E_x$ field contains most of the energy of the TPD plasmons with a larger wave vector. One can see that the field energy associated with the TPD instability stays at about the same level (close to 70% of the energy of the incident laser electric field) after 2.5 ps, when one can assume that the saturation stage is reached. In the 2-D SP simulation, the energy of the $B_x$ field \[Fig. \[fig:figure1\](b)\] is used as an indicator for the level of SRS instability. The energy of the scattered light saturates at a level of about 8% of the energy of the incident laser magnetic field.
In the 3-D simulation, the diagram for TPD and SRS is shown in Fig. \[fig:figure1\](c), where the incident light ($\vec{k}_0$) decays into a plasmon ($\vec{k}_{\text{SRS},1}$) and a light wave ($\vec{k}_{\text{SRS},2}$) in the case of SRS and into two plasmons ($\vec{k}_{\text{TPD},1}$ and $\vec{k}_{\text{TPD},2}$) in the case of TPD. The $E_x$ field energy \[red line in Fig. \[fig:figure1\](d)\] now includes the energy of the TPD plasmons and the SRS plasmons. The red line is overlaid with the dotted black line that represents the maximum TPD growth rate [@simon83] ($7.7\times10^{-4}~\omega_0$) minus the damping rate of plasma waves ($2.1\times10^{-4}~\omega_0$) measured in the 3-D simulation. The $E_x$ field saturates at a level of about 40% of the laser field energy, which is much lower than the saturation level in the 2-D PP simulation. The green line in Fig. \[fig:figure1\](d) corresponds to the energy of the scattered light wave (propagating in the $z$ direction) from the SRS and is overlaid with the dashed purple line representing the maximum growth rate of the SRS [@liu74]($8.2\times10^{-4}~\omega_0$) minus the damping rate of the plasma waves. One can see from Fig. \[fig:figure1\](d) that the growth of the $B_x$ field energy in time is in reasonable agreement with the theoretical result [@liu74]. The $B_x$ field energy is under 10% of the incident laser field energy after reaching its peak value, which is consistent with the 2-D SP simulation result.
![The integrated energy of different field components in the simulation region as a function of time for the (a) 2-D $p$-polarized simulation,(b) 2-D $s$-polarized simulation, and (d) 3-D simulation. The wave-vector diagram for TPD and SRS is shown in (c).[]{data-label="fig:figure1"}](figure1-eps-converted-to.pdf){width="3.375in"}
The spectra of plasma waves ($|\vec{E}_\text{L}|$) obtained at a time interval between 0.3 ps and 1.0 ps in the 2-D PP and SP simulations are plotted in Figs. \[fig:figure2\](a) and \[fig:figure2\](b), respectively. From the 3-D simulation, the spectra of plasma waves at a time interval between 1.3 ps and 2.0 ps are plotted in Fig. \[fig:figure2\](c) (close to $k_z=0$ plane, where TPD dominates) and in Fig. \[fig:figure2\](d) (far away from $k_z=0$ plane, where SRS dominates). One can see from Figs. \[fig:figure2\](c) and \[fig:figure2\](d) that TPD and SRS co-exist near $\sfrac{1}{4}$ $n_\text{c}$. The spectra of the unstable modes for TPD and SRS are close to the linear theory results (see overlaid lines in Fig. \[fig:figure2\]).
![(a) Plasma-wave spectra in the linear instability stage as a function of plasma density and the wave frequency normalized to laser frequency in the 2-D PP simulation, (b) 2-D SP simulation, and (c) 3-D simulation for modes with $0 \leq k_z/k_0 <0.2$ and (d) $0.2 \leq k_z/k_0 < 3$. The overlaid solid black lines and the dashed black lines represent the dispersion relations satisfying the matching conditions for TPD and SRS, respectively. []{data-label="fig:figure2"}](figure2-eps-converted-to.pdf){width="3.375in"}
As the instability evolves from the linear stage to the saturation stage, the frequency spectra shown in Fig. \[fig:figure2\] evolve into the spectra shown in Fig. \[fig:figure3\]. One can see that the spectra in all these simulations are broader in the saturation stage compared to the linear stage. The density in Fig. \[fig:figure3\] is calculated using the initial density profile. Compared to the 2-D PP simulation \[Fig. \[fig:figure3\](a)\], the TPD is much weaker at densities lower than 0.23 $n_\text{c}$ in the 3-D simulation \[Fig. \[fig:figure3\](c)\]. The weakening of the TPD modes at these densities is also illustrated in Fig. \[fig:figure4\](a) \[and Fig. \[fig:figure4\](b)\], where the spectrum of plasma waves at densities below 0.23 $n_\text{c}$ in the saturation stage is integrated over $k_z$ (and $k_y$). There are no prominent modes along the TPD hyperbola [@meyer93] \[black solid line in Fig. \[fig:figure4\](a)\] at $k_x > k_0$, which corresponds to the TPD daughter waves with larger wave vectors. Two types of low-frequency density fluctuations are identified in our simulations \[see Fig. \[fig:figure4\](c)\]. One type are the ion acoustic waves driven by the Langmuir-decay instability (LDI) [@DuBois67; @DuBois96] and the other type are driven with the beating of the same-frequency daughter waves of SRS and TPD. The LDI modes form a broad feature at $k_x \approx 1.7~k_0$ (about $2\times$ the laser wave vector in plasma) in the spectrum of the ion density fluctuations shown in Fig. \[fig:figure4\](c). The beating of the SRS plasmons with wave vector $(k_x,~k_y,~k_z) = (0.87~k_0,~0,~\pm 0.2~k_0)$ creates density perturbations with wave vector $(k_x,~k_z) = (0,~\pm 0.4~k_0)$. The coupling between SRS plasmons and density perturbations generates higher-order modes in the field at $k_z=\pm (0.2 + m 0.4~k_0)$, \[$m=1,~2,~3~...$, as shown in Fig. \[fig:figure4\](b)\] and in the density perturbation at $(k_x,~k_z)=[0,~\pm (0.4 + m 0.4~k_0)]$ \[see Fig. \[fig:figure4\](c)\].
Although SRS and TPD grow independently in the linear stage, in the nonlinear stage they interact through low-frequency density perturbations. TPD growth starts from the region near $\sfrac{1}{4}$ $n_\text{c}$ and spreads to lower densities [@yan12] before being saturated by ion density perturbations. Compared to TPD saturation in 2-D (without SRS) ion density perturbations are much larger in 3-D (with both SRS and TPD), especially near the plasma region where the frequencies of TPD and SRS plasmons are close. In this region where the dispersion lines for TPD and SRS plasmons intersect \[near 0.23 $n_\text{c}$ in our simulations, see Fig. \[fig:figure3\](c) and \[fig:figure3\](d)\] multiple pairs of SRS and TPD daughter waves have close frequencies and can drive ion density perturbations through the ponderomotive force to much higher levels compared to other density regions \[see the black line in Fig. \[fig:figure4\](d)\]. The growth of TPD plasmons at densities below 0.23 $n_\text{c}$ is disrupted by these enhanced ion density perturbations, as illustrated by a decrease in the level of TPD-driven plasmons below 0.23 $n_\text{c}$ in Fig. \[fig:figure3\](c).
The correlation between the local plasmon intensity $|E_L|^2$ and the density fluctuations $\delta n$ is captured using the caviton correlator [@vu12] $ C_{E,n} = \langle - \delta n
|E_L|^2 \rangle / (\langle (\delta n)^2 \rangle^{1/2}
\langle |E_L|^2 \rangle)$. As shown in the lower panel of Fig. \[fig:figure4\](d), the plasma waves and the density fluctuations are weakly correlated between 0.255 $n_\text{c}$ and 0.235 $n_\text{c}$ : $C_{E,n} = 0.1-0.2$ in spite of a significant level of plasmons in this density range. At densities close to 0.23 $n_\text{c}$, lower panel of Fig. \[fig:figure4\](d) shows the increase not only in the plasmon intensity and density fluctuations, but also in the correlation between them with $C_{E,n}$ reaching up to 0.6. The large caviton correlator indicates that the plasma waves are strongest in areas where density is depleted. The ponderomotive force of multiple pairs of SRS and TPD daughter waves with close frequencies is responsible for driving the enhanced density perturbations. The nonlinear coupling of TPD and SRS through ion perturbations leads to a lower TPD saturation level in the 3-D simulation compared to the 2-D PP simulation, which is illustrated in the upper panel of Fig. \[fig:figure4\](d).
![Plasma-wave spectra in the saturation stage in the 2-D (from 3.3 ps to 4.1 ps) and 3-D (from 2.3 ps to 3.1 ps) simulations as a function of plasma density and the wave frequency. Each panel displays the same quantity as in Fig. \[fig:figure2\]. []{data-label="fig:figure3"}](figure3-eps-converted-to.pdf){width="3.375in"}
The fast-electron flux is defined as the energy flux carried by electrons with kinetic energy above 55 keV leaving the simulation box minus the energy flux carried by the thermal electrons injected into the simulation region from the thermal boundaries (in the $x$ direction). Information about the hot electrons is collected during the saturation stage in each simulation for 0.5 ps. In the 3-D simulation, the fast-electron flux associated with the forward- and backward-going hot electrons was found to be 1.7% and 0.8%, respectively. The plasma-wave spectrum in the 3-D simulation corresponds to a smaller $k$-space domain than the spectrum in the 2-D PP simulation, which makes the staged acceleration mechanism less efficient in 3-D than in 2-D and explains a smaller number of hot electrons in the 3-D simulation compared to the 2-D PP simulation (6.6% and 3.4% in the forward and backward direction, respectively). The influence of wave-breaking on the fast-electron generation is small as the maximal electric field amplitude (0.04 $m_e \omega_0 c / e$) is below the wave-break limit(0.1 $m_e \omega_0 c / e$) [@coffey71].
![(a) The spectrum of plasmons in the saturation stage of 3-D simulation at densities lower than $0.23~n_c$ plotted in the $k_x$–$k_y$ plane and (b) the $k_x$–$k_z$ plane. (c) The spectrum of ion density fluctuation plotted in the $k_x$–$k_z$ plane on a logarithmic scale. (d) Lower panel: Ion density fluctuations RMS (root-mean-square average over the transverse direction and time) normalized to background density (black solid line), longitudinal electric field RMS (blue dashed line) and caviton correlator $C_{E,n}$ (blue dotted line). Upper panel: the ratio of the electric field amplitude of the TPD plasmons with larger wave vector between 2-D PP and 3D simulations.[]{data-label="fig:figure4"}](figure4-eps-converted-to.pdf){width="3.375in"}
The nonlinear regime including both TPD and SRS is also observed in simulations with the speckled laser beam [@kato84; @skupsky89] and electron–ion collision effects included. The speckled laser beam is modeled by a single speckle in the simulation region that mirrors itself in the transverse direction because of the periodic boundary conditions. A series of simulations has been performed to study how the speckles affect the generation of hot electrons. All parameters are the same as the simulations described previously except for the temperatures of electrons and ions being 1.5 times higher. The peak intensities in the laser speckles are $1.8 \times 10^{15}~\text{W/cm}^2$ (twice of the average intensities). A collision package (CP) is available for the PIC code OSIRIS [@fonseca02]. The main physics processes are observed to be the same in simulations with plane-wave beams and speckled beams.
The fast-electron flux values in simulations are listed in Table \[tab:eflux\] for different incident laser beams as well as with CP turned on and off. By comparing the left and right columns of Table \[tab:eflux\], one can see that adding collisions can reduce the fast-electron flux by about 50% and in the case of plane-wave 2-D PP simulation by almost 70%. Also note that the reduction of the fast-electron flux caused by collisions affects both the forward-going electrons and backward-going electrons since the collisional damping rate affects all the plasma waves. The fast-electron flux generated in the 2-D SP simulations is much smaller than the fast-electron flux generated in the 2-D PP simulations, which indicates that the plasma waves driven by TPD are the main source of the electron acceleration.
The hot electron fraction observed in the ICF experiments on the OMEGA laser system does not exceed few percent [@Michel13]. At the same time, in the previous PIC simulations of TPD in 2-D the hot electron fraction was close to an order of magnitude larger than in the experiments. The 3-D PIC simulations presented in this Letter for the first time produce the results for the hot electron fraction that are close to the experimental levels.
Fast-electron flux
-------------------- ------------------ ------------------
Collision package On Off
Plane wave 2-D PP $1.6\%/1.3\%$ $5.5\%/3.8\%$
Plane wave 2-D SP $(<0.1\%)/0.2\%$ $(<0.1\%)/0.5\%$
Speckle 2-D PP $6.8\%/1.7\%$ $9.4\%/3.8\%$
Speckle 2-D SP $(<0.1\%)/0.3\%$ $(<0.1\%)/0.7\%$
Speckle 3-D $0.4\%/0.3\%$ $0.8\%/0.5\%$
: Fast-electron flux normalized to the incident laser energy flux.[]{data-label="tab:eflux"}
laser–plasma interaction near $\sfrac{1}{4}$ $n_\text{c}$ determines the generation of fast electrons that are crucial for the performance of ICF targets. The fast-electron flux in simulations is found to be closely related to the plasma-wave spectra. The TPD-driven plasma waves with large wave vectors are very important for accelerating electrons. At the same time, the SRS-driven plasma waves are less effective in accelerating electrons. Therefore the modeling including the nonlinear coupling of TPD and SRS in 3-D is the only way to correctly describe the generation of fast electrons in laser-driven ICF.
Our 3-D PIC simulations have shown the large decrease (up to an order of magnitude) in the fast-electron flux compared to 2-D TPD modeling. The reason is the nonlinear coupling between SRS and TPD which is especially pronounced at densities lower and around 0.23 $n_\text{c}$. In this region plasma waves and growing density perturbations are localized in same areas as illustrated by the caviton correlator. Enhanced density perturbations detune and weaken the TPD-driven plasmons effective in the fast electron generation. In addition to the TPD suppresion, the plasma wave spectra in 3-D simulations are much more narrow compared to the spectra in 2-D TPD modeling. To conclude, 3-D PIC simulations presented in this Letter fully model the laser-plasma interaction near $\sfrac{1}{4}$ $n_\text{c}$ including SRS and TPD, and obtain the fast electron fraction level close to experimental results, resolving the large discrepancy between ICF experiments and PIC simulations that existed for many years before.
This work was supported by the Department of Energy National Nuclear Security Administration under Award Number DE-NA0001944, the University of Rochester, and the New York State Energy Research and Development Authority. We also acknowledge the support by the DOE under grant No. DE-SC0012316, and by the NSF under grant No. PHY-1314734. This report was prepared as an account of work sponsored by an agency of the U.S. Government. Neither the U.S. Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the U.S. Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the U.S. Government or any agency thereof.
[21]{}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 [****, ()](\doibase 10.1063/1.4934714) [****, ()](\doibase 10.1063/1.4878623) [****, ()](\doibase
10.1103/PhysRevLett.103.175002) [****, ()](\doibase
10.1103/PhysRevLett.108.175002) [****, ()](\doibase 10.1063/1.4764075) [****, ()](\doibase 10.1063/1.4757978) [****, ()](\doibase 10.1063/1.1693620) [****, ()](\doibase 10.1103/RevModPhys.55.403) [****, ()](\doibase 10.1063/1.864037) [****, ()](\doibase 10.1063/1.1694867) [****, ()](\doibase doi:10.1063/1.872505) [****, ()](\doibase 10.1103/PhysRevLett.31.1197) [****, ()](\doibase doi:10.1063/1.872507) [****, ()](\doibase
10.1063/1.3125242) [****, ()](\doibase 10.1103/PhysRevLett.71.2915) [****, ()](\doibase 10.1103/PhysRev.164.207) [****, ()](\doibase 10.1088/0031-8949/1996/T63/002) [****, ()](\doibase 10.1103/PhysRevLett.53.1057) [****, ()](\doibase
10.1063/1.344101) in *Computational Science – ICCS 2002*, edited by , Lecture Notes in Computer Science, Vol. 2331 () p. . [****, ()](\doibase
10.1063/1.4803090)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that violation of genuine multipartite Bell inequalities can be obtained with sampled, probabilistic phase space methods. These genuine Bell violations cannot be replicated if any part of the system is described by a local hidden variable theory. The Bell violations are simulated probabilistically using quantum phase-space representations. We treat mesoscopically large Greenberger-Horne-Zeilinger (GHZ) states having up to $60$ qubits, using both a multipartite SU(2) Q-representation and the positive P-representation. Surprisingly, we find that sampling with phase-space distributions can be exponentially *faster* than experiment. This is due to the classical parallelism inherent in the simulation of quantum measurements using phase-space methods. Our probabilistic sampling method predicts a contradiction with local realism of “Schrödinger-cat” states that can be realized as a GHZ spin state, either in ion traps or with photonic qubits. We also present a quantum simulation of the observed super-decoherence of the ion-trap “cat” state, using a phenomenological noise model.'
author:
- 'M. D. Reid$^{1}$, B. Opanchuk$^{1}$, L. Rosales-Zárate$^{1}$, P. D. Drummond$^{1}$'
bibliography:
- 'MultiQsims.bib'
title: 'Quantum probabilistic sampling of multipartite 60-qubit Bell inequality violations'
---
Introduction
============
Quantum simulation of systems with many degrees of freedom is a difficult and interesting problem of much topical interest. Calculating the dynamics of many-body quantum systems is hard, since the Hilbert space dimension increases exponentially with the number of modes or degrees of freedom [@Dirac1929; @Feynman1982]. There are two main approaches: one can do a computational simulation [@Haake1979; @Corney2006; @Deuar2007; @Alon2008; @Gambetta2008; @Trotzky2012], or else a physical simulation with another quantum system [@Cirac2003; @Jaksch2005; @Buluta2009; @Islam2011; @Georgescu2014]. Universal quantum computers provide a third option [@Lanyon2011], but these are limited in size.
One path to solving this problem is to use probabilistic simulations whose correlations correspond to quantum averages. For large problems, this approach was pioneered by Glauber and co-authors [@Glauber1978; @Haake1979], who studied quantum statistics of super-fluorescence. Later, their approximate method was generalized to an exact probabilistic representation of arbitrary quantum states [@Drummond1980; @Carter1987]. Quantum simulation predictions were experimentally verified for multi-mode optical fields displaying squeezing and quantum entanglement [@Rosenbluh1991; @Drummond1993-solitons; @Heersink2005; @Corney2006]. More recently, the method has been applied to colliding BEC systems [@Deuar2007; @Lewis-Swan2014], and to Bell violations in parametric down-conversion experiments [@Rosales-Zarate2014].
Here we study how efficiently such probabilistic methods can be used to simulate the most extreme quantum superposition states $-$ or “Schrödinger cat” states. The “cat” state is often represented as a GHZ state for $M$ particles [@Greenberger1989]: $$\vert\Phi\rangle=\frac{1}{\sqrt{2}}\left(\vert\uparrow\ldots\uparrow\rangle+e^{i\phi}\vert\downarrow\ldots\downarrow\rangle\right).\label{eq:ghz-full-1}$$ where $|\uparrow\ldots\uparrow\rangle=\bigotimes_{j=1}^{M}|\uparrow\rangle_{j}$, $|\downarrow\ldots\downarrow\rangle=\bigotimes_{j=1}^{M}|\downarrow\rangle_{j}$ and $|\uparrow\rangle_{j}$, $|\downarrow\rangle_{j}$ are the eigenstates of the spin $\hat{\sigma}^{z}$ of the $j$-th particle. A powerful signature of the “cat” state is its $M$-qubit nonlocality. These have been explored in photonic [@Lu2007] and ion trap experiments [@Leibfried2005-creation], which demonstrated Bell-Mermin violations and genuine $M$-particle entanglement for up to $M=14$ions [@Lanyon2014].
We investigate probabilistic methods for simulating both Bell-Mermin violations and the more challenging Svetlichny-Collins genuine Bell violations in these multipartite “cat” states. The latter inequality allows us to demonstrate the genuine $M$-partite nonlocality of the multipartite GHZ state (\[eq:ghz-full-1\]) for up to $M=60$ ions or modes. This is a true Schrödinger cat signature: it cannot be obtained if any subset has a local hidden variable (LHV) description. It is often thought that probabilistic sampling would be extraordinarily difficult for a “cat” state of large size. Probabilistic methods using measured eigenvalues are impossible, since this would amount to an LHV theory, which cannot violate a Bell inequality.
Importantly, Mermin showed that the difference between LHV predictions and quantum predictions scales exponentially with increasing system size $M$ [@Mermin1990-entanglement], making this a significant challenge for probabilistic methods. However, the techniques used in this paper do *not* rely on LHV theories, but instead sample over stochastic variables whose values are permitted to go beyond the eigenvalue spectrum. The ability to simulate quantum mechanics in this way gives a beautiful analogy to the theory of weak values and measurements [@Aharonov1988], as has been explained elsewhere [@Drummond2014-bell-sim; @Opanchuk2014-bell-sim; @Rosales-Zarate2014].
We find that sampling errors of high-order correlations are larger than for low-order correlations. We note that $M$-order correlations are needed to display the signature of an $M$-partite Bell-Svetlichny nonlocality. However, for these results, phase-space simulations have a classical parallelism not available in any quantum experiment. This parallelism occurs because exponentially many non-commuting measurements can be calculated at once. The result is an exponential *speedup* for simulating multipartite Bell violations. The simulation is $e^{M/3}$ times faster than experiment with methods used here.
The general advantage of probabilistic sampling in quantum simulations compared to wave-function methods [@Islam2011] is that the required computational memory scales linearly, not exponentially, with the number of qubits. This eliminates the problem of exponential scaling in memory size found in direct, orthogonal basis calculations. The time taken, or equivalently the number of samples required, depends on the type of measurement and the resulting sampling error in a finite ensemble. We consider the Greenberger-Horne-Zeilinger (GHZ) state (\[eq:ghz-full-1\]), and simulate spin correlations as well as multipartite Bell violations.
The utility of phase-space representations is that they provide a route to performing such probabilistic sampling. We employ two common positive phase-space distributions, namely, the SU(2) Q function [@Husimi1940; @Arecchi1972; @Gilmore1975] and the positive P-distribution [@Drummond1980]. The latter method has already been used to obtain analytic results for probabilistic Bell violation [@Drummond1983]. To focus on the sampling issue, we mostly treat static cases with known probability distributions. We also treat dynamical simulations of decoherence. A summary of the results is published elsewhere [@Opanchuk2014-bell-sim].
In these multipartite investigations, we use Bell-like inequalities that extend the usual bipartite inequalities to many qubits. We test the MABK (Mermin-Ardehaly-Belinski-Klyshko) inequality [@Mermin1990-entanglement; @Ardehali1992; @Belinskii1993-interference; @Belinsky1993-N-particle] which is a Bell inequality generalized to multipartite qubit systems, and the Collins-Svetlichny inequalities [@Svetlichny1987; @Collins2002], which are sufficient conditions for genuine multipartite Bell violations [@Ghose2009; @Ajoy2010; @Bancal2011; @Chen2011; @Grandjean2012]. Genuine multipartite Bell violations prove that Bell violations are a macroscopic property.
We find different behavior depending on the order of the correlation function. There is no growth in sampling error with the number of qubits when simulating low-order correlations in our calculations. Thus, fixed order correlations do not have an exponential increase in simulation time. However, correlations with a growing order equal to the number of qubits take an exponentially long time to simulate. Yet even these calculations scale only as a *fractional* power of the number of qubits. This allows us to simulate genuine multipartite Bell violations of GHZ states with 60 qubits, corresponding to a Hilbert space of a quintillion ($10^{18}$) dimensions.
Bell violations as large as this would also require a quintillion different measurement settings in the laboratory. At around 1 ms per measurement in an ion-trap experiment, full confirmation of a 60 qubit multipartite Bell violation would take over $30$ million years, even with just one measurement per laboratory setting. For multipartite Bell inequalities, our simulations took less than $48$ hours, so the exponential speedup obtained through phase space quantum simulations is a highly practical computational tool. To demonstrate applications for decoherence dynamics, we use the method to simulate the observed super-decoherence in ion traps.
The paper is organized as follows. In Section \[sec:MABK and N Partite Bell Ineq\] we discuss the multipartite Bell inequalities. The sampling of the GHZ states using the positive P and the Q section are described in Sections \[sec:Sampling-GHZ-states\] and \[sec:Sampling-GHZ-Qfunction\], respectively. The computational results are shown in Section \[sec:Results\]. In Section \[sec:Decoherence-model\] we describe a decoherence model that shows the dynamical decay of the Bell inequality, as observed experimentally. Finally, Section \[sec:Conclusions\] gives a summary of our results and conclusions.
\[sec:MABK and N Partite Bell Ineq\] Multipartite Bell inequalities
====================================================================
The challenge for quantum simulation is to simulate very large systems where quantum effects can still manifest themselves. The best example is a macroscopic superposition state of the type considered in the “Schrödinger cat” paradox. With this objective, we will analyze how to simulate the genuine multipartite Bell inequality violations of $M$ entangled particles. Our goal is to determine whether this is possible, using probabilistic sampling. We also wish to understand the relevant scaling properties, as they depend on the measurements themselves. A detailed treatment of the bipartite case, including dynamical simulations, is presented elsewhere [@Rosales-Zarate2014].
MABK Bell inequalities for $M$ sites
------------------------------------
First, we summarize well-known Bell inequalities that test local hidden variable (LHV) theories involving $M$ spin-$1/2$ particles at different sites. We label the sites by $j$, where $j=1,\ldots,M$.
In the case of $M$ particles emitted from a common source, measurements of $M$ spatially separated observers are modeled in the LHV theory by taking random samples of a common set of parameters (the hidden variables) symbolized by $\lambda$. Measured values are then functions of some local detector/analyzer settings and the hidden parameters $\lambda$.
We use the notation that $X_{m}(\lambda)\equiv X_{m}(\theta_{m},\lambda)$ for the $m$-th observer with the detector analyzer setting $\theta_{m}$, denoting the measurement value by $X_{m}$. Here, the measurement event includes the selection of the measurement setting $\theta_{m}$ at each site. The $M$ measurement events are assumed to be space-like separated. In an LHV theory the correlations are thus obtained from a probabilistic calculation of the form: $$\begin{aligned}
E(X_{1},X_{2},\ldots X_{M}) & \equiv & \left\langle \left[\prod_{m=1}^{M}X_{m}\right]\right\rangle \nonumber \\
& = & \int\left[\prod_{m=1}^{M}X_{m}(\lambda)\right]P(\lambda)d\lambda.\label{eq:LHV}\end{aligned}$$ where $P(\lambda)$ is a probability distribution for the hidden variables $\lambda$.
One can consider that at each site the experimentalist makes one of two choices for the measurement. Here, we denote these two choices by the quantum observables $\hat{x}_{j}$ and $\hat{y}_{j}$, and denote the outcomes associated with these measurements by $X_{j}$, $Y_{j}$ respectively. Experimentally, one uses an adjustable polarizer or Rabi rotation at each site to determine which of the choices to make, and there are $2^{M}$ possible combinations. For each of these $2^{M}$ choices, an ensemble of measurements is necessary to obtain the relevant correlations.
Following Mermin [@Mermin1990-entanglement], we can construct for mathematical convenience the operator $$\hat{A}_{j}=\hat{x}_{j}+i\hat{y}_{j},\label{eq:f}$$ bearing in mind that this is not a measured observable. We can also define the complex function $F_{j}=X_{j}+iY_{j}$. We now examine the dichotomic case using qubits. We follow Mermin and choose: $$\begin{aligned}
\hat{x}_{j} & = & \hat{\sigma}_{j}^{\theta_{j}}\nonumber \\
\hat{y}_{j} & = & \hat{\sigma}_{j}^{\theta_{j}+\pi/2},\end{aligned}$$ where $\hat{\sigma}_{j}^{\theta}=\hat{\sigma}_{j}^{x}\cos\theta_{j}+\hat{\sigma}_{j}^{y}\sin\theta_{j}$, and $\hat{\sigma}_{j}^{x/y}$ are the Pauli spin operators. Therefore: $$\hat{A}_{j}=\left(\hat{\sigma}_{j}^{x}+i\hat{\sigma}_{j}^{y}\right)e^{-i\theta_{j}}.$$
Next, we consider the measurable moments given by the expression: $$A_{\mathrm{QM}}=\langle\prod_{j=1}^{M}\hat{A}_{j}\rangle\equiv\left\langle \hat{A}\right\rangle ,\qquad\hat{A}\equiv\prod_{j=1}^{M}\hat{A}_{j}$$ and the corresponding LHV prediction for this moment
$$A_{\lambda}=\langle\prod_{j=1}^{M}\hat{A}_{j}\rangle_{\lambda}\equiv\langle\prod_{j=1}^{M}F_{j}\rangle.$$ where $\Pi$ denotes the product (standard notation). One can expand the terms of the product, and write as a real and imaginary part: So, we define the real and imaginary parts by: $$A_{\mathrm{QM}/\lambda}=\mathrm{Re}A_{\mathrm{QM}/\lambda}+i\mathrm{Im}A_{\mathrm{QM}/\lambda}.\label{eq:expfreim}$$
It is known that LHV theories place a constraint on what should be observed for these quantities. These are the Mermin-Ardehali-Belinskiĭ-Klyshko (MABK) Bell inequalities. Mermin [@Mermin1990-entanglement] originally derived the following Bell inequality (which we will call Mermin’s inequality),
$$\mathrm{Im}A_{\lambda}\le\begin{cases}
2^{(M-1)/2}, & M\ \mathrm{is\ odd},\\
2^{M/2}, & M\ \mathrm{is\ even}.
\end{cases}\label{eq:MABKMermin}$$
The same inequalities hold for the $\mathrm{Re}A_{\lambda}$. Mermin’s inequality for even $M$ is weak, and is not violated by the Bell state (\[eq:ghz\]) for $M=2$. Therefore for the case of even $M$ we will follow Ardehali, Belinskiĭ and Klyshko (ABK) [@Ardehali1992; @Belinskii1993-interference; @Belinsky1993-N-particle], who derived the following inequalities:
$$\mathrm{Re}A_{\lambda}+\mathrm{Im}A_{\lambda}\le\begin{cases}
2^{M/2}, & M\ \mathrm{is\ even},\\
2^{(M+1)/2}, & M\ \mathrm{is\ odd}.
\end{cases}\label{eq:MABKArdehali}$$
ABK inequalities are stronger for even $M$, but not for odd $M$, and thus the MABK Bell inequalities [@Belinskii1993-interference; @Belinsky1993-N-particle] are the combination of (\[eq:MABKMermin\]) for odd $M$, and (\[eq:MABKArdehali\]) for even $M$.
We can expand these inequalities explicitly to see what they are. For $M=2$, $\theta_{j}=0$, the MABK inequality is: $$\langle\sigma_{1}^{x}\sigma_{2}^{y}\rangle_{\lambda}+\langle\sigma_{1}^{y}\sigma_{2}^{x}\rangle_{\lambda}+\langle\sigma_{1}^{x}\sigma_{2}^{x}\rangle_{\lambda}-\langle\sigma_{1}^{y}\sigma_{2}^{y}\rangle_{\lambda}\leq\sqrt{2}\label{eq:sigma2}$$ which is the famous Clauser-Horne-Shimony-Holt (CHSH) Bell inequality. For $M=3$, $\theta_{j}=0$ the resulting inequality is: $$\langle\sigma_{1}^{y}\sigma_{2}^{x}\sigma_{3}^{x}\rangle_{\lambda}+\langle\sigma_{1}^{x}\sigma_{2}^{y}\sigma_{3}^{x}\rangle_{\lambda}+\langle\sigma_{1}^{x}\sigma_{2}^{x}\sigma_{3}^{y}\rangle_{\lambda}-\langle\sigma_{1}^{y}\sigma_{2}^{y}\sigma_{3}^{y}\rangle_{\lambda}\leq2\label{eq:sigma2-1}$$ as derived by Mermin. We note by defining $F_{j}=X_{j}-iY_{j}$ a different set of MABK inequalities with different signs can be derived.
MABK violations with a GHZ state
--------------------------------
All of the MABK inequalities are predicted by LHV theories, but only for the right quantum state are they maximally violated. Let us consider the GHZ state: $$|\psi\rangle=\frac{1}{\sqrt{2}}\left(\bigotimes_{j=1}^{M}|\uparrow\rangle_{j}+e^{i\phi}\bigotimes_{j=1}^{M}|\downarrow\rangle_{j}\right),\label{eq:ghz}$$ where $|\uparrow\rangle_{j}$ $|\downarrow\rangle_{j}$ are the eigenstates of $\hat{\sigma}_{j}^{z}$. It is known that the state (\[eq:ghz\]) with $r=M$ violates (\[eq:MABKArdehali\]) by the *maximum* amount predictable by Quantum Mechanics (QM) [@Werner2001]. For the Mermin-type inequalities (\[eq:MABKMermin\]), this maximal violation occurs for the angle $\phi=\pi/2$ and the measurement choice $\theta_{j}=0$: $$A_{j}=\sigma_{j}^{x}+i\sigma_{j}^{y},\quad j=1,\ldots,M.\label{eq:angleperfect}$$ where we have now denoted the results $X_{j}$, $Y_{j}$ of the measurements $\hat{\sigma}_{j}^{x}$, $\hat{\sigma}_{j}^{y}$ by $\sigma_{j}^{x}$, $\sigma_{j}^{y}$ written without the operators. This orthogonal angle choice corresponds to the famous cases of the EPR-Bohm and GHZ paradoxes [@Einstein1935; @Greenberger1989; @Mermin1990-reality], that yield perfect correlations between spatially separated spins. The quantum prediction for the choice of measurement orientations (\[eq:angleperfect\]) is [@Mermin1990-entanglement]: $$\mathrm{Im}A_{\mathrm{QM}}=2^{M-1}.\label{eq:QMMermin}$$ On the other hand, the Ardehali-Bell-CHSH-type inequalities (\[eq:MABKArdehali\]) give a maximum when $\phi=\pi$ and one site has a shifted measurement angle: $$\begin{aligned}
F_{j} & = & \sigma_{j}^{x}-i\sigma_{j}^{y},\,\, j\neq M\label{eq:anglenonper}\\
F_{M} & = & \sigma^{-\pi/4}+i\sigma^{\pi/4}.\nonumber \end{aligned}$$ We note this corresponds for $M=2$ to the case of Bell and CHSH [@Bell1964; @Clauser1969; @Clauser1978; @D'Espagnat1971]. Here, the measurement choice does not allow perfect correlation between spatially separated measurements for a fixed setting, and the violation is obtained statistically. The quantum prediction in this case is [@Ardehali1992; @Mermin1990-entanglement]: $$\mathrm{Re}A_{\mathrm{QM}}+\mathrm{Im}A_{\mathrm{QM}}=2^{M-1/2}.\label{eq:QMArdehali}$$
It is convenient to join the odd- and even-$M$ inequalities using an operator
$$\hat{V}=\begin{cases}
\mathrm{Re}\hat{A}+\mathrm{Im}\hat{A}, & M\ \mathrm{is\ even},\\
\sqrt{2}\,\mathrm{Im}\hat{A}, & M\ \mathrm{is\ odd}.
\end{cases}\label{eq:V_Op}$$
In this case the MABK inequality for all $M$ is, in the case of an LHV theory:
$$V_{\lambda}\equiv\vert\langle\hat{V}\rangle_{\lambda}\vert\le2^{M/2}.\label{eq:V}$$
This is violated by quantum mechanics with the state and measurement choices above, since:
$$V_{\mathrm{QM}}\equiv\langle\hat{V}\rangle=2^{M-1/2}>V_{\lambda}.$$
The ratio between the LHV limit and the QM result is thus:
$$\frac{V_{\mathrm{QM}}}{V_{\lambda}}\ge2^{\left(M-1\right)/2},$$
which grows exponentially with $M$.
Genuine $M$-partite Bell nonlocality
------------------------------------
Svetlichny [@Svetlichny1987] introduced the idea of genuine multipartite nonlocality. He derived inequalities that if violated indicate a three-body (rather than two-body) nonlocality. The inequalities have been generalized to $M-$partite cases by Collins *et al* [@Collins2002] and by Seevinck and Svetlichny [@Seevinck2002]. We point out that other recent works [@Aolita2012; @Gallego2012; @Bancal2013] have improved Svetlichny’s approach further.
The Svetlichny-CGPRS inequality is: $$\begin{aligned}
V_{{\cal S}}\equiv\mathrm{Re}A_{\lambda}+\mathrm{Im}A_{\lambda} & \leq & 2^{M-1}\,,\label{eq:merminsteerstat-1}\end{aligned}$$ the violation of which is sufficient to confirm *genuine $M$-partite Bell nonlocality*. For $M=3$ this means that the violation cannot be explained using product states or mixtures with Bell nonlocality between only two sites. More generally, for arbitrary $M$, this terminology means that the violation cannot be explained using states with a genuine $ $$m$-partite Bell nonlocality, where $m<M$. The quantum prediction maximizes at (\[eq:QMArdehali\]) to predict violation, for even $M$, by a *constant* amount: $$\frac{V_{QM}}{V_{S}}=\sqrt{2}\,.$$
This constant violation ratio differs from the exponential violation predicted for the MABK inequalities, which makes the effect both harder to measure experimentally, and more difficult to simulate than the usual Bell inequality. However it is necessary to achieve this stronger correlation if one wishes to assert that a given superposition is truly macroscopic to a given level, i.e., if one wishes to exclude the possibility that there are only microscopic violations of local realism present in a quantum system.
\[sec:Sampling-GHZ-states\]Sampling GHZ states with positive phase-space distributions
======================================================================================
The states we wish to sample are GHZ states (\[eq:ghz\]), which are experimentally prepared in a number of photonic and ion-trap experiments. We rewrite these as: $$\vert\Phi\rangle=\frac{1}{\sqrt{2}}\left(\vert\uparrow\ldots\uparrow\rangle+e^{i\phi}\vert\downarrow\ldots\downarrow\rangle\right).\label{eq:ghz-full}$$ Of course, any experiment will inevitably also include other states owing to decoherence effects. Here we wish to start by considering the pure state, which is a worst-case scenario from the point of view of phase-space simulations. The up- and down-states can be represented differently, depending on the underlying physical system, which will in turn affect the sampling. We will consider different sampling techniques using different operator representations, in order to compare their efficiency.
Phase-space methods
-------------------
In general terms, a phase-space representation is a mapping from a c-number distribution function $P\left(\vec{\lambda}\right)$ to a density matrix $\hat{\rho}$, defined by
$$\hat{\rho}=\int P\left(\vec{\lambda}\right)\hat{\Lambda}\left(\vec{\lambda}\right)d\vec{\lambda}.\label{eq:general-phase-space}$$
Here $\hat{\Lambda}\left(\vec{\lambda}\right)$ is a complete operator basis, which is parametrized with a phase-space variable $\vec{\lambda}$, and $P\left(\vec{\lambda}\right)$ is a distribution over $\vec{\lambda}$ which typically allows one to calculate observables as moments. For our present purposes, we will focus on mappings that involve a positive-definite distribution $P\left(\vec{\lambda}\right)$. This allows probabilistic sampling, which is a very scalable route for calculating high-dimensional integrals and correlations. It also removes the need to have a numerical representation of an exponentially large matrix. This approach results in efficient scaling for low-order correlations, even for highly nonclassical states like the GHZ state, and can be sampled for high-order correlations with somewhat lower efficiency.
There are many such mappings known. The earliest methods developed were for the Wigner function [@Wigner1932], Q-function [@Husimi1940] and P-function [@Glauber1963-states; @Sudarshan1963]. These are all for bosonic Hilbert spaces, are defined for a real phase-space $\vec{\lambda}$, and correspond to different operator orderings. Of these, only the Q-function is positive-definite. Subsequently, positive-definite extensions of these were developed that use complex instead of real phase-spaces, including the positive P-representation [@Drummond1980], the positive Wigner representation [@Chaturvedi1994; @DeOliveira1992] and the Gaussian representation [@Corney2003]. The positive P-representation is useful, as it combines stochastic time-evolution with simple observables.
All these bosonic methods involve a Hilbert space of too large a dimension for optimum sampling of the GHZ state, as we explain below. It is most efficient to only represent those parts of a Hilbert space that are measured. Hence, it is better to use a phase-space representations that is specifically matched to a finite dimensional Hilbert space. The earliest of these were the SU(2) based continuous representations [@Radcliffe1971; @Arecchi1972; @Zhang1990], which employ Lie group methods. These have a similar form to the bosonic case. Once again, there are both positive and non-positive distributions, as well as complex phase-space methods [@Barry2008]. A widely used positive form is the SU(2) Q-function [@Arecchi1972; @Gilmore1975], which we analyze in detail in the next section.
Recently, a number of interesting and innovative methods have been introduced that treat finite Hilbert spaces in a different way. These replace the integral in Eq (\[eq:general-phase-space\]) with a summation over a finite set of points. Using this technique, it is possible to develop a discrete Wigner distribution [@Wootters1987; @Wootters1989; @Gibbons2004; @Wootters2004; @Wootters2006; @Bjork2008], which uses hermitian matrices instead of distributions to represent the Hilbert space. In the standard construction of such methods, certain specific quantum states have positive representations, but this is not true in general. In other words, the generic case for the discrete Wigner distribution is that the mapping is non-positive.
These discrete approaches have the property that the underlying discrete Wigner distribution is a $2^{M}\times2^{M}$ matrix for $M$ qubits [@Wootters2004; @Gibbons2004; @Bjork2008]. In the largest case treated here, with $M=60$, this involves $10^{36}$ matrix elements. These do not all have to be stored in memory, which is impossible with current computers. Nevertheless, calculating observables with $10^{36}$ elements requires sampling to reduce the computation time. As the elements are not all positive, this would presumably involve a sign or phase term, which can lead to inefficiencies.
Accordingly, we do not investigate the discrete Wigner function here. Yet such methods may also be useful. The main challenge is that the resulting large matrix representations are not probabilistic. The question of how to sample these efficiently is an open question at present. However, extending such discrete techniques to allow probabilistic sampling may not be impossible. This is outside the scope of the present paper, so we now return to the question of efficient sampling using continuous, positive phase-space distributions.
Positive P-representation
-------------------------
We first consider the positive P-representation [@Drummond1980]. This is a probabilistic phase-space representation widely used in quantum optics. It is most suitable when using photonic methods to obtain qubit observables, as it can represent any multi-mode bosonic quantum state. With this representation, a general quantum density matrix $\hat{\rho}$ is represented using a positive distribution $P\left(\vec{\alpha},\vec{\beta}\right)$, where:
$$\hat{\rho}=\int P\left(\vec{\alpha},\vec{\beta}\right)\hat{\Lambda}\left(\vec{\alpha},\vec{\beta}\right)d^{2M}\vec{\alpha}d^{2M}\vec{\beta}.$$
Here the projector $\hat{\Lambda}$ is: $$\hat{\Lambda}\left(\vec{\alpha},\vec{\beta}\right)=\frac{\left|\vec{\alpha}\right\rangle \langle\vec{\beta}^{*}\vert}{\langle\vec{\beta}^{*}\vert\vec{\alpha}\rangle},$$ where $\left|\vec{\alpha}\right\rangle =\left|\alpha_{1}.\ldots\alpha_{n}\right\rangle $ is a multi-mode coherent state.
This representation maps quantum states into $4M$ real coordinates: **$\vec{\alpha},\vec{\beta}$**, which is twice the dimension of a classical phase-space. The expectation of any normally ordered observable $\hat{O}\equiv O(\hat{a}_{1}^{\dagger},\hat{a}_{1},\ldots)$ is then: $$\left\langle \hat{O}\right\rangle =\int O(\beta_{1},\alpha_{1},\ldots)P(\vec{\alpha},\vec{\beta})d^{2M}\vec{\alpha}\, d^{2M}\vec{\beta}.\label{+Pcorrels}$$
A general, although non-unique positive construction is: $$\begin{split}P(\vec{\alpha},\vec{\beta})= & \frac{\left\langle \vec{\mu}\right|\widehat{\rho}\left|\vec{\mu}\right\rangle }{\left(2\pi\right)^{2M}}e^{-\left|\vec{\nu}\right|^{2}}\end{split}
\,,\label{eq:P-from-rho}$$ where we have made a variable change to sum and difference variables: $$\vec{\nu}=\left(\vec{\alpha}-\vec{\beta}^{*}\right)/2,\quad\vec{\mu}=\left(\vec{\alpha}+\vec{\beta}^{*}\right)/2\,.\label{eq:P-variable-change}$$
Spin state representation
-------------------------
The natural choice for up- and down-states are spin states $\vert\uparrow\rangle\equiv\vert10\rangle$, $\vert\downarrow\rangle\equiv\vert01\rangle$. Spin operators can be mapped into bosons with the Schwinger representation [@Biedenharn1965]: $$\begin{aligned}
\sigma_{j}^{x} & = & \hat{a}_{j}^{\prime\dagger}\hat{a}_{j}^{\prime\prime}+\hat{a}_{j}^{\prime\prime\dagger}\hat{a}_{j}^{\prime},\nonumber \\
\sigma_{j}^{y} & = & \frac{1}{i}\left(\hat{a}_{j}^{\prime\dagger}\hat{a}_{j}^{\prime\prime}-\hat{a}_{j}^{\prime\prime\dagger}\hat{a}_{j}^{\prime}\right),\nonumber \\
\sigma_{j}^{z} & = & \hat{a}_{j}^{\prime\dagger}\hat{a}_{j}^{\prime}-\hat{a}_{j}^{\prime\prime\dagger}\hat{a}_{j}^{\prime\prime},\end{aligned}$$ where $\hat{a}_{j}^{\prime\dagger}$ creates a particle in the first position of the $j$-th spin operator, and $\hat{a}_{j}^{\prime\prime\dagger}$ creates one in the second position. Substituting $\hat{\rho}=\vert\Phi\rangle\langle\Phi\vert$ into (\[eq:P-from-rho\]) and performing the substitution (\[eq:P-variable-change\]), we get the following positive-P function: $$P=\frac{1}{2\pi^{4M}}e^{-\vert\vec{\nu}\vert^{2}}e^{-\vert\vec{\mu}^{\prime}\vert^{2}-\vert\vec{\mu}^{\prime\prime}\vert^{2}}\left|\prod_{j=1}^{M}\mu_{j}^{\prime}+e^{-i\phi}\prod_{j=1}^{M}\mu_{j}^{\prime\prime}\right|^{2}.$$
To sample this distribution, we use the von Neumann rejection method, which requires a known reference distribution as an upper bound. This distribution is bounded above by the following expression: $$P\le2G(\vec{\nu})P_{0}(\vec{\mu})\,,$$ where: $$G(\vec{\nu})=\frac{1}{\pi^{2M}}e^{-\vert\vec{\nu}\vert^{2}},$$ and $$P_{0}=\frac{1}{2\pi^{2M}}e^{-\vert\vec{\mu}^{\prime}\vert^{2}-\vert\vec{\mu}^{\prime\prime}\vert^{2}}\left(\prod_{j=1}^{M}\left|\mu_{j}^{\prime}\right|^{2}+\prod_{j=1}^{M}\left|\mu_{j}^{\prime\prime}\right|^{2}\right).$$
These two reference distributions can be sampled exactly using a combination of Gamma and Gaussian variates. The expectation of the Mermin operator $\hat{A}$ of interest here is then given by: $$\begin{aligned}
\langle\hat{A}\rangle & = & \int d^{4M}\vec{\alpha}d^{4M}\vec{\beta}P\left(\vec{\mu}(\vec{\alpha},\vec{\beta}),\vec{\nu}(\vec{\alpha},\vec{\beta})\right)\nonumber \\
& & \times\prod_{j=1}^{M}\left(\left(\beta_{j}^{\prime}\alpha_{j}^{\prime\prime}+is_{j}\beta_{j}^{\prime\prime}\alpha_{j}^{\prime}\right)e^{-is_{j}\theta_{j}}\right).\end{aligned}$$
While this method is able to sample the required GHZ state, the sampling is rather inefficient. We can improve the results using a more compact Hilbert space mapping technique, described in the next subsection.
Number state representation
---------------------------
Sampling is generally improved if the Hilbert space dimension is reduced as far as possible, to eliminate samples that overlap the unused part of the space. We can decrease the number of dimensions in the required phase space by half, by using number states instead of spin states. This is possible because we really only need the fact that occupations are binary.
![Correlations for the different parts of the quantity (\[eq:F-2-particles-1\]) in the positive-P representation, with the number state method and $2^{26}$ samples.\[fig:Spin-moments-P\]](ghz_distribution_P1 "fig:")\
![Correlations for the different parts of the quantity (\[eq:F-2-particles-1\]) in the positive-P representation, with the number state method and $2^{26}$ samples.\[fig:Spin-moments-P\]](ghz_distribution_P2 "fig:")
This can be done because our operators of interest $\hat{A}$ depend on $\sigma_{j}^{x}$ and $\sigma_{j}^{y}$ linearly. Therefore if we denote $\vert\uparrow\rangle\equiv\vert1\rangle$, $\vert\downarrow\rangle\equiv\vert0\rangle$, we can formally write: $$\begin{aligned}
\sigma_{x}^{j} & = & \hat{a}_{j}+\hat{a}_{j}^{\dagger},\nonumber \\
\sigma_{y}^{j} & = & \frac{1}{i}\left(\hat{a}_{j}^{\dagger}-\hat{a}_{j}\right),\nonumber \\
\sigma_{z}^{j} & = & \hat{a}_{j}-\hat{a}_{j}^{\dagger}.\end{aligned}$$ One can verify that, for instance, $\vert0\rangle+\vert1\rangle$ is an eigenstate of $\sigma_{x}$: $$\langle\Phi\vert\sigma_{x}\left(\vert0\rangle+\vert1\rangle\right)=\langle\Phi\vert\left(\vert0\rangle+\vert1\rangle+\vert2\rangle\right)=\langle\Phi\vert\left(\vert0\rangle+\vert1\rangle\right)\,.$$
Just as in the previous subsection, substituting $\hat{\rho}=\vert\Phi\rangle\langle\Phi\vert$ into (\[eq:P-from-rho\]) and performing the substitution (\[eq:P-variable-change\]), we get the positive-P function: $$\begin{aligned}
P & = & \frac{1}{2\pi^{2M}}e^{-\vert\vec{\nu}\vert^{2}}e^{-\vert\vec{\mu}\vert^{2}}\nonumber \\
& & \times\left(1+\prod_{j=1}^{M}\mu_{j}^{*}\mu_{j}+e^{-i\phi}\prod_{j=1}^{M}\mu_{j}^{*}+e^{i\phi}\prod_{j=1}^{M}\mu_{j}\right)\nonumber \\
& = & \frac{1}{2\pi^{2M}}e^{-\vert\vec{\nu}\vert^{2}}e^{-\vert\vec{\mu}\vert^{2}}\left|\prod_{j=1}^{M}\mu_{j}+e^{-i\phi}\right|^{2}.\end{aligned}$$
The target distribution can be sampled using von Neumann rejection sampling: $$P\le2G(\vec{\nu})P_{0}(\vec{\mu}),$$ where the reference distributions are now: $$G(\vec{\nu})=\frac{1}{\pi^{2M}}e^{-\vert\vec{\nu}\vert^{2}},$$
$$P_{0}=\frac{1}{2\pi^{M}}e^{-\vert\vec{\mu}\vert^{2}}\left(\left|\prod_{j=1}^{M}\mu_{j}\right|^{2}+1\right).$$
In this representation the expectation of the target operator is: $$\begin{aligned}
\langle\hat{A}\rangle & = & \int d^{2M}\vec{\alpha}d^{2M}\vec{\beta}P\left(\vec{\mu}(\vec{\alpha},\vec{\beta}),\vec{\nu}(\vec{\alpha},\vec{\beta})\right)\nonumber \\
& & \times\prod_{j=1}^{M}\left(\left(\alpha_{j}+is_{j}\beta_{j}\right)e^{-is_{j}\theta_{j}}\right).\end{aligned}$$
In Figure (\[fig:Spin-moments-P\]), we show the distribution of results with the positive-P representations, for a portion of the Ardehali inequality for the case $M=2$, given by: $$F_{XY}=-\langle\hat{\sigma}_{1}^{x}\hat{\sigma}_{2}^{x}\rangle+\langle\hat{\sigma}_{1}^{y}\hat{\sigma}_{2}^{y}\rangle.\label{eq:F-2-particles-1}$$ In a LHV theory the values of $\mathrm{Re}\,\sigma_{x}^{1}$ and $\mathrm{Re}\,\sigma_{x}^{2}$ are limited to the range $[-1,1]$, but clearly our results are not limited to that range. This essential feature means that Bell’s theorem does not limit our results, because the sampled values are not the same as their physical eigenvalues [@Reid1986-violations]. The connection with weak values [@Aharonov1988]has been discussed in a previous paper [@Rosales-Zarate2014]. This demonstrates an essential feature of this phase-space representation: it is analogous to a weak-value measurement, giving results outside the normal range of the eigenvalues.
We note that this is still a positive phase-space representation, valid in a more limited subspace than before, but certainly able to represent the GHZ state. However, both approaches have the drawback that they use methods designed to represent infinite dimensional Hilbert spaces, which is not a good match to the GHZ state requirements.
\[sec:Sampling-GHZ-Qfunction\]Sampling GHZ states with the Q-function
=====================================================================
The Hilbert space occupied by the GHZ state is a finite-dimensional Hilbert space for which such an infinite-dimensional bosonic mapping is not strictly necessary. Next we turn to methods that are more suited to the task of representing finite dimensional states. Our interest in doing this is to determine if this can improve the sampling properties.
The Q-function for bosons was first introduced by Husimi [@Husimi1940] as an expectation value of the density matrix in an over-complete coherent-state basis. It gives a mapping of a general many-body density matrix into a unique, positive distribution. This method has been widely used as a method to probabilistically represent statistical properties in quantum optics. It has had a diverse range of applications, mostly in tomography.
The same technique can be used to define a multipartite Q-function based on SU(2) coherent states, as an alternative and more efficient means of phase-space sampling for qubits.
The SU(2) Q-function
--------------------
For purposes of calculations, we will consider as the basis set an un-normalized version of the SU(2) coherent states [@Arecchi1972; @Radcliffe1971; @Zhang1990] defined as: $$\left\Vert \vec{z}\,\right\rangle =\prod_{j=1}^{M}\left(\left|0\right\rangle _{j}+z_{j}\left|1\right\rangle _{j}\right).\label{eq:un-normalized_SU2CS-1}$$
In terms of this un-normalized state the resolution of unity is given by: $$\int d^{2}\vec{z}\left(\prod_{j=1}^{M}{\cal N}\left(\left|z_{j}\right|^{2}\right)\right)\left\Vert \vec{z}\,\right\rangle \left\langle \vec{z}\,\right\Vert =\hat{1}.\label{eq:un-normResUnity_SU2CS-1}$$ Here we have defined the normalization factor ${\cal N}\left(\left|z_{j}\right|^{2}\right)$ as: $${\cal N}\left(\left|z_{j}\right|^{2}\right)=\frac{2}{\pi}\frac{1}{\left(1+\left|z_{j}\right|^{2}\right)^{3}}.$$ Using the resolution of unity for the un-normalized SU(2) coherent states (\[eq:un-normResUnity\_SU2CS-1\]), we can define a Q-function: $$Q(\vec{z})=\left[\prod_{j=1}^{M}{\cal N}\left(\left|z_{j}\right|^{2}\right)\right]\left\langle \vec{z}\,\right\Vert \hat{\rho}\left\Vert \vec{z}\,\right\rangle ,\label{eq:Qf_un-norm_SU2CS-1}$$ which has the property that: $$\int d^{2}\vec{z}Q(\vec{z})=1.$$ This Q-function is positive definite and is defined for any quantum density matrix and is normalized to one.
In our GHZ state of interest (\[eq:ghz-full\]) we denote $\left|\downarrow\right\rangle =\left|0\right\rangle $ and $\left|\uparrow\right\rangle =\left|1\right\rangle $. Hence for the density matrix $\hat{\rho}=\vert\Phi\rangle\langle\Phi\vert$ we obtain: $$\begin{aligned}
\left\langle \vec{z}\,\right\Vert \hat{\rho}\left\Vert \vec{z}\,\right\rangle & = & \left\langle \vec{z}\,\right\Vert \vert\Phi\rangle\langle\Phi\vert\left\Vert \vec{z}\,\right\rangle =\left|\left\langle \vec{z}\,\right\Vert \vert\Phi\rangle\right|^{2}\nonumber \\
& = & \frac{1}{2}\left|\prod_{j}\left(_{j}\left\langle 0\right|+z_{j}^{*}\,_{j}\left\langle 1\right|\right)\left(\vert1\ldots1\rangle+e^{i\phi}\vert0\ldots0\rangle\right)\right|^{2}\nonumber \\
& = & \frac{1}{2}\left|\prod_{j}z_{j}+e^{-i\phi}\right|^{2}.\end{aligned}$$ Therefore the Q-function for our states of interest is: $$\begin{aligned}
Q(\vec{z}) & = & \frac{1}{2}\left(\frac{2}{\pi}\right)^{M}\prod_{j=1}^{M}\frac{1}{\left(1+\left|z_{j}\right|^{2}\right)^{3}}\left|\prod_{j}z_{j}+e^{-i\phi}\right|^{2}.\end{aligned}$$
The expectation value of $\hat{A}$ can be expressed in terms of the Q-function using (\[eq:Q-spin-moments\]), the details of the evaluations are shown in the next section, and the fact that $\sigma_{x}^{j}=2\hat{S}_{x}^{j}=\hat{S}_{+}^{j}+\hat{S}_{-}^{j}$ and $\sigma_{y}^{j}=2\hat{S}_{y}^{j}=(\hat{S}_{+}^{j}-\hat{S}_{-}^{j})/i$, hence: $$\hat{A}=\prod_{j=1}^{M}\left(\left((1+s_{j})\hat{S}_{+}^{j}+(1-s_{j})\hat{S}_{-}^{j}\right)e^{-is_{j}\theta_{j}}\right).$$ Therefore the expectation value of the target operator using the Q-function is: $$\begin{aligned}
\langle\hat{A}\rangle & = & \langle\Phi\vert\hat{A}\vert\Phi\rangle\nonumber \\
& = & \int d\vec{z}Q(\vec{z})\frac{3^{M}}{\prod_{j}\left(1+\left|z_{j}\right|^{2}\right)}\\
& & \times\prod_{j=1}^{M}\left(\left((1+s_{j})z_{j}^{*}+(1-s_{j})z_{j}\right)e^{-is_{j}\theta_{j}}\right).\nonumber \end{aligned}$$
Evaluation of moments\[sec:Appendix\_EvaluationMoments-1\]
----------------------------------------------------------
In this section we show the evaluation of the moments of the form $\left\langle \prod_{j}\hat{S}_{d_{j}}^{j}\right\rangle $ with directions $d_{j}\in\left\{ -,+\right\} $, in terms of the SU(2) Q-function. In order to evaluate the moments, we notice that we can express the action of the raising spin operators on the un-normalized SU(2) coherent state $\left\Vert \mathbf{z}\right\rangle $ as a derivative of the SU(2) coherent state $\left\Vert \vec{z}\right\rangle $, so that: $$\begin{aligned}
\hat{S}_{+}^{j}\left\Vert \vec{z}\right\rangle & = & \hat{S}_{+}^{j}\left(\prod_{j}e^{\hat{S}_{+}^{j}z_{j}}\left|0\right\rangle _{j}\right)\nonumber \\
& = & \frac{\partial}{\partial z_{j}}\left(\prod_{j}e^{\hat{S}_{+}^{j}z_{j}}\left|0\right\rangle _{j}\right)\nonumber \\
& = & \frac{\partial}{\partial z_{j}}\left\Vert \vec{z}\right\rangle .\end{aligned}$$
Similarly, there is a conjugate expression: $$\begin{aligned}
\left\langle \vec{z}\right\Vert \hat{S}_{-}^{j} & = & \left(\left\langle 0\right|e^{\hat{\bm{S}}_{-}\cdot\vec{z}}\right)\hat{S}_{-}^{j}\nonumber \\
& = & \frac{\partial}{\partial z_{j}^{*}}\left(\prod_{j}\,_{j}\left\langle 0\right|e^{\hat{S}_{-}^{j}z_{j}}\right)\nonumber \\
& = & \frac{\partial}{\partial z_{j}^{*}}\left\langle \vec{z}\right\Vert ,\end{aligned}$$ while for the $z$-direction one obtains: $$\begin{aligned}
\hat{S}_{z}^{j}\left\Vert \mathbf{z}\right\rangle & = & \hat{S}_{z}^{j}\left(\prod_{j}e^{\hat{S}_{+}^{j}z_{j}}\left|0\right\rangle _{j}\right)\nonumber \\
& = & \prod_{j}\frac{1}{2}\left(-\left|0\right\rangle _{j}+z_{j}\left|1\right\rangle _{j}\right)\nonumber \\
& = & \prod_{j}\frac{1}{2}\left(2z_{j}\frac{\partial}{\partial z_{j}}-1\right)\left\Vert \mathbf{z}\right\rangle .\end{aligned}$$ Here we have used that: $$\begin{aligned}
z_{j}\frac{\partial}{\partial z_{j}}\left(\left|0\right\rangle _{j}+z_{j}\left|1\right\rangle _{j}\right) & = & z_{j}\left|1\right\rangle _{j},\end{aligned}$$ and hence the last identity above is obtained from: $$\begin{aligned}
2z_{j}\frac{\partial}{\partial z_{j}}\left\Vert \mathbf{z}\right\rangle -\left\Vert \mathbf{z}\right\rangle & = & z_{j}\left|1\right\rangle _{j}-\left|0\right\rangle _{j}.\end{aligned}$$
Next, we evaluate the moments of the spin operators $\hat{S}_{+}^{j}$, $\hat{S}_{-}^{j}$ and $\hat{S}_{z}^{j}$ using the resolution of unity (\[eq:un-normResUnity\_SU2CS-1\]) as well as the definition of the Q-function (\[eq:Qf\_un-norm\_SU2CS-1\]) so that: $$\begin{aligned}
\left\langle \prod_{j}\hat{S}_{d_{j}}^{j}\right\rangle & = & {\rm Tr}\left[\hat{\rho}\prod_{j}\hat{S}_{d_{j}}^{j}\right]\\
& = & \int d^{2}\vec{z}\prod_{j}{\cal N}\left(\left|z_{j}\right|^{2}\right)\nonumber \\
& & \times\left(\prod_{d_{j}=-}\frac{\partial}{\partial z_{j}^{*}}\right)\left\langle \vec{z}\right\Vert \hat{\rho}\left(\prod_{d_{j}=+}\frac{\partial}{\partial z_{j}}\right)\left\Vert \vec{z}\right\rangle \nonumber \end{aligned}$$ Integrating by parts for each $j$, providing that the boundary terms vanish, we get: $$\begin{aligned}
\left\langle \prod_{j}\hat{S}_{d_{j}}^{j}\right\rangle & = & (-1)^{M}\int d^{2}\vec{z}\left\langle \vec{z}\right\Vert \hat{\rho}\left\Vert \vec{z}\right\rangle \label{eq:Q-spin-moments}\\
& & \times\prod_{j,d_{j}=-}\frac{\partial{\cal N}\left(\left|z_{j}\right|^{2}\right)}{\partial z_{j}^{*}}\prod_{j,d_{j}=+}\frac{\partial{\cal N}\left(\left|z_{j}\right|^{2}\right)}{\partial z_{j}}\nonumber \\
& = & \int d^{2}\vec{z}Q(\vec{z})\frac{3^{M}}{\prod_{j}\left(1+\left|z_{j}\right|^{2}\right)}\prod_{j,d_{j}=-}z_{j}\prod_{j,d_{j}=+}z_{j}^{*}.\nonumber \\
& & \,\nonumber \end{aligned}$$ Here we have used that the derivative of ${\cal N}\left(\left|z_{j}\right|^{2}\right)$ is: $$\begin{aligned}
\frac{1}{{\cal N}\left(\left|z_{j}\right|^{2}\right)}\frac{\partial}{\partial z_{j}}{\cal N}\left(\left|z_{j}\right|^{2}\right) & = & \frac{2}{\pi{\cal N}\left(\left|z_{j}\right|^{2}\right)}\frac{\partial}{\partial z_{j}}\frac{1}{\left(1+\left|z_{j}\right|^{2}\right)^{3}}\nonumber \\
& = & \frac{-3z_{j}^{*}}{\left(1+\left|z_{j}\right|^{2}\right)}.\label{eq:derivative_Norm_nmode-1}\\
& & \,\nonumber \end{aligned}$$
Here, the results of Q-function sampling of the GHZ state are presented. Firstly we show the results of the difference between the calculations with the positive-P and SU(2)-Q representations, using a portion of the Ardehali inequality for the case $M=2$, given as previously by: $$F_{XY}=-\langle\hat{\sigma}_{1}^{x}\hat{\sigma}_{2}^{x}\rangle+\langle\hat{\sigma}_{1}^{y}\hat{\sigma}_{2}^{y}\rangle.\label{eq:F-2-particles}$$
![Correlations for the different parts of the quantity (\[eq:F-2-particles\]) in the SU(2)-Q representation, $2^{26}$ samples.\[fig:Spin-moments-Q\]](ghz_distribution_Q1 "fig:")\
![Correlations for the different parts of the quantity (\[eq:F-2-particles\]) in the SU(2)-Q representation, $2^{26}$ samples.\[fig:Spin-moments-Q\]](ghz_distribution_Q2 "fig:")
In Fig. \[fig:Spin-moments-Q\](a) we show the correlation between the real parts of $\hat{\sigma}_{i}^{x}$, $i=1,\:2$ of the quantity $F_{XY}$, while the correlation between the two terms of (\[eq:F-2-particles\]) is plotted in Fig. \[fig:Spin-moments-Q\](b). Once again, this method is analogous to a weak-value measurement, giving results outside the normal range of the eigenvalues.
For the spin operator $\hat{S}_{z}^{j}$ we obtain: $$\begin{aligned}
\left\langle \prod_{j}\hat{S}_{z}^{j}\right\rangle & = & {\rm Tr}\left[\prod_{j}\hat{\rho}\hat{S}_{z}^{j}\right]\\
& = & \prod_{j}\int d^{2}z_{j}{\cal N}\left(\left|z_{j}\right|^{2}\right)\left\langle \mathbf{z}\right\Vert \hat{\rho}\left(z_{j}\frac{\partial}{\partial z}-\frac{1}{2}\right)\left\Vert \mathbf{z}\right\rangle \,.\nonumber \end{aligned}$$ Next we use the following result, which also involves partial integration: $$\begin{aligned}
& & \prod_{j}\int d^{2}z_{j}{\cal N}\left(\left|z_{j}\right|^{2}\right)\left\langle \mathbf{z}\right\Vert \hat{\rho}z_{j}\frac{\partial}{\partial z}\left\Vert \mathbf{z}\right\rangle \nonumber \\
& & =-\prod_{j}\int d^{2}z_{j}\left\langle \mathbf{z}\right\Vert \hat{\rho}\left\Vert \mathbf{z}\right\rangle \frac{\partial}{\partial z_{j}}\left(z_{j}{\cal N}\left(\left|z_{j}\right|^{2}\right)\right)\nonumber \\
& & =\prod_{j}\int d^{2}z_{j}Q\left(\mathbf{z}\right)\left(\frac{3\left|z_{j}\right|^{2}}{\left(1+\left|z_{j}\right|^{2}\right)}-1\right)\,,\end{aligned}$$ and leads to our final spin operator identity, $$\left\langle \prod_{j}\hat{S}_{z}^{j}\right\rangle =\prod_{j}\frac{3}{2}\int d^{2}z_{j}Q\left(\mathbf{z}\right)\left(\frac{\left|z_{j}\right|^{2}-1}{\left|z_{j}\right|^{2}+1}\right)\,.$$
\[sec:Results\]Multipartite Bell violation results
==================================================
To simulate multipartite Bell violations, the GHZ state (\[eq:ghz-full\]) was sampled using probabilistic random number generators using both Q-function and positive P-distribution methods. Of the two positive P-distribution mappings, the Schwinger representation method is less compact, and has a larger sampling error. For the results graphed here, we therefore chose the number state positive-P distribution. Although more efficient than the Schwinger representation, this still has a large basis set that corresponds to an infinite dimensional Hilbert space, with a much larger dimension than is needed for the GHZ state.
The lowest sampling errors were obtained with the $SU(2)$ Q-distribution method, which uses a much more compact Hilbert space, having a dimension equal to the physical qubit dimension.
Multipartite sampling error properties
--------------------------------------
We initially investigate the scaling properties of the sampling errors as the number of qubits $M$ is varied. This also determines the time taken for the simulation to reach a predetermined error, since one can include more parallel samples to reduce the simulated errors to any desired level.
First we consider the scaling with system-size of the sampling errors for *single* measurements of a low-order spin correlation (Fig. \[fig:GHZ-violations-1\]). For low-order correlation we have chosen the total number of “spin-ups” $N=\langle\sum_{j=1}^{M}\left(\hat{\sigma}_{z}^{j}+1\right)/2\rangle$. In this case we noticed that the sampling errors *decreases* as $M$ increases.
![Scaling properties for sampled correlations of multi-particle GHZ states. Relative errors are plotted for high order ($V$) correlations, (blue line) and first order correlations, or total number of “spin-ups” (green dashed line) using the SU(2)-Q representation with $2^{40}$ samples. The dotted reference line shows the point at which the sampling errors would give scaling properties of an experimental measurement. The red dotted line shows the scaling of the $V$ correlations using the less efficient positive-P representation.\[fig:GHZ-violations-1\]](ghz_errors)
In contrast to this, high-order correlations showed exponentially increasing sampling error. The relative error in $V$ scales as $2^{M/3}$, meaning that the time taken at constant error scales as $2^{2M/3}$. This means that probabilistic sampling scales more favorably than experiment, which would take time in proportion to $2^{M}$. Therefore, the sampling takes place in times that scale $2^{M/3}$ times faster than any possible experiment.
In practical terms, such laboratory measurements would be highly nontrivial, due to the need to eliminate background noise for high-order correlations. No correlation measurements of this size have been reported to date. Experimentally, it is possible that such high-order correlations will be reported in future.
Even then, it is likely that one may only be able to measure a subset of all the high-order correlations possible for large $M$ values. This is because of the enormous time required to make all possible correlation measurements for these inequalities, which is exponentially slower than the phase-space simulation.
Simulations of multipartite genuine Bell violations
---------------------------------------------------
In Fig. \[fig:GHZ-violations\] we show the expectation value of the multipartite, multi-measurement quantity $V$ compared with the quantum mechanical prediction $\left\langle V\right\rangle _{QM}$ from sampling the $SU(2)$ Q-distribution. The dashed line is the minimum correlation required to demonstrate a Bell violation, with a number of qubits ranging from $M=2$ to $M=60$. For all cases we verified clear Bell violations to at least $12$ standard deviations from the classical limit.
![Violations for multi-particle GHZ states. Simulated Mermin violation using SU(2)-Q representation with $2^{43}$ samples. The values of expectations and errors are normalized by the quantum mechanical prediction for the corresponding $M$. The horizontal grey dashed line gives the quantum prediction. The error bars show the sampled result and estimated sampling errors at each value of $M$. The dash-dotted line is the LHV prediction, which gives a Bell violation when above this line. Genuine multipartite Bell violations occur for even $M$ when $V/V_{\mathrm{QM}}>1/\sqrt{2}$.\[fig:GHZ-violations\]](ghz_violations "fig:")\
Genuine multipartite violations of LHV, requiring all $M$ observers to participate, were verified for even $M$ to at least $4$ standard deviations. These cases all satisfied the more stringent requirement that: $$V/V_{\mathrm{QM}}>1/\sqrt{2}\,.$$
The simulations were carried out using graphical processor unit (GPU) technology at a clock speed of $1.2$ GHz, which allowed calculations with $50$ GPUs on $22,000$ parallel computational cores. The plotted results correspond in the $60$ qubit case to simulating the results of a quintillion ($10^{18}$) distinct sixtieth order correlation functions. This took less than $48$ hours. A reasonable estimate of the laboratory time-scale for carrying out all possible correlation measurements, at $10^{-3}s$ per measurement setting, is $3\times10^{7}$ years. This is more than $10^{9}$ times slower than the simulations.
\[sec:Decoherence-model\]Decoherence simulations
=================================================
We have shown that it is possible to simulate genuine Bell violations, as well as obtaining scaling laws for GHZ states using phase space methods. But we can also ask whether it is also possible to use the positive phase-space methods to simulate decoherence processes? In order to answer this question, here we will focus on the question of the study of dynamical noise and decoherence in ion traps, which is an important issue in the observation of mesoscopic quantum effects [@Brune1996]. Ion traps have been widely used in order to create entangled states and also to investigate the decay rate of GHZ states [@Monz2011].
Here we will follow the noise model of Monz et al. [@Monz2011], which was used to explain the observed super-decoherence found in ion-trap experiments. This is physically due to the fact that the magnetic field noise reservoir is correlated over all the qubits. As a result, they do not decohere with independent noise or error sources, as is often assumed theoretically. To model this, we assume a delta-correlated magnetic field noise which is shared by all the ions, such that $$\left\langle \Delta B(t)\Delta B(t')\right\rangle =\Delta B_{0}^{2}\delta(t-t')\,.$$ In this case we assume that the interaction or noise Hamiltonian is: $$\hat{H}=\frac{\mu\Delta B(t)}{2}\sum_{j=1}^{M}\hat{\sigma}_{z}^{j}.\label{eq:H-decoherence}$$ This model can be simulated dynamically multiplying, in each of the samples after every time step $\Delta t$, an independent noise term $\exp\left(i\epsilon N\zeta_{j}\right)$ by the value corresponding to the operator $\prod_{j}^{M}\left(\hat{\sigma}_{x}^{j}+\hat{\sigma}_{y}^{j}\right)$. We use the respectively measurement choice $V$ of (\[eq:V\_Op\]) for odd $M$ and even $M$. Here $$\epsilon=\mu\Delta B_{0}\sqrt{\Delta t}/\hbar$$ defines the speed of the decoherence, and $\zeta_{j}$ is a Gaussian random number such that $$\left\langle \zeta_{j}\zeta_{j'}\right\rangle =\delta_{jj'}\,.$$ The results of the simulation are shown in Fig. \[fig:Decoherence\]. This demonstrates the experimentally observed quadratic decoherence, with decay times scaling with $1/M^{2}$ as $M$ increases over a range comparable to current experiments, therefore showing the effect of super-decoherence.
![Decay of the sampled quantity $V$ using the model of super-decoherence of Eq (\[eq:H-decoherence\]), for 2 (solid blue line), 3 (red dashed line), 4 (green dash-dotted line) and 6 (yellow dotted line) particles, with decoherence rate $\epsilon=0.1$. The horizontal axis is the dimensionless time, $\tau=t/\Delta t$.\[fig:Decoherence\]](ghz_decoherence)
\[sec:Conclusions\]Conclusions
==============================
Our main result is that it is possible to sample quantum events probabilistically, even when they display macroscopic quantum paradoxes and Schrödinger cat behavior. This is not prohibited by the Bell inequality, although we use standard digital computers. Our calculations generate a distribution equivalent to the observables predicted by quantum mechanics. These results demonstrate the potential for phase-space methods to simulate macroscopic quantum superpositions.
We have demonstrated genuine multipartite Bell inequalities with up to $60$ qubits. In all cases we have shown violations of these inequalities using positive phase space distributions. We interested in the question of whether or not probabilistic sampling can be carried out for GHZ states. This is not obvious a-priori, since one might expect highly nonclassical states to be hard to sample probabilistically. We also have performed dynamical simulations of super-decoherence.
These results demonstrate that the simulation of both low and high order correlations is feasible, despite Bell violations. Some reasonable conclusions about the advantages and limitations of these methods are as follows. Probabilistic phase-space algorithms appear well suited to low order correlations, including fundamentally nonclassical low order Bell inequality violations. Higher order correlations generate larger sampling errors with a probabilistic approach.
We also find a classical parallelism which gives an an unexpected exponential speed-up for qubit sampling, when calculating all the high order correlations required for multipartite Bell violations. Here the speed-up is relative to the corresponding experimental times, and is exponential in the qubit number. This uniquely useful feature of probabilistic phase-space methods is due to their ability to simultaneously calculate many non-commuting observables in parallel.
Such classical *measurement* parallelism is complementary to the *state* parallelism of quantum mechanics. We have utilized this in the calculation of the MABK function, to indicate genuine violation of multipartite Bell inequalities.
We wish to acknowledge research funding from the Australian Research Council, as well as useful discussions with P. Deuar, R. Blatt and B. Lanyon.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'All binary self-dual $[44,22,8]$ codes with an automorphism of order $3$ or $7$ are classified. In this way we complete the classification of extremal self-dual codes of length 44 having an automorphism of odd prime order.'
author:
- |
Stefka Bouyuklieva,\
Faculty of Mathematics and Informatics,\
Veliko Tarnovo University, Bulgaria,\
Nikolay Yankov and Radka Russeva,\
Faculty of Mathematics and Informatics,\
Shumen University, Bulgaria\
title: ' On the classification of binary self-dual $[44,22,8]$ codes with an automorphism of order $3$ or $7$'
---
Introduction
============
Let $\mathbb{F}_q^n$ be the $n$-dimensional vector space over the field $\mathbb{F}_q$ of $q$ elements. A *linear $[n,k]$ code* $C$ is a $k$-dimensional subspace of $\mathbb{F}_q^n$. The elements of $C$ are called *codewords*. The *weight* of a vector $v\in\mathbb{F}_q^n$ (denoted by $\mbox{wt}(v)$) is the number of its non-zero coordinates. The *minimum weight* $d$ of $C$ is the smallest weight among all nonzero weights of codewords of $C$; a code $C$ with minimum weight $d$ is called an $[n,k,d]$ code. A matrix whose rows form a basis of $C$ is called a *generator matrix* of this code. The weight enumerator $W(y)$ of a code $C$ is given by $W(y)=\sum_{i=0}^n A_iy^i$ where $A_i$ is the number of codewords of weight $i$ in $C$. Two binary codes are called *equivalent* if one can be obtained from the other by a permutation of coordinates. The permutation $\sigma\in S_n$ is an *automorphism* of $C$, if $C=\sigma(C)$ and the set of all automorphisms of $C$ forms a group called the *automorphism group* of $C$, which is denoted by $Aut(C)$ in this paper.
Let $(u,v)\in\mathbb{F}_q$ for $u,v\in\mathbb{F}_q^n$ be an inner product in $\mathbb{F}_q^n$. The *dual code* of an $[n,k]$ code $C$ is $C^{\perp}=\{u \in \mathbb{F}_q^n \mid (u,v)=0$ for all $v \in C \}$ and $C^{\perp}$ is a linear $[n,n-k]$ code. If $C
\subseteq C^{\perp}$, $C$ is termed *self-orthogonal*, and if $C = C^{\perp}$, $C$ is self-dual. We call a binary code *self-complementary* if it contains the all-ones vector. Every binary self-dual code is self-complementary. If $u=(u_1,\cdots, u_n)$, $v = (v_1,\cdots,v_n)\in\mathbb{F}_2^n$ then $(u,v)=\sum_{i=1}^{n}{u_iv_i}\in\mathbb{F}_2$. It was shown in [@Rains98] that the minimum weight $d$ of a binary self-dual code of length $n$ is bounded by $d \le 4[n/24]+4$, unless $n \equiv 22 \pmod{24}$ when $d \le 4[n/24]+6$. We call a self-dual code meeting this upper bound [*extremal*]{}.
In this paper, we consider extremal binary self-dual $[44,22,8]$ codes. All the odd primes $p$ dividing the order of the automorphism group of such a code are 11, 7, 5, and 3 [@160]. The codes with automorphisms of order 11 and 5 are classified in [@160], [@164], [@St42-44], and [@Stef-SO]. Unfortunately we noticed that there are some omissions in the classification of the codes with automorphisms of order 7 given in [@RN-SMB'08]. That’s why we focus on the automorphisms of orders 3 and 7, and we complete the classification of $[44,22,8]$ self-dual codes having an automorphism of odd prime order.
As in the case of binary self-dual $[42,21,8]$ codes with an automorphism of order 3, there are five different possibilities for the number of independent cycles in the decomposition of the automorphism, namely 6, 8, 10, 12, and 14 [@BYR]. Codes with automorphisms of order 3 with 6 and 14 independent 3-cycles are considered but not classified in [@Stef-SO] and [@Nikolay3-14], respectively. In this paper, we give the classification of all self-dual $[44,22,8]$ codes having an automorphism of order 3 or 7. To do that we apply the method for constructing binary self-dual codes via an automorphism of odd prime order developed in [@Huff] and [@Yo]. We give a short description of this method in Section 2. In Section 3 and Section 4 we classify the extremal self-dual codes of length 44 with an automorphism of order 3 and 7, respectively. In Section 5 we present the full classification of the self-dual $[44,22,8]$ codes having automorphisms of odd prime order, and offer some open problems.
The weight enumerators of the extremal self-dual codes of length $44$ are known (see [@CS]): $$W_{44,1}(y) =1 + (44 + 4\beta)y^8 + (976-8\beta)y^{10} +
(12289-20\beta)y^{12}+\dots$$ for $10\leq \beta\leq 122$ and $$W_{44,2}(y) =1 + (44 + 4\beta)y^8 + (1232-8\beta)y^{10} +
(10241-20\beta)y^{12}+\dots$$ for $0\leq \beta\leq 154$.
Codes exist for $W_{44,1}$ when $\beta= 10,\dots, 68$, 70, 72, 74, 82, 86, 90, 122 and for $W_{44,2}$ when $\beta = 0, \dots , 56$, 58, …, 62, 64, 66, 68, 70, 72, 74, 76, 82, 86, 90, 104, 154 (see [@FFTAHuff]).
Construction Method
===================
Let $C$ be a binary self-dual code of length $n=44$ with an automorphism $\sigma $ of prime order $p\geq 3$ with exactly $c$ independent $p$-cycles and $f=44-cp$ fixed points in its decomposition. We may assume that $$\sigma=(1,2,\cdots,p)(p+1,p+2,\cdots,2p)\cdots(p(c-1)+1,p(c-1)+2,\cdots,pc),$$ and say that $\sigma$ is of type $p$-$(c,f)$.
Denote the cycles of $\sigma $ by $\Omega_1,\ldots, \Omega_c$, and the fixed points by $\Omega_{c+1},\ldots, \Omega_{c+f}$. Let $F_\sigma(C)=\{ v \in C\mid v \sigma=v \}$ and $$E_\sigma(C)=\{ v\in C\mid\mbox{wt}(v\vert \Omega_i)\equiv 0\pmod
2, i=1,\cdots,c+f\},$$ where $v\vert \Omega_i $ is the restriction of $v$ on $\Omega_i $.
[[@Huff]]{} The self-dual code $C$ is a direct sum of the subcodes $F_\sigma(C)$ and $E_\sigma(C)$. These subcodes have dimensions $\frac{c+f}{2}$ and $\frac{c(p-1)}{2}$, respectively.
Thus each choice of the codes $F_\sigma(C)$ and $E_\sigma(C)$ determines a self-dual code $C$. So for a given length all self-dual codes with an automorphism $\sigma$ can be obtained.
We have that $v\in F_\sigma (C)$ if and only if $v\in C$ and $v$ is constant on each cycle. Let $\pi :F_\sigma (C)\to
\mathbb{F}_{2}^{c+f} $ be the projection map where if $v\in
F_\sigma (C)$, $(\pi(v))_i =v_j $ for some $j\in\Omega_i$, $i=1,2,\ldots,c+f $. Denote by $E_\sigma(C)^{*}$ the code $E_\sigma(C)$ with the last $f$ coordinates deleted. So $E_\sigma(C)^{*}$ is a self-orthogonal binary code of length $pc$ and dimension $c(p-1)/2$. For $v\in
E_\sigma (C)^*$ we let $v\vert\Omega_i=(v_0,v_1,\cdots,v_{p-1})$ correspond to the polynomial $v_0+v_1 x+\cdots +v_{p-1}x^{p-1}$ from ${\cal P}$, where ${\cal P}$ is the set of even-weight polynomials in ${\cal R}_p=\mathbb{F}_2[x]/\langle x^p-1\rangle$. Thus we obtain the map $\varphi:E_\sigma(C)^{*}\to {\cal P}^c $. ${\cal P}$ is a cyclic code of length $p$ with generator polynomial $x-1$. It is known that $\varphi(E_{\sigma}(C)^{*})$ is a submodule of the ${\cal P}$-module ${\cal P}^c$ [@Huff; @Yorus].
[[@Yorus]]{}\[thm2\] A binary $[n,n/2]$ code $C$ with an automorphism $\sigma$ is self-dual if and only if the following two conditions hold:
- $C_{\pi}=\pi(F_\sigma(C))$ is a binary self-dual code of length $c+f$,
- for every two vectors $u, v$ from $C_{\varphi}=\varphi(E_{\sigma}(C)^{*})$ we have $$u_1(x)v_1(x^{-1})+ u_2(x)v_2(x^{-1})+ \cdots + u_c(x)v_c(x^{-1}) =
0.$$
Let $x^p-1=(x-1)h_1(x)\dots h_s(x)$, where $h_1,\dots,h_s$ are irreducible binary polynomials. If $g_j(x)=(x^p-1)/h_j(x)$, and $I_j=\langle g_j(x)\rangle$ is the ideal in ${\cal R}_p$, generated by $g_j(x)$, then $I_j$ is a fields with $2^{deg(h_j(x))}$ elements, $j=1, 2, \dots, s$, and ${\cal P} =
I_1\oplus I_2\oplus \cdots \oplus I_s$ [@Handbook].
\[M1M2M3\] [[@Yorus]]{} Let $M_j=\{ u\in \varphi(E_\sigma(C)^{*})\vert
u_i\in I_j,i=1,2,\ldots,c\}$, $j=1, 2, \dots, s$. Then
1\) $M_j$ is a linear space over $I_j$, $j=1, 2, \dots, s$;
2\) $C_\varphi=\varphi(E_\sigma (C)^{*})=M_1\oplus M_2\oplus \cdots
\oplus M_s$ (direct sum of ${\cal P}$-submodules);
3\) If $C$ is a self-dual code, then $\sum\limits_{j =
1}^s{dim_{I_j}M_j = cs/2}$.
In the case, when 2 is a primitive root modulo $p$, ${\cal P}$ is a field with $2^{p-1}$ elements and the following theorem holds
[[@Huff]]{} \[primroot\] Let $2$ be a primitive root modulo $p$. Then the binary code $C$ with an automorphism $\sigma$ is self-dual iff the following two conditions hold:
- $C_\pi$ is a self-dual binary code of length $c + f$;
- $C_\varphi$ is a self-dual code of length $c$ over the field ${\cal P}$ under the inner product $(u,v)=\sum_{i=1}^{c}{u_iv_i^{(p-1)/2}}.$
Let ${\cal B}$, respectively ${\cal D}$, be the largest subcode of $C_\pi$ whose support is contained entirely in the left $c$, respectively, right $f$, coordinates. Suppose ${\cal B}$ and ${\cal D}$ have dimensions $k_1$ and $k_2$, respectively. Let $k_3
= k - k_1 - k_2$. Then there exists a generator matrix for $C_\pi$ in the form $$\label{matGpi}
G_{\pi} =\left(\begin{array}{cc}
B & \ O\\
O & \ D\\
E & \ F
\end{array}\right)$$ where $B$ is a $k_1\times c$ matrix with $gen ({\cal B}) = [B \
O]$, $D$ is a $k_2\times f$ matrix with $gen ({\cal D}) = [O \
D]$, $O$ is the appropriate size zero matrix, and $[E \ F]$ is a $k_3\times n$ matrix. Let ${\cal B}^*$ be the code of length $c$ generated by $B$, ${\cal B}_E$ the code of length $c$ generated by the rows of $B$ and $E$, ${\cal D}^*$ the code of length $f$ generated by $D$, and ${\cal D}_F$ the code of length $f$ generated by the rows of $D$ and $F$. The following theorem is a modification of Theorem 2 from [@Handbook-Pless].
\[structure\] With the notations of the previous paragraph
- $k_3 = \mbox{rank}(E) = \mbox{rank}(F)$,
- $k_2 = k + k_1 - c=k_1+\frac{f-c}{2}$, and
- ${\cal B}_E^{\perp} = {\cal B}^*$ and ${\cal D}_F^{\perp} = {\cal D}^*$.
Extremal Self-Dual Codes of Length 44 with an Automorphism of Order 3
=====================================================================
Using Theorem \[primroot\], as 2 is a primitive root modulo 3, ${\cal P}$ is a field with 4 elements. We have that ${\cal P}=\{0,
e=x+x^2, w=1+x^2, w^2=1+x\}\cong\mathbb{F}_4$ where $e$ is the identity of ${\cal P}$. In this case $C_\varphi$ is a (Hermitian) self-dual code of length $c$ over the quaternary field ${\cal P}$ under the inner product $(u,v)=\sum_{i=1}^{c}{u_iv_i^2}.$ Since the minimum distance of $E_\sigma(C)$ is at least 8, this Hermitian code should have minimum distance at least 4.
To classify the codes, we need additional conditions for equivalence. That’s why we use the following theorem:
[[@Yo]]{}\[thm:eq\] The following transformations preserve the decomposition and send the code $C$ to an equivalent one:
- a permutation of the fixed coordinates;
- a permutation of the 3-cycles coordinates;
- a substitution $x\rightarrow x^2$ in $C_{\varphi}$ and
- a cyclic shift to each 3-cycle independently.
Codes with an automorphism of type $3$-$(6,26)$
-----------------------------------------------
The extremal self-dual $[44,22,8]$ codes having an automorphism of type $3$-$(6,26)$ are considered in [@Stef-SO] but the author didn’t succeed to classify all codes. We do this classification now. Generator matrices of the codes $C_\varphi$ and $E_{\sigma}(C)^*$ are presented in [@Stef-SO]. In the same paper, it is also proved that $C_\pi$ is a binary self-dual $[32,16,\ge 4]$ code with a generator matrix $$G_{\pi} = \left(
\begin{array}{c c}
0&D\\ I_6& F\\
\end{array} \right)$$ where $D$ generates a $[26,10,8]$ self-orthogonal code ${\cal D}^*$, and ${\cal D} _F$ is its dual code. The code ${\cal D}^*$ cannot be self-complementary (see [@Stef-SO]). According to [@BBGO], there are 1768 inequivalent $[26,10,8]$ self-orthogonal codes. Using as $D$ generator matrices of those codes which are not self-complementary, we obtain the self-dual $[44,22,8]$ codes invariant under the given permutation. To test them for equivalence, we use the program <span style="font-variant:small-caps;">Q-Extension</span> [@Iliya-aut]. The weight enumerators of the constructed codes are listed in Table \[Table:c6\].
There are exactly $15621$ self-dual $[44,22,8]$ codes having an automorphism of type $3$-$(6,26)$.
\[Table:c6\]
[|c|ccccccccccccc|]{} $\beta$&14&15&16&17&18&19&20&21&22&23&24&25&26\
$W_1$&-&-&-&-&4&16&33&31&59&62&82&79&72\
$W_2$&11&26&58&201&342&433&505&462&677&685&717&599& 611\
$\beta$&27&28&29&30&31&32&33&34&35&36&37&38&39\
$W_1$&47&72&48&51&51&68&54&64&39&54&38&38&29\
$W_2$& 463& 490& 452& 485&654&724&674&851&558&530&430&438&327\
$\beta$&40&41&42&43&44&45&46&47&48&49&50&51&52\
$W_1$&32&32&28&35&66&49&51&41&40&33&39&29&33\
$W_2$&328&238&194&120&140&72&89&43&85&13&46&5&27\
$\beta$&53&54&55&56&57&58&59&60&61&62&63&64&65\
$W_1$&17&24&8&18&4&15&4&7&1&5&1&2&3\
$W_2$&5&21&6&11&-&15&1&6&1&7&-&2&-\
$\beta$&66&67&68&70&72&74&76&82&86&90&104&122&154\
$W_1$&5& 2& 1& 2& 1&2&-&1& 1&1&-&1&-\
$W_2$&1&-&1&1&3&4&2&2&1&1&1&-&1\
Codes with an automorphism of type $3$-$(8,20)$
-----------------------------------------------
Up to equivalence, a unique Hermitian quaternary $[8,4,4]$ code exists (see [@MWOSW]). So up to equivalence we have a unique subcode $E_{\sigma}(C)^*$. The code $C_{\pi}$ is a binary self-dual $[28,14,\ge 4]$ code with a generator matrix $G_{\pi}$ given in (\[matGpi\]) where $B$ and $D$ generate self-orthogonal $[8,k_1,\ge 4]$ and $[20,k_1+6,\ge 8]$ codes, respectively. Since $0\le k_1\le 4$, ${\cal D}^*$ is a binary self-orthogonal $[20,6\le k_2\le 10,\ge
8]$ code. All optimal binary self-orthogonal codes of length 20 are classified in [@Stef-SO]. There are exactly 23 inequivalent $[20,6,8]$ self-orthogonal codes, four inequivalent $[20,7,8]$ self-orthogonal codes, and a unique $[20,8,8]$ self-orthogonal code. Hence $k_1\leq 2$.
In the case $k_1=2$ we obtain only two inequivalent extremal codes of length 44, both with weight enumerator $W_{44,2}$, respectively for $\beta=68$ and $\beta=76$. For $k_1=1$, there exist 52 self-dual $[44,22,8]$ codes, and for $k_1=0$, the inequivalent codes number 5399. Their weight enumerators are listed in Table \[Table:c8\].
There are exactly $5453$ self-dual $[44,22,8]$ codes having an automorphism of type $3$-$(8,20)$.
**Remark:** The extremal self-dual $[44,22,8]$ codes invariant under a permutation of type 3-(8,20) are considered independently in [@Kim-c8]. The author of that paper has classified all extremal self-dual codes which have an automorphism of order 3 with 8 independent 3-cycles.
\[Table:c8\]
[|c|ccccccccccccccc|]{} $\beta$&8&9&10&11&12&13&14&15&16&17&18&19&20&21&22\
$W_1$&-&-&-&-&-&-&2&-& -& 5&5&3& 9&16&8\
$W_2$&2&-& 3&10&8& 27&47&81& 157&174&330&395&442&481&560\
$\beta$&23&24&25&26&27&28&29&30&31&32&33&34&35&36&37\
$W_1$&16& 28&16&69&36&39&27&60&29&55&26&34&15&25&15\
$W_2$&442&432&307&298&140&172&79&69& 41&56&13&25&9&9 &6\
$\beta$&38&39&40&41&42&43&44&45&46&49&50&52&53&68&76\
$W_1$&15&3&8& 4&6&1& 4& -&3&1&1&-&1&-&-\
$W_2$& 18&-&9& 4&6&3&4& -& 3&-&3&1&-&1&1\
Codes with an automorphism of type $3$-$(10,14)$ {#sect:c10}
-------------------------------------------------
In this case $C_{\varphi}$ is a Hermitian self-dual $[10,5,4]$ code and by [@MWOSW] is equivalent to either $E_{10}$ or $B_{10}$. As in [@BYR], we can fix the generator matrix of the subcode $E_{\sigma}(C)^*$ in the following two forms, respectively: $${\footnotesize
\begin{pmatrix}
011011011011000000000000000000\cr
101101101101000000000000000000\cr
000000011011011011000000000000\cr
000000101101101101000000000000\cr
000000000000011011011011000000\cr
000000000000101101101101000000\cr
000000000000000000011011011011\cr
000000000000000000101101101101\cr
011000011000011000011000101110\cr
101000101000101000101000110011\cr
\end{pmatrix} \ {\rm and} \
\begin{pmatrix}
011011011011000000000000000000\cr
101101101101000000000000000000\cr
000011101110011000000000000000\cr
000101110011101000000000000000\cr
000000000000000011011011011000\cr
000000000000000101101101101000\cr
000000000000000000011101110011\cr
000000000000000000101110011101\cr
000011110101000000011110101000\cr
000101011110000000101011110000\cr
\end{pmatrix}}.$$ The code $C_\pi$ has parameters $[24,12,\geq 4]$. There are exactly thirty inequivalent such codes, namely $E_8^3$, $E_{16}\oplus E_8$, $F_{16}\oplus E_8$, $E_{12}^2$, and the indecomposable codes denoted by $A_{24}, B_{24},\dots, Z_{24}$ in [@CPS32]. All codes have minimum weight 4 except the extended Golay code $G_{24}$ with minimum weight 8 and the code $Z_{24}$ with minimum weight 6. We use the generator matrices of the codes given in [@PlessSl]. For any weight 4 vector in $C_\pi$ at most two nonzero coordinates may be fixed points. An examination of the vectors of weight 4 in the listed codes eliminates 23 of them. By investigation of all alternatives for a choice of the 3-cycle coordinates in the remaining codes $G_{24}$, $R_{24}$, $U_{24}$, $W_{24}$, $X_{24}$, $Y_{24}$ and $Z_{24}$ we obtain, up to equivalence, all possibilities for the generator matrix of the code $C_\pi$.
Let $C_\pi$ be $R_{24}$. There is a unique possibility for the choice of the 3-cycle coordinates up to equivalence. The generator matrix of $C_\pi$ in this case can be fixed in the form $$G_{\pi}(R_{24})={\footnotesize
\begin{pmatrix}
1100000000 \ 11000000000000\cr
0110000000 \ 01100000000000\cr 0001100000 \ 00011000000000\cr
0000110000 \ 00001100000000\cr 0000001100 \ 00000011000000\cr
0000000110 \ 00000001100000\cr 0000000011 \ 00000000110000\cr
1001000000 \ 10010000001111\cr 1000001000 \ 10000010001100\cr
0000001110 \ 00000000010110\cr 1110000000 \ 00000000000111\cr
0001110000 \ 00000000001110\cr
\end{pmatrix}}.$$
Let $\tau$ be a permutation of the ten cycle coordinates in $G_{\pi}(R_{24})$. Denote by $C^{\tau }$ the self-dual $[44,22]$ code determined by $C_{\varphi}$ and the matrix $ \tau (G_{\pi}(R_{24}))$. We consider the products of transformations (ii), (iii) and (iv) from Theorem \[thm:eq\] which preserve the quaternary code $C_{\varphi}$. Their permutation parts form a subgroup of the symmetric group $S_{10}$ which we denote by $L$. Let $S=Stab(R_{24})$ be the stabilizer of the automorphism group of the code generated by $G_{\pi}(R_{24})$ on the set of the fixed points. It is easy to prove that if $\tau_1$ and $\tau_2$ are permutations from the group $S_{10}$, the codes $C^{\tau_1}$ and $C^{\tau_2}$ are equivalent iff the double cosets $S\tau_1L$ and $S\tau_2L$ coincide. In our case $Stab(R_{24})=\left<(7, 8)(9,10)\right.$, $(7,9,10)$, $(7,9)(8,10)$, $(7,10)$, $(5,6)$, $(4,6,5)$, $(2,3)$, $(1,3,2)$, $\left.(1,4)(2,5)(3,6)\right>$. When $C_{\varphi}=B_{10}$ we found in [@BYR] a subgroup of the group $L$ generated by the permutations $(3,4)(8,9)$, $(1,2)(3,4)$, $(1,3)(2,4)$, $(6,7)(8,9)$, $(6,8)(7,9)$ and $(1,6)(2,7)(3,8)(4,9)(5,10)$. So we obtain four $[44,22,8]$ self-dual codes: $C_{B_{10}}^{id }$, $C_{B_{10}}^{(567)}$, $C_{B_{10}}^{(36754)}$ and $C_{B_{10}}^{(368574)}$. These codes have weight enumerator $W_{44,1}$ with $\beta = 60, 33,$ 30 and 21 and automorphism groups of orders $2^7\cdot 3^4$, $2^4\cdot 3^3$, 72 and 48, respectively.
When $C_{\varphi}=E_{10}$ the group $L=\left<(1,3,5,7,9)(2,4,6,8,10), (1,2)(3,4)\right.$, $\left.
(1,3)(2,4), (9,10)\right>.$ We obtain seven $[44,22,8]$ self-dual codes $C_{E_{10}}^{\tau }$ for $\tau \in \{$id$, (4,5,6,7),
$(4,5,7)(6,9,8), $(2,3,5,4)$, $(2,3,5,4)(6,7)$, $(2,3,5,7,4)(6,9,8)$, $(6,7)\}.$ These codes have also weight enumerator $W_{44,1}$ with $\beta = 42, 30,$ 36, 24, 42, 30 and 21 and automorphism groups of orders $2^{10}\cdot3$, 24, 192, 36, $2^7\cdot3^2$, again 24 and 720, respectively. In this way from all the cases for $C_\pi$ we constructed 1865 inequivalent $[44,22,8]$ self-dual codes with weight enumerator $W_{44,1}$ for $\beta = 10,\dots,52$, 54, 55, 60, 62, 65 and 6873 codes with weight enumerator $W_{44,2}$ for $\beta = 3,\dots,36$, 38, 42, 45, 46, 50 and 52. The calculations for these results were done with the GAP Version 4r4 software system and the program <span style="font-variant:small-caps;">Q-Extension</span> [@Iliya-aut]. The results are summarized in Tables \[Table:c10a\] and \[Table:c10\].
\[Table:c10a\]
[|c|c|c||c|c|c||c|c|c|]{} & $W_{44,1}$ & $W_{44,2}$& & $W_{44,1}$ & $W_{44,2}$& & $W_{44,1}$ & $W_{44,2}$\
$G_{24}, B_{10}$ & 3 & 12 & $U_{24}, E_{10}$ & 74 & 49 & $Y_{24}, B_{10}$ & 136 & 746\
$G_{24}, E_{10}$ & 6 & 25 & $W_{24}, B_{10}$ & 71 & 11 & $Y_{24}, E_{10}$ & 456 & 2764\
$R_{24}, B_{10}$ & 4 & - & $W_{24}, E_{10}$ & 188 & 33 & $Z_{24}, B_{10}$ & 71 & 541\
$R_{24}, E_{10}$ & 7 & - & $X_{24}, B_{10}$ & 161 & 224 & $Z_{24}, E_{10}$ & 207 & 1824\
$U_{24}, B_{10}$ & 29 & 19 & $X_{24}, E_{10}$ & 459 & 635 & & &\
There are exactly $8738$ inequivalent self-dual $[44,22,8]$ codes having an automorphism of type $3$-$(10,14)$.
\[Table:c10\]
[|c|cccccccccccccc|]{} $\beta$&3&4&5&6&7&8&9&10&11&12&13&14&15&16\
$W_1$&-&-&-&-&-&-&-&1&2&11&49&63&25&114\
$W_2$&1 &3&31&31&93&143&183&377&428&560&622&552&755&510\
$\beta$&17&18&19&20&21&22&23&24&25&26&27&28&29&30\
$W_1$&97&51&159&134&71&157&99&63&129&81&49&90&61&41\
$W_2$&411&585&270&223&321&145&96&176&35& 71& 64&32&13& 57\
$\beta$&31&32&33&34&35&36&37&38&39&40&41&42&43&44\
$W_1$&55&31&28&41&21&16&22&21&11&14&11&10&12&4\
$W_2$& 7&23&16&11& 3&8&-&9& -& -&-&4&-&-\
$\beta$&45&46&47&48&49&50&51&52&54&55&60&62&65&\
$W_1$ &4& 4&1&1& 1& 2& 1& 2&1&1&1&1&1&\
$W_2$&1&1&-&-&-&1&-&1&-&-&-&-&-&\
Codes with an automorphism of type $3$-$(12,8)$
-----------------------------------------------
In this case $C_{\varphi}$ is a quaternary Hermitian self-dual code of length 12 with minimum weight at least 4. There exist exactly five inequivalent quaternary self-dual $[12,6,4]$ codes, denoted by $d_{12}$, $2d_6$, $3d_4$, $e_6\oplus e_6$, and $e_7+e_5$ in [@MWOSW].
The code $C_\pi$ is a binary self-dual $[20, 10, \geq 4]$ code. There are exactly seven such codes, namely $d_{12}+d_8$, $d_{12}+e_8$, $d_{20}$, $d_4^5$, $d_6^3+f_2$, $d_8^2+d_4$, and $e_7^2+d_6$ [@CPS32]. Each choice for the fixed points can lead to a different subcode $F_\sigma(C)$. We have considered all possibilities for each of these seven codes, and found exactly 7 inequivalent codes for $d_{12}+d_8$, one code for $d_{12}+e_8$, one code for $d_{20}$, 10 codes for $d_4^5$, 26 codes for $d_6^3+f_2$, 18 codes for $d_8^2+d_4$, and 3 codes for $e_7^2+d_6$. Denote these codes by $H_{i,j}$, for $i=1, 2\dots, 7$.
By the method used in Section \[sect:c10\], considering the permutation parts of the products of transformations (ii), (iii) and (iv) from Theorem \[thm:eq\] and the stabilizer of the automorphism group of the codes $H_{i,j}$ on the fixed points, we classified all codes up to equivalence. There are exactly 122787 inequivalent codes. Their weight enumerators are of type $W_{44,1}$ for $\beta=$ 10, …, 68, 70, 72, 74, 82, 86, 90, 122 and of type $W_{44,2}$ for $\beta=$ 0, …, 56, 58, …, 62, 64, 66, 68, 70, 72, 74, 76, 82, 86, 90, 104, 154. The values obtained for $\beta$ are listed in Table \[Table:c12\].
There are exactly $122787$ inequivalent self-dual $[44,22,8]$ codes having an automorphism of type $3$-$(12,8)$.
\[Table:c12\]
[|c|cccccccccccc|]{} $\beta$&0&1&2&3&4&5&6&7&8&9&10&11\
$W_1$&-&-&-&-&-&-&-&-&-&-&789&556\
$W_2$&7&151&594&1434&2178&3468&5793&7034&6881&9434&10031&6906\
$\beta$&12&13&14&15&16&17&18&19&20&21&22&23\
$W_1$&313&1915&1072&655&2141&1105&912&1770&1029&736&1338&666\
$W_2$&8502&7975&5072&4805&5111&2549&2552&2438&1692&1176&1609&778\
$\beta$&24&25&26&27&28&29&30&31&32&33&34&35\
$W_1$&642&731&511&382&568&286&286&263&236&161&179&99\
$W_2$&773&745&532&311&484&204&242&169&217&65&176&32\
$\beta$&36&37&38&39&40&41&42&43&44&45&46&47\
$W_1$&126&87&88&55&69&38&52&28&48&17&32&10\
$W_2$&73&42&68&30&44&29&30&21&21&9&26&10\
$\beta$&48&49&50&51&52&53&54&55&56&57&58&59\
$W_1$&18&7&19&9&15&5&7&3&11&4&9&5\
$W_2$&14&7&17&3&15&4&9&6&13&-&11&1\
$\beta$&60&61&62&63&64&65&66&67&68&70&72&74\
$W_1$&3&1&2&1&2&3&4&2&1&2&1&2\
$W_2$&6&1&5&-&2&-&1&-&1&1&3&4\
$\beta$&76&82&86&90&104&122&154&&&&&\
$W_1$&-&1&1&1&-&1&-&&&&&\
$W_2$&2&2&1&1&1&-&1&&&&&\
Codes with an automorphism of type $3$-$(14,2)$
-----------------------------------------------
The code $C_\pi$ in this case is a self-dual $[16, 8, 4]$ code. There are exactly three such codes, namely $d_8^{2}$, $d_{16}$, and $e_8^2$ [@CPS32]. We consider their generator matrices in the form $$G_1=gen(d_8^{2})={\small \left(\begin{array}{@{}c@{}}
1000000011100000\\
0100000011010000\\
0010000000001110\\
0001000000001101\\
0000100011001011\\
0000010011000111\\
0000001010111100\\
0000000101111100\\
\end{array}\right)}, \ \ G_2=gen(d_{16})={\small \left(\begin{array}{@{}c@{}}
1111000000000000\\
0011110000000000\\
0000111100000000\\
0000001111000000\\
0000000011110000\\
0000000000111100\\
0000000000001111\\
0101010101010101
\end{array}\right)},$$ and $G_3=gen(e_8^2)=\left(\begin{array}{@{}c@{}}
HO\\OH\end{array}\right)$, where $H=(I_4\vert J+I_4)$, $J$ is the all-ones $4\times 4$ matrix and $O$ is the $8\times8$ zero matrix. We have to consider permutations on these generator matrices that can lead to different subcodes $F_\sigma(C)$. From all possibilities for each of these codes we have found exactly 7 different cases for $C_\pi$ which can produce inequivalent codes $C$, namely $G_1$, $G_1^{(1,16)}$, $G_1^{(3,16)}$, $G_2$, $G_2^{(1,16)}$, $G_3$, and $G_3^{(1,16)}$.
The code $C_\varphi$ is a quaternary Hermitian self-dual $[14,7,4]$ code. There are exactly 10 such codes, namely $d_{14}$, $2e_7$, $d_8+e_5+f_1$, $2e_5+d_4$, $d_8+d_6$, $2d_6+f_2$, $d_6+2d_4$, $3d_4+f_2$, $2d_4+1_8$, and $q_{14}$ [@MWOSW].
Again, considering the permutation parts of the products of transformations (ii), (iii) and (iv) from Theorem \[thm:eq\], and the stabilizer of the automorphism group of the codes $C_\pi$ on the fixed points, we classified all codes up to equivalence.
When $C_\pi=d_{16}$ all codes have weight enumerators $W_{44,1}$ for $\beta$=11, 14, 17, 20, 23, 26, 29, 32, 35, 38, 41, 44, 53, 62, and 65. When $C_\pi=e_8^2$ the weight enumerators are $W_{44,1}$ for $\beta$=10, 13, 16, 19, 22, 25, 28, 31, 34, 37, 40, 43, 46, 49, 52, and 58. Lastly, when $C_\pi=d_8^{2}$ we constructed codes with weight enumerator $W _{44,2}$ for $\beta=$1, 2, 4, 5, 7, 8, 10, 11, 13, 14, 16, 17, 19, 20, 22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43, 44, 46, 52, and 55. The total number of all self-dual $[44, 22, 8]$ codes, having an automorphism of type $3$-$(14,2)$ is 243927. The results are presented in Tables \[Table:c14a\] and \[Table:c14\].
\[Table:c14a\]
[|c|c|c|c|c|c|]{} &$d_{14}$&$2e_7$&$d_8+e_5+f_1$&$2e_5+d_4$&$d_8+d_6$\
$d_{16}$ & 7 & 33 & 66 & 26 & 144\
$ e_8^2$ & 9 & 20 & 77 & 26 & 197\
$d_8^{2+}$ & 114 & 876 & 2907 & 490 & 6148\
&$2d_6+f_2$&$d_6+2d_4$&$3d_4+f_2$&$2d_4+1_8$&$q_{14}$\
$d_{16}$ & 573 & 384 & 2040 & 1663 & 1191\
$e_8^2$ & 735 & 496 & 2830 & 2225 & 1561\
$d_8^{2+}$ & 25841 & 14639 & 84081 & 60246 & 34520\
There are exactly $243927$ inequivalent self-dual $[44,22,8]$ codes having an automorphism of type $3$-$(14,2)$.
\[Table:c14\]
[|c|cccccccccc|]{} $\beta$&1&2&4&5&7&8&10&11&13&14\
$W_1$&-&-&-&-&-&-&704&984&1912&1537\
$W_2$&4565&4374&21709&15796&35653&26242&33236&22914&21064&14322\
$\beta$&16&17&19&20&22&23&25&26&28&29\
$W_1$&2006&1281&1447&1008&978&493&480&384&295&147\
$W_2$&10879&6663&4407&3053&1866&992&621&521&344&152\
$\beta$&31&32&34&35&37&38&40&41&43&44\
$W_1$&123&124&98&29&17&54&21&8&9&18\
$W_2$&109&88&85&19&16&24&14&9&4&2\
$\beta$&46&49&52&53&55&58&62&65&&\
$W_1$&7&2&3&1&-&1&1&2&&\
$W_2$&4&-&4&-&2&-&-&-&&\
All self-dual $[44,22,8]$ codes with an automorphism of order $3$
-----------------------------------------------------------------
Here we summarize all obtained results for the extremal self-dual codes of length 44 having an automorphism of order 3. To test the codes for equivalence, we used the program <span style="font-variant:small-caps;">Q-Extension</span>. The classification result is given in the following theorem.
There are exactly $394916$ inequivalent self-dual $[44,22,8]$ codes having an automorphism of order $3$.
We list the number of the codes with different weight enumerators in Table \[Table:p=3\]. For $\beta\ge 67$, all codes have simultaneously automorphisms of type 3-(12,8) and also automorphisms of type 3-(6,26). This proves that the orders of the automorphism groups of these codes are multiples of 9. We give these orders in Table \[Table:beta68\]. All codes with $\beta\ge 63$ have automorphisms of type 3-(6,26). All seven codes with $\beta=0$ have automorphisms of type 3-(12,8). The full automorphism group for four of them is the cyclic group of order 3, and the other three codes have automorphism groups of order 12.
\[Table:p=3\]
[|c|ccccccccccc|]{} $\beta$&0&1&2&3&4&5&6&7&8&9&10\
$W_1$&-&-&-&-&-&-&-&-&-&-&1487\
$W_2$&7&4713&4968&1435&23881&19271&5824&42768&33242&9617&43614\
$\beta$&11&12&13&14&15&16&17&18&19&20&21\
$W_1$&1539&324&3860&2659&680&4248&2471&972&3385&2182&851\
$W_2$&30231&9070&29668&19954&5666&16669&9965&3804&7898&5880&2440\
$\beta$&22&23&24&25&26&27&28&29&30&31&32\
$W_1$&2523&1327&807&1428&1080&512&1051&558&431&515&504\
$W_2$&4798&2963&2095&2250&1985&978&1465&870&849&965&1048\
$\beta$&33&34&35&36&37&38&39&40&41&42&43\
$W_1$&266&396&201&213&176&205&95&131&86&92&79\
$W_2$&761&1082&606&609&477&526&348&366&258&221&134\
$\beta$&44&45&46&47&48&49&50&51&52&53&54\
$W_1$&115&66&86&47&55&39&55&34&43&18&28\
$W_2$&151&73&102&44&87&15&51&5&30&5&23\
$\beta$&55&56&57&58&59&60&61&62&63&64&65\
$W_1$&8&25&5&20&5&7&1&6&1&2&3\
$W_2$&6&15&-&17&1&6&1&7&-&2&-\
$\beta$&66&67&68&70&72&74&76&82&86&90&104\
$W_1$&5&2&1&2&1&2&-&1&1&1&-\
$W_2$&1&-&1&1&3&4&2&2&1&1&1\
$\beta$&122&154&&&&&&&&&\
$W_1$&1&-&&&&&&&&&\
$W_2$&-&1&&&&&&&&&\
Extremal Self-Dual Codes of Length 44 with an automorphism of order 7
=====================================================================
If $\sigma$ is an automorphism of a binary self-dual $[44,22,8]$ code of order 7, then $\sigma$ is of type $7$-$(3,23)$ or $7$-$(6,2)$ [@FFTAHuff].
Let $h_1(x)=(x^3+x+1)$ and $h_2(x)=(x^3+x^2+1)$. As $x^7-1=(x-1)h_1(x)h_2(x)$, we have ${\cal P}=I_1\oplus I_2$, where $I_j$ is an irreducible cyclic code of length 7 with parity-check polynomial $h_j(x)$, $j=1,2$. According Lemma \[M1M2M3\], $C_{\varphi}=M_1\oplus M_2$, where $M_j=\{u\in C_{\varphi}\mid
u_i\in I_j, i=1,\dots,c\}$ is a linear code over the field $I_j$, $j=1,2$, and $\dim_{I_1}M_1+\dim_{I_2}M_2=c$. The polynomials $e_1=x^4+x^2+x+1$ and $e_2=x^6+x^5+x^3+1$ generate the ideals $I_1$ and $I_2$ defined above. Any nonzero element of $I_j=\{0,
e_j, xe_j\dots, x^6e_j\}, j=1,2$ generates a binary cyclic $[7,4,3]$ code. Since the minimum weight of the code $C$ is 8, every vector of $C_{\varphi}$ must contain at least two nonzero coordinates. Hence the minimum weight of $M_j$ is at least 2, $j=1,2$.
The transformation $x\rightarrow x^{-1}$ interchanges $e_1$ and $e_2$. The orthogonal condition (ii) from Theorem \[thm2\] implies that once chosen, $M_1$ determines $M_2$ and the whole $C_{\varphi}$. So we can assume, without loss of generality, that $\dim_{I_1}M_1\leq \dim_{I_2}M_2$, and we can examine only $M_1$.
Codes with an automorphism of type $7$-$(3,23)$
------------------------------------------------
Let $C$ be a binary self-dual $[44,22,8]$ code having an automorphism of type $7$-$(3,23)$. Then we have $\dim_{I_1}M_1+\dim_{I_2}M_2=3$. Since the minimum weight of $M_2$ is at least 2, we have $1\leq\dim_{I_1}M_1\leq\dim_{I_2}M_2\leq
2$. Hence $\dim_{I_1} M_1=1$ and $\dim_{I_2}M_2=2$. Then $M_2$ is an MDS $[3,2,2]$ code over the field $I_2$ and according to condition (ii) from Theorem \[thm2\], $M_1=\langle(e_1,e_1,e_1)\rangle$ and $M_2=\langle(e_2,e_2,0),(0,e_2,e_2)\rangle$.
In this case $C_\pi$ is a binary self-dual code of length 26. If $v=(1100\ldots0)\in C_\pi$ then $\pi^{-1}(v)+(\phi^{-1}(e_2,e_2,0),00\ldots0)$ will be a codeword from $C$ of weight 6 which contradicts the minimum weight of $C$. Hence in the notations of Theorem \[structure\], $k_1=0, k_2=10, k_3=3$, and $gen~C_\pi=\left(\begin{array}{cc}
0 & D \\
E & F \\
\end{array}\right)$, where the matrix $D$ generates a $[23,10,\geq 8]$ binary self-orthogonal code. There are three such codes and their generator matrices are given in [@BBGO]. We take $E=I_3,$ and we determine the matrix $F$ using the condition (iii) of Theorem \[structure\]. For each of the three codes there is a unique possibility for the matrix $F$, up to equivalence. We obtain the codes $C_{7,1}$ with weight enumerator $W_{44,1}$ for $\beta=122$, $C_{7,2}$ with weight enumerator $W_{44,2}$ for $\beta=104$, and $C_{7,3}$ with weight enumerator $W_{44,2}$ for $\beta=154$. The orders of their automorphism groups are $3251404800=2^{15}\cdot3^{4}\cdot5^{2}\cdot7^{2}$, $116121600=2^{13}\cdot3^{4}\cdot5^{2}\cdot7$, and $786839961600=2^{16}\cdot3^{4}\cdot5^{2}\cdot7^{2}\cdot11^{2}$, respectively. All of these codes have automorphisms of order 5 and are known from [@St42-44].
There are exactly three inequivalent binary $[44,22,8]$ codes having an automorphism of type $7$-$(3,23)$.
Codes with an automorphism of type $7$-$(6,2)$.
-----------------------------------------------
Let $C$ be a binary self-dual $[44,22,8]$ code having an automorphism of type $7$-$(6,2)$. Now $C_\pi$ is a binary $[8,4]$ self-dual code equivalent either to $C_2^4$ or the extended Hamming code $E_8$, generated by the matrices $G_1=\left(I_4|I_4\right)$ and $G_2 =\left(I_4|J+I_4\right)$, respectively where $I_4$ is the $4\times 4$ identity matrix and $J$ is the all-ones $4\times 4$ matrix.
In this case $\dim_{I_1}
M_1+\dim_{I_2}M_2=6$ and $1\leq\dim_{I_1} M_1\leq\dim_{I_2} M_2\leq 5$. Hence $\dim_{I_1} M_1=1,2,$ or 3.
**Case I:** $\dim_{I_1}M_1=1, \dim_{I_2}M_2=5$. It follows that $M_2$ is an MDS $[6,5,2]$ code, and $M_1=\langle(e_1,e_1,e_1,e_1,e_1,e_1)\rangle$. If $C_\pi= C_2^4$, then $C_\pi$ contains a codeword $v=(v_1,00)$ such that $\mbox{wt}(v_1)=2$. Since $M_2$ is an MDS code, it contains a codeword $w$ of weight 2 with the same support as $v_1$. But then the codeword $\pi^{-1}(v)+(\phi^{-1}(w),00)\in C$ has weight 6 - a contradiction. Therefore $C_\pi= E_8$. Fixing the codes $M_1$ and $M_2$ and considering all binary codes equivalent to $E_8$, we found only one $[44,22,8]$ code with weight enumerator $W_{44,1}$ for $\beta=38$ and $|Aut(C)|=8064$.
**Case II:** $\dim_{I_1}M_1=2, \dim_{I_2}M_2=4$. We can take $$gen(M_1)=\left(\begin{array}{cccccc}
e_1&0&\alpha_1&\alpha_2&\alpha_3&\alpha_4\\
0&e_1&\alpha_5&\beta_1&\beta_2&\beta_3\\
\end{array}\right),$$ where $\alpha_i\in\{0,e_1\}, i=1,\cdots,5$, and $\beta_i\in I_1, i=1,2,3$. Considering all such matrices we obtain nine possibilities such that the minimum weight of $M_1$ is $\geq
2$, up to equivalence. Here $gen(M_1)$ is written in the form $(I_2|A)$, where $A$ is one of the following matrices:\
$$A_1=\left(
\begin{array}{cccc}
e_1 & 0 & 0 & 0 \\
e_1 & e_1 & e_1 & e_1 \\
\end{array}
\right), \ \ A_4=\left(
\begin{array}{cccc}
e_1 & e_1 & 0 & 0 \\
e_1 & xe_1 & e_1 & e_1 \\
\end{array}
\right), \ \ A_7=\left(
\begin{array}{cccc}
e_1 & e_1 & e_1 & 0 \\
e_1 & xe_1 & x^2e_1 & e_1 \\
\end{array}
\right),$$ $$A_2=\left(
\begin{array}{cccc}
e_1 & e_1 & 0 & 0 \\
0 & 0 & e_1 & e_1 \\
\end{array}
\right), \ \ A_5=\left(
\begin{array}{cccc}
e_1 & e_1 & e_1 & 0 \\
0 & e_1 & e_1 & e_1 \\
\end{array}
\right), \ \ \ A_8=\left(
\begin{array}{cccc}
e_1 & e_1 & e_1 & 0 \\
e_1 & xe_1 & x^3e_1 & e_1 \\
\end{array}
\right),$$ $$A_3=\left(
\begin{array}{cccc}
e_1 & e_1 & 0 & 0 \\
0 & e_1 & e_1 & e_1 \\
\end{array}
\right), \ \ A_6=\left(
\begin{array}{cccc}
e_1 & e_1 & e_1 & 0 \\
0 & e_1 & xe_1 & e_1 \\
\end{array}
\right), \ \ A_9=\left(
\begin{array}{cccc}
e_1 & e_1 & e_1 & e_1 \\
e_1 & xe_1 & x^2e_1 & x^3e_1 \\
\end{array}
\right).$$
In the case $C_\pi= C_2^4$, denote by $A_i^{\tau}$ the $[44,22,8]$ code determined by $(I_2|A_i)$ and $C_\pi=\tau(G_1)$. There are 21 inequivalent codes, namely $A_1^{id}$, $A_2^{(2,5,7,3,6)}$, $A_3^{(2,5,6)}$, $A_3^{(2,5,7,3,6)}$, $A_3^{(2, 5, 4, 6)}$, $A_4^{(2, 5, 7, 3, 6)}$, $A_5^{id}$, $A_6^{(2, 5, 6)}$, $A_6^{(4, 5, 6)}$, $A_6^{(4, 5, 7, 8, 6)}$, $A_6^{(3, 5)(4, 6)}$, $A_7^{id}$, $A_7^{(2, 5, 7, 3, 6)}$, $A_7^{(4, 5, 7, 8)}$, $A_7^{(3, 6, 4, 5, 7)}$, $A_8^{id}$, $A_8^{(2, 5, 7, 3, 6)}$, $A_9^{id}$, $A_9^{(3, 6, 7)}$, $A_9^{(4,
6, 7, 8)}$, and $A_9^{(2, 5, 6)}$. The code $A_2^{(2,5,7,3,6)}$ has an automorphism group of order $786839961600$ and is equivalent to the code $C_{7,3}$ constructed above.
In the case $C_\pi= E_8$, denote by $B_i^{\tau}$ the $[44,22,8]$ code determined by $(I_2|A_i)$ and $C_\pi=\tau(G_2)$. There are 19 inequivalent codes, namely $B_1^{id}$, $B_2^{id}$, $B_3^{id}$, $B_3^{(5,6)}$, $B_4^{id}$, $B_4^{(3, 7, 6, 8, 5)}$, $B_5^{id}$, $B_6^{id}$, $B_6^{(6, 7, 8)}$, $B_6^{(5, 6)}$, $B_6^{(4, 5, 6, 7,
8)}$, $B_6^{(3, 7, 8, 6, 4, 5)}$, $B_7^{id}$, $B_7^{(6, 7, 8)}$, $B_7^{(5, 6)}$, $B_8^{id}$, $B_9^{id}$, $B_9^{(5, 6)}$, and $B_9^{(5, 6, 7)}$. The code $B_2^{id}$ is equivalent to $C_{7,2}$, constructed in the previous section.
**Case III:** $\dim_{I_1}M_1=\dim_{I_2}M_2=3$. Then $$gen(M_1)=\left(\begin{array}{cccccc}
e_1&0&0&\alpha_1&\alpha_2&\alpha_3\\
0&e_1&0&\alpha_4&\beta_1&\beta_2\\
0&0&e_1&\alpha_5&\beta_3&\beta_4\\
\end{array}\right),$$ where $\alpha_i\in\{0,e_1\}, i=1,\cdots,5$, and $\beta_i\in I_1, i=1,2,3,4$. There are 18 inequivalent codes $M_1$ with minimum weight $\geq 2$. We can fix the generator matrices for $M_1$ and $M_2$ and consider all possibilities for $C_\pi$.
When $C_\pi= E_8$ we obtain 64 inequivalent codes with $W_{44,1}$ for $\beta=10$, 17, 24, 31, 38, 52, and 122. In the case $C_\pi=
C_2^4$ we obtain 87 inequivalent codes with $W_{44,2}$ for $\beta=0$, 7, 14, 21, 28, 35, 42, 56, and 154. The codes with $\beta=122$ and $\beta=154$ are equivalent to $C_{7,1}$ and $C_{7,3}$, respectively.
\[Table:p7c6h\]
[|c|cccccccccc|]{} $|Aut(C)|$&7&14&21&28&42&56&84&112&126&168\
$\sharp$ codes&13&35&1&5&9&2&2&2&1&1\
$|Aut(C)|$&252&336&672&1344&2688&5040&5376&8064&64512&3251404800\
$\sharp$ codes&1&2&2&1&1&1&1&1&1&1\
\[Table:p7c6c\]
[|c|cccccccc|]{} $|Aut(C)|$&7&14&28&42&56&112&168&336\
Number of codes&49&33&4&1&3&2&1&2\
$|Aut(C)|$&672&1344&2688&10752&21504&43008&786839961600&\
Number of codes&1&2&3&1&1&3&1&\
\[Table:p7all\]
[|c|cccccccccc|]{} $\beta$ in $W_{44,1}$&&&10&17&24&31&38&52&59&122\
Number of codes&&&23&19&14&12&9&4&1&1\
$\beta$ in $W_{44,2}$&0&7&14&21&28&35&42&56&104&154\
Number of codes&27&29&32&5&7&1&1&4&1&1\
There are exactly $191$ inequivalent $[44,22,8]$ codes having an automorphism of order $7$.
The orders of the automorphism groups of these codes are presented in Tables \[Table:p7c6h\] and \[Table:p7c6c\]. The weight enumerators of the constructed codes are listed in Table \[Table:p7all\].
Summary
=======
The self-dual $[44,22,8]$ codes having automorphisms of order 11 are classified in [@164] and [@160]. The codes invariant under an automorphism of order 5 are presented in [@St42-44] and [@Stef-SO]. Summarizing these classifications and the results from the previous sections, we obtain the following theorem.
There are exactly $395555$ inequivalent self-dual $[44,22,8]$ codes having an automorphism of odd prime order.
All constructed codes with $\beta\ge 43$ have automorphisms of order 3. In Table \[Table:all\] we list the number of codes having an automorphism of odd prime order according to their weight enumerators but only for these values of $\beta$ for which there are also codes having automorphisms of order 5, 7 or 11, but not 3. For the other values of $\beta$ the number of all extremal self-dual codes having an automorphism of odd prime order is the same as in Table \[Table:p=3\]. We can send the generator matrices of the obtained codes by e-mail to everybody who is interested.
\[Table:all\]
[|c|cccccccc|]{} $\beta$&0&4&5&7&9&10&11&12\
$W_1$&-&-&-&-&-&[**1506**]{}&1539&[**397**]{}\
$W_2$&[**54**]{}&[**23926**]{}&[**19293**]{}&[**42796**]{}&[**9658**]{}&[**43639**]{}&[**30237**]{}&9070\
$\beta$&14&15&17&19&20&21&22&24\
$W_1$&2659&680&[**2549**]{}&3385&2182&851&[**2561**]{}&[**820**]{}\
$W_2$&[**20026**]{}&[**5672**]{}&9965&[**7909**]{}&[**5888**]{}&[**2445**]{}&[**4802**]{}&[**2117**]{}\
$\beta$&25&27&28&29&30&31&32&\
$W_1$&1428&[**528**]{}&1051&558&431&[**525**]{}&[**523**]{}&\
$W_2$&[**2251**]{}&978&[**1470**]{}&[**872**]{}&[**852**]{}&965&1048&\
$\beta$&34&35&37&38&42&44&&\
$W_1$&396&201&[**179**]{}&[**207**]{}&[**96**]{}&115&&\
$W_2$&[**1090**]{}&[**607**]{}&477&526&221&[**153**]{}&&\
Looking at the tables, one can notice that there is only one code for $\beta=154$. This code has a large automorphism group - its order is $2^{16}\cdot3^{4}\cdot5^{2}\cdot7^{2}\cdot11^{2}=786839961600$. The same is the situation with the codes for $\beta=122$ and $\beta=104$. These two codes have automorphism groups of orders $2^{15}\cdot3^{4}\cdot5^{2}\cdot7^{2}=3251404800$ and $2^{13}\cdot3^{4}\cdot5^{2}\cdot7=116121600$, respectively. In Table \[Table:beta68\] we present the orders of the automorphism groups of the codes with $\beta\geq 67$. All these orders are multiples of $288=9\cdot 2^5$. Actually, all 12 codes with automorphism groups of orders bigger than 400000 have weight enumerators of both types with $\beta\ge 72$ and are given in Table \[Table:beta68\]. We list the number of codes $C$ with full automorphism groups of orders $6000<|Aut(C)|<400000$, $|Aut(C)|\neq 2^s$, in Table \[Table:paut\]. The code with the largest automorphism group (order 368640) which is not listed in Table \[Table:beta68\] has weight enumerator $W_{44,1}$ with $\beta=42$. Actually, the full automorphism group for most of the codes (exactly 309666) is the cyclic group of order 3. These codes have weight enumerators of both types with $\beta\le 42$.
\[Table:beta68\]
[|c|c|c|c|c|c|c|]{} $\beta$ &67&68&70&72&74&76\
$\sharp$ codes&&&&&&\
with $W_{44,1}$&2&1&2&1&2&-\
$|Aut(C)|$&2592&5184&13824&6912 &6912&-\
&2304&&18432&&73728&\
$\sharp$ codes&&&&&&\
with $W_{44,2}$&-&1&1&3&4&2\
&&&&92160&69120&207360\
$|Aut(C)|$&-&207360&69120&1105920&331776-2&165888\
&&&&184320&14745600&\
$\beta$ &82&86&90&104&122&154\
$\sharp$ codes&&&&&&\
with $W_{44,1}$&1&1&1&-&1&-\
$|Aut(C)|$&7372800&1105920&2211840&-&3251404800&\
$\sharp$ codes&&&&&&\
with $W_{44,2}$&2&1&1&1&-&1\
&663552&&&&&\
$|Aut(C)|$&2211840&1105920&14745600&116121600&-&786839961600\
\[Table:paut\]
[|c|ccccccc|]{} $|Aut(C)|$&368640&331776&207360&184320&165888&98304&92160\
Number of codes&1&2&2&1&1&1&1\
$|Aut(C)|$&73728&69120&64512&61440&55296&46080&43008\
Number of codes&2&2&1&2&1&1&3\
$|Aut(C)|$&36864&34560&21504&18432&15552&13824&12288\
Number of codes&4&2&1&8&1&2&11\
$|Aut(C)|$&11520&10752&10368&9216&8064&6912&6144\
Number of codes&1&1&1&6&1&6&35\
Looking at the weight enumerators of the extremal codes of length 44 constructed up to now, the following open problems arise:
1. Prove that there are not extremal self-dual $[44,22,8]$ codes with weight enumerator $W_{44,1}$ for $\beta=69$, $71$, $73$, $75,\dots,81$, 83, 84, 85, 87, 88, 89, $91,\dots,121$, or $W_{44,2}$ for $\beta=57$, $63$, $65$, $67$, $69$, $71$, $73$, $75$, $77,\dots,81$, 83, 84, 85, 87, 88, 89, $91,\dots,103$, $105,\dots,153$.
2. Are the constructed codes with weight enumerators $W_{44,1}$ for $\beta=61$, 63, 68, 72, 82, 86, 90, 122, and $W_{44,2}$ for $\beta=59$, 61, 66, 68, 70, 86, 90, 104, 154, the unique examples for their weight enumerators?
3. Which of these codes have connections with combinatorial designs?
Acknowledgements {#acknowledgements .unnumbered}
================
The authors would like to acknowledge the many helpful suggestions of the anonymous reviewers. We also thank the Editor of this Journal.
The first author thanks the Department of Algebra and Geometry at Magdeburg University for its hospitality while this work was completed, and the Alexander von Humboldt Foundation for its support.
[10]{}
Bouyukliev, I. (2007) ’About the code equivalence’, in *Advances in Coding Theory and Cryptology*, T. Shaska, W. C. Huffman, D. Joyner, V. Ustimenko: Series on Coding Theory and Cryptology, World Scientific Publishing, Hackensack, NJ.
Bouyukliev, I., Bouyuklieva, S., Gulliver, T.A. and Östergård, P. (2006) ’Classification of optimal binary self-orthogonal codes’, [*J. Combin. Math. and Combin. Comput.*]{}, Vol. 59, pp.33–87.
Bouyuklieva, S. (1997) ’New extremal self-dual codes of lengths $42$ and $44$’, *IEEE Trans. Inform. Theory*, Vol. 43, pp.1607–1612.
Bouyuklieva, S. (2004) ’Some optimal self-orthogonal and self-dual codes’, *Discrete Mathematics*, Vol. 287, pp.1–10.
Bouyuklieva, S., Yankov, N. and Russeva, R. (2007) ’Classification of the binary self-dual $[42,21,8]$ codes having an automorphism of order $3$’, *Finite Fields and Their Applications*, Vol.13, pp.605–615.
Conway, J.H., Pless, V. and Sloane, N.J.A. (1992) ’The binary self-dual codes of length up to 32: a revised enumeration’, [*Journal of Comb. Theory*]{}, Ser.A, Vol. 60, pp.183–195.
Conway, J.H. and Sloane, N.J.A. (1990) ’A new upper bound on the minimal distance of self-dual codes’, *IEEE Trans. Info. Theory*, Vol. 36, pp.1319–1333.
Huffman, W.C. (1982) ’Automorphisms of codes with application to extremal doubly-even codes of length 48’, *IEEE Trans. Info. Theory*, Vol. 28, pp.511–521.
Huffman, W.C. (2005) ’On the classification and enumeration of self-dual codes’, [*Finite Fields Appl.*]{}, Vol. 11, pp.451–490.
Hyun Jin Kim (2010) ’Self-dual codes with automorphism of order 3 having 8 cycles’, [*Designs, Codes and Cryptography*]{}, Vol. 57, pp.329–346.
MacWilliams, F.J., Odlyzko, A.M., Sloane, N.J.A. and Ward, H.N. (1997) ’Self-Dual Codes over $GF(4)$’, [*Journal of Comb. Theory*]{}, Ser.A, Vol. 25, pp.288–318.
Pless, V. (1998) ’Coding constructions’, in [*Handbook of Coding Theory*]{}, V.S.Pless and W.C. Huffman, eds., Elsevier, Amsterdam, pp.141–176.
Pless, V. and Huffman, W.C. (1998) *Handbook of Coding Theory*, Elsevier, Amsterdam.
Pless, V. and Sloane, N.J.A. (1975) ’On the classification and enumeration of self-dual codes’, *Journ. Combin. Theory*, ser. A, Vol. 18, pp.313–335.
Rains, E.M. (1998) ’Shadow bounds for self-dual-codes’, [*IEEE Trans. Inform. Theory*]{}, Vol. 44, pp.134–139.
Yankov, N. and Russeva, R. (2008) ’Classification of the Binary Self-Dual $[44,22,8]$ Codes with Automorphisms of Order 7’, *Proceedings of the 37th Conference of the UBM*, Bulgaria, pp.239–244.
Yankov, N. (2007) ’Extremal self-dual \[44,22,8\] codes with automorphism of order 3 with 14 cycles’, *Proceedings of the International Workshop on Optimal Codes and Related Topics (OCRT)*, Bulgaria, pp.249–254.
Yorgov, V. (1987) ’A method for constructing inequivalent self-dual codes with applications to length 56’, [*IEEE Trans. Info. Theory*]{}, Vol. 33, pp.77–82.
Yorgov, V. (1983) ’Binary self-dual codes with an automorphism of odd order’, [*Problems Inform.Transm.*]{}, Vol. 4, pp.13–24 (in Russian).
Yorgov, V. (1993) ’New extremal singly-even self-dual codes of lenght 44’, [*Procedings of the Sixth Joint Swedish -Russian Intern. Workshop on Inform. Theory (Sweden)*]{}, pp.372–375.
Yorgov, V. and Russeva, R. (1993) ’Two extremal codes of length 42 and 44’, [*Problems Inform.Transm.*]{}, Vol. 29, pp.385–388.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider different renormalizable models of Lorentz invariance violation. We show that the limits on birefringence of the propagation of cosmic microwave background photons from the five year data of the Wilkinson microwave anisotropy probe (WMAP) can be translated into a limit of Lorentz symmetry violation. The obtained limits on Lorentz invariance violation are stronger than other published limits. We also cast them in terms of limits on a birefringent effective photo “mass” and on a polarization dependence of the speed of light.'
author:
- Tina Kahniashvili
- Ruth Durrer
- Yurii Maravin
title: Testing Lorentz Invariance Violation with the WMAP Five Year Data
---
Introduction
============
The principal spacetime symmetry of particle interactions in the standard model is Lorentz invariance. Experiments confirm Lorentz symmetry at all currently accessible energy scales of up to 2 TeV. This scale will be extended shortly to 14 TeV with the Large Hadron Collider (LHC) at CERN. Although present experiments confirm Lorenz invariance to a good precision, it can be broken in the very early Universe when energies approach the Planck scale. There are a number of extensions of the standard model of particle physics and cosmology that violate Lorentz invariance (for reviews see Refs. [@k05b; @shore; @m05; @jlm05; @recent]).
As it can be expected, Lorentz invariance violation (LV) affects photon propagation (the dispersion relation), and generically results in a rotation of linear polarization (birefringence). Other effects include new particle interactions such as a photon decay and vacuum Cherenkov radiation [@jlm05]. All these effects can be used to probe Lorentz invariance. The dispersion measure test is based on a phenomenological energy dependence of the photon velocity [@a98] (see also Refs. [@sarkar] for reviews and Refs. [@jlm03; @bwhc04; @MP06] for recent studies of this effect; early discussions include Refs. [@16']; Refs. [@a98; @bwhc04; @MP06] consider Lorentz symmetry violating models which preserve rotational and translational invariance but break boost invariance).
Several models of LV predict frequency dependent effects. Such high energy Lorentz invariance breaking are discussed in Refs. [@km01; @at01; @myers]. Refs. [@GLP05] study generalizations of electromagnetism, motivated by this kind of Lorentz invariance violation. On the other hand, LV associated with a Chern-Simons interaction [@rj98; @cfj90] affects the entire spectrum of electromagnetic radiation, not just the high frequency part, and induces a frequency independent rotation of polarization (see Sec. 4 of Ref. [@shore] and Sec. III of this work).
To determine the effects induced by Lorentz symmetry violation, it is useful to consider the analogy with the propagation of electromagnetic waves in a magnetized plasma as outlined in Refs. [@cfj90; @km01; @jlm03; @mmu05; @kgr07]. Using the well known formalism for the propagation of light in a magnetized plasma, is easy to see that for Lorentz symmetry violating models which depend also on polarization and not only on frequency, the rotation measure constrains the symmetry breaking scale more tightly than the dispersion measure, see Refs. [@GLP05; @myers; @kgr07].
The propagation of ultra-high energy photons represents a promising possibility to probe Lorentz symmetry [@sigl08]. Gamma Ray Bursts (GRB) are astrophysical objects located at cosmological distances which emit very energetic photons [@a98]; reviews describing cosmological tests involving GRBs are e.g. Refs. [@piran; @m05], for recent studies see [@Albert:2007qk]. After the observation of highly linearly-polarized $\gamma$-rays from GRB021206 has been reported [@polarimetry], Refs. [@mitrofanov; @jlms04] have proposed to test Lorentz symmetry violation with the rotation measure by analysis of GRB polarization. Even though this measurement has been strongly contested [@pol2], there is evidence that the $\gamma$-ray flux from GRB930131 and GRB960924 is consistent with more than $35\%$ and $50\%$ polarization, respectively [@Willis]. However, the issue of polarization of GRB $\gamma$-rays is still under debate and additional $X$-ray studies are needed to either confirm or disprove polarization of $\gamma$-rays from GRB’s [@25].
In this paper we mainly consider renormalizable models of LV as described in Ref. [@shore]. We use the very well understood and measured temperature anisotropy and polarization of the cosmic microwave background (CMB) to constrain Lorentz symmetry violation. These data have been proposed as a probe of Lorentz invariance in the Universe in Refs. [@Lue; @Feng; @mewes; @kam]. In our study we use the WMAP 5 year limits on birefringence [@WMAP] and obtain limits which are significantly more stringent than those obtained from radio galaxy polarimetry [@cfj90].
As we shall see below, generically Lorentz symmetry violation leads to birefringence, i.e. a photon dispersion relation which depends on polarization. This leads to a rotation of the CMB polarization which induces parity-odd cross correlations, such as Temperature-$B$-polarization and $E$-$B$-polarization [@Lue]. These correlators vanish in models which preserve parity. Generally speaking, the effect is similar to that induced by a homogeneous magnetic field [@Bfield; @Bconst]. In this paper we use the WMAP-5 year limit on the rotation measure [@WMAP] to contrain Lorentz invariance violating theories.
Lorentz invariance violation: general description
=================================================
For methodological purpose let us first briefly summarize the usual Faraday rotation effect. We consider an electromagnetic wave with frequency $\omega$ and spatial wave vector ${\bf k}$, $k\equiv |{\bf k}|$ propagating in a magnetized plasma. A linearly polarized wave can be expressed as superposition of left ($-$) and right ($+$) circularly polarized waves. In a magnetized plasma, a homogeneous magnetic field induces a difference in the phase velocity of left and right handed waves. This causes a rotation of the polarization, called Faraday rotation [@krall]. The group velocity of the wave also differs from $c$. These two effects can be expressed in terms of the refractive indices defined by $k_{\pm} =n_{\pm}\omega$ where $k_{\pm}$ denotes the wave number for right and left handed waves. The indices $ n_{{\pm}}$ are [@krall] $$n_{\pm}^2 = 1-\frac{{\omega}_p^2}{{\omega}({\omega}\pm{\omega}_c)} \simeq
1-\frac{{\omega}_p^2}{{\omega}^2} \pm \frac{{\omega}_p^2{\omega}_c}{{\omega}^3} ~,
\label{eq:05}$$ Here ${\omega}_p = 4\pi e^2n_e/m_e$ is the plasma frequency and ${\omega}_c
=eB/m_e$ is the electron cyclotron frequency for the magnetic field $B$ (see Sec. 4.9 of Ref. [@krall]).
The magnitude of both the dispersion measure, due to the different group velocities, and the rotation measure, i.e., the rotation of polarization, are proportional to the photon travel distance $\Delta l$, $$\begin{aligned}
\Delta t_{\pm} &=& {\Delta l} \left( 1 - \frac{\partial
k_{\pm}}{\partial \omega} \right), \label{time-delay}
\\
\Delta \alpha &=& \frac{1}{2}(k_{+} - k_{-}) \Delta l.
\label{angle}\end{aligned}$$ Here, $\Delta t_{\pm}$ is the difference between the travel time of a right-handed (left-handed) photon and that of a photon traveling at the speed of light, and $\Delta \alpha$ is the rotation of the angle of polarization.
Faraday rotation is widely used in astrophysics to measure magnetic fields in galaxies and clusters (see Ref. [@Valee] for a review and references therein). In cosmology, Faraday rotation of CMB photons [@Bfield; @faraday] has been used to constrain the amplitude of a homogeneous as well as a stochastic cosmological magnetic field [@fine].
In the following, we show that Lorentz symmetry violation leads to a modification of Maxwell’s equations [@myers; @GLP05] analogous to the modifications described above.
Following Ref. [@shore], the most general renormalizable form of Lorentz symmetry violation can be expressed by two additional terms in the action (we set $\hbar=c=1$) $$\Gamma_{\rm LV} = \int d^4 x\sqrt{-g} \left[ K_{\mu\nu\lambda\rho}
F^{\mu\nu}F^{\lambda\rho} - \frac{1}{4} L^\mu A^\nu \tilde
F_{\mu\nu}\right], \label{Gamma}$$ where Greek indices ($\mu,\nu,\lambda,\rho$) denote time-space coordinates, $F^{\mu\nu}$ is the electromagnetic field strength tensor, $\tilde F_{\mu\nu}
={\epsilon_{\mu\nu}}^{\lambda\rho}F_{{\lambda}\rho}$ is its dual, $\epsilon_{\mu\nu\lambda\rho}$ is the totally antisymmetric tensor normalized such that $\epsilon_{0123} = \sqrt{-g}$ and $A^\nu$ is the vector potential. The four-vector $(L_\mu) = (L_0, {\bf L})$ has the dimension of mass and describes a super-renormalizable (dimension 3) coupling and $K_{\mu\nu\lambda\rho}$ is a renormalizable, dimensionless coupling giving raise to a dimension 4 operator. We want to break Lorentz symmetry, but keep conformal invariance of electrodynamics in this work. For this we have to ask that the components of ${K_{\mu\nu}}^{{\lambda}\rho}$ and $L_\mu$ be independent of conformal transformations of the metric. In the cosmological setup with $g_{\mu\nu}
=a^2(t)\eta_{\mu\nu}$, the above action is then independent of the scale factor $a(t)$. I.e. in a conformally flat spacetime the action is like in flat space. To see this note that the forms $A_\mu$ and $F_{\mu\nu}$ are independent of the metric hence $K_{\mu\nu\lambda\rho}F^{\mu\nu}F^{\lambda\rho}$ and $L^\mu A^\nu
\tilde F_{\mu\nu}$ scale like $a^{-4}$ which is canceled by $\sqrt{-g}=a^4$.
The tensor $K_{\mu\nu\lambda\rho}$ has the same symmetries as the Riemann tensor and we only consider its trace-free part which is analog to the Weyl tensor (the trace part also leads to dispersion measure but not to birefringence, we therefore do not consider it here). Even though we apply the formalism used for the Weyl tensor below, we do not consider $K_{\mu\nu\lambda\rho}$ to be the Weyl tensor which of course vanishes in a (unperturbed) Friedmann universe. The most plausible origin for the Lorentz violating terms in (\[Gamma\]) is that $K_{\mu\nu\lambda\rho}$ or the vector $L_\mu$ stem from the non-vanishing vacuum expectation value of some dynamical field and the action (\[Gamma\]) therefore represents a spontaneous rather than explicit breaking of Lorentz symmetry. However, for the following discussion the origin of the Lorentz violating terms is not relevant.
Both terms in Eq. (\[Gamma\]) lead to birefringence but the frequency dependence is different. The first term in the action $\Gamma_{\rm LV} $ can be computed within the Newman-Penrose formalism, which is usually applied for the Weyl tensor [@shore]. We consider a plane wave with conformal wave vector $(k_{\mu})=({\omega},{\bf k})$. We normalize the scale factor to unity today, $a_0=a(t_0)=1$, so that conformal frequencies or length scales correspond to physical scales today. In terms of the conformal wave-vector, the dispersion relation is like in flat space where it has been derived in Ref. [@shore], $$\omega^2 = k^2 \mp {8} \omega^2 |\Psi_0| \,. \label{term1}$$ Here $\Psi_0$ is the analog of the Newman-Penrose scalar (for more details see [@shore]), $$\Psi_0 =-a^{-4}\left[K_{0i0j} - K_{0ilj} n^l - K_{kilj}n^k n^l
\right] m^im^j\,,$$ where ${\bf m}$ and $\bar{\bf m}$ represent the left and right circular polarization basis vectors and ${\bf n}={\bf k}/k$ is the photon propagation direction. We normalize ${\bf n}$ and ${\bf m}$ with the flat metric, $n^in^j{\delta}_{ij}=m^i\bar
m^j{\delta}_{ij}=1$, and multiply the expression with the correct power of the scale factor, $a^{-4}$, so that, given the scaling of the tensor $K$, one sees explicitly that $\Psi_0$ is independent of the scale factor. (Latin indices indicate spatial components of a vector or tensor.)
The second term in the action $\Gamma_{\rm LV}$ leads to the dispersion relation [@cfj90; @shore] $$(k_\mu k^\mu)^2 + (k_\mu k^\mu) (L_\nu L^\nu) - (L_\mu k^\mu)^2
=0, \label{second}$$ To first order in the small parameters $L_0$ and $\sqrt{{\delta}^{ij}L_iL_j}\equiv L$ one has $$\omega^2 = k^2 \mp {\omega}(L_0 - L\cos\phi),
\label{second1}$$ where $\phi$ is an angle between the photon propagation direction and the vector ${\bf L}$, $\cos\phi=({\bf L}\cdot{\bf n})/L$. Note the similarity of the expressions (\[term1\]) and (\[second1\]) with the corresponding ones following from Eq. (\[eq:05\]).
To be as general as possible, we rewrite the dispersion relation for both types of Lorentz symmetry violation in the form (see also [@kgr07]), $$k^2 = \omega^2 \left[ 1\pm \left(\frac{M}{M_{\rm PL}}\right)
\left(\frac{\omega}{M_{\rm PL}}\right)^{N-4} \right],
\label{dispersion-gen}$$ where $M_{\rm PL}$ is the Planck mass, $M_{\rm PL} \simeq 1.2 \times 10^{19}$ GeV, $N$ is the dimension of the Lorentz symmetry violating operator and $M$ is a mass scale of the model. For $N=4$, the birefringent part is independent of the photon energy and we have $8\Psi_0=M/M_{\rm
PL}$. For $N=3$ the Planck mass cancels out and the birefringent term is inversely proportional to the photon energy. The mass scale is $M=L_0-L\cos\phi$. Generally speaking, the smaller $M$, the weaker LV. For $N=4$, LV is frequency independent and the amplitude of the effect is of order $M/M_{\rm Pl}$, while for the super-renormalizable case, $N=3$, LV is strongest at low frequencies, ${\omega}<M$. Our aim is to limit the function $$\gamma({\omega})
\equiv \left(\frac{M}{M_{\rm PL}}\right) \left(\frac{{\omega}}{M_{\rm
PL}}\right)^{N-4}$$ from CMB birefringence. This ansatz can also be applied to non-renormalizable models with higher dimension operators. For $N\ge 5$, $M\neq 0$ indicates that there is LV at frequencies ${\omega}\gsim M_{\rm Pl}(M_{\rm Pl}/M)^{\frac{1}{N-4}}$.
Results
=======
To compute the CMB polarization rotation angle induced by Lorentz symmetry violation, we follow the analogy with photon propagation in a magnetized medium which yields $n_{\pm} = 1 \pm
\gamma({\omega})/2$. Using Eq. (\[angle\]), we obtain $$\Delta \alpha^{(LV)} = \frac{1}{2} \omega \gamma({\omega}) \Delta l
\label{angle1}.$$ In the case $N=4$, $\gamma$ is frequency independent, hence $\Delta \alpha^{(LV)}$ grows linearly with frequency. In this case, and for all models with higher dimension operators, the best limits can in principle be obtained from high frequency photons (for example GRB $\gamma$-rays [@mitrofanov; @jlms04]), while CMB photons are less affected. However, the fact that the theory of CMB anisotropies and polarization yields that both $TB$ and $EB$ polarization have to vanish in standard cosmology, while the polarization of GRB’s is still under debate, at present, a test using CMB data is to be preferred. Another advantage is that for the CMB the distance $\Delta l \simeq H_0^{-1}$ is maximal.
In the dimension 3 model, $\Delta \alpha^{(LV)} = -\frac{1}{2}
(L_0-L\cos\phi) \Delta l$, is frequency-independent. In Ref. [@cfj90] the above result is applied to polarization data from distant radio galaxies, $\Delta \alpha < 6^{\rm o}$ at $95\%$ C.L. at redshift $z \sim 0.4$. The constraint obtained if Ref. [@cfj90] is $|L_0 - L\cos\phi | \leq 1.7 \times 10^{-42}
h_0$ GeV, where $h_0\simeq 0.7$ is the present Hubble parameter in units of $100$ km ${\rm s}^{-1}$ ${\rm Mpc}^{-1}$.
We use the recent WMAP-5 year constraints on the rotation angle of the CMB polarization plane (combined constraints from the low and high multipole number, $l$, data assuming a constant $\Delta
\alpha $ across the entire multipole range), $-5.9^{\rm o} <
\Delta \alpha < 2.4^{\rm o}$ at 95% C.L. and $\Delta \alpha =
-1.7^{\rm o} \pm 2.1^{\rm o}$ at 68% C.L. [@WMAP] (Sec. 4.3). Assuming Gaussian errors, it is straightforward to convert this to the following limits on the absolute value of rotation angle, $$\begin{aligned}
|\Delta \alpha |_{\rm obs} &\leq & 4.90^{\rm o}~~~~~~~{\rm at}
~~~~95\%~~{\rm C.L.}~, \label{95}\\| \Delta \alpha |_{\rm obs}
&\leq& 2.52^{\rm o}~~~~~~~{\rm at}~~~~68\%~~{\rm C.L.}~.
\label{68}\end{aligned}$$ We adopt $\Delta l \simeq 9.8 \times 10^9 h_0^{-1}$ years. We express our results in terms of $\nu_{100}=\nu/100{\rm GHz}$ to keep them as independent of the CMB band frequency as possible.
Using Eq. (\[angle1\]), we find the following limit on the function $\gamma(\nu)$ with ${\omega}=2\pi\nu$: $$\begin{aligned}
\label{e:gamma}
\gamma(\nu) &\leq& 8.6 \times 10^{-31}
\nu^{-1}_{ 100}h_0 ~ \mbox{ at 95\% C.L.,}\\
\gamma(\nu) &\leq& 4.4 \times 10^{-31} \nu^{-1}_{ 100}h_0 ~
\mbox{ at 68\% C.L.,}~.\end{aligned}$$ We can also express the limit on $\gamma$ in terms of a limit for the mass scale $M$ or the dimensionless parameter $M/M_{\rm Pl}$ $$\frac{M}{M_{\rm Pl}} \lsim 8.6\times10^{-31}
\left(3\times 10^{31}\right)^{(N-4)}\nu_{100}^{3-N}h_0 ~
\mbox{ at 95\% C.L.}~.$$ For $N>4$, these limits are not very interesting, while for $N=4$ or $N=3$ ’naturally expected’ values of the parameters are ruled out. More precisely, for the models considered we constrain the dimensionless scalar $\Psi_0$ for the $N=4$ case, $$|\Psi_0| \leq 1.1 \times 10^{-31} h_0 \nu^{-1}_{100} ~ \mbox{ at
95\% C.L.,}$$ while we find for $N=3$ $$|L_0-L\cos\phi| \leq 3.6 \times 10^{-43} h_0\mbox{GeV \qquad
at 95\% C.L.}~.$$ This is almost an order of magnitude better than the limit obtained in Ref. [@cfj90].
We can also introduce an effective photon “mass” by writing the modified dispersion relation in the form $\omega^2 = k^2 \pm
m_\gamma^2$ with $$m_\gamma^2 = \omega^2 \gamma({\omega}) = M \omega
\left(\frac{\omega}{M_{\rm PL}}\right)^{N-3} =
2\frac{\Delta\alpha}{\Delta l}\omega\,.$$ For $N> 2$ this is not a mass in the usual sense of the energy of the particle at rest, but rather a measure for the modification of the dispersion relation which tends to zero with frequency. For the renormalizable dimension 4 and 3 operators considered in this work we have $ m_{\gamma}^{(4D)}({\omega}) = 2 \omega |2\Psi_0|^{1/2}$ and $m_\gamma^{(3D)} = \left[\omega(|L_0 - L\cos\phi|)\right]^{1/2}$ respectively. As in Ref. [@mitrofanov] we can interpret our result also in terms of a polarization dependent group velocity, $$v_{\pm} = 1\pm \frac{N-2}{2}\frac{M}{M_{\rm Pl}}
\left(\frac{\omega}{M_{\rm Pl}}\right)^{N-4} = 1\pm
\frac{N-2}{2}{\gamma}({\omega})~.$$ Ref. [@mitrofanov] only studied the cases $N\ge 5$. From Eq. (\[e:gamma\]) we derive the constraint on the effective birefringent mass, $$\label{e:mass}
m_\gamma \leq 3.8 \times 10^{-19}\left( h_0
\nu_{100}\right)^{1/2} {\rm eV~~ at}~ 95\%~{\rm C.L.}$$ Note that left and right handed photons have effective square masses of opposite sign. For the velocity difference this implies $$\label{e:vel}
|v_+-v_-| \leq \left\{\begin{array}{lll}
8.6\times 10^{-31} h_0v_{100}^{-1} & \mbox{ at 95\% C.L.,} & \mbox{for } N=3\\
1.7\times 10^{-30} h_0v_{100}^{-1} & \mbox{ at 95\% C.L.,} & \mbox{for } N=4~.
\end{array} \right.$$ The limits on $m_\gamma$ are model independent because $m_\gamma$ only depends on the directly measured rotation angle $\Delta\alpha$ and on the frequency.
If $L \ll L_0$, we can safely neglect the angular dependence, and assume that $m_\gamma^{(3D)} = \sqrt{\omega L_0}$. However, if $L
\gg L_0$, the modification of the photon dispersion becomes direction dependent, and must be averaged over all sky for the CMB photons. Then, the rotation angle can be estimated by the two-point correlation function, $i.e.$, $\Delta \alpha_{\rm eff}
= \langle |\Delta \alpha|^2\rangle^{1/2}$. A rough estimate leads to a pre-factor $\sim1/\sqrt{2}$. In a more detailed analysis the presence of ${\bf L}$ breaks rotational symmetry and leads to off-diagonal correlations in the temperature anisotropy and polarization spectra analog to the effects on the CMB by a constant magnetic field [@Bconst; @faraday]. To take this fully into account requires to estimate the CMB Temperature-B polarization, E- and B-polarization cross correlations, as well as B-polarization spectra due to the Lorentz symmetry violating vector field ${\bf L}$, and to compare theoretical estimates with the corresponding CMB anisotropy and polarization data. Also the scalar $|\Psi_0|$ of the 4D model breaks rotational symmetry and taking the direction dependence of $\Delta\alpha$ into account is relatively complicated. This breakdown of statistical isotropy can also be tested using the bipolar power spectrum introduced in Ref. [@soura].
We shall address this issue in future work, but even though the limits may improve somewhat, we do not expect them to change significantly. $$$$
Conclusions
===========
The obtained bound on a birefringent effective photon mass is below the limit for a standard photon mass given by the particle data group [@particle], $m_\gamma \leq 3\times 10^{-19}< 10^{-18}$ eV, but is less stringent than the limit from galactic magnetic fields which is, however model dependent [@dvali]. Of course our photon mass would be measured only when measuring the dispersion relation of a polarized photon beam and would disappear when averaging over polarizations. It is not an ordinary mass.
Another useful bound is the departure of the refraction index in vacuum from unity, $i.e.$, $|\Delta n| =
|1-k/\omega|=|\gamma({\omega})|/2$. In the 4D model, $|\Delta n^{(4D)}|
\simeq 4 |\Psi_0|$, In the 3D model, $|\Delta n^{(3D)}| \simeq
L_0/ 2\omega $ (when $L\ll L_0$). Generically Eq. (\[e:gamma\]) implies $|\Delta n| \leq 4.3 \times 10^{-31} h_0
\nu^{-1}_{100}$. The difference of the refractive index from $1$ can be viewed as a difference of the photon speed from 1, $\Delta
c$ at the level of $10^{-30}$, which is much more stringent than the (more general) limit obtained in Ref. [@coleman], which is $\Delta c < 10^{-23}$. The formalism given here is applicable also for higher dimension operators, but due to the frequency dependence $|\alpha^{(LV)}| \propto \omega^{N-3}$, the CMB based limits on the amplitudes for higher dimension operators become much weaker that those given from high energy photons ($\gamma$ or $X$-rays). Even the bounds obtained from the nearby Crab Nebulae are more promising [@crab] if $N\geq 5$.
We appreciate useful comments and discussions with A. Kosowsky, G. Lavrelashvili, and B. Ratra. R.D. acknowledges support from the Swiss National Science Foundation. T.K. acknowledges hospitality of Geneva University, partial supports from GNSF grant ST06/4-096, and from the ICTP associate membership program. R.D. and T.K. acknowledge the INTAS grant 061000017-9258. T.K. and Y.M. have received partial support from DOE grant E-FG02-99ER41093.
V. A. Kostelecký, [*Third meeting on CPT and Lorentz Symmetry*]{} (World Scientific, Singapore, 2005).
G. M. Shore, Nucl. Phys. B [**717**]{}, 86 (2005).
D. Mattingly, Living Rev. Rel. [**8**]{}, 5 (2005).
T. Jacobson, S. Liberati, and D. Mattingly, Ann. Phys. [**321**]{}, 150 (2006).
W. Bietenholz, arXiv:0806.3713.
G. Amelino-Camelia, et al., Nature, [**393**]{}, 763, (1998).
S. Sarkar, Mod. Phys. Lett. A [**17**]{}, 1025 (2002); T. Piran, in [*Planck Scales Effects in Astrophysics and Cosmology*]{}, eds. J. Kowalski-Glikman and G. Amelino-Camelia (Springer, Berlin, 2005), p. 351.
T. Jacobson, S. Liberati, and D. Mattingly, Nature [**424**]{}, 1019 (2003).
e.g., S. D. Biller, et al. Phys. Rev. Lett. [**83**]{}, 2108 (1999); B. E. Schaefer, Phys. Rev. Lett. [**82**]{}, 4964 (1999); P. Kaaret, Astron. Astrophys. [**345**]{}, L32 (1999); J. R. Ellis, et al., Astrophys. J. [**535**]{}, 139 (2000); J. R. Ellis, et al., Astron. Astrophys. [**402**]{}, 409 (2003); J. R. Ellis, et al., Int. J. Mod. Phys. A [**19**]{}, 4413 (2004); E. Boggs, C. B. Wunderer, K. Hurley, and W. Coburn, Astrophys. J. Lett. [**611**]{}, L77 (2004); T. G. Pavlopoulos, Phys. Lett. B [**625**]{}, 13, (2005); J. R. Ellis, et al., Astropart. Phys. [**25**]{}, 402 (2006) \[Astropart. Phys. [**29**]{}, 158 (2008)\]; J. Ellis, et al., arXiv:0712.2781 \[astro-ph\]; M. Gogberashvili, A. S. Sakharov and E. K. G. Sarkisyan, Phys. Lett. B [**644**]{}, 179 (2007)
M. Rodríguez Martínez and T. Piran, JCAP [**4**]{}, 006 (2006).
B. Lovell, F.L. Whipple, and L. H. Solomon, Nature [**202**]{}, 377 (1964); B. Warner and R. E. Nather, Nature [**222**]{}, 157 (1969); G. Feinberg, Science, [**166**]{}, 879 (1969); Z. Bay and J. A. White, Phys. Rev. D [**5**]{}, 796 (1972); J. M. Rawls, Phys. Rev. D [**5**]{}, 487 (1972).
V. A. Kostelecký and M. Mewes, Phys. Rev. Lett. [**87**]{}, 251304 (2001).
G. Amelino-Camelia and T. Piran, Phys. Rev. D [**64**]{}, 036005 (2001); J. Magueijo and L. Smolin, Phys. Rev. D [**67**]{}, 044017 (2003).
R. C. Myers and M. Pospelov, Phys. Rev. Lett. [**90**]{}, 211001 (2003); R. Montemayor and L. F. Urrutia, Phys. Rev. D [**72**]{}, 045018 (2005).
J. M. Carmona, J. L. Cortés, J. Gamboa, and P. Méndez, Phys. Lett. B [**565**]{}, 222 (2003); J. Gamboa and J. López-Sarrión, Phys. Rev. D [**71**]{}, 067702 (2005); J. Gamboa, J. López-Sarrión and A. P. Polychronakos, Phys. Lett. B [**634**]{}, 471 (2006); T. Mariz, J. R. Nascimento and V. O. Rivelles, Phys. Rev. D [**75**]{}, 025020 (2007).
S. M. Carroll, G. B. Field, and R. Jackiw, Phys. Rev. D [**41**]{}, 141601 (1990).
B. Nodland and J. P. Ralston, Phys. Rev. Lett. [**78**]{}, 3043 (1997); D. Colladay and A. Kostelecký, Phys. Rev. D [**58**]{} (1998).
S. A. Martínez, R. Montemayor, and L. F. Urrutia, arXiv:gr-qc/0511117.
T. Kahniashvili, G. Gogoberidze and B. Ratra, Phys. Lett. B [**643**]{}, 81 (2006)
M. Galaverni and G. Sigl, Phys. Rev. D [**78**]{}, 063003 (2008).
e.g., T. Piran, Rev. Mod. Phys. [**76**]{}, 1143 (2004).
J. Albert [*et al.*]{} \[MAGIC Collaboration and Other Contributors\] Collaboration\], Phys. Lett. B [**668**]{}, 253 (2008); F. Aharonian [*et al.*]{}, Phys. Rev. Lett. [**101**]{}, 170402 (2008); X. J. Bi, Z. Cao, Y. Li and Q. Yuan, arXiv:0812.0121 \[astro-ph\].
I. Mitrofanov, Nature [**426**]{}, 139 (2003).
T. Jacobson, et al., Phys. Rev. Lett. [**93**]{}, 021101 (2004).
W. Coburn and S. E. Boggs, Nature [**423**]{}, 415 (2003); S. E. Boggs and W. Coburn arXiv:astro-ph/0310515.
R. E. Rutledge and D. B. Fox, Mon. Not. R. Astron. Soc. [**350**]{}, 1288 (2004); C. Wigger, et al., Astrophys. J. [**613**]{}, 1088 (2004).
D. R. Willis, et al., Astron. Astrophys. [**439**]{}, 245 (2005).
e.g., S. K. Solanki, M. P. Haugan, and R. B. Mann, Phys. Rev. D [**59**]{}, 047101 (1999); R. Gambini and J. Pullin, Phys. Rev. D [**59**]{}, 124021 (1999); N. F. Lepora, arXiv:gr-qc/9812077; R. J. Gleiser and C. N. Kozameh, Phys. Rev. D [**64**]{}, 083007 (2001); J. Alfaro, H. A. Morales-Técotl, and L. F. Urrutia, Phys. Rev. D [**65**]{}, 103509 (2002); K. R. Balaji, R. H. Brandenberger, and D. A. Easson, JCAP [**12**]{}, 008 (2003); Y. Z. Fan, D. M. Wei and D. Xu, Mon. Not. Roy. Astron. Soc. [**376**]{}, 1857 (2006); R. Lamon, N. Produit and F. Steiner, Gen. Rel. Grav. [**40**]{}, 1731 (2008); R. Lamon, JCAP [**0808**]{}, 022 (2008).
A. Lue, L. Wang, and M. Kamionkowski, Phys. Rev. Lett. [**83**]{}, 1506 (1999).
B. Feng, et al., Phys. Rev. Lett. [**96**]{}, 221302 (2006); S. H. S. Alexander, arXiv:hep-th/0601034; E. Y. Wu [*et al.*]{} \[QUaD Collaboration\], arXiv:0811.0618.
V. A. Kostelecky and M. Mewes, Phys. Rev. Lett. [**99**]{}, 011601 (2007); M. Mewes, arXiv:0804.0269.
M. Kamionkowski, arXiv:0810.1286.
E. Komatsu [ et al.]{} \[WMAP Collaboration\], arXiv:0803.0547.
E. S. Scannapieco and P. G. Ferreira, Phys. Rev. D [**56**]{}, 7493 (1997); C. Scoccola, D. Harari and S. Mollerach, Phys. Rev. D [**70**]{}, 063003 (2004); J. R. Kristiansen and P. G. Ferreira, Phys. Rev. D [**77**]{}, 123004 (2008).
R. Durrer, T. Kahniashvili and A. Yates, Phys. Rev. D58 123004 (1998); T. Kahniashvili, G. Lavrelashvili and B. Ratra, Phys. Rev. D [**78**]{}, 063012 (2008); A. Bernui and W. S. Hipolito-Ricaldi, Mon. Not. R. Astron. Soc. [**389**]{}, 1453 (2008).
N. A. Krall and A. W. Trivelpiece, [*Principles of Plasma Physics*]{} (McGraw-Hill, New York, 1973).
J. P. Vall[é]{}e, New Astron. Rev. [**48**]{}, 763, (2004).
A. Kosowsky and A. Loeb, Astrophys. J. [**469**]{}, 1 (1996);\
L. Campanelli, et al.,Astrophys. J. [**616**]{}, 1 (2004);\
A. Kosowsky, et al., Phys. Rev. D [**71**]{}, 043006 (2005).
F. Finelli and M. Galaverni, arXiv:0802.4210; M. Giovannini and K. E. Kunze, Phys. Rev. D [**78**]{}, 023010 (2008); T. Kahniashvili, Y. Maravin and A. Kosowsky, arXiv:0806.1876.
D. Grasso and H. R. Rubinstein, Phys. Rept. [**348**]{}, 163 (2001).
C. Amsler [*et al.*]{} \[Particle Data Group\], Phys. Lett. B [**667**]{}, 1 (2008).
G.V. Chibisov, Sov. Phys. Usp. [**19**]{}, 624 (1974); E. Adelberger, G. Dvali and A. Gruzinov, Phys. Rev. Lett. [**98**]{}, 010402 (2007).
S. R. Coleman and S. L. Glashow, Phys. Rev. D [**59**]{}, 116008 (1999).
P. Kaaret, Nature [**427**]{}, 287 (2004); L. Maccione, et al., JCAP [**0710**]{}, 013 (2007).
A. Hajian and T. Souradeep, Astrophys. J. [**597**]{}, L5 (2003).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Carnot heat engine sets an upper bound on the efficiency of a heat engine. As an ideal, reversible engine, a single cycle must be performed in infinite time, and so the Carnot engine has zero power. However, there is nothing in principle forbidding the existence of a heat engine whose efficiency approaches that of Carnot while maintaining finite power. Such an engine must have very special properties, some of which have been discussed in the literature, in various limits. While recent theorems rule out a large class of engines from maintaining finite power at exactly the Carnot efficiency, the approach to the limit still merits close study. Presented here is an exactly solvable model of such an approach that may serve as a laboratory for exploration of the underlying mechanisms. The equations of state have their origins in the extended thermodynamics of electrically charged black holes.'
author:
- 'Clifford V. Johnson'
bibliography:
- 'johnson\_carnot.bib'
title: |
An Exact Model of the Power/Efficiency Trade-Off\
While Approaching the Carnot Limit
---
It is well known that a heat engine, regardless of working substance and the details of the thermodynamic cycle, has a fundamental limit on its efficiency given by the Carnot efficiency $\eta_{\rm C}^{\phantom{C}}$. If $T_H$ is the highest operating temperature in the engine and $T_C$ the lowest, (the temperatures at which the input heat $Q_H$ and exhaust heat $Q_C$ are exchanged with the hot and cold reservoirs, respectively), the efficiency $\eta$ is bounded as follows: $$\eta=1-\frac{Q_C}{Q_H} \leq \eta_{\rm C}^{\phantom{C}}=1-\frac{T_C}{T_H}\ .$$ It is also familiar that the Carnot engine itself is an idealized reversible engine, with a cycle that is composed of two isotherms and two adiabats, with the expansions and compressions performed quasi–statically, in order to maintain reversibility. In other words, it takes an infinite amount of time to perform one cycle of the Carnot engine: It has zero power.
While most typical heat engines, working at finite power, operate well below the Carnot efficiency, there is no issue of principle that prevents their efficiency from approaching that of Carnot, but it becomes increasingly difficult for typical working substances and choices of operating cycle. The question naturally arises as to what kind of engine is needed to approach the Carnot efficiency while maintaining finite power. (This is a separate issue from the Curzon–Ahlborn bound on $\eta$ when working at maximum power [@1975AmJPh..43...22C].) There have been recent discussions of this in the thermodynamics and statistical mechanics literature [@2011PhRvL.106w0602B; @2013PhRvL.111e0601A; @2015PhRvL.115i0601P; @PhysRevX.5.031019; @2015PhRvL.114e0601P; @2016NatCo...711895C; @2016EPJB...89..248K; @2016PhRvL.117s0601S; @2017arXiv170101914S], and two papers in particular [@2015PhRvL.114e0601P; @2016NatCo...711895C] consider having the working substance near criticality as an approach to the problem, exploiting either fluctuations, or a diverging heat capacity to argue for the maintenance of finite power as $\eta$ grows closer to $\eta_{\rm C}^{\phantom{C}}$. It has been argued in refs. [@2016PhRvL.117s0601S; @2017arXiv170101914S] that it is forbidden (for widely applicable assumptions) to be exactly at the Carnot efficiency while at finite power, but the issue of the approach to the limit is still of considerable interest, for both practical and theoretical reasons. This paper presents an exactly solvable model of such an approach that may be of use in gaining better understanding of how various models (perhaps less computationally accessible) work. A critical system will also feature in the present work, but its role appears to be of quite a different character from what was argued for in the systems of refs. [@2015PhRvL.114e0601P; @2016NatCo...711895C]. Fluctuations and diverging specific heat do not explicitly play an essential role in the core construction. This can be determined because the system employed is can be readily queried with a computation: All the needed properties of the working substance are available [*via*]{} a full set of exact defining equations.
The system to be used here has its origins in [*extended*]{} gravitational thermodynamics: The traditional treatment [@Gibbons:1976ue] of black holes in semi–classical quantum gravity supplies them with a temperature $T$ and an entropy $S$, which depend upon the mass $M$ and the horizon radius $r_+$. This treatment is extended [@Kastor:2009wy] by making dynamical the cosmological constant ($\Lambda$) of the gravity theory [^1], which yields a pressure variable $p=-\Lambda/8\pi$ [^2] and its conjugate volume $V\equiv (\partial H/\partial p)_S$. The enthalpy $H$ is the black hole’s mass, and the First Law in terms of all these quantities is $dH=TdS+Vdp$. Studies of the extended thermodynamics of gravitational systems have uncovered many phenomena that are familiar from statistical physics and thermodynamics. (For a recent review see ref. [@Kubiznak:2016qmn].)
Notice that for negative $\Lambda$ the pressure is positive. Ref. [@Johnson:2014yja] presented the idea of defining heat engines that do traditional mechanical work $W=\int pdV$ in this context [^3]. The heat flows $Q_H$ and $Q_C$ into and out of the engine can be considered as from and to non–backreacting heat baths of radiation filling the spacetime, as is traditional in black hole thermodynamics. (See [*e.g.*]{} ref. [@Hawking:1982dh].) Such engines were called holographic heat engines since gravitational physics in spacetimes with negative $\Lambda$ is known to have a dual description in terms of strongly coupled non–gravitational physics (in one dimension fewer). These are examples of a broader phenomenon in quantum gravity known as holography. (For a review, see ref. [@Aharony:1999ti]). Such dualities are not needed here, but it is worth noting that if the gravitational language is not to a reader’s taste, this could all in principle be translated to the language of a class of strongly coupled gauge theories. In other words, the gravitational aspect of this example is not essential, but it is more economical to use that simpler language.
The context will be a $(3{+}1)$–dimensional Einstein–Maxwell system with action: $$I=-\frac{1}{16\pi }\int \! d^4x \sqrt{-g} \left(R-2\Lambda -F^2\right)\ ,
\label{eq:action}$$ where $\Lambda{=}-3/l^2$ sets a length scale $l$. The black hole spacetimes of interest here are Reissner–Nordstrom–like solutions of charge $q$. The metric and gauge potential are: $$\begin{aligned}
ds^2 &=& -Y( r)dt^2
+ {dr^2\over Y(r)} + r^2 (d\theta^2+\sin^2\theta d\varphi^2)\ ,\nonumber \\
Y( r) &\equiv& 1-\frac{2M}{r}+\frac{q^2}{r^2}+\frac{r^2}{l^2}\ , A_t = q\left(\frac{1}{r_+}-\frac{1}{r}\right)\! .\end{aligned}$$ The potential is chosen to vanish on the horizon at $r=r_+$, the largest positive real root of $Y(r)$.
Several aspects of the thermodynamics of these solutions were studied in refs. [@Chamblin:1999tk; @Chamblin:1999hg]. There, a rich phase structure was uncovered, a van der Waals–like nature was elucidated, including a second order critical point. By including variable $\Lambda$ (and hence a pressure $p$), ref. [@Kubiznak:2012wp] clarified the resemblance to van der Waals and showed that the system has the same universal behaviour near the critical point as the van der Waals gas.
The standard semi–classical quantum gravity procedures [@Gibbons:1976ue] yield a temperature for each black hole solution, which depends on $r_+$, $q$, and $\Lambda$. The entropy is given by a quarter of the area of the horizon: $S=\pi r_+^2$. The extended thermodynamics [@Kastor:2009wy] allows for all appearances of the length scale $l$ to be replaced by the pressure $p$ using the relation $p=3/(8\pi l^2)$, and the thermodynamic volume turns out to be $V=4\pi r_+^3/3$. So all occurrences of $r_+$ can be traded in for either an $S$ or a $V$, as they are not independent. All of this results in an equation of state $p(V,T)$: $$%p=\frac{T}{v}-\frac{1}{2\pi v^2}+\frac{2q^2}{\pi v^4}
p=\frac{1}{8\pi}\left(\frac{4\pi}{3}\right)^\frac43\left(\frac{3T}{V^\frac13} - \left(\frac{3}{4\pi}\right)^{\frac23}\frac{1}{V^\frac23}+\frac{q^2}{V^\frac43}\right)\ .
\label{eq:equationofstate}$$ Some sample isotherms are plotted in figure \[fig:isotherms\]. Note that there is a wedge–shaped exclusion region extending from the $V=0$ axis, bounded on the right by the $T=0$ curve (the dashed line) and on the bottom by the $p=0$ line. Points inside that region are unphysical, having $T<0$. (The black hole at $(T=0, p=0)$ is the extremal Reissner–Nordstrom black hole of volume $V_0=4\pi q^3/3$.) Below a critical isotherm the isotherms yield multiple values for $p$, and (in full parallel with the classic van der Waals system) are “repaired" by an isobar (not shown in figure \[fig:isotherms\]) at a value of the pressure determined by a study of the free energy. This results in a family of first order phase transitions between large and small black holes terminating in a second order critical point at the critical isotherm [@Chamblin:1999tk; @Chamblin:1999hg]. The details of the first order transitions will not affect the main issue being addressed here, and so they won’t be explored further.
![Main: Sample isotherms for $q{=}4$. The temperature is higher for the curves further away from the origin. The central (blue) isotherm is at the critical temperature, and the (blue) cross marks the critical point. The isotherms at lower temperatures get modified, as discussed in the main text, but this is not shown here. The dotted green rectangle is an example of the special engine cycle discussed in the text (with $L{=}1$). The dashed curve is the $T{=}0$ isotherm. Inset: The labelling of the engine cycle.[]{data-label="fig:isotherms"}](isotherms_inset){width="3.0in"}
An equivalent expression to eq. (\[eq:equationofstate\]) is: $$T=\frac{1}{4\sqrt{\pi}}S^{-\frac32}\left(8pS^2+S-{\pi q^2}\right)\ .
\label{eq:temperature}$$ Meanwhile the mass (enthalpy) $H(S,p)$ is [@Dolan:2011xt]: $$M\equiv H=\frac{1}{6\sqrt{\pi}} S^{-\frac12}\left(8pS^2+3S+3\pi q^2\right) \ ,
\label{eq:enthalpy}$$ and the constant $V$ and $p$ specific heats are [@Kubiznak:2012wp]: $$C_V=0 \ ; C_p=2S\left(\frac{8pS^2+S-\pi q^2}{8pS^2-S+3\pi q^2}\right)\ .$$
It is these black holes that were the working substance in the prototype holographic heat engine of ref. [@Johnson:2014yja], using a rectangular cycle in the $(p,V)$–plane made of isobars and adiabats (which are equivalent to isochors for static black holes since $S$ and $V$ both depend only on $r_+$). The inset of figure \[fig:isotherms\] shows the labelling of the cycle to be used below. Later, in ref. [@Johnson:2016pfa], an exact equation for the efficiency was derived for the cycle, and it will be extremely useful here. Key is that the heat flows can be written as mass/enthalpy differences, giving: $$\eta=1-\frac{Q_C}{Q_H}=1-\frac{M_3-M_4}{M_2-M_1}\ ,
\label{eq:efficiency-prototype}$$ where $M_i$ is the black hole mass evaluated at the $i$th corner. Its simplicity means that there is no need to make the kinds of approximations ([*e.g.*]{} high temperature or pressure) usually needed to write explicit efficiency formulae for some particular choice of location of this cycle in the $(p,V)$ plane.
The next step is to decide where to place the cycle in the $(p,V)$ plane. In previous work in this area, $q$ has been treated essentially as a label for a family of solutions, and was conveniently set to a positive non–zero value and forgotten about, since the key features don’t depend upon its actual value. This will not be the case here. Consider making $q$ large, for reasons that will become clear shortly. For a given choice of the position of the cycle (choosing a range for $p$ and $V$ (or $S$)), a sensible engine can be defined, but for large enough $q$ eventually the system will become unphysical: $T$ (see eq. (\[eq:temperature\])) on some parts of the cycle starts becoming negative because the exclusion region moves to the right with increasing $q$. This can be avoided by seeking choices for the $p,V$ (or $S$) coordinates of the cycle variables that scale with $q$ in such a way as to stay physical. Eq. (\[eq:temperature\]) shows that the scaling is $S\sim q^2$, $p\sim q^{-2}$, and hence $T\sim q^{-1}$. There is a very distinguished point exhibiting this exact scaling behaviour. It is the second order critical point, defined by the $p(V,T)$ curve with a point of inflection: $\partial p/\partial V=\partial^2 p /\partial V^2=0$: $$\label{eq:critical}
p_{\rm cr}=\frac{1}{96\pi q^2}\ , \quad S_{\rm cr}=6\pi q^2\ , \quad T_{\rm cr}= \frac{1}{3\sqrt{6}\pi q}\ ,$$ with $V_{\rm cr}= 8\sqrt{6}\pi q^3.$ (See fig. \[fig:isotherms\] for the case of $q=4$.)
So if the engine cycle is chosen to be in the neighbourhood of this point, and of a size that does not extend into the exclusion region, it will be physical. There are several ways of making such a choice, and one family will be chosen here for illustration. Place the critical point at corner 3: $p_3=p_4=p_{\rm cr}$, and choose the upper isobar [^4] to be at some multiple of $p_{\rm cr}$: $p_1=p_2=3p_{\rm cr}/2$. (See the dotted rectangle in fig. \[fig:isotherms\].) In preparation for large $q$ it is to be noted that since $p\sim q^{-2}$, the cycle is in danger of shrinking to zero area, giving vanishing work and hence vanishing $\eta$. But if $V_2-V_1=V_3-V_4$ are chosen to scale as $q^2$, then the work will be finite at any $q$. So $V_2=V_3=V_{\rm cr}$ while $V_1=V_4= V_{\rm cr}-V_{\rm cr}L/q$, where $L$ is a constant. This gives $W=p_{\rm cr}V_{\rm cr}L/2q=L/4\sqrt{6}$. So as $q$ is increased, the whole cycle shrinks vertically, but grows horizontally in such a way as to keep the work done finite. It is now a matter of studying the $q$ dependence of the input heat $Q_H$. It is simply the mass (enthalpy) difference $M_2-M_1$, with $p=3p_{\rm cr}/2$, $S_2=S_{\rm cr}$ and $S_1=S_{\rm cr}(1-L/q)^\frac23$ placed into eq. (\[eq:enthalpy\]). Interestingly, the large $q$ expansion of $Q_H$ decreases to a limiting value: $$Q_H= \frac{19\sqrt {6}}{72}L+\frac{\sqrt {6}}{27}{\frac {{L}^{2}}{q}}+{\frac {4\,{L}^{3}
\sqrt {6}}{243\,{q}^{2}}}+O \left( {q}^{-3} \right)\ , %+{\frac {25\,{L}^{4}\sqrt {6}}{2916\,{q}^{3}}
%}+O \left( {q}^{-4} \right)$$ and hence the efficiency $\eta=W/Q_H$ is, at large $q$: $$\eta=\frac{3}{19}-\frac{8}{361}{\frac {L}{q}}-{\frac {416\,{L}^{2}}{61731\,{q}^{2}}}-{\frac {3286{
L}^{3}}{1172889\,{q}^{3}}}+O \left( {q}^{-4} \right) \ .$$ The next thing to do is compute the Carnot efficiency for the engine. Directly inserting the chosen values for ($S_2,p_2$) and ($S_4,p_4$) into eq. (\[eq:temperature\]) gives $T_H{=}(19\sqrt{6}/288)(\pi q)^{-1}$ exactly, while the large $q$ expansion for $T_C$ begins: $$T_C= \frac{1}{18}\,{\frac {\sqrt {6}}{\pi\,q}}-{\frac {{L}^{3}\sqrt {6}}{972\,\pi\,
{q}^{4}}}-{\frac {7\,{L}^{4}\sqrt {6}}{3888\,\pi\,{q}^{5}}}+O \left( {
q}^{-6} \right) \ ,$$ and so: $$\label{eq:carnot_critical}
\eta_{\rm C}^{\phantom{C}}=1-\frac{T_C}{T_H} = {\frac{3}{19}}+{\frac {8\,{L}^{3}}{513\,{q}^{3}}}+{\frac {14\,{L}^{4}
}{513\,{q}^{4}}}+O \left( {q}^{-
5} \right)
\ .$$
These simple but striking results constitute the main demonstration promised for this paper. (Fig. \[fig:ratio\] is a plot of the ratio $\eta/\eta_{\rm C}^{\phantom{C}}$ [*vs.*]{} $q$, showing the rise to unity at large $q$.) This is a heat engine that does finite work at any $q$, and $\eta\to\eta_{\rm C}^{\phantom{C}}$ as $q\to\infty$. This is atypical, since usually going to the Carnot limit for one of the classic heat engine cycles (or variants thereof) translates into vanishing or infinite work. As an example, the Otto cycle has $\eta=1-r^{1-\gamma}$, where $\gamma=C_p/C_V>1$ and $r$ is the ratio of largest to smallest volumes, and so $\eta$ is maximized for $r\to\infty$. As another, the (Brayton–like) rectangular cycle defined for black holes at high pressures and temperatures well away from the critical point has [@Johnson:2014yja] $\eta=(1-p_4/p_1)(1+O(1/p_1)\cdots)$ (using the labelling in figure \[fig:isotherms\]), which may be written as: $$\eta=1-\frac{T_C}{T_H}\left(\frac{V_2}{V_4}\right)^{1/3}+\cdots \ ,$$ This is an analogue of an ideal gas limit [@Johnson:2015ekr], and Carnot efficiency is approached if $V_2\to V_4$, resulting in no work. In the case under consideration however, the region of interest is far from an ideal gas regime and in fact as $q$ grows $T$ decreases. This is appealing since most of the current discussions in the literature about physical realizations of finite power efficient engines are about low temperature experiments. Finite work here as $\eta\to\eta_{\rm C}^{\phantom{C}}$ is a useful feature to have under control on the way to studying finite power.
The time taken to do a cycle, $\tau$, is all that needs to be examined next. On the face of it, that seems to be finite at any finite $q$ (but see further discussion below) and so this is indeed a model of an [*approach*]{} to Carnot while maintaining finite power in the following precise sense: An efficiency as close to the Carnot efficiency as desired can be achieved by choosing large enough $q$, and picking the cycle according to the prescription above. Precisely at the limit $\eta=\eta_{\rm C}^{\phantom{C}}$, however, while the work is finite, the power has vanished, since the pressures in the isobars are proportional to $p_{\rm cr}\sim q^{-2}$, which vanishes in the limit, meaning that the time it takes to perform the isobaric expansions and compressions diverges. So the inequality of [*e.g.*]{} refs.[@2016PhRvL.117s0601S; @2017arXiv170101914S] showing the unattainability of finite power exactly at Carnot efficiency is easily satisfied. It is: $$\frac{W}{\tau}\leq{\bar\Theta} \frac{\eta(\eta_{\rm C}^{\phantom{C}}-\eta)}{T_C}\ ,$$ where ${\bar\Theta}$ is a model dependent constant capturing the characteristics of the engine. For the current model, the right hand side (divided by ${\bar\Theta}$) has a large $q$ expansion that begins as: $%\eta(\eta_{\rm C}-\eta)/T_C=
{ {72\pi\sqrt {6}}L/{6859}}+{ {224\pi\sqrt {6}}L^2/{(130321
q)}}%+{\frac {19358\,\pi\,\sqrt {6}}{2476099\,{q}^{2}}}
+O \left( {q}^{-2}\right).$ Meanwhile, the work $W{=}L/4\sqrt{6}$, but an optimistic estimate of the $q$ dependence of $\tau$ based on the behaviour of the pressures (discussed above) is $\tau\sim q^2$.
The above cycle is just one example of the kind of scheme that will work. Variations were studied, and some are worth reporting the results for. One is to put the critical point at a different point on the lower isobar. This results in qualitatively similar results at large $q$. The difference is that both $T_H$ and $T_C$ have higher order corrections to their leading $1/q$ form at large $q$. Again, $\eta$ and $\eta_{\rm C}^{\phantom{C}}$ converge at large $q$ to $3/19$. It should also be noted that the approach $\eta\to\eta_{\rm C}$ at large $q$ is achieved even if the critical point is not anywhere on the cycle itself. It suffices to be near enough to the critical region, in the manner outlined.
![The behaviour of the ratio $\eta/\eta_{\rm C}$ as a function of $q$ for the prototype scheme. Variations of the scheme discussed in the text behave qualitatively similarly.[]{data-label="fig:ratio"}](efficiency_ratio){width="2.0in"}
A concern that might be raised is whether the presence of a critical point somewhere on the cycle might invalidate the claim to be able to achieve finite power (at large but finite $q$) due to possible critical slowing of the system. As mentioned above, similar results were achieved by avoiding the critical point, only having the cycle be near it, with $\eta$ approaching $\eta_{\rm C}^{\phantom{C}}$ at large $q$ as before. So the critical point’s presence directly on the cycle is not crucial. There might have been an expectation that certain aspects of the physics near the critical point itself are responsible for the (finite power) approach to the Carnot efficiency at large $q$. For example, in the discussion in ref. [@2016NatCo...711895C] it is argued that the divergence of the specific heat produces an enhancement in the power’s scaling with the effective system size, $N$, through an enhancement in the heat flows. (There are $N$ coupled quantum Otto engines constituting the system.) Here, $q$ acts as a system size parameter analogous to that paper’s $N$. Indeed, near criticality $C_p$’s peak (inverse) width and height are enhanced with increasing $q$, as happens in ref. [@2016NatCo...711895C]. But the explicit expressions for $Q_H$ and $Q_C{=}W{-}Q_H$ show that they actually decrease with increasing $q$, even with the critical point on the isobar. Also, the fact that the same qualitative behaviour happens away from the critical point suggests that in this model the peak in $C_p$ plays no crucial role in driving the efficiency toward Carnot. On the other hand, since the construction presented here removes the $q$ dependence (analogously, $N$ dependence) from the work and places it all into the heat flows, it is difficult to compare the approaches further.
Nevertheless, the critical point itself [*is*]{} important in the whole scheme, since as shown, its neigbourhood (which depends on $q$, see eq. (\[eq:critical\])) is key in determining the coordinates of the cycle needed to approach the Carnot efficiency as $q$ increases. A qualitative reason why this all works so well is as follows: The neighbourhood of the critical point on the critical isotherm, being a region containing a point of inflection, is locally quite horizontal. Other isotherms in the region will inherit some of this behaviour, and this is even more true at higher $q$. Close to horizontal means that they do not deviate too far from the isobar shape of the 1–2 and 3–4 parts of the cycle. As discussed earlier, the vertical 2–3 and 4–1 isochoric parts are also adiabats (because of the properties of static black holes). So a prescription for picking a cycle that stays in the neighbourhood of the critical point therefore ensures that the cycle itself becomes an increasingly better approximation to a Carnot cycle (two isotherms and two adiabats) as $q$ grows. The behaviour of the pressures resulting from this is such that they will vanish in the limit and result in diverging $\tau$ as expected for Carnot.
The underlying system controlling the physics at large $q$ is worth further investigation: It has low pressure and temperature, and high volume and entropy. In the gravitational model it originates as a special family of large charge black holes, but there might be analogues of such equations of state in other, non–gravitational, systems. They would be interesting to identify.
CVJ thanks the US Department of Energy for support under grant DE-FG03-84ER-40168, the Simons Foundation for a Simons Fellowship (2017), and Amelia for her support and patience.
[^1]: A dynamical $\Lambda$ can be naturally implemented if the gravity theory is embedded within a theory with other dynamical sectors. An example is when there are dynamical scalars $\varphi_i$ with a potential $V(\varphi_i)$. Moving between fixed points of the potential, where the scalars take on fixed values $\varphi^{\rm c}_i$, that set a non–zero constant $V(\varphi^{\rm c}_i)$, setting the value of $\Lambda$. See [*e.g.*]{} the review in ref. [@Aharony:1999ti].
[^2]: Here we are using geometrical units where $G,c,\hbar,k_{\rm B}$ have been set to unity.
[^3]: Work can be interpreted here results in a change of the overall energy of a spacetime since the volume removes removes portions of it from the standard energy integral. Recall that $p$ sets an energy density [*via*]{} $\Lambda$’s equation of state $\rho=-p$ and so a positive change $dV$ results in an energy gain $|\rho| dV$. See ref. [@Kastor:2009wy].
[^4]: The choice of placing the critical point on the top isobar can also be made. The cycle dips into the repaired region where there are first order transitions. However, it is easier to determine the temperatures on the isobars away from that region, and so for the sake of simplicity, the bottom isobar was chosen.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $G \simeq M \rtimes C$ be an $n$-generator group which is a split extension of an Abelian group $M$ by a cyclic group $C$. We study the Nielsen equivalence classes and T-systems of generating $n$-tuples of $G$. The subgroup $M$ can be turned into a finitely generated faithful module over a suitable quotient $R$ of the integral group ring of $C$. When $C$ is infinite, we show that the Nielsen equivalence classes of the generating $n$-tuples of $G$ correspond bijectively to the orbits of unimodular rows in $M^{n -1}$ under the action of a subgroup of ${\textnormal{GL}}_{n - 1}(R)$. Making no assumption on the cardinality of $C$, we exhibit a complete invariant of Nielsen equivalence in the case $M \simeq R$. As an application, we classify Nielsen equivalence classes and ${\operatorname{T}}$-systems of soluble Baumslag-Solitar groups, split metacyclic groups and lamplighter groups. We also exhibit, for every $N \ge 1$, a two-generated Abelian-by-(infinite cyclic) group with exactly $N$ Nielsen equivalence classes and $N$ ${\operatorname{T}}$-systems of generating triples.'
address: |
EPFL ENT CBS BBP/HBP. Campus Biotech. B1 Building, Chemin des mines, 9\
Geneva 1202, Switzerland
author:
- Luc Guyot
bibliography:
- 'Biblio.bib'
title: Generators of split extensions of Abelian groups by cyclic groups
---
Introduction {#SecIntro}
============
Given a finitely generated group $G$, we denote by ${\textnormal{rk}}(G)$ the minimal number of its generators. For $n \ge {\textnormal{rk}}(G)$, we let ${\operatorname{V}}_n(G)$ be the set of *generating $n$-vectors* of $G$, i.e., the set of elements in $G^n$ whose components generate $G$. In order to classify generating vectors, we can rely on a well-studied equivalence relation on ${\operatorname{V}}_n(G)$, namely the *Nielsen equivalence relation*: two generating $n$-vectors are said to be *Nielsen equivalent* if they can be related by a finite sequence of transformations taken in the set $\{ L_{ij}, I_i; 1 \le i \neq j \le n\}$ where $L_{ij}$ and $I_i$ replace the component $g_i$ of ${\mathbf{g}}= (g_1, \dots, g_n) \in {\operatorname{V}}_n(G)$ by $g_j g_i$ and $g_i^{-1}$ respectively and leave the other components unchanged. We recommend [@Lub11; @Eva07; @Pak01; @LM93] to the reader interested in Nielsen equivalence and its applications. Let $F_n$ be the free group with basis ${\mathbf{x}}= (x_1,\dots,x_n)$. The Nielsen equivalence relation turns out to be generated by an ${\textnormal{Aut}}(F_n)$-action. Indeed, the set ${\operatorname{V}}_n(G)$ identifies with the set $\operatorname{Epi}(F_n,G)$ of epimorphisms from $F_n$ onto $G$ via the bijection ${\mathbf{g}}\mapsto \pi_{{\mathbf{g}}}$ with $\pi_{{\mathbf{g}}}$ defined by $\pi_{{\mathbf{g}}}({\mathbf{x}}) = {\mathbf{g}}$. Therefore defining ${\mathbf{g}}\psi$ for $\psi \in {\textnormal{Aut}}(F_n)$ through $\pi_{{\mathbf{g}}\psi} \Doteq \pi_{{\mathbf{g}}} \circ \psi$ yields a right group action of ${\textnormal{Aut}}(F_n)$ on ${\operatorname{V}}_n(G)$. Because ${\textnormal{Aut}}(F_n)$ has a set of generators which induce the elementary Nielsen transformations $L_{ij}$ and $I_i$ [@LS77 Proposition 4.1], this action generates the Nielsen equivalence relation.
In this article, we are concerned with finitely generated groups $G$ containing an Abelian normal subgroup $M$ and a cyclic subgroup $C$ such that $G = MC$ and $M \cap C = 1$. Denoting by $\sigma$ the natural map $G \twoheadrightarrow G/M \simeq C$, such a group $G$ fits into the split exact sequence $$\label{EqExt}
0 \longrightarrow M \longrightarrow G \xlongrightarrow{\sigma} C \longrightarrow 1$$ where the arrow from $M$ to $G$ is the inclusion $M \subset G$. The cyclic group $C = \langle a \rangle$ is finite or infinite and is given together with a generator $a$. The action of ${\textnormal{Aut}}(F_n)$ on ${\operatorname{V}}_n(G)$ is known to be transitive if $n > {\textnormal{rk}}(G) + 2$ [@Eva93 Theorem 4.9]. Our goal is to describe the ${\textnormal{Aut}}(F_n)$-orbits for the three exceptional values of $n$, namely ${\textnormal{rk}}(G)$, ${\textnormal{rk}}(G) + 1$ and ${\textnormal{rk}}(G) + 2$. To this end, we relate the problem of classifying Nielsen equivalence classes to a pure module-theoretic problem involving $M$. The conjugacy action of $C$ on $M$ defined by ${{\vphantom{m}}^{c}{m}} \Doteq cmc^{-1}$, with $m \in M$ and $c \in C$ extends linearily to ${\mathbb Z}{\lbrack C \rbrack}$, turning $M$ into a module over ${\mathbb Z}{\lbrack C \rbrack}$. Let ${\operatorname{ann}}(M)$ be the annihilator of $M$. Then $M$ is a faithful module over $$R \Doteq { {\mathbb Z}\lbrack C \rbrack}/{\operatorname{ann}}(M).$$
Let ${\textnormal{rk}}_R(M)$ be the minimal number of generators of $M$ considered as an $R$-module. For $n \ge {\textnormal{rk}}_R(M)$, we denote by ${\textnormal{Um}}_n(M)$ the set of elements in $M^n$ whose components generate $M$ as an $R$-module. The group ${\textnormal{GL}}_n(R)$ acts on ${\textnormal{Um}}_n(M)$ by matrix right-multiplication. There are two subgroups of ${\textnormal{GL}}_n(R)$ which are relevant to us. The first is ${\textnormal{E}}_n(R)$, the subgroup generated by the elementary matrices, i.e., the matrices that differ from the identity by a single off-diagonal element (agreeing that ${\textnormal{E}}_1(R) = \{1\}$). The second is ${\textnormal{D}}_n(T)$, the subgroup of diagonal matrices whose diagonal coefficients belong to $T$, the *group of trivial units*. We call a unit in $R^{\times} \Doteq {\textnormal{GL}}_1(R)$ a *trivial unit*, if it lies in the image of $\pm C$ by the natural map ${ {\mathbb Z}\lbrack C \rbrack}\twoheadrightarrow R$. Our first result establishes a connection between the ${\textnormal{Aut}}(F_n)$-orbits of generating $n$-vectors and the orbits of unimodular rows in $M$ with size $n -1$ under the action of $$\Gamma_{n -1}(R) \Doteq {\textnormal{D}}_{n- 1}(T) {\textnormal{E}}_{n -1}(R).$$
Additional definitions are needed to state this result. Denoting by $\vert C \vert$ the cardinal of $C$, we define the *norm element* of ${ {\mathbb Z}\lbrack C \rbrack}$ by $0$ if $C$ is infinite, and by $1 + a + \cdots + a^{{\vert C \vert}- 1}$ otherwise. Let $\nu(G)$ be the image in $R$ of the norm element of ${ {\mathbb Z}\lbrack C \rbrack}$ via the natural map. Let $\pi_{ab} :G \twoheadrightarrow G_{ab}$ be the abelianization homomorphism of $G$ and let $M_C$ be the largest quotient of $M$ with a trivial $C$-action. We assume throughout this paper that $n \ge \max({\textnormal{rk}}(G), 2)$ whenever the integer $n$ refers to the size of generating vectors of $G$. Let $\varphi_a: {\textnormal{Um}}_{n - 1}(M) \rightarrow {\operatorname{V}}_n(G)$ be defined by $\varphi_a({\mathbf{m}}) = ({\mathbf{m}}, a)$. This is elementary to check that $\varphi_a$ induces a map $$\Phi_a: {\textnormal{Um}}_{n - 1}(M)/\Gamma_{n - 1}(R) \rightarrow {\operatorname{V}}_n(G)/{\textnormal{Aut}}(F_n)$$If $\nu(G) = 0$, e.g, $C$ is infinite, then Lemma \[LemNu\] below shows that $n > {\textnormal{rk}}_R(M)$ holds true and $\Phi_a$ is surjective. Our first result fully characterizes the latter two conditions.
\[ThMTimesC\] The inequality $n > {\textnormal{rk}}_R(M)$ holds and the map $\Phi_a$ is surjective if, and only if, at least one of the following holds:
- $n > {\textnormal{rk}}(G_{ab})$.
- $C$ is infinite.
- ${\textnormal{rk}}(M_C) < {\textnormal{rk}}(G)$ and $M_C$ is not isomorphic to ${\mathbb Z}^{{\textnormal{rk}}(G) - 1}$.
- ${\vert C \vert}\in \{2, 3, 4, 6\}$ and $M_C$ is isomorphic to ${\mathbb Z}^{{\textnormal{rk}}(G) - 1}$.
In addition, the map $\Phi_a$ is bijective if $C$ is infinite.
Evidently, Theorem \[ThMTimesC\] has no bearing on the case $n = {\textnormal{rk}}(G) = {\textnormal{rk}}_R(M)$. Proposition \[PropCyclicDecMaxRank\] below handles this situation only when $M$ is a free module and most of our results assume that $n > {\textnormal{rk}}_R(M) $. Combined with various assumptions on $C$, $M$ or $R$ (e.g., $C$ is infinite and $R$ is Euclidean), Theorem \[ThMTimesC\] provides a complete description of Nielsen equivalence classes of generating $n$-vectors for all $n > {\textnormal{rk}}_R(M)$. All applications to groups with arbitrary ranks are gathered in Corollaries \[CorGE\] and \[CorGEMFree\] below. We present now a specific example that will help understand to which groups Theorem \[ThMTimesC\] can apply. For $n \ge 2$, let us denote by ${\mathfrak{n}}_n(G)$ the cardinality of the set of Nielsen equivalence classes of generating $n$-vectors of $G$.
\[CorFpToDRtimesZ\] Let $p$ be a prime and let $d \ge 1$. Let $G = {\mathbb F}_p^d \rtimes_A {\mathbb Z}$ where ${\mathbb F}_p$ denotes the field with $p$ elements and where the canonical generator of ${\mathbb Z}$ acts on ${\mathbb F}_p^d$ as a matrix $A \in {\textnormal{GL}}_d({\mathbb F}_p)$. Then ${\mathfrak{n}}_{{\textnormal{rk}}(G)}(G) = \vert R^{\times}/T \vert$ where $R = {\mathbb F}_p {\lbrack X^{\pm 1} \rbrack} / (P(X))$, $P(X) \in {\mathbb F}_p{\lbrack X \rbrack}$ is the first invariant factor of $A$ and $T$ is the subgroup of $R^{\times}$ generated by the images of $-1$ and $X$. Moreover, ${\mathfrak{n}}_n(G) = 1$ if $n > {\textnormal{rk}}(G)$.
In the above example, the polynomial $P(X)$ can be computed by means of the Smith Normal Form algorithm [@DF04 Section 12.2] and, from there, an explicit formula can be derived for ${\mathfrak{n}}_{{\textnormal{rk}}(G)}(G)$. Indeed, if $P(X)$ is of degree $k$ and has $l$ irreducible factors with degrees $d_1, \dots, d_l$, then $\vert R^{\times} \vert =
p^{k} \prod_{i = 1}^l (1 - p^{d_i - k})$ (use for instance Lemma \[LemUnitsOfAFiniteRing\]) while $\vert T \vert$ can be deduced from the computation of the order in ${\textnormal{GL}}_k({\mathbb F}_p)$ of the companion matrix of $P$.
For the next results, we make no assumption on the cardinality of $C$ but suppose that $M \simeq R$. Therefore $G \simeq R \rtimes C$ is generated by $a$ and the identity $b$ of the ring $R$. At this stage, few more examples may help understand the kind of two-generated groups we want to address. Assume that $C$ is the cyclic subgroup of ${\textnormal{GL}}_2({\mathbb Z})$ generated by an invertible matrix $a$. Let $b$ denote the $2$-by-$2$ identity matrix and let $G$ be the semi-direct product ${\mathbb Z}^2 \rtimes_a C$ where $a$ acts on ${\mathbb Z}^2$ via matrix multiplication. It is readily checked that ${\textnormal{rk}}(G) = 2$ if and only if $M \Doteq {\mathbb Z}^2$ is a cyclic ${ {\mathbb Z}\lbrack C \rbrack}$-module. If this holds, then $M$ naturally identifies with the subring $R = {\mathbb Z}a + {\mathbb Z}b$ of the ring of $2$-by-$2$ matrices over ${\mathbb Z}$ and we can certainly write $G \simeq R \rtimes_a C$. If the minimal polynomial of $a$ is moreover irreducible and if $\alpha \in {\mathbb C}$ is one of its roots, then $G$ identifies in turn with the semi-direct product $G(\alpha) = {\mathbb Z}{\lbrack \alpha^{\pm 1} \rbrack} \rtimes_{\alpha} \langle \alpha \rangle$ where $\alpha$ acts on ${\mathbb Z}{\lbrack \alpha^{\pm 1} \rbrack} \subset {\mathbb C}$ via complex multiplication. For arbitrary choices of $\alpha \in {\mathbb C}$, the family $G(\alpha)$ provides us with countably many interesting non-isomorphic examples. For instance, if $\alpha \in {\mathbb Z}\setminus \{0\}$, then $G(\alpha)$ is the Baumslag-Solitar group ${\left\langle a,b \ \big\vert\ aba^{-1} = b^{\alpha} \right\rangle}$, which is handled in Corollary \[CorBS\] below. If $\alpha$ is transcendental over ${\mathbb Q}$, then ${\mathbb Z}{\lbrack \alpha^{\pm 1} \rbrack}$ is isomorphic to the ring ${ {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$ of univariate Laurent polynomials over ${\mathbb Z}$. In this case, the group $G(\alpha)$ is isomorphic to the restricted wreath product ${\mathbb Z}\wr {\mathbb Z}$, the subject of Corollary \[CorZwrZ\] below.
Let us return to the presentation of our results. Under the assumption $M \simeq R$, we shall exhibit a complete invariant of Nielsen equivalence for generating pairs. In addition, if $n = 3$ and $C$ is finite, or if $n = 4$, we prove that ${\textnormal{Aut}}(F_n)$ acts transitively on ${\operatorname{V}}_n(G)$. Note that if $n = 3$ and $C$ is infinite, then Theorem \[ThMTimesC\] reduces the study of the Nielsen equivalent triples to the description of the orbit set ${\textnormal{Um}}_2(R)/{\textnormal{E}}_2(R)$. Our invariant is based on a map $D$ defined as follows. If $\nu(G) = 0$, there is a unique derivation $d \in {\textnormal{Der}}(C, R)$ satisfying $d(a)=1$ (see Section \[SecTsys\]). For ${\mathbf{g}}= (rc, r'c') \in G^2$ with $(r,r') \in R^2$ and $(c,c') \in C^2$, we set then $$\label{EqD1}
D({\mathbf{g}}) \Doteq r' d(c) - r d(c') \in R.$$ If $\nu(G) \neq 0$, we set furthermore $$\label{EqD2}
D({\mathbf{g}}) \Doteq D(\pi_{\nu(G) R}({\mathbf{g}})) \in R/\nu(G) R,$$ where $\pi_{\nu(G) R}$ stands for the natural map $R \rtimes C \twoheadrightarrow R/\nu(G) R \rtimes C$ and the right-hand side of (\[EqD2\]) is defined as in (\[EqD1\]). The following observations will enable us to construct the desired invariant.
\[PropDa\] Let $G \simeq R \rtimes_{\alpha} C$ and let ${\mathbf{g}}\in G^2$.
- If ${\mathbf{g}}\in {\operatorname{V}}_2(G)$, then $D({\mathbf{g}}) \in (R/\nu(G) R)^{\times}$.
- Assume $\nu(G)$ is nilpotent. Then ${\mathbf{g}}$ generates $G$ if and only if $\sigma({\mathbf{g}})$ generates $C$ and $D({\mathbf{g}}) \in (R/\nu(G) R)^{\times}$.
Setting $$\Lambda = R / \nu(G) R, \quad T_{\Lambda} = \pi_{\nu(G) R}(T),$$ we define the map $$\begin{array}{cccc}
\Delta:& {\operatorname{V}}_2(G) & \rightarrow & \Lambda^{\times}/ T_{\Lambda} \\
& {\mathbf{g}}& \mapsto & T_{\Lambda} D({\mathbf{g}})
\end{array}$$ The map $\Delta$ is the invariant we need and we are now in position to describe the Nielsen equivalence classes of generating $n$-vectors of $G \simeq R \rtimes C$ for $n = 2, 3$ and $4$.
\[ThN2\] Let $G \simeq R \rtimes_{\alpha} C$. Let ${\mathfrak{n}}_n(G)$ ($n =2, 3$ and $4$) be defined as in Corollary \[CorFpToDRtimesZ\]. Then the following hold:
- Two generating pairs ${\mathbf{g}}, {\mathbf{g}}'$ of $G$ are Nielsen equivalent if and only if $ \pi_{ab}({\mathbf{g}})$ and $\pi_{ab}({\mathbf{g}}')$ are Nielsen equivalent and $
\Delta({\mathbf{g}}) = \Delta({\mathbf{g}}').
$
- If $C$ is infinite or $G_{ab}$ is finite, then $\Delta$ induces a bijection from\
${\operatorname{V}}_2(G)/{\textnormal{Aut}}(F_2)$ onto $\Lambda^{\times}/T_{\Lambda}$. In particular ${\mathfrak{n}}_2(G) = \vert \Lambda^{\times}/T_{\Lambda} \vert$.
- If ${\textnormal{SL}}_2(R) = {\textnormal{E}}_2(R)$, e.g., $C$ is finite, then ${\mathfrak{n}}_3(G) = 1$.
- ${\mathfrak{n}}_4(G) = 1$.
Assertion $(i)$ of Theorem \[ThN2\] provides us with an algorithm which decides whether or not two generating pairs of $G$ are Nielsen equivalent. Indeed, the first condition in $(i)$ can be determined by means of the Diaconis-Graham determinant [@DG99] while the second condition can be reduced to the ideal membership problem in ${ {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$ which is solvable [@PU99; @Asc04].
Consider now the left group action of ${\textnormal{Aut}}(G)$ on ${\operatorname{V}}_n(G)$ where we define $\phi {\mathbf{g}}$ for $\phi \in {\textnormal{Aut}}(G)$ by $\pi_{\phi {\mathbf{g}}} \Doteq \phi \circ \pi_{{\mathbf{g}}}$, using the identification of ${\operatorname{V}}_n(G)$ with ${\operatorname{Epi}}(F_n , G)$. This action clearly commutes with the right ${\textnormal{Aut}}(F_n)$-action introduced earlier so that $
(\phi, \psi) {\mathbf{g}}\Doteq \phi {\mathbf{g}}\psi^{-1}
$ is an action of ${\textnormal{Aut}}(G) \times {\textnormal{Aut}}(F_n)$ on ${\operatorname{V}}_n(G)$. Following B.H. Neumann and H. Neumann [@NN51], we call the orbits of this action the *${\operatorname{T}}$-systems* of generating $n$-vectors of $G$, or concisely, the *${\operatorname{T}}_n$-systems* of $G$. We denote by ${\mathfrak{t}}_n(G)$ the cardinality of the ${\operatorname{T}}_n$-systems of $G$.
\[ThT2\] Let $G \simeq R \rtimes_{\alpha} C$. Let $A(C)$ be set of the automorphisms of $C$ which are induced by automorphisms of $G$ preserving $M$. Let $A'(C)$ be the subgroup of ${\textnormal{Aut}}(C)$ generated by $A(C)$ and the involution $c \mapsto c^{-1}$. Then the following hold.
- The cardinality ${\mathfrak{t}}_2(G)$ is finite and we have ${\mathfrak{t}}_2(G) \le \vert {\textnormal{Aut}}(C) / A'(C) \vert$, with equality if $R$ is a characteristic subgroup of $G$. If $C$ is infinite or $G_{ab}$ is finite, then ${\mathfrak{t}}_2(G) = 1$.
- If $C$ is infinite and $R$ is characteristic in $G$, then we have
- $\vert A(C) \vert \le 2$,
- ${\mathfrak{n}}_3(G) \le \vert A(C) \vert {\mathfrak{n}}_2(G) {\mathfrak{t}}_3(G),$
- $\vert A(C) \vert {\mathfrak{t}}_3(G) \ge \vert {\operatorname{SK}}_1(R)\vert,$
where ${\mathfrak{n}}_2(G)$ and ${\mathfrak{n}}_3(G)$ are as in Corollary \[CorFpToDRtimesZ\] and ${\operatorname{SK}}_1(R)$ denotes the special Whitehead group of $R$.
Note that assertion $(i)$ of Theorem \[ThT2\] generalizes Brunner’s theorem [@Bru74 Theorem 2.4] according to which $2$-generated Abelian-by-(infinite cyclic) groups have only one ${\operatorname{T}}_2$-system; another kind of generalization is given in Theorem \[ThOneTSystem\] below. The next theorem shows that there is no upper bound for ${\mathfrak{t}}_2(G)$ and ${\mathfrak{t}}_3(G)$ when $G$ ranges in the class of two-generated Abelian-by-cyclic groups.
\[ThT2GN\] For every integer $N \ge 1$, there exist
- a group $G_N$ of the form $R \rtimes_{\alpha} C$ with $C$ finite such that ${\mathfrak{t}}_2(G_N) \ge N$,
- a two-generated Abelian-by-(infinite cyclic) group $H_N$ such that\
${\mathfrak{n}}_3(H_N) = {\mathfrak{t}}_3(H_N) = N$,
where ${\mathfrak{t}}_2$, ${\mathfrak{n}}_3$ and ${\mathfrak{t}}_3$ are defined as in Corollary \[CorFpToDRtimesZ\] and Theorem \[ThT2\].
For comparison, Dunwoody constructed for every $N \ge 1$ a two-step nilpotent $2$-group $D_N$ on two generators such that ${\mathfrak{t}}_2(D_N) \ge N$ [@Dun63]. Note that assertion $(ii)$ of Theorem \[ThT2GN\] contrasts with the results of Evans [@Eva93 Lemma 4.3.ii] and Dunwoody [@Dun70] according to which ${\mathfrak{n}}_n(G) = 1$ holds if $G$ is a finitely generated nilpotent group or a finite solvable group and $n > {\textnormal{rk}}(G)$. We know by Theorem \[ThT2\] that ${\mathfrak{t}}_2(G)$ is finite whenever $G$ is of the form $R \rtimes_{\alpha} C$, but we were not able to settle the question as to whether ${\mathfrak{t}}_3(G)$ can be infinite for such a group $G$. We now turn to applications of Theorems \[ThN2\] and \[ThT2\] for three different classes of two-generated groups, namely the soluble Baumslag-Solitar groups, the split metacyclic groups and the lamplighter groups. A *Baumslag-Solitar group* is a group with a presentation of the form $$BS(k, l) ={\left\langle a,b \ \big\vert\ ab^ka^{-1} = b^l \right\rangle}$$ for $k, l \in {\mathbb Z}\setminus \{0\}$. Brunner proved that $BS(2,3)$ has infinitely many ${\operatorname{T}}_2$-systems whereas its largest metabelian quotient, namely $G(2/3) = {\mathbb Z}{\lbrack 1/6 \rbrack} \rtimes_{2/3} {\mathbb Z}$, has only one ${\operatorname{T}}_2$-system [@Bru74 Theorem 3.2] . The group $BS(k, l)$ is soluble if and only if $\vert k \vert = 1$ or $\vert l \vert = 1$. As a result, a soluble Baumslag-Solitar group is isomorphic to $BS(1, l)$ for some $l \in {\mathbb Z}\setminus \{0\}$ and hence admits a semi-direct decomposition ${\mathbb Z}{\lbrack 1/l \rbrack} \rtimes_l {\mathbb Z}$ where the canonical generator of ${\mathbb Z}$ acts as the multiplication by $l$ on ${\mathbb Z}{\lbrack 1/l \rbrack} = \{\frac{z}{l^i}; \,z \in {\mathbb Z}, i \in {\mathbb N}\}$.
\[CorBS\] Let $G = BS(1, l)$ with $l \in {\mathbb Z}\setminus \{0\}$ and let ${\mathfrak{n}}_n, {\mathfrak{t}}_n$ ($n = 2,3$) be defined as in Corollary \[CorFpToDRtimesZ\] and Theorem \[ThT2\]. Then the following hold:
- ${\mathfrak{n}}_2(G)$ is finite if and only if $l = \pm p^d$ for some prime number $p \in {\mathbb N}$ and some non-negative integer $d$. In this case, ${\mathfrak{n}}_2(G) = \max(d, 1)$.
- ${\mathfrak{t}}_2(G) = {\mathfrak{n}}_3(G) = 1$.
Define ${\mathbb Z}_d = {\mathbb Z}/d{\mathbb Z}$ for $d \ge 0$ (thus ${\mathbb Z}_0 = {\mathbb Z}$) and let $\varphi(d)$ be the cardinal of ${\mathbb Z}_d^{\times}$. A *split metacyclic group* is a semi-direct product of the form ${ {\mathbb Z}_k \rtimes_{\alpha} {\mathbb Z}_l}$ with $k, l \ge 0$. Here the canonical generator of ${\mathbb Z}_l$ is denoted by $a$ and acts on ${\mathbb Z}_k$ as the multiplication by $\alpha \in {\mathbb Z}_k^{\times}$.
\[CorMetacyclic\] Let $G = { {\mathbb Z}_k \rtimes_{\alpha} {\mathbb Z}_l}$ ($k, l \ge 0$) and let ${\mathfrak{n}}_n, {\mathfrak{t}}_n$ ($n = 2,3$) be defined as in Corollary \[CorFpToDRtimesZ\] and Theorem \[ThT2\].. Then the following hold.
- ${\mathfrak{n}}_2(G) = \frac{\varphi(\lambda)}{\omega}$, where $\lambda \ge 0$ is such that ${\mathbb Z}_{\lambda} \simeq {\mathbb Z}_k/\nu(G) {\mathbb Z}_k$ and $\omega$ is the order of the subgroup of ${\mathbb Z}_{\lambda}^{\times}$ generated by $-1$ and the image of $\alpha$.
- ${\mathfrak{t}}_2(G) = {\mathfrak{n}}_3(G) = 1$.
- If $G$ is finite, then $\left\vert {\operatorname{V}}_2(G) \right\vert = \frac{k\varphi(k)}{e\varphi(e)} \left\vert {\operatorname{V}}_2({\mathbb Z}_e \times {\mathbb Z}_l) \right\vert$, where $e \ge 1$ is such that ${\mathbb Z}_e \simeq {\mathbb Z}_k/(1 - \alpha) {\mathbb Z}_k$. In addition, every Nielsen equivalence class of generating pairs has the same number of elements.
In assertion $(iii)$ above, the cardinality $\left\vert {\operatorname{V}}_2({\mathbb Z}_e \times {\mathbb Z}_l) \right\vert$ of the set of generating pairs of the abelianization ${G_{ab}}\simeq {\mathbb Z}_e \times {\mathbb Z}_l$ of $G$ can be computed using the formulas of [@DG99 Remark 1]. Our Corollary \[CorMetacyclic\] applies for instance to dihedral groups and also to almost all $p$-groups with a cyclic subgroup of index $p$ [@Bro82 Theorem IV.4.1]. A two-generated *lamplighter group* is a restricted wreath product of the form ${\mathbb Z}_k \wr {\mathbb Z}_l$ with $k, l \ge 0$. Such a group reads also as $R \rtimes_a C$ with $C = {\mathbb Z}_l$ and $R = {\mathbb Z}_k {\lbrack C \rbrack}$, the integral group ring of $C$ over ${\mathbb Z}_k$. We are able to determine the number of Nielsen classes and ${\operatorname{T}}$-systems of any two-generated lamplighter groups with the exception of ${\mathbb Z}\wr {\mathbb Z}$.
\[CorWreath\] Let $G = {\mathbb Z}_k \wr {\mathbb Z}_l$ ($k, l \ge 0$ and $k, l \neq 1$) and let ${\mathfrak{n}}_n$ and ${\mathfrak{t}}_n$ ($n = 2,3$) be defined as in Corollary \[CorFpToDRtimesZ\] and Theorem \[ThT2\]. Then the following hold.
- ${\mathfrak{t}}_2(G) = 1$.
- If ${\mathbb Z}_k$ or ${\mathbb Z}_l$ is finite, then ${\mathfrak{n}}_3(G) = 1$.
- Assume that ${\mathbb Z}_k$ is finite and ${\mathbb Z}_l$ is infinite. Then ${\mathfrak{n}}_2(G)$ is finite if and only if $k$ is prime; in this case ${\mathfrak{n}}_2(G) = \max(\frac{k - 1}{2}, 1)$.
- Assume that ${\mathbb Z}_k$ is infinite and ${\mathbb Z}_l$ is finite. Then ${\mathfrak{n}}_2(G)$ is finite if and only if $l \in \{2, 3, 4, 6\}$; in this case ${\mathfrak{n}}_2(G) = 1$.
The case of finite two-generated lamplighter groups is addressed by Corollary \[CorFiniteWreathProduct\] below and a formula for $\left\vert {\operatorname{V}}_2({\mathbb Z}_k \wr {\mathbb Z}_l)\right\vert$ is derived. For the restricted wreath product ${\mathbb Z}\wr {\mathbb Z}$, we show that classifying Nielsen equivalence classes and ${\operatorname{T}}$-systems of generating triples tightly relates to an open problem in ring theory.
\[CorZwrZ\] Let $G = {\mathbb Z}\wr {\mathbb Z}$ and let ${\mathfrak{n}}_2, {\mathfrak{n}}_3$ and ${\mathfrak{t}}_3$ be defined as in Corollary \[CorFpToDRtimesZ\] and Theorem \[ThT2\]. Then we have ${\mathfrak{n}}_2(G) = 1$ and ${\mathfrak{n}}_3(G) \le 2 {\mathfrak{t}}_3(G)$. In addition, the following are equivalent:
- ${\mathfrak{n}}_3(G) = 1$.
- ${\mathfrak{t}}_3(G) = 1$.
- The ring $R$ of univariate Laurent polynomials over ${\mathbb Z}$ satisfies ${\textnormal{SL}}_2(R) = {\textnormal{E}}_2(R)$ (cf. [@Abr08 Conjecture 5.3], [@BMS02 Open Problem MA1], [@BM82 Open problem]).
The paper is organized as follows. Section \[SecPreliminary\] deals with notation and gathers known facts on rings which are ubiquitous in our presentation: the generalized Euclidean rings and the quotients of the ring of univariate Laurent polynomials over ${\mathbb Z}$. Section \[SecMC\] is dedicated to the proof of Theorem \[ThMTimesC\]. In Section \[SecARow\] we determine the conditions under which the map $\Phi_a$ of Theorem \[ThMTimesC\] is surjective while Section \[SecCIsZ\] addresses the case $C \simeq {\mathbb Z}$ for which it is shown that $\Phi_a$ is bijective. Section \[SecRC\] is dedicated to the proofs of Theorems \[ThN2\], \[ThT2\] and \[ThT2GN\]. Section \[SecLamplighterGroups\] presents applications to Baumslag-Solitar groups, split metacyclic groups and lamplighter groups.\
#### **Acknowledgments**.
The author is grateful to Pierre de la Harpe, Tatiana Smirnova-Nagnibeda and Laurent Bartholdi for encouragements, useful references and comments made on preliminary versions of this paper.
Preliminary results {#SecPreliminary}
===================
Notation and definitions {#SecNotation}
------------------------
We set in this section the notation and the definitions used throughout the article. A parallel is drawn between generating vectors of a group and unimodular rows of a module.
#### **Rings**
All considered rings are commutative rings with identity. Given a ring $R$, we denote by ${\mathcal{J}}(R)$ its *Jacobson radical*, i.e., the intersection of all its maximal ideals. We denote by ${\operatorname{nil}}(R)$ the *nilradical* of $R$, i.e., the intersection of all its prime ideals. The nilradical coincides with the set of nilpotent elements [@Mat89 Theorem 1.2]. Let $M$ be a finitely generated $R$-module. Then an $R$-epimorphism of $M$ is an $R$-automorphism [@Mat89 Theorem 2.4], a fact that we will use without further notice. Let $N \subset M$ be finitely generated $R$-modules and $I \subset {\mathcal{J}}(R)$ an ideal of $R$. Then the identity $N + IM = M$ implies $N = M$. We refer to the latter fact as Nakayama’s lemma [@Mat89 Theorem 2.2’s corollary]. Apart from Section \[SecGE\], the ring $R$ will always be a quotient of ${ {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$, the ring of univariate Laurent polynomials over ${\mathbb Z}$. In this case, we have ${\operatorname{nil}}(R) = {\mathcal{J}}(R)$ [@Eis95 Theorem 4.19] and we shall consistently denote by $\alpha$ the image of $X$ by the quotient map. We set ${\mathbb Z}_d \Doteq {\mathbb Z}/d{\mathbb Z}$ for $d \ge 0$. Thus the additive group of ${\mathbb Z}_d$ is the cyclic group with $d$ elements if $d > 0$ whereas ${\mathbb Z}_0 = {\mathbb Z}$. We denote by $\varphi$ the Euler totient function, so that $\varphi(d) = \vert {\mathbb Z}_d^{\times}\vert$ if $d > 0$. We set furthermore $\varphi(0) \Doteq 2$.
#### **Orbits of generating vectors**
Let $G$, $H$ be groups and let $f \in {\operatorname{Hom}}(G, H)$. We denote by $1_G$ the trivial element of $G$. For ${\mathbf{g}}=(g_i) \in G^n$, we set $f({\mathbf{g}}) \Doteq (f(g_i))$. Thus the component-wise left action of ${\textnormal{Aut}}(G)$ on ${\operatorname{V}}_n(G)$ reads as $\phi {\mathbf{g}}= \phi({\mathbf{g}})$ for $(\phi ,{\mathbf{g}}) \in {\textnormal{Aut}}(G) \times{\operatorname{V}}_n(G)$. This action clearly coincides with the ${\textnormal{Aut}}(G)$-action introduced earlier. Let us examine the component-wise counterpart of the ${\textnormal{Aut}}(F_n)$-action we previously defined via the identification of ${\operatorname{V}}_n(G)$ with ${\operatorname{Epi}}(F_n, G)$. For $\psi \in {\textnormal{Aut}}(F_n)$, set $w_i({\mathbf{x}}) \Doteq \psi(x_i) \in F_n$, with $i=1,\dots,n$. Then the ${\textnormal{Aut}}(F_n)$-action also reads as $ {\mathbf{g}}\psi = (w_i({\mathbf{g}}))$.
#### **Orbits of unimodular rows**
For $r \in R$ and $i \neq j$, we denote by $E_{ij}(r) \in {\textnormal{GL}}_n(R)$ the elementary matrix with ones on the diagonal and whose $(i,j)$-entry is $r$, all other entries being $0$. For $u \in R^{\times}$, we denote by $D_i(u) \in {\textnormal{GL}}_n(R)$ the diagonal matrix with ones on the diagonal except at the $(i,i)$-entry, which is set to $u$. Recall that ${\textnormal{E}}_n(R)$ is the subgroup of ${\textnormal{GL}}_n(R)$ generated by elementary matrices. Given a subgroup $U$ of $R^{\times}$, we define ${\textnormal{D}}_n(U)$ as the subgroup of ${\textnormal{GL}}_n(R)$ generated by the matrices $D_i(u)$ with $u \in U$. Following P.M. Cohn [@Coh66], we denote by ${\textnormal{GE}}_n(R)$ the subgroup generated by ${\textnormal{E}}_n(R)$ and ${\textnormal{D}}_n(R^{\times})$. Let $\Gamma$ be a subgroup of ${\textnormal{GL}}_n(R)$ and let $M$ be an $R$-module. Two rows ${\mathbf{r}},{\mathbf{r}}' \in M^n$ are said to be *$\Gamma$-equivalent* if there is $\gamma \in \Gamma$ such that ${\mathbf{r}}' = {\mathbf{r}}\gamma$. A a row is termed *unimodular* if its components generate $M$ as an $R$-module. By definition, ${\textnormal{Um}}_n(M)$ is the set of unimodular rows of $M$.
#### **Elementary rank and stable rank**
We say that $n \ge 1$ belongs to the *elementary range* of $R$ if ${\textnormal{E}}_{n + 1}(R)$ acts transitively on ${\textnormal{Um}}_{n + 1}(R)$. The *elementary rank* of $R$ is the least integer ${\textnormal{er}}(R)$ such that $n$ lies in the elementary range of $R$ for every $n \ge {\textnormal{er}}(R)$. We say that $k$ lies in the *stable range* of $R$ if for every $n \ge k$ and every $(r_i) \in {\textnormal{Um}}_{n + 1}(R)$ there is $(s_i) \in R^n$ such that $(r_1 + s_1 r_{n + 1},r_2 + s_2 r_{n + 1}, \dots, r_n + s_n r_{n + 1}) \in {\textnormal{Um}}_n(M)$. The *stable rank* of $R$ is the least integer ${\textnormal{sr}}(M)$ lying in the stable range of $M$. By [@McCR87 Proposition 11.3.11] we have: $$\label{EqStableRank}
1 \le {\textnormal{er}}(R) \le {\textnormal{sr}}(R).$$ If $R$ is moreover Noetherian, the Bass Cancellation Theorem asserts [@McCR87 Corollary 6.7.4] that $$\label{EqBassCancellationTh}
{\textnormal{sr}}(R) \le {\dim_{\textnormal{Krull}}}(R) + 1.$$
${\textnormal{GE}}$-rings {#SecGE}
-------------------------
Our study of Nielsen equivalence is significantly simplified when restricting to a class of rings $R$ which are similar to Euclidean rings in a specific sense. This is why we introduce the following definitions and elementary results. A ring $R$ is termed a *${\textnormal{GE}}_n$-ring* if ${\textnormal{GE}}_n(R) = {\textnormal{GL}}_n(R)$, which is equivalent to say that ${\textnormal{SL}}_n(R) = {\textnormal{E}}_n(R)$. Indeed, we have $
{\textnormal{GE}}_n(R) = {\textnormal{D}}_n(R^{\times}) {\textnormal{E}}_n(R)
$ and a matrix $D \in {\textnormal{D}}_n(R^{\times})$ lies in ${\textnormal{SL}}_n(R)$ if and only if it lies in ${\textnormal{E}}_n(R)$ by Whitehead’s lemma [@Wei13 Lemma 1.3.3]. Thus the latter equality implies the former, the converse being obvious. A ring $R$ is called a *generalized Euclidean ring* in the sense of P. M. Cohn [@Coh66], or a ${\textnormal{GE}}$-ring for brevity, if it is a ${\textnormal{GE}}_n$-ring for every $n \ge 1$. Euclidean rings are known to satisfy this property [@HO89 Theorem 4.3.9]. The reader is invited to check the two following elementary lemmas.
\[LemGECriteria\] The following assertions hold:
- $R$ is a ${\textnormal{GE}}_2$-ring if and only if $1$ lies in the elementary range of $R$, i.e, ${\textnormal{E}}_2(R)$ acts transitively on ${\textnormal{Um}}_2(R)$.
- If ${\textnormal{er}}(R) = 1$ then $R$ is a ${\textnormal{GE}}$-ring. In particular, $R$ is a ${\textnormal{GE}}$-ring if its stable rank is $1$.
A semilocal ring, i.e., a ring with only finitely many maximal ideals, has stable rank $1$ [@Bas64 Corollary 6.5]. As a result, semilocal rings, and Artinian rings in particular, are ${\textnormal{GE}}$-rings.
\[LemStabilityOfGE\] The following assertions hold:
- Let $J$ be an ideal contained in ${\mathcal{J}}(R)$. Then $R$ is a ${\textnormal{GE}}$-ring if and only if $R/J$ is a ${\textnormal{GE}}$-ring [@Gel77 Proposition 5].
- Assume $R$ is a direct sum of rings. Then $R$ is a ${\textnormal{GE}}$-ring if and only if each of its direct factors is a ${\textnormal{GE}}$-ring. [@Coh66 Theorem 3.1]
\[LemArtinianCoefficientAndGE\] Let $R$ be an Artinian ring. Then every homomorphic image of $R\lbrack X^{\pm1} \rbrack$ is a ${\textnormal{GE}}$-ring.
[Proof of Lemma \[LemArtinianCoefficientAndGE\]]{} Since ${\mathcal{J}}(R) = {\operatorname{nil}}(R)$, we have ${\mathcal{J}}(R) \lbrack X^{\pm1} \rbrack \subset {\mathcal{J}}(R\lbrack X^{\pm1} \rbrack)$. As the factor ring $P \Doteq R\lbrack X^{\pm1} \rbrack / {\mathcal{J}}(R) \lbrack X^{\pm1} \rbrack$ is isomorphic to a direct sum of Euclidean rings we deduce from Lemma \[LemArtinianCoefficientAndGE\] that $R\lbrack X^{\pm1} \rbrack$ is a ${\textnormal{GE}}$-ring. Let us consider a quotient $Q$ of $R\lbrack X^{\pm1} \rbrack$. Then $Q/{\mathcal{J}}(Q)$ is a quotient of $P$ and is therefore a direct sum whose factors are Euclidean or Artinian. As a result $Q$ is a ${\textnormal{GE}}$-ring.
\[RemStableRankReduction\] If ${\textnormal{sr}}(R) = r < \infty$, then it easy to prove that any matrix in ${\textnormal{GL}}_n(R)$ for $n > r$ can be reduced to a matrix of the form $
\begin{pmatrix}
A & 0 \\
0 & I_{n - r}
\end{pmatrix}
$ with $A \in {\textnormal{GL}}_r(R)$ by elementary row transformations. Thus $R$ is a ${\textnormal{GE}}$-ring if it is a ${\textnormal{GE}}_n$-ring for every $n \le r$.
The ring of univariate Laurent polynomials and its quotients {#SecZC}
------------------------------------------------------------
Because the structure of the $R$-module $M$ fully determines $G$, and because $R$ is a quotient of ${\mathbb Z}{\lbrack {\mathbb Z}\rbrack} \simeq { {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$, the ring of univariate Laurent polynomials over ${\mathbb Z}$, this ring plays a prominent role in the article. In this section, we collect preliminary facts about ${ {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$ and its quotients. These facts relate to row reduction of matrices over $R$ and unit group description; they impact directly the count of Nielsen equivalence classes in our applications.
A ring $R$ is said to be *completable* if every unimodular row over $R$ can be completed into an invertible square matrix over $R$, or equivalently, if ${\textnormal{GL}}_n(R)$ acts transitively on ${\textnormal{Um}}_n(R)$ for every $n \ge 1$.
\[LemCompletable\] Every homomorphic image of ${ {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$ is completable.
If $R$ is isomorphic to ${ {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$, then $R$ is completable by [@Sus77 Theorem 7.2]. So we can assume that ${\dim_{\textnormal{Krull}}}(R) \le 1$. Let $n \ge 2$ and ${\mathbf{r}}\in {\textnormal{Um}}_n(R)$. Since ${\textnormal{sr}}(R) \le 2$, we can find $E \in {\textnormal{E}}_n(R)$ such that ${\mathbf{r}}E = (r_1, r_2, 0, \dots, 0)$. Let $A \in {\textnormal{SL}}_2(R)$ be such that $(r_1, r_2)A = (1, 0)$ and set $B = \begin{pmatrix}
A & 0 \\
0 & I_{n -2}
\end{pmatrix}
$. Then we have ${\mathbf{r}}EB = (1,0,\dots,0)$.
The following is at the core of Theorem \[ThN2\].$iii$.
\[ThZCGE\] [@Guy16c Theorem A] Let $C$ be a finite cyclic group. Then every homomorphic image of ${\mathbb Z}{\lbrack C \rbrack}$ is a ${\textnormal{GE}}$-ring.
The last lemma of this section will come in handy in Section \[SecLamplighterGroups\] when scrutinizing lamplighter groups. Before we can state this lemma, we need to introduce some notation. Given a rational integer $d > 0$ we set $\zeta_d \Doteq e^{\frac{2i\pi}{d}}$ and let $$\lambda_d: {\mathbb Z}{\lbrack X \rbrack} \rightarrow { {\mathbb Z}\lbrack \zeta_d \rbrack}$$ be the ring homomorphism induced by the map $X \mapsto \zeta_d$. Given a set ${\mathcal{D}}$ of positive rational integers, we define $$\lambda_{{\mathcal{D}}} : {\mathbb Z}{\lbrack X \rbrack} \rightarrow {\prod_{d \in {\mathcal{D}}} {\mathbb Z}\lbrack \zeta_d \rbrack}$$ by $\lambda_{{\mathcal{D}}} \Doteq \prod_{d \in {\mathcal{D}}} \lambda_d$ and set ${\mathcal{O}(\mathcal{D})}\Doteq \lambda_{{\mathcal{D}}}({ {\mathbb Z}\lbrack X^{\pm 1} \rbrack})$. Let $\alpha \Doteq \lambda_{{\mathcal{D}}}(X) \in {\mathcal{O}(\mathcal{D})}$. Recall that a unit in ${\mathcal{O}(\mathcal{D})}^{\times}$ is said to be trivial if it lies in $T$, the subgroup generated by $-1$ and $\alpha$.
\[LemUnitsOfOD\] Let ${\mathcal{D}}$ be a non-empty set of positive rational integers.
1. The torsion-free rank of ${\mathcal{O}(\mathcal{D})}^{\times}$ is $\sum_{d \in {\mathcal{D}},\, d > 2} (\frac{\varphi(d)}{2} - 1)$.
2. Assume ${\mathcal{D}}$ is the set of divisors of $l$, with $l \ge 2$.
- Any non-trivial unit of finite order in ${\mathcal{O}(\mathcal{D} \setminus \{1\})}$ is of the form $u\left(1 + \sum_{i \in E} \alpha^i\right)$ for some trivial unit $u$ and some non-empty subset $E \subset \{ 1, 2, \dots, l - 1\}$.
- If $l \in \{2, 3, 4, 6\}$, then the units of ${\mathcal{O}(\mathcal{D} \setminus \{1\})}$ are trivial.
The proofs of $1$ and $2$.i essentially adapts [@AA69 Theorems 3 and 4] to the rings ${\mathcal{O}(\mathcal{D})}$ under consideration; we provide them for completeness.
$(1)$. Since the additive groups of $R = {\prod_{d \in {\mathcal{D}}} {\mathbb Z}\lbrack \zeta_d \rbrack}$ and ${\mathcal{O}(\mathcal{D})}$ are free Abelian groups of the same rank $r = \sum_{d \in {\mathcal{O}(\mathcal{D})}} \varphi(d)$, the latter group is of finite index $k$ in the former for some $k \ge 1$. By Dirichlet’s Unit Theorem, the group $R^{\times}$ is finitely generated and its torsion-free rank is $\sum_{d \in {\mathcal{D}},\, d > 2} (\frac{\varphi(d)}{2} - 1)$. Therefore, it is sufficient to prove that ${\mathcal{O}(\mathcal{D})}^{\times}$ is of finite index in $R^{\times}$. This certainly holds if every unit $u \in { {\mathbb Z}\lbrack \zeta_d \rbrack}$ for $d \in {\mathcal{D}}$ is of finite order modulo ${\mathcal{O}(\mathcal{D})}$. To see this, consider the principal ideal $I$ of ${ {\mathbb Z}\lbrack \zeta_d \rbrack}$ generated by $k$. Since ${ {\mathbb Z}\lbrack \zeta_d \rbrack}/ I$ is finite, there is $k' \ge 1$ such that $u^{k'} \equiv 1 \mod I$. Therefore $u^{k'} = 1 + k \zeta$ for some $\zeta \in { {\mathbb Z}\lbrack \zeta_d \rbrack}$. As $k \zeta \in {\mathcal{O}(\mathcal{D})}$, we deduce that $u^{k'} \in {\mathcal{O}(\mathcal{D})}$, which completes the proof.
$(2).i$. Let $g \in {\mathcal{O}(\mathcal{D})}$ such that the projection ${\operatorname{pr}}_1: {\mathcal{O}(\mathcal{D})}\twoheadrightarrow {\mathcal{O}(\mathcal{D} \setminus \{1\})}$ maps $g$ to a unit of finite order. Identifying ${\mathcal{O}(\mathcal{D})}$ with ${ {\mathbb Z}\lbrack C \rbrack}$ for $C = {\mathbb Z}/l{\mathbb Z}$, we can write $g = \sum_{c \in C} a_c c$ with $a_c \in {\mathbb Z}$. For $d$ dividing $n$, let $\pi_d$ be the natural projection of ${\mathcal{O}(\mathcal{D})}$ onto ${ {\mathbb Z}\lbrack \zeta_d \rbrack}$, let $\chi_d = {\pi_d}_{\vert C}$ and $\rho_d = \pi_d(g) = \sum_{c \in C} a_c \chi_d(c)$. Since ${\operatorname{pr}}_1(g)$ is of finite order, $\rho_d$ is a root of unity for every divisor $d > 1$. The characters $\chi_d$ form a complete set of inequivalent characters of $C$ by [@AA69 Lemma 2]. Therefore, we have $$\sum_{c \in C} a_c \chi(c) = \rho_{\chi}$$ for every $\chi \in \hat{C}$, the character group of $C$, where $\rho_{\chi}$ is a root of unity if $\chi \neq 1$. Using the orthogonality relation of characters, we obtain $$n a_c = \sum_{\chi \in \hat{G}} \rho_{\chi} {\overline{\chi(c)}}$$ for every $c \in C$. Hence $\vert a_c - a_{c'} \vert \le \frac{2(l - 1)}{n} < 2$ for every $c, c' \in C$. Replacing $g$ by $\varepsilon (g - k \sum_{c \in C}c)$ for a suitable choice of $\varepsilon \in \{\pm 1\}$ and $k \in {\mathbb Z}$, we can assume that $a_c \in \{0, 1\}$ for every $c$. Replacing $g$ by $cg$ for some suitable choice of $c \in C$, we can assume that $a_{1_C} = 1$. This ensures eventually that the image of $g$ in ${\mathcal{O}(\mathcal{D} \setminus \{1\})}$ has the desired form.
$(2).ii$. If $l \in \{2, 3\}$, then ${\mathcal{O}(\mathcal{D} \setminus \{1\})}= {\mathbb Z}{\lbrack \zeta_l \rbrack}$ and this ring has only trivial units. Assume now $l = 4$. Since $R = \mathcal{O}(\{2, 4\})$ embeds into ${\mathbb Z}\times {\mathbb Z}{\lbrack i \rbrack}$, it has at most $8$ units. It is easily checked that there are exactly $8$ trivial units in $R$. Therefore all units are trivial. Assume eventually that $l = 6$. Since $R = \mathcal{O}(\{2, 3, 6\})$ embeds into ${\mathbb Z}\times {\mathbb Z}{\lbrack \zeta_3 \rbrack} \times {\mathbb Z}{\lbrack \zeta_3 \rbrack}$, it has only units of finite orders. Considering projections on each of the three factors, it is routine to check that no element of the form $1 + \sum_{i \in E} \alpha^i$ with $\emptyset \neq E \subset \{1, 2, 3, 4 ,5\}$ is a unit in $R$. This proves that $R$ has only trivial units by 2.$i$.
Nielsen equivalence in finitely generated Abelian groups {#SecNielsenAbel}
--------------------------------------------------------
We present in this section the classification of generating tuples modulo Nielsen equivalence in finitely generated Abelian groups. This result is instrumental in Section \[SecARow\] when reducing generating vectors to a standard form. Different parts of the aforementioned classification were obtained by different authors: [@NN51 Satz 7.5], [@Dun63 Section 2’s lemmas], [@Eva93 Lemma 4.2] [@LM93 Example 1.6] [@DG99]. The classification reaches its complete form with
[@Oan11 Theorem 1.1] \[ThNielsenAbel\] Let $G$ be a finitely generated Abelian group whose invariant factor decomposition is $${\mathbb Z}_{d_1} \times \cdots \times {\mathbb Z}_{d_k}, \quad 1 \neq d_1 \, \vert \, d_2 \, \vert \, \cdots \, \vert \, d_k, \, d_i \ge 0$$
Then every generating $n$-vector ${\mathbf{g}}$ with $n \ge k = {\textnormal{rk}}(G)$ is Nielsen equivalent to $(\delta e_1, e_2, \dots, e_k, 0, \dots, 0)$ for some $\delta \in {\mathbb Z}_{d_1}^{\times}$ and where ${\mathbf{e}}= (e_i) \in G^k$ is defined by $(e_i)_i = 1 \in {\mathbb Z}_{d_i}$ and $(e_i)_j = 0$ for $j \neq i$.
- If $n > k$, then we can take $\delta = 1$.
- If $n = k$, then $\delta$ is unique, up to multiplication by $-1$.
In particular, $G$ has only one Nielsen equivalence class of generating $n$-vectors for $n > k$ and only one ${\operatorname{T}}_k$-system while it has $\max(\varphi(d_1)/2, 1)$ Nielsen equivalence classes of generating $k$-vectors where $\varphi$ denotes the Euler totient function extended by $\varphi(0) = 2$.
In the remainder of this section, we consider decompositions with cyclic factors which might differ from the invariant factor decomposition of $G$.
\[CorDet\] [@Guy17 Corollary C] Let $G = {\mathbb Z}_{d_1} \times \cdots \times {\mathbb Z}_{d_k}$ with $d_i \ge 0$ for $i = 1, \dots, k$. Let $d$ be the greatest common divisors of the integers $d_i$. For ${\mathbf{g}}\in {\operatorname{V}}_k(G)$, denote by $\det({\mathbf{g}})$ the determinant of the matrix whose coefficients are the images in ${\mathbb Z}_d$ of the $(g_i)_j$’s via the natural maps ${\mathbb Z}_{d_j} \twoheadrightarrow {\mathbb Z}_d$. Then ${\mathbf{g}}, {\mathbf{g}}' \in {\operatorname{V}}_k(G)$ are Nielsen equivalent if and only if $\det({\mathbf{g}}) = \pm \det({\mathbf{g}}')$.
Let ${\mathbb Z}_{d_1} \times \cdots \times {\mathbb Z}_{d_k}$ be a decomposition of $G$ in cyclic factors such that $k = {\textnormal{rk}}(G)$. The identity elements of each factor ring in such decomposition form a generating vector of $G$. We refer to this vector as the *generating vector naturally associated* to the given decomposition. If ${\mathbf{e}}$ is such a vector, we define $\det_{{\mathbf{e}}}$ as the determinant function of Corollary \[CorDet\].
In the case $k = {\textnormal{rk}}(G)$, Corollary \[CorDet\] shows in particular that ${\mathbf{g}}\in {\operatorname{V}}_k(G)$ is Nielsen equivalent to $(\det_{{\mathbf{e}}}({\mathbf{g}})e_1, e_2, \dots, e_k)$ where ${\mathbf{e}}= (e_i)$ is the generating vector naturally associated to the given decomposition of $G$ in cyclic factors.
Nielsen equivalence classes and ${\operatorname{T}}$-systems of $M \rtimes_{\alpha} C$ {#SecMC}
======================================================================================
Throughout this section, we assume that $G$ is a group which fits in the exact sequence (\[EqExt\]) and $n \ge \max({\textnormal{rk}}(G), 2)$. Recall that $\alpha$ denotes the image of a favored generator $a$ of $C$ via the natural map ${ {\mathbb Z}\lbrack C \rbrack}\twoheadrightarrow R$. Computing with powers of group elements in $G$ shall be facilitated by the following notation. For $u \in R^{\times}$ and $l \in {\mathbb Z}\cup \{ \infty \}$, let $$\partial_u(l)=\left\{
\begin{array}{cc}
1 + u + \cdots + u^{l - 1} & \text{ if } l > 0, \\
0 & \text{ if } l = 0, \,\infty, \\
- u^{-1} \partial_{u^{-1}}(-l) & \text{ if } l < 0.
\end{array}
\right.$$ For every $l \in {\mathbb Z}$ we have then $(1- u) \partial_u(l) = 1 - u^l$. In particular $(1 - \alpha)\nu(G) = 0$ for $\nu(G) = \partial_{\alpha}({\vert C \vert})$. If $C$ is infinite then $\partial_{\alpha}$ is the composition of the Fox derivative over $C$ [@Fox53] with the natural embedding of $C$ into $R = { {\mathbb Z}\lbrack C \rbrack}$. For $k,\,l \in {\mathbb Z}$ and $m \in M$, we have the identity $
(ma^k)^l = \left(\partial_{\alpha^k}(l)m\right) a^{kl}.
$
The description of Bachmuth’s ${\textnormal{IA}}$ automorphisms will considerably ease off the study of Nielsen equivalence in $G \simeq M \rtimes_{\alpha} C$. Recall that $F_n$ denotes the free group on ${\mathbf{x}}= (x_1,\dots,x_n)$. For $\psi \in {\textnormal{Aut}}(F_n)$, let ${\overline{\psi}}$ be the automorphism of ${\mathbb Z}^n$ induced by $\psi$. We denote by ${\mathbf{e}}= (e_i)$ the image of ${\mathbf{x}}$ by the abelianization homomorphism $
F_n \twoheadrightarrow (F_n)_{ab}={\mathbb Z}^n.
$ The map $\psi \mapsto {\overline{\psi}}$ is an epimorphism from ${\textnormal{Aut}}(F_n)$ onto ${\textnormal{GL}}_n( {\mathbb Z})$ [@LS77 Proposition I.4.4] whose kernel is denoted by ${\textnormal{IA}}(F_n)$. This group clearly contains the isomorphisms $\varepsilon_{ij}$ defined by $\varepsilon_{ij}(x_i) = x_j^{-1} x_i x_j$ and $\varepsilon_{ij}(x_k) = x_k$ if $k \neq i$. In turn, ${\textnormal{IA}}(F_n)$ is generated by the automorphisms $\varepsilon_{ijk}$ defined by $\varepsilon_{ijk}(x_i) = x_i \lbrack x_j, x_k \rbrack$ and $\varepsilon_{ijk}(x_l) = x_l$ for $l \neq i$ [@LS77 Chapter I.4] where ${\lbrack x, y \rbrack} \Doteq xyx^{-1}y^{-1}$.
Reduction to an $a$-row {#SecARow}
-----------------------
We discuss here circumstances under which a generating $n$-vector ${\mathbf{g}}$ of $G$ can be Nielsen reduced to an *$a$-row*, i.e., a generating $n$-vector of the form $({\mathbf{m}}, a)$ with ${\mathbf{m}}\in {\textnormal{Um}}_{n - 1}(R)$ and $a$ a favored generator of $C$ fixed beforehand. We first observe that any generating $n$-vector ${\mathbf{g}}$ is Nielsen equivalent to $({\mathbf{m}}, ma)$ for some ${\mathbf{m}}\in M^{n - 1}$ and some $m \in M$. Indeed, we can find $\psi \in {\textnormal{Aut}}(F_n)$ such that $\sigma({\mathbf{g}}) \psi = (1_{C^{n - 1}}, a)$ using Theorem \[ThNielsenAbel\]. This proves the claim. We shall establish conditions under which the element $m$ can be cancelled by a subsequent Nielsen transformation.
\[LemRankOfM\] Let ${\mathbf{g}}=({\mathbf{m}}, ma) \in G^n$ with ${\mathbf{m}}\in M^{n - 1}$ and $m \in M$. Then ${\mathbf{g}}$ generates $G$ if and only if $({\mathbf{m}}, \nu(G) m)$ generates $M$ as an $R$-module.
Assume first that ${\mathbf{g}}\in {\operatorname{V}}_n(G)$. Given $m' \in M$ there exists $w \in F_n$ such that $m' = w({\mathbf{m}}, ma)$. We can write $w = vx_n^s$ with $v$ lying in the normal closure of $\{x_1, \dots, x_{n - 1}\}$ in $F_n$ and $s \in {\mathbb Z}$. Since conjugation by $ma$ induces multiplication by $\alpha$ on $M$, $v({\mathbf{g}})$ lies in the $R$-submodule of $M$ generated by ${\mathbf{m}}$. As $(m a)^s = \partial_{\alpha}(s)m a^s$, we deduce that $s = 0$ if $C$ is infinite or $s \equiv 0 \mod {\vert C \vert}$ if $C$ is finite. Therefore $(m a)^s$ belongs to ${\mathbb Z}\nu(G) m = R \nu(G) m$ in both cases, which completes the proof of the ’only if’ part.
Assume now that $({\mathbf{m}}, \nu(G) m)$ generates $M$ as an $R$-module. Let $H$ be the subgroup of $G$ generated by ${\mathbf{g}}$ and write ${\mathbf{m}}= (m_i)$. The subgroup $H$ contains the conjugates of the elements $m_i$ by powers of $a$, hence it contains the submodule of $M$ generated by ${\mathbf{m}}$. It also contains the powers of $m a$, hence the submodule of $M$ generated by $\nu(G) m$. Thus it contains both $M$ and $a$, so that it is equal to $G$.
Lemma \[LemRankOfM\] implies the following inequalities: $${\textnormal{rk}}(G) - 1 \le {\textnormal{rk}}_R(M) \le {\textnormal{rk}}(G)$$ When every generating $n$-vector of $G$ can be Nielsen reduced to an $a$-row, we say that $G$ enjoys property ${\mathcal{N}}_n(a)$. If ${\mathcal{N}}_n(a)$ holds for $n = {\textnormal{rk}}(G)$, then the equality ${\textnormal{rk}}_R(M) = {\textnormal{rk}}(G) - 1$ must be satisfied. The converse does not hold, as the latter is equivalent to the weaker property ${\mathcal{N}}_n(C)$ according to which every generating $n$-vector can be Nielsen reduced to a $c$-row with $c$ ranging among generators of $C$, see Theorem \[ThReductionToARow\] below.
\[LemGeneratorsOfM\] Let ${\mathbf{g}}=({\mathbf{m}}, ma) \in {\operatorname{V}}_n(G)$ with ${\mathbf{m}}\in M^{n - 1}$ and $m \in M$. Then the following hold:
- If $m \in (1 - \alpha)M$ then ${\mathbf{m}}$ generates $M$ as an $R$-module.
- If $\nu(G)$ is nilpotent then ${\mathbf{m}}$ generates $M$ as an $R$-module.
- If ${\mathbf{m}}$ generates $M$ as an $R$-module then ${\mathbf{g}}$ is Nielsen equivalent to $({\mathbf{m}}, a)$.
We know that $({\mathbf{m}}, \nu(G) m)$ generates $M$ as an $R$-module by Lemma \[LemRankOfM\]. If $m \in (1 - \alpha)M$ then $\nu(G) m = 0$. Hence ${\mathbf{m}}$ generates $M$ as an $R$-module, which proves $(i)$. If $\nu(G) \in {\mathcal{J}}(R)$, the same conclusion follows from Nakayama’s lemma, which proves $(ii)$. Let us prove $(iii)$. If ${\mathbf{m}}= (m_i)$ generates $M$ as an $R$-module then $m$ is a sum of elements of the form $k \alpha^l m'$ with $k,l \in {\mathbb Z}$ and $m' \in \{m_1, \dots, m_{n - 1}\}$. We can subtract each of these terms from $m$ in the last entry of ${\mathbf{g}}$ by applying transformations of the form $\varepsilon_{i, n}^l$ and $L_{n, i}^{-k}$.
Combining assertions $(ii)$ and $(iii)$ of Lemma \[LemGeneratorsOfM\] yields:
\[LemNu\] If $\nu(G)$ is nilpotent then ${\mathcal{N}}_n(a)$ holds for every $n \ge {\textnormal{rk}}(G)$.
Let $\pi_{ab} :G \twoheadrightarrow G_{ab}$ be the abelianization homomorphism of $G$. Let $\pi_C$ be the natural map $M \twoheadrightarrow M_{C} \Doteq M/(1-\alpha)M$. Then we have $G_{ab} = M_C \times C$ and $\pi_{ab} = \pi_C \times \sigma$.
\[PropReductionToRow\] Let ${\mathbf{g}}\in {\operatorname{V}}_n(G)$ and assume that at least one of the following holds:
- $n > {\textnormal{rk}}(G_{ab})$,
- $n > {\textnormal{rk}}(M_C)$ and $M_C$ is not free over ${\mathbb Z}$.
Then ${\mathbf{g}}$ is Nielsen equivalent to a vector $({\mathbf{m}}, a)$ with ${\mathbf{m}}\in {\textnormal{Um}}_{n - 1}(M)$.
Let $k = {\textnormal{rk}}(M_C)$. Observe first that both assumptions imply $n > k$. Let ${\mathbb Z}_{d_1} \times \cdots \times {\mathbb Z}_{d_k}$ be the invariant factor decomposition of $M_C$. Let then ${\mathbb Z}_{d_1} \times \cdots \times {\mathbb Z}_{d_{n -1}} \times C$ be the decomposition of $G_{ab}$ where $d_i = 1$ if $i > k$. Define ${\mathbf{e}}= (e_i) \in G_{ab}^{n - 1}$ by $e_i = 1 \in {\mathbb Z}_{d_i}$ if $i \le k$, $e_i = 0$ otherwise. Set $\overline{{\mathbf{g}}} = \pi_{ab}({\mathbf{g}})$. Suppose now that $(ii)$ holds so that ${\mathbb Z}_{d_1}$ must be finite. Let $\tilde{\delta} \in {\mathbb Z}_{d_1}^{\times}$ be a lift of $\delta \Doteq \det_{{\mathbf{e}}}(\overline{{\mathbf{g}}})$ and let $\tilde{{\mathbf{g}}} = (\tilde{\delta}e_1, e_2, \dots, e_{n - 1}, a)$. Since $\det_{{\mathbf{e}}}(\tilde{{\mathbf{g}}}) = \delta = \det_{{\mathbf{e}}}(\overline{\mathbf{g}})$, the vectors $\overline{{\mathbf{g}}}$ and $\tilde{{\mathbf{g}}}$ are Nielsen equivalent by Corollary \[CorDet\]. By Theorem \[ThNielsenAbel\] this is also true if we assume $(i)$ and set $\tilde{\delta} = 1 \in {\mathbb Z}_{d_1}$. Hence under assumption $(i)$ or $(ii)$ there is $\psi \in {\textnormal{Aut}}(F_n)$ such that $\tilde{{\mathbf{g}}} = \overline{{\mathbf{g}}} \psi$. Then ${\mathbf{g}}' \Doteq {\mathbf{g}}\psi$ is of the form $({\mathbf{m}}, ma)$ with ${\mathbf{m}}= (m_1,\dots, m_{n - 1}) \in M^{n - 1}$ and $m \in M$ such that $\pi_C(m) = 0$. Applying Lemma \[LemGeneratorsOfM\] gives the result for any of the two hypotheses.
\[PropReductionToRowFree\] Assume $C$ is finite and let $n = {\textnormal{rk}}(G)$. Suppose moreover that $M_C$ is isomorphic to ${\mathbb Z}^{n - 1}$. Let ${\mathbf{g}}\in {\operatorname{V}}_n(G)$. Then ${\mathbf{g}}$ is Nielsen equivalent to a vector $({\mathbf{m}}, a^k)$ with ${\mathbf{m}}\in {\textnormal{Um}}_{n - 1}( M)$ and $k = \pm \det_{({\mathbf{e}}, a)}({\overline{{\mathbf{g}}}})$ where ${\overline{{\mathbf{g}}}} = \pi_{ab}({\mathbf{g}})$ and ${\mathbf{e}}$ is any basis of $M_C$. In particular, ${\textnormal{rk}}_R(M) = {\textnormal{rk}}(G) - 1$.
Let $\mathbf{e}$ be a basis of $M_C$ over ${\mathbb Z}$. By Theorem \[ThNielsenAbel\], we can assume that ${\overline{{\mathbf{g}}}} = (\mathbf{e}, a^k)$ for some $k$ such that $k = \pm \det_{{\mathbf{e}}}(\overline{{\mathbf{g}}})$. Hence ${\mathbf{g}}= ({\mathbf{m}}, ma^k)$ for some ${\mathbf{m}}\in M^{n -1}$ and $m \in (1 - \alpha)M$. Then ${\mathbf{g}}$ is Nielsen equivalent to $({\mathbf{m}}, a^k)$ by Lemma \[LemGeneratorsOfM\].
Eventually, we present a characterization of property ${\mathcal{N}}_n(a)$ which establishes the first part of Theorem \[ThN2\].
\[ThReductionToARow\] Property ${\mathcal{N}}_n(a)$ holds if and only if at least one of the following does:
- $n > {\textnormal{rk}}(G_{ab})$.
- $C$ is infinite.
- ${\textnormal{rk}}(G) > {\textnormal{rk}}(M_C)$ and $M_C$ is not isomorphic to ${\mathbb Z}^{{\textnormal{rk}}(G) - 1}$.
- ${\vert C \vert}\in \{2, 3, 4, 6\}$ and $M_C$ is isomorphic to ${\mathbb Z}^{{\textnormal{rk}}(G) - 1}$.
Let us show first that any of the assertions $(i)$ to $(iv)$ implies that property ${\mathcal{N}}_n(a)$ holds. For assertion $(i)$, it is Proposition \[PropReductionToRow\]. For assertion $(ii)$, it follows from Lemma \[LemNu\]. For the remaining assertions, we can assume that assertion $(i)$ doesn’t hold, so that $n = {\textnormal{rk}}(G)$. Then assertion $(iii)$ implies ${\mathcal{N}}_n(a)$, by Proposition \[PropReductionToRow\]. So does assertion $(iv)$ by Proposition \[PropReductionToRowFree\].
Let us assume now that none of the assertions $(i)$ to $(iv)$ hold. We shall prove that property ${\mathcal{N}}_n(a)$ doesn’t hold. Since assertion $(i)$ is assumed not to hold, we infer that $n = {\textnormal{rk}}(G)$. We can assume moreover that ${\textnormal{rk}}_R(M) = n - 1$, since otherwise ${\mathcal{N}}_n(a)$ would fail to be true. Because none of the assertions $(ii)$, $(iii)$ and $(iv)$ hold, the group $M_C$ is isomorphic to ${\mathbb Z}^{{\textnormal{rk}}(G) -1}$ and $C$ is finite and such that $\varphi({\vert C \vert}) > 2$. By Proposition \[PropReductionToRowFree\], we can find ${\mathbf{m}}\in {\textnormal{Um}}_{n - 1}(M)$ and $k > 1$ coprime with ${\vert C \vert}$ such that $({\mathbf{m}}, a^k)$ generates $G$ and cannot be Nielsen reduced to an $a$-row.
\[CorNC\] Property ${\mathcal{N}}_n(C)$ holds if and only if $n > {\textnormal{rk}}_R(M)$.
If ${\mathcal{N}}_n(C)$ holds, the inequality ${\textnormal{rk}}_R(M) \le n - 1$ is satisfied by definition. Let us assume now that ${\mathcal{N}}_n(C)$ doesn’t hold. Reasonning by contradiction, we assume furthermore that $n > {\textnormal{rk}}_R(M)$. Since property ${\mathcal{N}}_n(a)$ cannot hold it follows from Theorem \[ThReductionToARow\] that $n = {\textnormal{rk}}(G)$, $C$ is finite and $M_C$ is isomorphic to ${\mathbb Z}^{{\textnormal{rk}}(G) -1}$. The latter three conditions implie ${\mathcal{N}}_n(C)$ by Proposition \[PropReductionToRowFree\], a contradiction.
\[CorNc\] Property ${\mathcal{N}}_n(c)$ holds for some generator $c$ of $C$ if and only if it holds for all generators $c$ of $C$.
$\square$
Nielsen equivalence related to $\Gamma_{n - 1}(R)$-equivalence {#SecCIsZ}
--------------------------------------------------------------
In this section we scrutinize the relation between Nielsen equivalence of generating $n$-vectors and $\Gamma_{n - 1}(R)$-equivalence of unimodular rows. We prove there another part of Theorem \[ThMTimesC\], namely Proposition \[PropConverseZ\] below.
Recall that we denote by $T$ the subgroup of $R^{\times}$ generated by $-1$ and $\alpha$ and that $\Gamma_n(R)$ is the subgroup generated by ${\textnormal{E}}_n(R)$ and ${\textnormal{D}}_n(T)$. Since ${\textnormal{D}}_n(R^{\times})$ normalizes ${\textnormal{E}}_n(R)$, we have $
\Gamma_n(R) = {\textnormal{D}}_n(T) {\textnormal{E}}_n(R)
$.
\[LemGeneratorsOfGE\] For every $n \ge 2$, the group $\Gamma_n(R)$ is generated by ${\textnormal{D}}_n(T)$ together with the elementary matrices $E_{ij}(1)$ with $1 \le i \neq j \le n$.
For $1 \le i \neq j \le n$ and $(r,r') \in R^2, \, \beta \in \{\alpha^{\pm1}\}$, we have the following identities: $D_i(\beta)E_{ij}(r)D_i(\beta)^{-1} = E_{ij}(\beta r)$ and $E_{ij}(r)E_{ij}(r') = E_{ij}(r + r')$. Since $R$ is generated as a ring by $\alpha$ and $\alpha^{-1}$, the result follows.
\[LemGammaImpliesNielsen\] If ${\mathbf{m}}, {\mathbf{m}}' \in {\textnormal{Um}}_{n - 1}( M)$ are $\Gamma_{n - 1}(R)$-equivalent, then $({\mathbf{m}},\, a)$, $({\mathbf{m}}', \,a) \in {\operatorname{V}}_n(G)$ are Nielsen equivalent.
Since $({\mathbf{m}}E_{ij}(1), a) = ({\mathbf{m}}, a) L_{ij}$ for $1 \le i \neq j \le n -1$ and $({\mathbf{m}}D_{i}(\alpha), a) = ({\mathbf{m}}, a) \varepsilon_{i, n}$ for $1 \le i \neq j \le n - 1$, we deduce from Lemma \[LemGeneratorsOfGE\] that $({\mathbf{m}}, a)$ and $({\mathbf{m}}', a)$ are Nielsen equivalent.
We establish now a partial converse to Lemma \[LemGammaImpliesNielsen\].
\[PropConverseZ\] Assume $C$ is infinite. If $({\mathbf{m}}, a), ({\mathbf{m}}', a) \in {\operatorname{V}}_n(G)$ are Nielsen equivalent then ${\mathbf{m}}, {\mathbf{m}}' \in {\textnormal{Um}}_{n - 1}(M)$ are $\Gamma_{n - 1}(R)$-equivalent.
Suppose that $({\mathbf{m}}', a) = ({\mathbf{m}}, a) \psi$ for some $\psi \in {\textnormal{Aut}}(F_n)$. We claim that $\psi$ is of the form $\psi_0 \psi_1 L$ where
- $\psi_0 \in {\textnormal{IA}}(F_n)$,
- $\psi_1 \in {\textnormal{Aut}}(F_{n - 1})$, i.e., $\psi_1(x_n) = x_n$ and $\psi_1$ leaves $F_{n - 1}=F(x_1, \dots, x_{n - 1})$ invariant,
- $L$ belongs to the group generated by the automorphisms $L_{n, j}$.
To see this, consider the automorphism ${\overline{\psi}}\in {\textnormal{GL}}_n({\mathbb Z})$ induced by $\psi$. Since ${\overline{\psi}}(e_n) = e_n$, we can find a product of lower elementary matrices $$\overline{L} \Doteq E_{n, 1}(\mu_1) \cdots E_{n , n -1}(\mu_{n - 1})$$ with $\mu_i \in {\mathbb Z}$ such that ${\overline{\psi}}\cdot \overline{L} \in {\textnormal{GL}}_{n - 1}({\mathbb Z})$. Let $\psi_1 \in {\textnormal{Aut}}(F_{n - 1})$ be an automorphism inducing ${\overline{\psi}}\cdot \overline{L}$ on ${\mathbb Z}^{n - 1}$. Let $L$ be the product of automorphisms $L_{n, j}^{\mu_j}$ with $\mu_j \in {\mathbb Z}$. Then $L$ induces $\overline{L}$ and by construction we have $L \psi_1^{-1} \in {\textnormal{IA}}(F_n)$, which proves the claim.
The action of every ${\textnormal{IA}}$-automorphism $\varepsilon_{ijk}$ on $({\mathbf{m}}, a)$ leaves $\sigma({\mathbf{m}}, a)$ invariant and induces a transformation on ${\mathbf{m}}$ which lies in $\Gamma_{n - 1}(R)$. The same holds for every automorphism in ${\textnormal{Aut}}(F_{n - 1})$ and every automorphism $L_{ij}$ with $i>j$. Let ${\mathbf{g}}\Doteq ({\mathbf{m}}, a) \psi_0 \psi_1$. Then we can write ${\mathbf{g}}= (\mathbf{n}, ma)$ with $m \in M$ and where $\mathbf{n}$ is $\Gamma_{n - 1}(R)$-equivalent to ${\mathbf{m}}$. As $\sigma({\mathbf{g}}L) = \sigma({\mathbf{m}}', a) = (1_{C^{n -1}}, a)$ we deduce that $\mu_j = 0$ for every $j$, i.e., $L = 1$. Hence ${\mathbf{m}}' = \mathbf{n}$ which yields the result.
Nielsen equivalence classes {#SecNielsen}
---------------------------
In this section, we complete the proof of Theorem \[ThMTimesC\] by establishing Theorem \[ThNielsenCInfinite\] below. We subsequently discuss assumptions under which the latter theorem enables us to enumerate efficiently Nielsen equivalence classes. Recall that the map $\varphi_a: {\textnormal{Um}}_{n - 1}(M) \rightarrow {\operatorname{V}}_n(G)$ is defined by $\varphi_a({\mathbf{m}}) = ({\mathbf{m}}, a)$.
\[ThNielsenCInfinite\] The map $\varphi_a$ induces a map $$\Phi_a: {\textnormal{Um}}_{n - 1}(M)/\Gamma_{n - 1}(R) \rightarrow {\operatorname{V}}_n(G)/{\textnormal{Aut}}(F_n)$$
- Property ${\mathcal{N}}_n(a)$ holds if and only if $n > {\textnormal{rk}}_R(M)$ and $\Phi_a$ surjective.
- If $C$ is infinite then $\Phi_a$ is bijective.
It follows from Lemma \[LemGammaImpliesNielsen\] that $\Phi_a$ is well-defined. Assertion $(i)$ is trivial while assertion $(ii)$ results from Lemma \[LemGeneratorsOfM\].$ii$ and Proposition \[PropConverseZ\].
\[CorStableRank\] Assume that $M \simeq R$. If $n >{\textnormal{sr}}(R) + 1$, then $G = M \rtimes_{\alpha} C$ has only one Nielsen equivalence class of generating $n$-vectors.
The result follows from Theorem \[ThMTimesC\] and the inequality (\[EqStableRank\]).
We consider now several hypotheses under which the problem of counting Nielsen equivalence classes is particularly tractable. One of these hypotheses is that $R$ be *quasi-Euclidean*, i.e., $R$ enjoys the following row reduction property shared by Euclidean rings: for every $n \ge 2$ and every ${\mathbf{r}}= (r_1, \dots, r_n) \in R^n$, there exist $E \in {\textnormal{E}}_n(R)$ and $d \in R$ such that $
(d, 0,\dots,0) = {\mathbf{r}}E
$ (see [@AJLL14] for a comprehensive survey on quasi-Euclidean rings). If $R$ is a Noetherian quasi-Euclidean ring, then $M$ admits an *invariant factor decomposition*, i.e., a decomposition of the form $R /{\mathfrak{a}}_1 \times R/{\mathfrak{a}}_2 \times \cdots \times R /{\mathfrak{a}}_n$ with $R \neq {\mathfrak{a}}_1 \supset {\mathfrak{a}}_2 \supset \cdots \supset {\mathfrak{a}}_n$ where the ideals ${\mathfrak{a}}_i$ are referred to as the *invariant factors* of $M$ (see [@Guy17 Lemma 1]).
\[CorGE\] Let $G = M \rtimes_{\alpha} C$ and $n = {\textnormal{rk}}(G)$. Then the following hold:
- If $M$ is free over $R$, $C$ is infinite and $R$ is a ${\textnormal{GE}}_{n - 1}$-ring, then\
${\mathfrak{n}}_n(G) = \vert R^{\times}/T \vert$.
- If $C$ is infinite and $R$ is quasi-Euclidean, then ${\mathfrak{n}}_n(G) = \vert \Lambda^{\times}/T_{\Lambda} \vert$ where $\Lambda = R/{\mathfrak{a}}_1$, ${\mathfrak{a}}_1$ is the first invariant factor of $M$ and $T_{\Lambda}$ is the image of $T$ in $\Lambda$ by the natural map.
$(i)$. As $C$ is infinite, it follows from Theorem \[ThMTimesC\] that $M \simeq R^{n - 1}$. For ${\mathbf{m}}\in {\textnormal{Um}}_{n - 1}(M)$, let ${\textnormal{Mat}}({\mathbf{m}}) \in {\textnormal{GL}}_{n - 1}(R)$ be the matrix whose columns are the components of ${\mathbf{m}}$. For every $E \in \Gamma_{n - 1}(R)$ the identity ${\textnormal{Mat}}({\mathbf{m}}E) = {\textnormal{Mat}}({\mathbf{m}})E$ holds. As $R$ is a ${\textnormal{GE}}_{n - 1}$-ring, we have $\Gamma_{n - 1}(R) = {\textnormal{D}}_{n -1}(T) {\textnormal{SL}}_{n - 1}(R)$. We deduce from Whitehead’s lemma that ${\textnormal{Mat}}({\mathbf{m}})$ can be reduced to $D_{n - 1}(u)$ via right multiplication by some $E \in \Gamma_{n -1}(R)$, where $u$ is a member of a transversal of $R^{\times}/T$. Therefore ${\mathfrak{n}}_n(G) \le \vert R^{\times}/T \vert$. Since $uT = \det({\textnormal{Mat}}({\mathbf{m}})E)T$, we conclude that ${\mathfrak{n}}_n(G) = \vert R^{\times}/T \vert$.
$(ii)$. By [@Guy17 Theorem A and Corollary C], we have ${\textnormal{Um}}_n(M)/\Gamma_n(R) \simeq \Lambda^{\times}/T_{\Lambda}$. Theorem \[ThMTimesC\] implies ${\mathfrak{n}}_n(G) = \left\vert\Lambda^{\times}/T_{\Lambda}\right\vert$.
We examine in the next proposition the structural implication of $M$ being free over $R$ with $R$-rank equal to ${\textnormal{rk}}(G)$.
\[PropCyclicDecMaxRank\] Assume that ${\textnormal{rk}}_R(M) = n = {\textnormal{rk}}(G)$ and $M$ is the direct sum of $n$ cyclic factors, i.e., $$M = R /{\mathfrak{a}}_1 \times \cdots \times R /{\mathfrak{a}}_n$$ where the ${\mathfrak{a}}_i$ are ideals of $R$. Let ${\mathfrak{a}}= {\mathfrak{a}}_1 + \cdots + {\mathfrak{a}}_n$. Then $\nu(G)$ is invertible modulo ${\mathfrak{a}}$. In addition, $C$ is finite and $G/ {\mathfrak{a}}M = {\mathbb Z}_d^{{\textnormal{rk}}(G)} \times C$ where $d = \vert R/{\mathfrak{a}}\vert < \infty$ is prime to ${\vert C \vert}$.
We can assume without loss of generality that ${\mathfrak{a}}= \{0\}$. Let ${\mathbf{e}}= (e_i)$ be a basis of $M$ over $R$ and let ${\mathbf{g}}\in {\operatorname{V}}_n(G)$ for $n = {\textnormal{rk}}(G)$. We can also suppose that ${\mathbf{g}}= ({\mathbf{m}}, ma)$ with ${\mathbf{m}}\in M^{n - 1}$ and $m \in M$. By Lemma \[LemRankOfM\], the vector $({\mathbf{m}}, \nu(G) m)$ generates $M$ as an $R$-module. Therefore the map ${\mathbf{e}}\mapsto ({\mathbf{m}}, \nu(G) m)$ induces an $R$-automorphism of $M$. This shows that $e_n = \nu(G) m'$ for some $m' \in M$. Hence a relation of the form $\sum_{i = 1}^{n - 1} r_i e_i + (\nu(G) r_n - 1) e_n = 0$, with $r_i \in R$ holds in $M$. It follows that $\nu(G)$ is invertible. Thus $M = \nu(G) M$ is $C$-invariant so that $G = M_C \times C$. As a result $M = M_C$ is a free ${\mathbb Z}_d$-module with $d = \vert R \vert$ or $d = 0$. Since ${\textnormal{rk}}(M) = {\textnormal{rk}}(G)$, the group $C$ must be finite, $d$ must be non-zero and prime to ${\vert C \vert}$.
\[CorGEMFree\] Let $G = M \rtimes_{\alpha} C$. Assume that at least one of the following holds:
- $R$ is quasi-Euclidean.
- $M$ is free over $R$ and $R$ is a ${\textnormal{GE}}$-ring
Then ${\mathfrak{n}}_n(G) = 1$ for every $n > {\textnormal{rk}}(G)$.
Suppose first that $(i)$ holds. By [@Guy17 Corollary B], the set\
${\textnormal{Um}}_{n - 1}(M)/\Gamma_{n - 1}(R)$ is reduced to one element and Theorem \[ThMTimesC\] implies ${\mathfrak{n}}_n(G) = 1$. Suppose now that $(ii)$ holds. If ${\textnormal{rk}}_R(M) = {\textnormal{rk}}(G)$, then $G$ is Abelian by Proposition \[PropCyclicDecMaxRank\] and the result follows from Theorem \[ThNielsenAbel\]. Assume now that $M \simeq R^{{\textnormal{rk}}(G) - 1}$, let $k = {\textnormal{rk}}(G)$ and $n > k$. As the result certainly holds if $G$ is cyclic, we assume moreover that $k \ge 2$. Let us show that ${\textnormal{Um}}_{n - 1}(M)/{\textnormal{E}}_{n - 1}(R)$ is made of a single orbit. For ${\mathbf{m}}\in {\textnormal{Um}}_{n - 1}(M)$, let ${\textnormal{Mat}}({\mathbf{m}})$ be the $(k -1)$-by-$(n -1)$ matrix whose columns are the components of ${\mathbf{m}}$. Since $R$ is a ${\textnormal{GE}}$-ring, there is $E \in {\textnormal{E}}_{n - 1}(R)$ such that $$\label{EqMatM}
{\textnormal{Mat}}({\mathbf{m}})E =
\begin{pmatrix} 1 & 0 \\ A & B \end{pmatrix}$$ where $A$ is an $(k - 2)$-by-$1$ matrix and $B$ is a $(k - 2)$-by-$(n -2)$. If $k = 2$, then ${\mathbf{m}}$ has been reduced to a standard unimodular row. Otherwise, let ${\mathbf{m}}' = {\mathbf{m}}E$. Since ${\mathbf{m}}'$ generates $M$, we can find a column vector $V = \begin{pmatrix} v_1 \\ \vdots \\ v_{n - 1} \end{pmatrix}$ such that ${\textnormal{Mat}}({\mathbf{m}}')V = \begin{pmatrix} 1 \\ 0 \end{pmatrix}$. Combining the latter identity with (\[EqMatM\]), we deduce that $v_1 = 1$ and subsequently ${\textnormal{Mat}}({\mathbf{m}}')P =
\begin{pmatrix} 1 & 0 \\ 0 & B \end{pmatrix}$ where $P = E_{12}(v_2) \cdots E_{1 \, n -1}(v_{n - 1})$. By iterating this procedure, we reduce ${\textnormal{Mat}}({\mathbf{m}})$ to $\begin{pmatrix} I_{k - 1} & 0 \end{pmatrix}$ through elementary column reduction operations. Consequently, ${\textnormal{Um}}_{n -1}(M)/{\textnormal{E}}_{n - 1}$ contains a single orbit. Theorem \[ThMTimesC\] eventually implies ${\mathfrak{n}}_n(G) = 1$.
T-systems {#SecTsys}
---------
In this section, we prove results on the ${\operatorname{T}}$-systems of $G = M \rtimes_{\alpha} C$ under the assumption that $M$ is free over $R$. These results will be specialized in Section \[SecT2\] so as to prove Theorem \[ThT2\]. We present first all the definitions needed for describing ${\textnormal{Aut}}(G)$. Let $c \mapsto {\overline{c}}$ be the restriction to $C$ of the natural map ${ {\mathbb Z}\lbrack C \rbrack}\twoheadrightarrow R$. We call $d:C \longrightarrow M$ a *derivation* if $d(cc') = d(c) + \overline{c}d(c')$ holds for every $(c,c') \in C^2$. Given a derivation $d$, we denote by $X_d$ the automorphism of $G$ defined by $$mc \mapsto m d(c) c$$ For $t \in {\textnormal{Aut}}_R(M)$, denote by $Y_t$ the automorphism of $G$ defined by $$mc \mapsto t(m)c$$
The following lemma underlines the link between ${\textnormal{Der}}(C, M)$, the $R$-module of derivations, and the automorphisms of $G$ which leave $M$ point-wise invariant.
\[LemDerivation\] Let $m \in M$. Then the following are equivalent:
- $\nu(G) m = 0$.
- There exists $d \in {\textnormal{Der}}(C, M)$ such that $d(a) = m$.
- There exists $\phi \in {\textnormal{Aut}}(G)$ such that $\phi(a) = ma$.
If one of the above holds, then the derivation $d$ in $(ii)$ is uniquely defined by $d(a^k) = \partial_{\alpha}(k)m$ for every $k \in {\mathbb Z}$. If in addition the restriction to $M$ of the automorphism $\phi$ in $(iii)$ is the identity, then $\phi(mc) = md(c)c$ for every $(m , c) \in M \times C$.
$\square$
We denote by $A(C)$ the subgroup of the automorphisms of $C$ induced by automorphisms of $G$ preserving $M$. The following result is referred to in [@Guy12 Proposition 4] where it is a key preliminary to the study of $2$-generated $G$-limits in the space of marked groups.
\[PropGenAut\] Assume $M$ is a free $R$-module and $\nu(G) = 0$. Let $n = {\textnormal{rk}}(G)$ and let ${\mathbf{g}}, {\mathbf{g}}' \in {\operatorname{V}}_n(G)$. Then the following are equivalent:
${\mathbf{g}}$ and ${\mathbf{g}}'$ are related by an automorphism of $G$ preserving $M$,
$\sigma({\mathbf{g}})$ and $\sigma({\mathbf{g}}')$ are related by an automorphism in $A(C)$.
Clearly, assertion $(i)$ implies $(ii)$. Let us prove the converse. Replacing, if needed, ${\mathbf{g}}$ by $\phi {\mathbf{g}}$ for some automorphism $\phi$ of $G$ that preserves $M$, we can assume without loss of generality that $\sigma({\mathbf{g}}) = \sigma({\mathbf{g}}')$. Replacing ${\mathbf{g}}$ and ${\mathbf{g}}'$ by ${\mathbf{g}}\psi$ and ${\mathbf{g}}\psi$ respectively for some $\psi \in {\textnormal{Aut}}(F_n)$, we can also assume that $\sigma({\mathbf{g}}) = \sigma({\mathbf{g}}') = (1_{C^{n - 1}}, a)$ and hence write ${\mathbf{g}}= ({\mathbf{m}}, ma)$ and ${\mathbf{g}}' = ({\mathbf{m}}', m' a)$ with ${\mathbf{m}}, {\mathbf{m}}' \in M^{n - 1}$ and $m, m' \in M$. By Lemma \[LemGeneratorsOfM\], the rows ${\mathbf{m}}$ and ${\mathbf{m}}'$ are bases of $M$. By Lemma \[LemDerivation\], there is $d' \in {\textnormal{Der}}(C, M)$ such that $d'(a) = m - m'$. Let $t \in {\textnormal{Aut}}_R(M)$ be defined by $t({\mathbf{m}})={\mathbf{m}}'$. Then we have ${\mathbf{g}}= Y_tX_{d'} {\mathbf{g}}'$, which proves the result.
Recall that $G$ is said to have property ${\mathcal{N}}_n(a)$ if every of its generating $n$-vectors can be Nielsen reduced to an $a$-row, i.e., a vector of the form $({\mathbf{m}}, a)$ with ${\mathbf{m}}\in {\textnormal{Um}}_{n -1}(M)$.
\[ThOneTSystem\] Let $n = {\textnormal{rk}}(G)$ and assume that $M$ is free over $R$. If moreover ${\mathcal{N}}_n(a)$ holds or ${\textnormal{rk}}_R(M) = n$, then $G$ has only one ${\operatorname{T}}_n$-system.
Assume ${\mathcal{N}}_n(a)$ holds. Given an $R$-basis $\mathbf{e}$ of $M$, we shall prove that any ${\mathbf{g}}\in {\operatorname{V}}_n(G)$ is in the same ${\operatorname{T}}_n$-system as $({\mathbf{e}}, a)$. By Theorem \[ThMTimesC\], we can assume without loss of generality that ${\mathbf{g}}$ is of the form $({\mathbf{m}}, a)$ with ${\mathbf{m}}\in {\textnormal{Um}}_{n -1}(M)$. The $R$-endomorphism $t$ mapping $\mathbf{e}$ to ${\mathbf{m}}$ is an $R$-isomorphism. Thus $({\mathbf{m}}, a) = Y_t({\mathbf{e}}, a)$. Let us assume now that ${\textnormal{rk}}_R(M) = n$ holds. The group $G$ is then a finite Abelian group by Proposition \[PropCyclicDecMaxRank\]. Thus the result follows from Theorem \[ThNielsenAbel\].
We denote by $A'(C)$ the subgroup of ${\textnormal{Aut}}(C)$ generated by $A(C)$ and the automorphism of $C$ which maps $a$ to $a^{-1}$. Here is the most general result we obtain regarding the count of ${\operatorname{T}}$-systems.
\[ThTSystems\] Let $n = {\textnormal{rk}}(G)$ and assume that $M$ is free over $R$. Then ${\mathfrak{t}}_n(G) \le \vert {\textnormal{Aut}}(C) / A'(C) \vert$, with equality if $M$ is a characteristic subgroup of $G$ and $M_C$ is isomorphic to ${\mathbb Z}^{n - 1}$.
In order to prove the above theorem, we will use this simple variation on results found in [@Sze04].
\[LemAutG\] Let ${\textnormal{Aut}}_{{\mathbb Z}}(M)$ be the group of ${\mathbb Z}$-automorphisms of $M$.
- Let $(\tau, \theta) \in {\textnormal{Aut}}_{{\mathbb Z}}(M) \times {\textnormal{Aut}}(C)$ such that $\theta(a) = a^k$ and $$\tau(cmc^{-1}) = \theta(c)\tau(m)\theta(c)^{-1} \text{ for all } m \in M,\, c \in C.$$ Then $Y_{\tau, \theta}: mc \mapsto \tau(m) \theta(c)$ is an automorphism of $G$ and the map $\alpha \mapsto \alpha^k$ induces a ring automorphism $\overline{\theta}$ of $R$ which satisfies $$\tau(rm) = \overline{\theta}(r)\tau(m) \text{ for all } m \in M,\, r \in R.$$
- Let $d \in {\textnormal{Der}}(C, M)$, $t \in {\textnormal{Aut}}_R(M)$ and $(\tau, \theta)$ as in $(i)$. Then we have $\tau^{-1} \circ d \circ \theta \in {\textnormal{Der}}(C, M)$, $\tau^{-1} \circ t \circ \tau \in {\textnormal{Aut}}_R(M)$ and $$Y_{\tau, \theta} X_d Y_{\tau, \theta}^{-1} = X_{\tau^{-1} \circ d \circ \theta}, \quad
Y_{\tau, \theta} Y_t Y_{\tau, \theta}^{-1} = Y_{\tau^{-1} \circ t \circ \tau}.$$
- Let $\phi \in {\textnormal{Aut}}(G)$ such that $\phi(M) = M$. Then there is $d \in {\textnormal{Der}}(C, M)$ and $(\tau, \theta)$ as in $(i)$, such that $\phi = X_d Y_{\tau, \theta}$. In particular, every automorphism in $A(C)$ is induced by some $Y_{\tau, \theta} \in {\textnormal{Aut}}(G)$.
The proofs of assertions $(i)$ and $(ii)$ are straightforward verifications.
$(iii)$. Let $\tau$ be the restriction of $\phi$ to $M$ and let $\theta$ the automorphism of $C$ induced by $\phi$. It is easy to check that $(\tau, \theta)$ satisfy the conditions of $(i)$. Let $\phi' = \phi Y_{\tau, \theta}^{-1}$. Then the restriction of $\phi'$ to $M$ is the identity and there is $m \in M$ such that $\phi'(a) = ma$. By Lemma \[LemDerivation\], there is $d \in {\textnormal{Der}}(C, M)$ such that $\phi' = X_d$.
[Proof of Theorem \[ThTSystems\]]{} If the property ${\mathcal{N}}_a(n)$ holds, or if ${\textnormal{rk}}_R(M) = n$, then Theorem \[ThOneTSystem\] implies that ${\mathfrak{t}}_n(G) = 1 \le \vert {\textnormal{Aut}}(C) / A'(C) \vert$. Therefore we can assume, without loss of generality that ${\mathcal{N}}_a(n)$ does not hold and ${\textnormal{rk}}_R(M) < n$.
By Theorem \[ThMTimesC\], the group $C$ is finite and $M_C$ is isomorphic to ${\mathbb Z}^{n - 1}$. Using Proposition \[PropReductionToRowFree\] and reasoning with a basis ${\mathbf{e}}$ of $M$ as in the proof of Theorem \[ThOneTSystem\].$i$, we see that every generating $n$-vector falls into the ${\operatorname{T}}_n$-system of $({\mathbf{e}}, a^k)$ for some $k$ coprime with ${\vert C \vert}$. It follows from Lemma \[LemAutG\].$ii$ that $({\mathbf{e}}, \theta(a^k))$ lies the ${\operatorname{T}}_n$-system of $({\mathbf{e}}, a^k)$ for every $\theta \in A'(C)$. Therefore ${\mathfrak{t}}_n(G) \le \vert {\textnormal{Aut}}(C) / A'(C) \vert$. Assume now that $M$ is a characteristic subgroup of $G$. If $({\mathbf{e}}, a)$ lies in the ${\operatorname{T}}_n$-system of $({\mathbf{e}}, a^k)$ for some $k$ coprime with ${\vert C \vert}$, then we can find $\phi \in {\textnormal{Aut}}(G)$ such that $\phi({\mathbf{e}}, a)$ is Nielsen equivalent to $({\mathbf{e}}, a^k)$. By Lemma \[LemAutG\].$ii$, we have $\phi({\mathbf{e}}, a) = ({\mathbf{e}}', m \theta(a))$ for some basis ${\mathbf{e}}'$ of $M$ and some $(m, \theta ) \in M \times A(C)$. By Lemma \[LemGeneratorsOfM\].$iii$, the vector $({\mathbf{e}}', \theta(a))$ is Nielsen equivalent to $({\mathbf{e}}, a^k)$. Proposition \[PropReductionToRowFree\] implies that $\theta(a) = a^{\pm k}$, hence there is $\theta' \in A'(C)$ such that $\theta'(a) = a^k$.
Nielsen equivalence classes and ${\operatorname{T}}$-systems of $R \rtimes C$ {#SecRC}
=============================================================================
In this section, we assume that $M \simeq R$, i.e., $G=\langle a,b\rangle$ is a split extension of the form $R \rtimes_{\alpha} C$ with $C = \langle a \rangle$, while $b$ is the identity of the ring $R$ and $a$ acts on $R$ as the multiplication by $\alpha \in R^{\times}$.
Nielsen equivalence of generating pairs {#SecNielsenPairs}
---------------------------------------
We prove here the first two assertions of Theorem \[ThN2\]. We begin with the definition of an invariant of Nielsen equivalence named $\Delta_a$. If $\nu(G) = 0$ there is a unique derivation $d_a \in {\textnormal{Der}}(C, R)$ satisfying $d_a(a)=1$. For ${\mathbf{g}}= (g, g') = (rc, r'c') \in G^2$ with $(r,r') \in R^2$ and $(c,c') \in C^2$, we set $$\label{EqDa}
D_a({\mathbf{g}}) = r d_a(c') - r' d_a(c) \in R$$ It is easily checked that ${\lbrack g, g' \rbrack}=(1 -\alpha)D_a({\mathbf{g}}).$ If $\nu(G) \neq 0$, we set further $$D_a({\mathbf{g}}) = D_a(\pi_{\nu(G) R}({\mathbf{g}})) \in R/\nu(G) R$$ where the right-hand side is defined as in (\[EqDa\]).
\[LemmaDelta\] We have $D_a({\mathbf{g}}) \in (R / \nu(G) R)^{\times}$ for every ${\mathbf{g}}\in {\operatorname{V}}_2(G)$.
We can assume without loss of generality that $\nu(G) = 0$.\
Let ${\mathbf{g}}= (ra^k, r'a^{k'}) \in {\operatorname{V}}_2(G)$ with $r,r' \in R$ and $k,k' \in {\mathbb Z}$. We first observe that $$\begin{array}{ccc}
D_a({\mathbf{g}}L_{12}) &=& \alpha^{k'} D_a({\mathbf{g}})\\
D_a({\mathbf{g}}L_{21}) &=& \alpha^{k} D_a({\mathbf{g}})\\
D_a({\mathbf{g}}I_1) &=& - \alpha^{-k}D_a({\mathbf{g}})
\end{array}$$ Thus $D_a({\mathbf{g}}{\textnormal{Aut}}(F_2)) = T D_a({\mathbf{g}})$. We know from Lemma \[LemGeneratorsOfM\] that ${\mathbf{g}}$ is Nielsen equivalent to $(r, a)$ for some $r \in R^{\times}$. Therefore $D_a({\mathbf{g}}) \in rT$ which shows that $D_a({\mathbf{g}})$ is invertible.
\[RemarkDaUnicity\] Assume $\nu(G) = 0$. Let $c$ be a generator of $C$ and let $d_c \in {\textnormal{Der}}(C, R)$ such that $d_c(c) = 1$. It is easily checked that $d_c = d_c(a) d_a$ and the identity $d_c(c) = 1$ implies that $d_c(a) \in R^{\times}$. For such elements $c$ there is thus only one map $D_c$ up to multiplication by a unit of $R$.
We set $$\Lambda = R / \nu(G) R, \quad T_{\Lambda} = \pi_{\nu(G) R}(T),$$ and define the map $$\begin{array}{cccc}
\Delta_a:& {\operatorname{V}}_2(G) & \rightarrow & \Lambda^{\times}/ T_{\Lambda} \\
& {\mathbf{g}}& \mapsto & T_{\Lambda} D({\mathbf{g}})
\end{array}$$
In the course of Lemma \[LemmaDelta\]’s proof we actually showed
\[LemDeltaAInvariant\] The map $\Delta_a$ is ${\textnormal{Aut}}(F_2)$-invariant.
The last stepping stone to the theorem of this section is the crucial
\[LemReductionModuloNuRank2\] Let ${\mathbf{g}}= (r, a)$, ${\mathbf{g}}' = (r', a)$ with $r, r' \in R^{\times}$. Then the following are equivalent:
- ${\mathbf{g}}$ and ${\mathbf{g}}'$ are Nielsen equivalent.
- $\pi_{\nu(G) R}({\mathbf{g}})$ and $\pi_{\nu(G) R}({\mathbf{g}}')$ are Nielsen equivalent.
- $\Delta_a({\mathbf{g}}) = \Delta_a({\mathbf{g}}')$.
$(i) \Rightarrow (ii)$. This follows from the ${\textnormal{Aut}}(F_2)$-equivariance of $\pi_{\nu(G) R}$.
$(ii) \Rightarrow (iii)$. This follows from Lemma \[LemDeltaAInvariant\].
$(iii) \Rightarrow (i)$. The result is trivial if $\nu(G) = 0$, thus we can assume that $C$ is finite. By hypothesis, there exist $k \in {\mathbb Z}$, $r_{\nu} \in R$ and $\epsilon \in \{\pm 1\}$ such that $r' = \epsilon \alpha^kr + r_{\nu}\nu(G)$. Replacing ${\mathbf{g}}'$ by a conjugate if needed, we can assume that $k = 0$. Taking the inverse of the first component of ${\mathbf{g}}$ if needed, we can moreover assume that $\epsilon = 1$, so that $r' = r + \nu(G) r_{\nu}$. Since $r' $ is a unit, we can argue as in the proof of Lemma \[LemRankOfM\].$iii$ to get $\psi \in {\textnormal{Aut}}(F_2)$ such that $(r', a) \psi = (r', r_{\nu} a)$. We have then $(r', a) \psi L_{1, 2}^{-{\vert C \vert}} = (r, r_{\nu} a)$. Since $r$ is a unit, we can cancel $r_{\nu}$ using another automorphism of $F_2$.
The next lemma will help us determine when $\Delta_a$ is a complete invariant of Nielsen equivalence.
\[LemSurjectivity\] Let $I \subset R$ be an ideal which is contained in all but finitely many maximal ideals of $R$. Then the natural map $R^{\times} \rightarrow (R/I)^{\times}$ is surjective.
Let ${\mathfrak{m}}_1, \dots, {\mathfrak{m}}_k$ be the maximal ideals of $R$ not containing $I$ and let $J = (\bigcap_i {\mathfrak{m}}_i) \cap I$. By the Chinese Remainder Theorem, the map $$\rho: r + J \mapsto (r + I, r + {\mathfrak{m}}_1, \dots, r + {\mathfrak{m}}_k)$$ is a ring isomorphism from $R/J$ onto $R/I \times R/{\mathfrak{m}}_1 \times \cdots \times R/{\mathfrak{m}}_k$. Given $u \in (R/I)^{\times}$ we can find $v = \tilde{u} + J \in (R/J)^{\times}$ such that $\rho(v) = (u, 1 + {\mathfrak{m}}_1, \dots, 1 + {\mathfrak{m}}_k)$. Hence we have $u = \tilde{u} + I$. As $J \subset {\mathcal{J}}(R)$, we also have $\tilde{u} \in R^{\times}$.
Before we can state the main result of this section, we need to introduce some notation. Given an ideal $I$ of $R$, we denote by $\pi_I$ the natural group epimorphism $R \rtimes_{\alpha} C \twoheadrightarrow R/I \rtimes_{\alpha + I} C$.
\[ThNielsenRankTwo\] Let ${\mathbf{g}}, {\mathbf{g}}' \in {\operatorname{V}}_2(G)$, ${\mathbf{e}}\Doteq \pi_{ab}(b, a)$ and $R_C \Doteq R/(1 - \alpha)R$.
1. The following are equivalent:
- The pairs ${\mathbf{g}}$ and ${\mathbf{g}}'$ are Nielsen equivalent.
- The pairs $\pi_I({\mathbf{g}})$ and $\pi_I({\mathbf{g}}')$ are Nielsen equivalent for every\
$I \in \left\{ (1 - \alpha)R, \nu(G) R\right\}$.
- $\det_{{\mathbf{e}}} \circ \pi_{ab}({\mathbf{g}}) = \pm \det_{{\mathbf{e}}} \circ \pi_{ab}({\mathbf{g}}')$ and $\Delta_a({\mathbf{g}}) = \Delta_a({\mathbf{g}}')$.
2. If $C$ is infinite or $R_C$ is finite then $\Delta_a$ is surjective and the above conditions are equivalent to $
\Delta_a({\mathbf{g}}) = \Delta_a({\mathbf{g}}').
$ In this case $\Delta_a$ is a complete invariant of Nielsen equivalence for generating pairs.
$1.(i) \Rightarrow (ii)$. This follows from the ${\textnormal{Aut}}(F_2)$-equivariance of $\pi_I$.\
$1.(ii) \Rightarrow (iii)$. We deduce the identity $\det_{{\mathbf{e}}} \circ \pi_{ab}({\mathbf{g}}) = \pm \det_{{\mathbf{e}}} \circ\pi_{ab}({\mathbf{g}}')$ from Theorem \[ThNielsenAbel\] and the identity $\Delta_a({\mathbf{g}}) = \Delta_a({\mathbf{g}}')$ from Lemma \[LemDeltaAInvariant\].
$1.(iii) \Rightarrow (i)$. Suppose first that $C$ is infinite or $R_C$ is finite. By Theorem \[ThMTimesC\], we know that ${\mathbf{g}}$ and ${\mathbf{g}}'$ can be Nielsen reduced to $(r, a)$ and $(r', a)$ for some $r, r' \in R^{\times}$. By Lemma \[LemReductionModuloNuRank2\] the pairs ${\mathbf{g}}$ and ${\mathbf{g}}'$ are Nielsen equivalent. Suppose now that $C$ is finite and $R_C$ is infinite. By Proposition \[PropReductionToRowFree\], we know that ${\mathbf{g}}$ and ${\mathbf{g}}'$ can be Nielsen reduced to $(r, a^k)$ and $(r', a^{k'})$ for some $r, r' \in R^{\times}$ and $k, k' \in {\mathbb Z}$ such that $k \equiv \pm \det_{{\mathbf{e}}} \circ \pi_{ab}({\mathbf{g}})$ and $k' \equiv \pm \det_{{\mathbf{e}}} \circ \pi_{ab}({\mathbf{g}}') $ modulo ${\vert C \vert}$. We deduce from Theorem \[ThNielsenAbel\] that $k' \equiv \pm k \mod {\vert C \vert}$. Replacing ${\mathbf{g}}$ by ${\mathbf{g}}I_2$ if needed, we can assume that $a^k = a^{k'}$. Thanks to Remark \[RemarkDaUnicity\], we can argue as in the first part of the proof where $a$ is replaced by $a^k$, which proves the Nielsen equivalence of ${\mathbf{g}}$ and ${\mathbf{g}}'$.
$2$. We already showed in the proof of $(1)$ that $\Delta_a$ is injective if $C$ is infinite or $R_C$ is finite. Thus we are left with the proof of $\Delta_a$’s surjectivity. Clearly, it suffices to show that the natural map $R^{\times} \rightarrow (R/\nu(G) R)^{\times}$ is surjective. This is trivial if $C$ is infinite since $\nu(G) = 0$ in this case. So let us assume that $R_C$ is finite. Since we have $(1 - \alpha)\nu(G) = 0$, the ring element $\nu(G)$ belongs to every maximal ideal of $R$ which doesn’t contain $1 - \alpha$. Hence it belongs to all but finitely many maximal ideals of $R$. Now Lemma \[LemSurjectivity\] yields the conclusion.
We end this section with an algorithmic characterization of generating pairs.
\[PropGeneratorsDelta\] Assume $\nu(G)$ is nilpotent and let ${\mathbf{g}}\in G^2$. Then the following are equivalent:
- ${\mathbf{g}}$ generates $G$.
- $\sigma({\mathbf{g}})$ generates $C$ and $D_a({\mathbf{g}}) \in \left(R/\nu(G) R\right)^{\times}$.
$(i) \Rightarrow (ii)$. This follows from $\sigma$’s surjectivity and Lemma \[LemmaDelta\].
$(ii) \Rightarrow (i)$. Since $\sigma({\mathbf{g}})$ generates $C$ there is $\psi \in {\textnormal{Aut}}(F_2)$ such that $\sigma({\mathbf{g}}) \psi = (1_C, a)$. Replacing ${\mathbf{g}}$ by ${\mathbf{g}}\psi$, we can then assume that ${\mathbf{g}}= (r, r' a)$ for some $r, r' \in R$. Since $\Delta_a({\mathbf{g}}) = r + \nu(G) R \in (R/\nu(G) R)^{\times}$ and $\nu(G) \in {\mathcal{J}}(R)$ we deduce that $r \in R^{\times}$. Therefore ${\mathbf{g}}$ generates $G$.
Nielsen equivalence of generating triples and quadruples
--------------------------------------------------------
In this section we prove the last two assertions of Theorem \[ThN2\]. Since ${\textnormal{rk}}(G) = 2$ and ${\dim_{\textnormal{Krull}}}(R) \le 2$, Corollary \[CorStableRank\] ensures that $G$ has only one Nielsen class of generating $n$-vectors for $n > 4$. Using Theorem \[ThMTimesC\] in combination with a theorem of Suslin [@Sus77 Theorem 7.2], we show in Theorem \[ThTriplesAndQuadruples\] below that this remains true if $n > 3$. Recall that the map $$\Phi_a: {\mathbf{r}}\Gamma_{n -1}(R) \mapsto ({\mathbf{r}}, a) {\textnormal{Aut}}(F_n)$$ defined in Theorem \[ThMTimesC\] is a bijection from ${\textnormal{Um}}_{n - 1}(R)/\Gamma_{n - 1}(R)$ onto\
${\operatorname{V}}_n(G)/{\textnormal{Aut}}(F_n)$ provided that $C$ is infinite.
\[LemGEImpliesOneNielsenClass\] Let $n \ge 3$. If $R$ is a ${\textnormal{GE}}_{n - 1}$-ring, then ${\mathfrak{n}}_n(G) = 1$.
Let ${\mathbf{g}}\in {\operatorname{V}}_n(G)$. We shall show that ${\mathbf{g}}$ is Nielsen equivalent to ${\mathbf{g}}_1 \Doteq ({\mathbf{r}}_1, a)$ with ${\mathbf{r}}_1 = (1, 0, \dots, 0) \in {\textnormal{Um}}_{n -1}(R)$. As $n > {\textnormal{rk}}(G)$, the property ${\mathcal{N}}_n(a)$ holds by Theorem \[ThMTimesC\]. Therefore ${\mathbf{g}}$ can be Nielsen reduced to a vector of the form $({\mathbf{r}}, a)$ with ${\mathbf{r}}\in {\textnormal{Um}}_{n - 1}(R)$. Since $R$ is a ${\textnormal{GE}}_{n - 1}$-ring and $R$ is completable by Lemma \[LemCompletable\], the group ${\textnormal{E}}_{n - 1}(R)$ acts transitively on ${\textnormal{Um}}_{n - 1}(R)$. Hence ${\mathbf{r}}$ can be transitioned to ${\mathbf{r}}_1$ under the action of ${\textnormal{E}}_{n - 1}(R)$. Lemma \[LemGammaImpliesNielsen\] implies that ${\mathbf{g}}$ is Nielsen equivalent to ${\mathbf{g}}_1$.
Our forthcoming result on generating triples involves the following definitions from algebraic $K$-theory. For every $n \ge 1$, the map $A \mapsto \begin{pmatrix} A & 0 \\0 & 1\end{pmatrix}$ defines an embedding from ${\textnormal{SL}}_n(R)$ into ${\textnormal{SL}}_{n + 1}(R)$, respectively from ${\textnormal{E}}_n(R)$ into ${\textnormal{E}}_{n + 1}(R)$. Denote by ${\textnormal{SL}}(R)$ and ${\textnormal{E}}(R)$ the respective ascending unions. Then ${\textnormal{E}}(R)$ is normal in ${\textnormal{SL}}(R)$ and the group ${\operatorname{SK}}_1(R)$, the *special Whitehead group of $R$*, is the quotient ${\textnormal{SL}}(R) /{\textnormal{E}}(R)$. The next lemma shows in particular that the image in ${\operatorname{SK}}_1(R)$ of a matrix in ${\textnormal{SL}}_2(R)$ depends only on its first row.
\[LemRho\] Let $R$ be any commutative ring with identity. Denote by $\hat{{\textnormal{E}}}_2(R)$ the normal closure of ${\textnormal{E}}_2(R)$ in ${\textnormal{SL}}_2(R)$. Let $\rho: {\textnormal{SL}}_2(R) \rightarrow {\textnormal{Um}}_2(R)$ be defined by $\begin{pmatrix} a & b \\ c & d \end{pmatrix} \mapsto (a, b)$. Then the map $\rho$ induces a bijection from ${\textnormal{SL}}_2(R)/\hat{{\textnormal{E}}}_2(R)$ onto ${\textnormal{Um}}_2(R)/\hat{{\textnormal{E}}}_2(R)$.
For every $A, B \in {\textnormal{SL}}_2(R)$ the identity $\rho(AB) = \rho(A)B$ holds. Therefore the map $\hat{\rho}: A \hat{{\textnormal{E}}}_2(R) \mapsto \rho(A) \hat{{\textnormal{E}}}_2(R)$ is well defined. Let $(a, b) \in {\textnormal{Um}}_2(R)$ and let $a', b' \in R$ be such that $a a' + b b' = 1$. Then $A \Doteq \begin{pmatrix} a & b \\ -b' & a' \end{pmatrix} \in {\textnormal{SL}}_2(R)$ and $(a, b) = \rho(A)$, so that $\rho$, and hence $\hat{\rho}$ is surjective. Let us prove that $\hat{\rho}$ is injective. Consider for this $A, B \in {\textnormal{SL}}_2(R)$ such that $\hat{\rho}(A) = \hat{\rho}(B)$. Multiplying $A$ on the right by a matrix in $\hat{{\textnormal{E}}}_2(R)$ if needed, we can assume that $\rho(A) = \rho(B)$. Thus $\rho(AB^{-1}) = \rho(A)B^{-1} = \rho(B)B^{-1} = (1, 0)$, which shows that $AB^{-1} \in {\textnormal{E}}_2(R)$. The result follows.
The *Mennicke symbol* ${\lbrack {\mathbf{r}}\rbrack}$ of ${\mathbf{r}}\in {\textnormal{Um}}_2(R)$ is the image in ${\operatorname{SK}}_1(R)$ of any matrix of ${\textnormal{SL}}_2(R)$ whose first row is ${\mathbf{r}}$. We are now in position to prove
\[ThTriplesAndQuadruples\] The following hold:
- If $C$ is infinite then ${\operatorname{V}}_3(G)/{\textnormal{Aut}}(F_3)$ surjects onto ${\operatorname{SK}}_1(R)$.
- If $R$ is a ${\textnormal{GE}}_2$-ring, e.g., $C$ is finite, then ${\mathfrak{n}}_3(G) = 1$.
- ${\mathfrak{n}}_4(G) = 1$.
$(i)$. By Theorem \[ThMTimesC\], we can identify the two orbit sets ${\operatorname{V}}_3(G)/{\textnormal{Aut}}(F_3)$ and ${\textnormal{Um}}_2(R)/\Gamma_2(R)$. The classical properties of the Mennicke symbol [@Lam06 Proposition 3.4] imply that the map ${\lbrack \cdot \rbrack}: {\textnormal{Um}}_2(R) \rightarrow {\operatorname{SK}}_1(R)$ is $\Gamma_2(R)$-invariant. This yields a map ${\textnormal{Um}}_2(R)/\Gamma_2(R) \rightarrow {\operatorname{SK}}_1(R)$. By Remark \[RemStableRankReduction\], the latter map is surjective.
$(ii)$. This is Lemma \[LemGEImpliesOneNielsenClass\] for $n = 3$.
$(iii)$. We can assume that $C$ is infinite since Lemma \[LemGEImpliesOneNielsenClass\] applies otherwise. If ${\dim_{\textnormal{Krull}}}(R) \le 1$, then ${\mathfrak{n}}_4(G) = 1$ by Corollary \[CorStableRank\]. Thus, we can also suppose that $R = { {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$. Since ${\operatorname{V}}_4(G)/{\textnormal{Aut}}(F_4)$ identifies with ${\textnormal{Um}}_3(R)/\Gamma_3(R)$ by Theorem \[ThMTimesC\] and since ${\textnormal{E}}_3(R)$ acts transitively on ${\textnormal{Um}}_3(R)$ by [@Sus77 Theorem 7.2], we deduce that ${\mathfrak{n}}_4(G) = 1$.
T-systems of generating pairs and triples {#SecT2}
-----------------------------------------
This section is dedicated to the proofs of Theorem \[ThT2\] and \[ThT2GN\]. Recall that we denote by ${\mathfrak{n}}_n(G)$ the number of Nielsen equivalence classes of generating $n$-vectors of $G$ and by ${\mathfrak{t}}_n$ the number its ${\operatorname{T}}_n$-systems of $G$, both numbers may be infinite.
[Proof of Theorem \[ThT2\]]{} $(i)$. Combine Theorems \[ThMTimesC\] and \[ThOneTSystem\].
$(ii)$. This is a specialization of Theorem \[ThTSystems\] to $G = R \rtimes_{\alpha} C$.
$(iii)$. Consider the action of ${\textnormal{Aut}}(G)$ on ${\operatorname{V}}_3(G)/{\textnormal{Aut}}(F_3)$ defined by $$\phi \cdot ({\mathbf{g}}{\textnormal{Aut}}(F_3)) = (\phi {\mathbf{g}}) {\textnormal{Aut}}(F_3), \quad {\mathbf{g}}\in {\operatorname{V}}_3(G), \phi \in {\textnormal{Aut}}(G).$$ Regarding the first inequality, it suffices to show that the stabilizer $S_{{\mathbf{g}}}$ of ${\mathbf{g}}{\textnormal{Aut}}(F_3)$ has index at most ${\vert A(C) \vert}{\mathfrak{n}}_2(G)$ in ${\textnormal{Aut}}(G)$ for every ${\mathbf{g}}\in {\operatorname{V}}_3(G)$. By Theorem \[ThMTimesC\].$i$, such a triple ${\mathbf{g}}$ is Nielsen equivalent to $(r, s, a)$ for some $r, s \in R$. Since $R = sR + tR = {\mathbb Z}{\lbrack \alpha^{\pm1} \rbrack}$, we easily see that every automorphism $X_d$ stabilizes ${\mathbf{g}}{\textnormal{Aut}}(F_3)$. For an automorphism $Y_{\tau, 1}$ of $G$, we observe that $\tau \in {\textnormal{Aut}}_{{\mathbb Z}}(R)$ is actually an $R$-automorphism, so that $\tau$ is the multiplication some unit of $ R$. If $u_{\tau}$ is a trivial unit, we see that $Y_{\tau} = Y_{\tau, 1}$ stabilizes ${\mathbf{g}}{\textnormal{Aut}}(F_3)$ considering conjugates of the first two components of ${\mathbf{g}}$. If $A(C)$ contains an automorphisms $\theta$ which maps $a$ to $a^{-1}$, we let $\phi_{-1}$ be an automorphism of the form $Y_{\tau, \theta}$ whose image is $\theta$ through the natural map ${\textnormal{Aut}}(G) \twoheadrightarrow A(C)$. Otherwise we set $\phi_{-1} = 1$. Let $V$ be a transversal of $R^{\times}/T$. It follows from Lemma \[LemAutG\], that $\{Y_{\tau} \phi^{\epsilon};\, \epsilon \in \{0, 1\}, \tau(b) \in V \}$ is a transversal of ${\textnormal{Aut}}(G)/S_{{\mathbf{g}}}$. Since ${\mathfrak{n}}_2(G) = \vert R^{\times}/T \vert$ by Theorem \[ThN2\].$2$, we deduce that $\left\vert {\textnormal{Aut}}(G)/S_{{\mathbf{g}}} \right\vert \le {\vert A(C) \vert}{\mathfrak{n}}_2(G)$, which completes the proof of the first inequality.
In order to prove the second inequality, consider the action of ${\textnormal{Aut}}(G)$ on ${\operatorname{SK}}_1(R)$ defined by $\phi \cdot {\lbrack {\mathbf{r}}\rbrack} = {\lbrack \phi({\mathbf{r}}) \rbrack}$ for $(\phi, {\mathbf{r}}) \in {\textnormal{Aut}}(G) \times {\textnormal{Um}}_2(R)$ and where ${\lbrack {\mathbf{r}}\rbrack}$ denotes the Mennicke symbol of ${\mathbf{r}}$. The map $({\mathbf{r}}, a) {\textnormal{Aut}}(F_3) \mapsto {\lbrack {\mathbf{r}}\rbrack}$ induces an ${\textnormal{Aut}}(G)$-equivariant map $\mu$ from ${\operatorname{V}}_3(G)/{\textnormal{Aut}}(F_3)$ onto ${\operatorname{SK}}_1(R)$. Every automorphism $\phi \in \{ X_d, Y_t \, \vert \, d \in {\textnormal{Der}}(C, R), \tau \in {\textnormal{Aut}}_R(R)\}$ fixes every symbol ${\lbrack {\mathbf{r}}\rbrack}$. It follows from Lemma \[LemAutG\] that the ${\textnormal{Aut}}(G)$-action on ${\operatorname{SK}}_1(R)$ factors through an $A(C)$-action and we have ${\textnormal{Aut}}(G)\backslash{\operatorname{SK}}_1(R) \simeq A(C)\backslash{\operatorname{SK}}_1(R)$. Therefore $\mu$ induces a surjective map from ${\textnormal{Aut}}(G) \backslash{\operatorname{V}}_3(G)/{\textnormal{Aut}}(F_3)$ onto $ A(C)\backslash{\operatorname{SK}}_1(R)$, which yields the result.
By Theorem \[ThT2\], we have ${\mathfrak{t}}_2(G) = 1$ if $C$ is infinite or $G_{ab}$ is finite. With Corollary \[CorTTwoSystemsExample\] below, we prove the first part of Theorem \[ThT2GN\], that is, $G$ can have arbitrarily many ${\operatorname{T}}_2$-systems when $C$ is finite but $G_{ab}$ isn’t.
\[CorTTwoSystemsExample\] Let $q = p^d$ and $N = q - 1$, with $p$ a prime integer and $d \ge 2$ an even integer. Let $\Phi_{N, p}(X)$ be the $N$-th cyclotomic polynomial over the field with $p$ elements and let $P \in {\mathbb Z}{\lbrack X \rbrack}$ a monic polynomial of degree $d$ whose reduction modulo $p$ is an irreducible factor of $\Phi_{N, p}(X)$. Let $R = {\mathbb Z}{\lbrack X \rbrack}/(X - 1)I$ where $I =(p, P(X))$ is the ideal generated by $p$ and $P(X)$. Then the image $\alpha$ of $X$ in $R$ is invertible. It generates a subgroup $C \subset R^{\times}$ with $N$ elements and the number of ${\operatorname{T}}_2$-systems of $G = R \rtimes_{\alpha} C$ is ${\mathfrak{t}}_2(G) = \varphi(N)/d$.
We will use the following straightforward consequence of Lemma \[LemAutG\]:
\[LemAutR\] Let $k \in {\mathbb Z}$. The following are equivalent:
- There is $\theta \in A(C)$ such that $\theta(a) = a^k$.
- The map $\alpha \mapsto \alpha^k$ induces a ring automorphism of $R$.
$\square$
\[LemRCharacteristic\] The two following hold:
- Let $g = r a^k \in G$ with $r \in R$, $k \in {\mathbb Z}$. Then $g$ centralizes its conjugacy class if and only if $
(1 - \alpha^k)^2 = 0 = (1 - \alpha)(1 - \alpha^k)r
$.
- Let $\omega$ be the order of $\alpha$ in $R^{\times}$. Assume that $\omega = {\vert C \vert}$ and for every $k \in {\mathbb Z}$, we have $(1 - \alpha^k)^2 \neq 0$ whenever $\alpha^k \neq 1$. Then $R$ is a characteristic subgroup of $G$.
Assertion $(i)$ is a direct consequence of the identity $${\lbrack g, hgh^{-1} \rbrack} = (1 - \alpha^k) \left( (1 - \alpha^k)r - (1 - \alpha^{k'})r'\right), \text{ where } h = r'a^{k'}.$$ In order to prove $(ii)$, consider $\phi \in {\textnormal{Aut}}(G)$ and write $\phi(b) = ra^k$ where $b$ is the identity of the ring $R$. Since $b$ centralizes its conjugacy class, so does $\phi(b)$. By $(i)$, we have $(1 - \alpha^k)^2 = 0$, which yields $\alpha^k = 1$. As $\omega = {\vert C \vert}$, we deduce that $\phi(b) = r$ and hence $\phi(R) = R$.
[Proof of Corollary \[CorTTwoSystemsExample\]]{} The existence of the polynomial $P(X)$ is guaranteed by [@LN97 Theorem 2.47]. Let ${\overline{P}}(X) \in {\mathbb Z}_p{\lbrack X \rbrack}$ be the reduction of $P(X)$ modulo $p$. By the Chinese Remainder Theorem, the ring $R$ identifies with ${\mathbb Z}\times {\mathbb F}_q$ where ${\mathbb F}_q = {\mathbb Z}_p{\lbrack X \rbrack} / ({\overline{P}}(X))$ is the field with $q$ elements. As a result, the element $\alpha$ identifies with $(1, x)$ where $x \in {\mathbb F}_q^{\times}$ is an $N$-th primitive root of unity. Thus $C \simeq {\mathbb F}_q^{\times}$ and the ring automorphisms of $R$ induced by maps of the form $\alpha \mapsto \alpha^k$ correspond bijectively to powers of the Frobenius endomorphism of ${\mathbb F}_q$. Lemma \[LemRCharacteristic\]’s hypotheses are easily checked so that $R$ is a characteristic subgroup of $G$. By Lemma \[LemAutR\], we have then $\vert A(C) \vert = d$ and hence $\vert A'(C) \vert = d$ for $d$ is even. By Theorem \[ThT2\], we obtain ${\mathfrak{t}}_2(G) = \vert {\textnormal{Aut}}(C) / A'(C) \vert = \varphi(N)/d$.
Using the $K$-theory of orders of arithmetic type, we establish now the second part of Theorem \[ThT2GN\].
\[ThOrder\] Let $N \ge 3$, $\zeta_N = e^{2i \pi/N}$ and $\alpha = 1 + N^2 \zeta_N$. Define the ring $S$ as the localization ${\mathbb Z}{\lbrack \zeta_N \rbrack}_{\alpha}$ of ${\mathbb Z}{\lbrack \zeta_N \rbrack}$ at $\alpha$ and let $R = {\mathbb Z}{\lbrack \alpha^{\pm 1} \rbrack} \subset S$. Let $G = R \rtimes_{\alpha} {\mathbb Z}$ where the canonical generator of ${\mathbb Z}$ acts on $R$ as the multiplication by $\alpha$. Then we have ${\mathfrak{n}}_3(G) = {\mathfrak{t}}_3(G) = N$.
We first need to show
\[PropOrder\] Let $R$ be as in Theorem \[ThOrder\]. Then ${\textnormal{E}}_2(R)$ is normal in ${\textnormal{SL}}_2(R)$ and we have ${\textnormal{SL}}_2(R) / {\textnormal{E}}_2(R) \simeq {\operatorname{SK}}_1(R) \simeq {\mathbb Z}_N$.
We redirect the reader to [@Mag02] for the definitions of the relative special Whitehead group and quadratic residue symbols.
Since $R$ contains $N^2S$ and $S$ is a Dedekind ring of arithmetic type, the ring $R$ is an order of arithmetic type in the sense of B. Liehl [@Lie81]. It has moreover infinitely many units. By [@Lie81 Formulas (5) and (19)], the group ${\textnormal{E}}_2(R)$ is normal in ${\textnormal{SL}}_2(R)$ and ${\textnormal{SL}}_2(R)/{\textnormal{E}}_2(R)$ is a quotient of ${\operatorname{SK}}_1(S, {\mathfrak{f}}^2)$ where ${\mathfrak{f}}= \{ x \in R \, \vert \, xS \subset R\}$ is the conductor of $R$ in $S$. Let us show that ${\mathfrak{f}}= N^2 S$, i.e., the exponent of the Abelian group $S/R$ is $N^2$. As $\{1, \zeta_N,\dots, \zeta_N^{\varphi(N) - 1}\}$ is a basis of ${ {\mathbb Z}\lbrack \zeta_N \rbrack}$ over ${\mathbb Z}$, the exponent of ${ {\mathbb Z}\lbrack \zeta_N \rbrack}/ ({\mathbb Z}+ N^2 { {\mathbb Z}\lbrack \zeta_N \rbrack})$ is $N^2$. Setting $\overline{\alpha} \Doteq \alpha + N^2{ {\mathbb Z}\lbrack \zeta_N \rbrack}= 1 + N^2 { {\mathbb Z}\lbrack \zeta_N \rbrack}$, we observe that $S/R = { {\mathbb Z}\lbrack \zeta_N \rbrack}_{\alpha} / ({\mathbb Z}+ N^2{ {\mathbb Z}\lbrack \zeta_N \rbrack})_{\alpha} =
\left({ {\mathbb Z}\lbrack \zeta_N \rbrack}/({\mathbb Z}+ N^2{ {\mathbb Z}\lbrack \zeta_N \rbrack})\right)_{\overline{\alpha}}
\simeq { {\mathbb Z}\lbrack \zeta_N \rbrack}/({\mathbb Z}+ N^2{ {\mathbb Z}\lbrack \zeta_N \rbrack})$. Therefore the exponent of $S/R$ is also $N^2$, which proves the claim.
We shall establish $$\label{EqIsoSK}
{\operatorname{SK}}_1(R) \simeq {\operatorname{SK}}_1(R, {\mathfrak{f}})$$ The inclusion map from $R$ into $S$ induces an epimorphism ${\operatorname{SK}}_1(R, {\mathfrak{f}}) \twoheadrightarrow {\operatorname{SK}}_1(S, {\mathfrak{f}})$. If (\[EqIsoSK\]) holds, then ${\operatorname{SK}}_1(R)$ maps homomorphically onto ${\operatorname{SK}}_1(S, {\mathfrak{f}})$, which is in turn isomorphic to both ${\operatorname{SK}}_1(S, {\mathfrak{f}}^2)$ and ${\mathbb Z}_N$ by the Bass-Milnor-Serre theorem [@Mag02 Theorem 11.33]. Using [@Lie81 Formula (19)], we see that there is an epimorphism from ${\operatorname{SK}}_1(R)$ onto ${\textnormal{SL}}_2(R)/{\textnormal{E}}_2(R)$. By Remark \[RemStableRankReduction\], there is also an epimorphism from ${\textnormal{SL}}_2(R) / {\textnormal{E}}_2(R)$ onto ${\operatorname{SK}}_1(R)$. Since both groups are finite, there are isomorphic.
It only remains to prove (\[EqIsoSK\]). To do so, we first observe that $R/{\mathfrak{f}}\simeq {\mathbb Z}_{N^2}$. Indeed, setting $Q = ({\mathbb Z}+ N^2{ {\mathbb Z}\lbrack \zeta_N \rbrack}) / N^2{ {\mathbb Z}\lbrack \zeta_N \rbrack}$, it is straightforward to show that $Q \simeq {\mathbb Z}_{N^2}$ and subsequently $R/{\mathfrak{f}}= ({\mathbb Z}+ N^2{ {\mathbb Z}\lbrack \zeta_N \rbrack})_{\alpha}/(N^2{ {\mathbb Z}\lbrack \zeta_N \rbrack})_{\alpha}
\simeq Q_{\overline{\alpha}} \simeq {\mathbb Z}_{N^2}$. In the exact sequence (see, e.g., [@Mag02 Theorem 13.20 and Example 13.22]) $${\operatorname{K}}_2(R/{\mathfrak{f}}) \rightarrow {\operatorname{SK}}_1(R, {\mathfrak{f}}) \rightarrow {\operatorname{SK}}_1(R) \rightarrow {\operatorname{SK}}_1(R/{\mathfrak{f}})$$ the last term, namely ${\operatorname{SK}}_1(R/{\mathfrak{f}})$, is trivial since $R/{\mathfrak{f}}$ is finite. In addition, the image of ${\operatorname{K}}_2(R/{\mathfrak{f}})$ in ${\operatorname{SK}}_1(R, {\mathfrak{f}})$ is also trivial. Indeed ${\operatorname{K}}_2(R/{\mathfrak{f}})$ is generated by the Steinberg symbol $\{-1 + {\mathfrak{f}}, -1 + {\mathfrak{f}}\}_{R/{\mathfrak{f}}}$ because $R/{\mathfrak{f}}\simeq {\mathbb Z}_{N^2}$ (see [@Mag02 Exercises 13A.10 and 15C.10]). As $\{-1, -1\}_R$ is a lift of the previous symbol in ${\operatorname{K}}_2(R)$ our claim follows and the proof is now complete.
[Proof of Theorem \[ThOrder\]]{} By Theorem \[ThNielsenCInfinite\], the set ${\operatorname{V}}_3(G)/{\textnormal{Aut}}(F_3)$ identifies with ${\textnormal{Um}}_2(R)/\Gamma_2(R)$ via $({\mathbf{r}}, a) {\textnormal{Aut}}(F_3) \mapsto {\mathbf{r}}\Gamma_2(R)$ . Since ${\textnormal{E}}_2(R)$ is normal in ${\textnormal{SL}}_2(R)$ by Proposition \[PropOrder\], the set ${\textnormal{Um}}_2(R)/{\textnormal{E}}_2(R)$ identifies in turn with ${\textnormal{SL}}_2(R)/{\textnormal{E}}_2(R) \simeq {\operatorname{SK}}_1(R)$ by Lemma \[LemRho\] and the identification is given by ${\mathbf{r}}\mapsto {\lbrack {\mathbf{r}}\rbrack}$. Consequently ${\textnormal{Um}}_2(R)/{\textnormal{E}}_2(R) \simeq {\textnormal{Um}}_2(R)/\Gamma_2(R)$ and it follows that ${\mathfrak{n}}_3(G) = \vert {\operatorname{SK}}_1(R) \vert$. Since $R$ is a domain, Lemma \[LemRCharacteristic\].$ii$ applies so that $R$ is a characteristic subgroup of $G$. As the complex modulus of $\alpha \in { {\mathbb Z}\lbrack \zeta_N \rbrack}$ is greater than $1$, the integer $\alpha$ is not a unit of ${ {\mathbb Z}\lbrack \zeta_N \rbrack}$. Therefore $\alpha$ and $\alpha^{-1}$ have distinct minimal polynomials over ${\mathbb Q}$ and hence $\alpha \mapsto \alpha^{-1}$ doesn’t induce a ring endomorphism of $R$. By Lemma \[LemAutR\], the group $A(C)$ is trivial and reasoning as in the proof of Theorem \[ThT2\].$iii$, we obtain a bijective map ${\textnormal{Aut}}(G) \backslash{\operatorname{V}}_3(G)/{\textnormal{Aut}}(F_3) \rightarrow {\operatorname{SK}}_1(R)$. In particular, ${\mathfrak{t}}_3(G) = \vert {\operatorname{SK}}_1(R) \vert$. We eventually observe that $\vert {\operatorname{SK}}_1(R) \vert = N$ by Proposition \[PropOrder\], which completes the proof.
\[ExOrder\] Let $\alpha = 1 + 3 \zeta_3$ and $R = {\mathbb Z}{\lbrack \alpha^{\pm 1} \rbrack}$. Reasoning as in the proof of Theorem \[ThOrder\], we can show that ${\textnormal{E}}_2(R)$ is normal in ${\textnormal{SL}}_2(R)$ and ${\textnormal{SL}}_2(R) / {\textnormal{E}}_2(R) \simeq {\operatorname{SK}}_1(R) \simeq {\mathbb Z}_2$. As the quadratic residue $$\binom{3}{1 + \frac{3 + 9 \sqrt{-3} }{2}}_2$$ is $-1$, the Mennicke symbol $\lbrack3, 1 + \frac{3 + 9 \sqrt{-3} }{2}\rbrack$ generates ${\operatorname{SK}}_1(R)$. Therefore $G = R \rtimes_{\alpha} {\mathbb Z}$ has exactly two Nielsen equivalence classes of generating triples which coincide with the distinct $T_3$-systems of $({\mathbf{r}}, a)$ for ${\mathbf{r}}\in \{ (1, 0), (3, 1 + \frac{3 + 9 \sqrt{-3} }{2}) \}$. In contrast $G$, has only one Nielsen class of generating pairs since the units of $R$ are easily seen to be generated by $-1$ and $\alpha$.
Baumslag-Solitar groups, split metacyclic groups and lamplighter groups {#SecLamplighterGroups}
=======================================================================
This section is dedicated to the proofs of the Corollaries \[CorBS\], \[CorMetacyclic\], \[CorWreath\] and \[CorZwrZ\].
[Proof of Corollary \[CorBS\]]{} $(i)$. Let $G = BS(1, l)$. By Theorem \[ThN2\].$ii$, we have ${\mathfrak{n}}_2(G) = \left\vert R^{\times} /\langle \pm \alpha \rangle\right\vert$ with $R = { {\mathbb Z}\lbrack 1/l \rbrack}$ and $\alpha = l$. The prime divisors of $l$ form a basis of a free Abelian subgroup of $R^{\times}$ of index $2$. Thus ${\mathfrak{n}}_2(G)$ is finite if only if $l = \pm p^d$ for some prime $p$ and some $d \ge 0$. If $d = 0$, then $R = {\mathbb Z}$ and clearly ${\mathfrak{n}}_2(G) = 1$. Otherwise, ${\mathfrak{n}}_2(G) = \left\vert \langle \pm p \rangle / \langle \pm p^d \rangle \right\vert = d$.
$(ii)$. By Theorem \[ThT2\].$i$ (equivalently Brunner’s theorem [@Bru74]) we have ${\mathfrak{t}}_2(G) = 1$. Since ${ {\mathbb Z}\lbrack 1/l \rbrack}$ is Euclidean, it follows from Theorem \[ThN2\].$iii$ that ${\mathfrak{n}}_3(G) = 1$.
The following lemma will be used for counting generating pairs in the proofs of Corollary \[CorMetacyclic\] and \[CorWreath\].
\[LemTSysAbel\] Let $G = R \rtimes_{\alpha} C$ be as in Section \[SecRC\]. Assume that the natural map $R^{\times} \twoheadrightarrow R_C^{\times}$ is surjective and that $C$ is infinite or $R_C \Doteq R/(1 - \alpha)R$ is finite. Then ${\textnormal{Aut}}(G) \times {\textnormal{Aut}}(F_2)$ acts transitively on ${\operatorname{V}}_2(G_{ab})$ where the action of ${\textnormal{Aut}}(G)$ is the action induced by the natural homomorphism ${\textnormal{Aut}}(G) \rightarrow {\textnormal{Aut}}(G_{ab})$.
Let us assume first that $C$ is infinite. By Theorem \[ThNielsenAbel\], every generating pair of $G_{ab} = R_C \times C$ is Nielsen equivalent to $(u, a)$ for some $u \in R_C^{\times}$. Let $t$ be the multiplication by $u^{-1}$ on $R_C$. By hypothesis, we can find a lift $\tilde{u}$ of $u$ in $R$ which is moreover a unit. Let $\tau$ be the multiplication by $\tilde{u}^{-1}$ on $R$. Then the automorphism $Y_t \in {\textnormal{Aut}}(G_{ab})$ is induced by $Y_{\tau} \in {\textnormal{Aut}}(G)$ and we have $Y_t(u, a) = ({\overline{b}}, a)$ where ${\overline{b}}$ denotes the identity of $R_C$. Therefore every generating pair of $G$ is in the orbit of $({\overline{b}}, a)$ under the action of ${\textnormal{Aut}}(G) \times {\textnormal{Aut}}(F_2)$.
Assume now that both $C$ and $R_C$ are finite and let $d$ be the greatest common divisor of ${\vert C \vert}$ and $\vert R_C \vert$. For ${\mathbf{g}}\in G_{ab}$, we define $\det({\mathbf{g}})$ as in Corollary \[CorDet\]. The latter corollary implies that generating pairs with the same determinant are Nielsen equivalent. Hence it suffices to prove that the orbit of an arbitrary generating pair $(x, y) \in {\operatorname{V}}_2(G_{ab})$ contains a pair of determinant $1 \in {\mathbb Z}_d$. By Lemma \[LemSurjectivity\], there is a lift $u$ of $\det({\mathbf{g}})$ in $R_C^{\times}$ and by hypothesis, there is in turn a lift $\tilde{u}$ of $u$ in $R^{\times}$. Reusing the notation of the previous paragraph, we see that $Y_t(x, y)$ is a generating pair of determinant $1$. The proof is then complete.
We consider now split metacyclic groups, i.e., semi-direct products of the form $G = { {\mathbb Z}_k \rtimes_{\alpha} {\mathbb Z}_l}$ with $k, l \ge 0$ and $\alpha \in {\mathbb Z}_k^{\times}$.
[Proof of Corollary \[CorMetacyclic\]]{} $(i)$. By Theorem \[ThN2\].$ii$, we have $${\mathfrak{n}}_2(G) = \left\vert (R/\nu(G) R)^{\times} /\langle \pm \alpha \nu(G)R \rangle\right\vert$$ with $R = {\mathbb Z}_k$. Thus ${\mathfrak{n}}_2(G) = \frac{\varphi(\lambda)}{\omega}$ follows from the definitions of $\lambda$ and $\omega$.
$(ii)$. By Theorem \[ThT2\].$i$ we have ${\mathfrak{t}}_2(G) = 1$. Since ${\mathbb Z}_k$ is a ${\textnormal{GE}}$-ring, it follows from Theorem \[ThN2\].$iii$ that ${\mathfrak{n}}_3(G) = 1$.
$(iii)$. Since by definition ${\mathbb Z}_k/(1 - \alpha){\mathbb Z}_k \simeq {\mathbb Z}_e$, with have $G_{ab} \simeq {\mathbb Z}_e \times {\mathbb Z}_l$. Let ${{\overline{\mathbf{g}}}} \in {\operatorname{V}}_2(G_{ab})$. Since ${\textnormal{Aut}}(G) \times {\textnormal{Aut}}(F_2)$ acts transitively on ${\operatorname{V}}_2(G_{ab})$ by Lemma \[LemTSysAbel\], the number of preimages of ${\overline{\mathbf{g}}}$ in ${\operatorname{V}}_2(G)$ with respect to the abelianization homomorphism $\pi_{ab}$ does not depend on ${\overline{\mathbf{g}}}$. Hence it suffices to compute this number for ${\overline{\mathbf{g}}}= (\overline{b}, a)$ where ${\overline{b}}$ denotes the image of $b$, the canonical generator of ${\mathbb Z}_k$, in $G_{ab}$. A generating pair ${\mathbf{g}}\in {\operatorname{V}}_2(G)$ which maps to $({\overline{b}}, a)$ via $\pi_{ab}$ is of the form $(r, sa)$ with $r \in 1 + (1 - \alpha) {\mathbb Z}_k$ and $s \in (1 - \alpha) {\mathbb Z}_k$. It follows from Lemma \[LemGeneratorsOfM\].$i$ that a pair of this form generates $G$ if and only if $r$ is, in addition, a unit. By Lemma \[LemSurjectivity\], the natural map ${\mathbb Z}_k \twoheadrightarrow {\mathbb Z}_k/(1 - \alpha){\mathbb Z}_k \simeq {\mathbb Z}_e$ induces an epimorphism ${\mathbb Z}_k^{\times} \twoheadrightarrow {\mathbb Z}_e^{\times}$. Therefore the number of preimages of ${\overline{\mathbf{g}}}$ is $\frac{k \varphi(k)}{e \varphi(e)}$. As a result, we obtain $\left\vert {\operatorname{V}}_2(G) \right\vert = \frac{k \varphi(k)}{e \varphi(e)} \left\vert {\operatorname{V}}_2(G_{ab}) \right\vert$. Since ${\mathfrak{t}}_2(G) = 1$, the group ${\textnormal{Aut}}(G)$ acts transitively on the set ${\operatorname{V}}_2(G) / {\textnormal{Aut}}(F_2)$. Therefore the Nielsen equivalence classes of generating pairs have the same number of elements.
We consider eventually the two-generated lamplighter groups, i.e., the restricted wreath products of the form $G = {\mathbb Z}_k \wr {\mathbb Z}_l$ with $k, l \ge 0$ and $k, l \neq 1$. Such a group $G$ reads also as $G = R \rtimes_a C$ with $C = {\mathbb Z}_k = \langle a \rangle$ and $R = {\mathbb Z}_k {\lbrack C \rbrack} \simeq {\mathbb Z}_k{\lbrack X \rbrack}/(X^l - 1)$. As before, we denote by $T$ the subgroup of $R^{\times}$ generated by $-1$ and $a$. We also set $
\Lambda \Doteq R/\nu(G) R
$ and $
T_{\Lambda} \Doteq \pi_{\nu(G) R}(T)
$, like in Section \[SecNielsenPairs\].
\[CorWreathOneTTwoSystem\] Let $k, l \ge 0$ with $k, l \neq 1$ and let $G = { {\mathbb Z}_k \wr {\mathbb Z}_l}$. Then the following hold.
- ${\mathfrak{t}}_2(G) = 1$.
- If ${\mathbb Z}_k$ is finite or ${\mathbb Z}_l$ is infinite, then $
{\mathfrak{n}}_2(G) = \left\vert \Lambda^{\times} /T_{\Lambda} \right\vert.
$
- If ${\mathbb Z}_k$ or ${\mathbb Z}_l$ is finite, then ${\mathfrak{n}}_3(G) = 1$.
$(i)$. If ${\mathbb Z}_k$ is finite, or ${\mathbb Z}_l$ is infinite, then ${\mathfrak{t}}_2(G) = 1$ by Theorem \[ThT2\].$i$. Otherwise, Theorem \[ThT2\].$ii$ applies and ${\mathfrak{t}}_2(G) \le \vert {\textnormal{Aut}}(C) / A'(C) \vert$. It is easy to see that the map $a \mapsto a^i$ induces a ring automorphism of $R$ for every $i$ coprime with $l$. Thus $A'(C) = {\textnormal{Aut}}(C)$ by Lemma \[LemAutR\], which implies ${\mathfrak{t}}_2(G) = 1$.
$(ii)$. This is an immediate consequence of Theorem \[ThN2\].$ii$.
$(iii)$. If ${\mathbb Z}_k$ is finite then $R$ is a ${\textnormal{GE}}$-ring by Lemma \[LemArtinianCoefficientAndGE\]. If ${\mathbb Z}_l$ is finite then $R$ is ${\textnormal{GE}}$-ring by Theorem \[ThZCGE\]. Therefore ${\mathfrak{n}}_3(G) = 1$ by Theorem \[ThTriplesAndQuadruples\].$ii$.
\[CorFiniteWreathProduct\] Assume that both ${\mathbb Z}_k$ and ${\mathbb Z}_l$ are finite and non-trivial. Given a prime divisor $p$ of $k$, we denote by $\nu_l(p, d)$ the number of distinct irreducible factors of $$1 + X + \cdots + X^{l - 1}$$ in ${\mathbb Z}_p {\lbrack X \rbrack}$ which are monic of degree $d$. Let $l' = 2l$ if $k \neq 2$, $l' = l$ otherwise. Then we have $${\mathfrak{n}}_2({ {\mathbb Z}_k \wr {\mathbb Z}_l}) = \frac{k^{l - 1}}{l'} \prod_{p, d} (1 - \frac{1}{p^d})^{\nu_l(p, d)}$$ where $p$ ranges over the prime divisors of $k$ and $d$ over the positive integers.
The following lemma makes easy the task of computing the cardinality of the unit group in each finite ring under consideration.
[@Ste12 Exercise 44] \[LemUnitsOfAFiniteRing\] Let $R$ be a finite ring. Then $$\vert R^{\times} \vert = \vert R \vert \prod_{{\mathfrak{m}}}(1 - \frac{1}{\vert R / {\mathfrak{m}}\vert})$$ where ${\mathfrak{m}}$ ranges over the maximal ideals of $R$.
[Proof of Corollary \[CorFiniteWreathProduct\]]{} Since $\nu(G) R = {\mathbb Z}_k \nu(G)$, the ring $\Lambda$ has $k^{l - 1}$ elements. Each maximal ideal ${\mathfrak{m}}$ is generated by a prime divisor $p$ of $k$ and the image in $\Lambda$ of a polynomial $P \in {\mathbb Z}_k{\lbrack X \rbrack}$ whose reduction modulo $p$ is an irreducible monic factor of $1 + X + \cdots + X^{l - 1}$. Hence $\Lambda /{\mathfrak{m}}= {\mathbb F}_{p^d}$ where $d$ is the degree of $P$. Thus $\vert \Lambda^{\times} \vert =
k^{l - 1}\prod_{p, d} (1 - \frac{1}{p^d})^{\nu_l(p, d)}$ by Lemma \[LemUnitsOfAFiniteRing\] and we conclude the proof in observing that $l' = \vert T_{\Lambda} \vert$.
Given a prime divisor $p$ of $k$, we denote by $\mu_l(p, d)$ the number of distinct irreducible factors of $1 - X^l$ in ${\mathbb Z}_p {\lbrack X \rbrack}$ which are monic of degree $d$. Reasoning as above, it is straightforward to establish the formula $$\left\vert ({\mathbb Z}_k{\lbrack {\mathbb Z}_l \rbrack})^{\times}\right\vert = k^l\prod_{p, d} (1- \frac{1}{p^d})^{\mu_l(p, d)}.$$ where $p$ ranges over the prime divisors of $k$.
\[CorFiniteWreathProductV2\] Assume that both ${\mathbb Z}_k$ and ${\mathbb Z}_l$ are finite and non-trivial. Then we have $$\left\vert {\operatorname{V}}_2({ {\mathbb Z}_k \wr {\mathbb Z}_l}) \right\vert = \frac{k^{l -1}}{\varphi(k)}\left\vert ({ {\mathbb Z}_k \lbrack {\mathbb Z}_l \rbrack})^{\times}\right\vert
\left\vert {\operatorname{V}}_2({\mathbb Z}_k \times {\mathbb Z}_l) \right\vert$$ and the number of elements in a Nielsen equivalence class of generating pairs is $$l' k^{l - 1}\left\vert {\operatorname{V}}_2({\mathbb Z}_k \times {\mathbb Z}_l) \right\vert.$$ where $l'$ is as in Corollary \[CorFiniteWreathProduct\].
Let $G = { {\mathbb Z}_k \wr {\mathbb Z}_l}$. As ${ {\mathbb Z}_k \lbrack {\mathbb Z}_l \rbrack}/(1 - \alpha) \simeq {\mathbb Z}_k$, we have $G_{ab} \simeq {\mathbb Z}_k \times {\mathbb Z}_l$. Let ${\overline{\mathbf{g}}}\in {\operatorname{V}}_2(G_{ab})$. By Lemma \[LemTSysAbel\], the number of preimages of ${\overline{\mathbf{g}}}$ in ${\operatorname{V}}_2(G)$ with respect to the abelianization homomorphism $\pi_{ab}$ does not depend on ${\overline{\mathbf{g}}}$. Let us compute this number for ${\overline{\mathbf{g}}}= (\overline{b}, a)$ where ${\overline{b}}$ denotes the image of $b$ in $G_{ab}$. Using Lemma \[LemGeneratorsOfM\].$i$ we see that a lift ${\mathbf{g}}\in G^2$ of $({\overline{b}}, a)$ generates $G$ if and only if it is of the form $(r, sa)$ with $r \in (1 + (1 - \alpha) { {\mathbb Z}_k \lbrack {\mathbb Z}_l \rbrack}) \cap ({ {\mathbb Z}_k \lbrack {\mathbb Z}_l \rbrack})^{\times}$ and $s \in (1 - \alpha) { {\mathbb Z}_k \lbrack {\mathbb Z}_l \rbrack}$. By Lemma \[LemSurjectivity\], the natural map ${ {\mathbb Z}_k \lbrack {\mathbb Z}_l \rbrack}\twoheadrightarrow {\mathbb Z}_k$ induces an epimorphism $({ {\mathbb Z}_k \lbrack {\mathbb Z}_l \rbrack})^{\times} \twoheadrightarrow {\mathbb Z}_k^{\times}$. Therefore the number of preimages of ${\overline{\mathbf{g}}}$ is $\frac{\left\vert ({ {\mathbb Z}_k \lbrack {\mathbb Z}_l \rbrack})^{\times}\right\vert}{\left\vert {\mathbb Z}_k^{\times}\right\vert} k^{l -1}$, which proves the first formula. Since ${\mathfrak{t}}_2(G) = 1$, the group ${\textnormal{Aut}}(G)$ acts transitively on the set ${\operatorname{V}}_2(G) / {\textnormal{Aut}}(F_2)$. Therefore the Nielsen equivalence classes of generating pairs have the same number of elements, that is $$\frac{\left\vert {\operatorname{V}}_2(G) \right\vert}{{\mathfrak{n}}_2(G)} =
\frac{l'k^l}{\varphi(k)} \prod_{p, d} (1- \frac{1}{p^d})^{\mu_l(p, d) - \nu_l(p, d)}
\left\vert {\operatorname{V}}_2(G_{ab}) \right\vert$$ where $p$ ranges over the prime divisors of $k$. The integer $\mu_l(p, d) - \nu_l(p, d)$ is the number of monic irreducible polynomials in ${\mathbb Z}_p {\lbrack X \rbrack}$ of degree $d$ which divides $1 - X^l$ but not $1 + X + \cdots + X^{l -1}$. Therefore $\mu_l(p, d) - \nu_l(p, d) = 1$ if $d = 1$ and it cancels otherwise. Thus we have $\prod_{p, d} (1- \frac{1}{p^d})^{\mu_l(p, d) - \nu_l(p, d)} = \prod_p (1- \frac{1}{p}) = \frac{\varphi(k)}{k}$, which gives the result
\[CorWreathNTwo\] Let $k, l \ge 0$ and $k, l \neq 1$.
- Assume that ${\mathbb Z}_k$ is finite and ${\mathbb Z}_l$ is infinite. Then ${\mathfrak{n}}_2(G)$ is finite if and only if $k$ is prime; in this case ${\mathfrak{n}}_2(G) = \max(\frac{k - 1}{2}, 1)$.
- Assume that ${\mathbb Z}_k$ is infinite and ${\mathbb Z}_l$ is finite. Then ${\mathfrak{n}}_2(G)$ is finite if and only if $l \in \{2, 3, 4, 6\}$; in this case ${\mathfrak{n}}_2(G) = 1$.
$(i)$. The result follows from Corollary \[CorWreathOneTTwoSystem\].$ii$ and the isomorphisms $$\begin{array}{lllllllll}
({ {\mathbb Z}_k \lbrack X^{\pm 1} \rbrack})^{\times} &\simeq& {\mathbb Z}_k^{\times} &\times & U_X &\times & U_{X^{-1}} &\times& {\mathbb Z}^{\rho}.\\
T &\simeq& \{\pm 1 \} &\times & \{1\} &\times & \{1\} &\times & {\mathbb Z}.
\end{array}$$ where $\rho$ is the number of prime divisors of $k$ and $U_Y = 1 + Y{\operatorname{nil}}({\mathbb Z}_k){\lbrack Y \rbrack}$ with $Y \in \{X^{\pm 1}\}$ (see e.g., [@Lam06 Exercise 3.17] where the units in the ring of Laurent polynomials are determined).
$(ii)$. As $\Lambda = {\mathbb Z}{\lbrack X \rbrack} / (1 + X + \cdots + X^{l - 1})$, the ring $\Lambda$ identifies with ${\mathcal{O}(\mathcal{D} \setminus \{1\})}$ as defined in Lemma \[LemUnitsOfOD\] and where ${\mathcal{D}}$ is the set of divisors of $l$. By Lemma \[LemUnitsOfOD\], the group $\Lambda^{\times}$ is finite if and only if $l \in \{2, 3, 4, 6 \}$; in this case the equality $\Lambda^{\times} = T_{\Lambda}$ holds. Since ${\mathfrak{n}}_2(G) = \max(\varphi(l)/2, 1) \left\vert \Lambda^{\times} /T_{\Lambda} \right\vert$ by Theorem \[ThN2\], the result follows.
We conclude with the group $G = {\mathbb Z}\wr {\mathbb Z}$, which isomorphic to ${ {\mathbb Z}\lbrack X^{\pm 1} \rbrack}\rtimes_X {\mathbb Z}$.
[Proof of Corollary \[CorZwrZ\]]{} It follows from Lemma \[LemRCharacteristic\] that $R = { {\mathbb Z}\lbrack X^{\pm 1} \rbrack}$ is characteristic in $G$. The inequality ${\mathfrak{n}}_3(G) \le 2 {\mathfrak{t}}_3(G)$ is then a consequence of Theorem \[ThT2\].$ii$. The implication $(i) \Rightarrow (ii)$ is obvious while the equivalence $(i) \Leftrightarrow (iii)$ results from Theorem \[ThMTimesC\] and Lemma \[LemGECriteria\].$i$. In order to prove $(ii) \Rightarrow (i)$, we assume that $(ii)$ holds true, fix ${\mathbf{g}}_0 \in {\operatorname{V}}_2(G)$ and let ${\mathbf{g}}$ be an arbitrary generating triples of $G$. As ${\mathfrak{t}}_3(G) = 1$ by hypothesis, we deduce that ${\mathbf{g}}$ is Nielsen equivalent to a triple of the form $(1_G, {\mathbf{g}}_1)$ with ${\mathbf{g}}_1 \in {\operatorname{V}}_2(G)$. By Corollary \[CorWreathOneTTwoSystem\].$ii$, we have ${\mathfrak{n}}_2(G) = 1$, so that $(1_G, {\mathbf{g}}_1)$ is Nielsen equivalent $(1_G, {\mathbf{g}}_0)$. Therefore ${\mathfrak{n}}_3(G) = 1$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Daniel M. Kane'
title: '$k$-Independent Gaussians Fool Polynomial Threshold Functions'
---
Introduction
============
In this paper we consider the ability of limited independence to fool polynomial threshold functions (PTFs). We recall that a (degree-$d$) polynomial threshold function is a function of the form $f(x) = {\textrm{sgn}}(p(x))$ for some $n$-dimensional polynomial $p$ of degree at most $d$. There has been recent interest in polynomial threshold functions in several areas of computer science. This paper expands on previous work in derandomizing polynomial threshold functions using limited independence.
We say that a random variables $X$ fools a family of functions with respect to some distribution $Y$ if for every function, $f$, in the family $$|{\textrm{E}}[f(X)]-{\textrm{E}}[f(Y)]| = O(\epsilon).$$ In this paper we will be interested in the case where the family is of all degree-$d$ polynomial threshold functions in $n$-variables, and $Y$ is either an $n$-dimension Gaussian distribution, and in particular the case where $X$ is an arbitrary family of $k$-independent Gaussian random variables. In particular, we prove that
\[mainTheorem\] Let $d>0$ be an integer and $\epsilon>0$ a real number, then there exists a $k=O_d\left( \epsilon^{-2^{O(d)}}\right)$, so that for any degree $d$ polynomial $p$ and any $k$-independent family of Gaussians $X$ and fully independent family of Gaussians $Y$ $$\left| {\textrm{E}}[{\textrm{sgn}}(p(X))]-{\textrm{E}}[{\textrm{sgn}}(p(Y))]\right| = O(\epsilon).$$
There has been a significant amount of recent work on the problem of fooling low degree polynomial threshold functions of Gaussian or Bernoulli random variables, especially via limited independence. It was shown in [@DGJSV] that $\tilde O(\epsilon^{-2})$-independence is sufficient to fool degree-1 polynomial threshold functions of Bernoulli random variables, and show that this is tight up to polylogarithmic factors. In [@DKN] it was shown that $\tilde O(\epsilon^{-9})$-independence sufficed for degree-2 polynomial threshold functions of Bernoullis and that $O(\epsilon^{-2})$ and $O(\epsilon^{-8})$ suffices for degree 1 and 2 polynomial threshold functions of Gaussians. The degree 1 case was also extended by [@BLY], who show that limited independence fools threshold functions of polynomials that can be written in terms of a small number of linear polynomials. Finally, in [@MZ] a more complicated pseudorandom generator for degree-$d$ polynomial threshold functions of Bernoulli variables is developed with seed length $2^{O(d)}\log(n) \epsilon^{-8d-3}$. As far as we are aware, our paper is the first result to show that degree-$d$ polynomial threshold functions are fooled by $k$-independence for any $k$ depending only on $\epsilon$ and $d$ for any $d\geq 3$.
Overview
========
We prove Theorem \[mainTheorem\] first by proving our result for multilinear polynomials, and then finding a reduction to the general case. In particular we prove
\[mainProp\] Let $d>0$ be an integer and $\epsilon>0$ a real number, then there exists a $k=O_d\left( \epsilon^{-2^{O(d)}}\right)$, so that for any degree $d$ multilinear polynomial $p:{\mathbb{R}}^n\rightarrow {\mathbb{R}}$ and any $k$-independent family of Gaussians $X$ and fully independent family of Gaussians $Y$ $$\left| {\textrm{E}}[{\textrm{sgn}}(p(X))]-{\textrm{E}}[{\textrm{sgn}}(p(Y))]\right| = O(\epsilon).$$
We define the notation $A\approx_\epsilon B$ to mean $|A-B|=O(\epsilon)$.
The proof of Proposition \[mainProp\] will be analogous to the proof of the main Theorem in [@DKN]. Our basic idea is as follows.
In Section \[momentSec\] we prove bounds on the moments of multilinear Gaussian polynomials. These results are essentially a reworking of the main result of [@Moments].
In Section \[structureSec\], we use these bounds to prove a structure Theorem for multilinear polynomials. In particular, we prove that we can write $p(X)$ in the form $h(P_1(X),P_2(X),\ldots,P_N(X))$ where $h$ is a polynomial and $P_i(X)$ are multilinear polynomials with relatively small higher moments. More specifically, the polynomials $P_i$ will be split into $d$ different classes, with the $i^{th}$ class consisting of $n_i$ polynomials each of whose $m_i^{th}$ moments are $O_d(m_i)^{m_i/2}$. This decomposition allows us to write $f(X)={\textrm{sgn}}(P(X))$ as ${\textrm{sgn}}(h(P_1(X),\ldots,P_N(X)))$.
From here we make use of the FT-Mollification method (see [@DKN] for another example of this technique). The basic idea will be to approximate ${\textrm{sgn}}\circ h$ by some smooth function $\tilde h$, and let $\tilde f(X) = \tilde h(P_1(X),\ldots,P_N(X))$, which we do in Section \[FTMSec\]. Our general strategy now will be to prove the sequence of approximations: $${\textrm{E}}[f(Y)] \approx_\epsilon {\textrm{E}}[\tilde f(Y)] \approx_\epsilon {\textrm{E}}[\tilde f(X)] \approx_\epsilon {\textrm{E}}[f(X)].$$ The middle equality will be proved by approximation $\tilde f$ by one of it’s Taylor polynomials. This is a polynomial, and hence its expectation is preserved under limited independence. The Taylor error can again be bounded by a polynomial, which will have small expectation since the $P_i$ have small moments. We cover this in Section \[TaylorSec\].
The first approximation above holds roughly because $\tilde f$ approximates $f$ everywhere except near places where $f$ changes sign. The result will hold due to anti-concentration results for $p(Y)$. The last approximation similarly holds because of anti-concentration of $p(X)$. Although anticoncentration of the $k$-independent $X$ can be proven using the above techniques applied to some other function $g$ for which $\tilde g$ is an upper bound for $f$, we deal with the problem indirectly. In particular, we show that ${\textrm{E}}[f(X)]$ can be bounded on either side by ${\textrm{E}}[{\textrm{sgn}}(p(Y)+c)]+O(\epsilon)$ for $c$ a small constant, and use anticoncentration of $p(Y)$. We cover this in Section \[AntiConSeq\].
Our application of FT-Mollification is complicated by the fact that our moment bounds on the $P_j$ are not uniform in $j$. To deal with this, we will construct $\tilde h$ to have different degrees of smoothness in different directions, and the parameter $C_i$ will describe the amount of smoothness along the $i^{th}$ set of coordinates (corresponding the the $i^{th}$ class of the $P_j$). This forces us to come up with modified techniques for producing $\tilde h$ and dealing with the Taylor polynomial and Taylor error.
In Section \[ReductionSec\], we reduce the general case to the case of multilinear polynomials by approximating $p(X)$ by a multilinear polynomial in some larger number of variables.
Finally, in Section \[kIndepConc\], we discuss the actual requirements for $k$ and the possibility of extended our results to the Bernoulli setting.
Moment Bounds {#momentSec}
=============
In this Section, we prove a bound on the moments of arbitrary degree-$d$ multilinear polynomials of Gaussians. Our bound is based on the main result of [@Moments]. It should be noted that this result is the only reason that we restrict ourselves for most of this paper to the case of multilinear polynomials, as it will make our bound easier to state and work with.
Throughout this Section, we will refer to two slightly different notions that of a multilinear polynomial and that of a multilinear form. For our purposes, a multilinear polynomial $p(X)$ ($X$ has $n$ coordinates) will be a polynomial so that the degree of $p$ with respect to any of the coordinates of $X$ is at most 1. A multilinear form will be a polynomial $q(X^1,X^2,\ldots,X^m)$ (here each of the $X^i$ may themselves have several coordinates) so that $q$ is linear (homogeneous degree 1) in each of the $X^i$. We call such a $q$ symmetric if it is symmetric with respect to interchanging the $X^i$. Finally, we note that to every homogeneous multilinear polynomial $p$ of degree $d$, there is an associated multilinear form $q(X^1,\ldots,X^d)$, which is the unique symmetric multilinear form so that $p(X)=q(X,\ldots,X).$
Before we can state our results we need a few more definitions.
Let $p:{\mathbb{R}}^n\rightarrow {\mathbb{R}}$ be a homogeneous degree-$d$ multilinear polynomial. Let $X_i, 1\leq i\leq n$ be independent standard Gaussians. For a integers $1\leq \ell \leq d$ define $M_\ell(p)$ in the following way. Consider all possible choices of: a partition of $\{1,\ldots,n\}$ into sets $S_1,S_2,\ldots,S_\ell$; a sequence of integers $d_i \geq 1, 1\leq i\leq \ell$ so that $d=\sum_{i=1}^\ell d_i$; a sequence of multilinear polynomials $p_i, 1\leq i \leq \ell$ so that $p_i$ depends only on the coordinates in $S_i$, $p_i$ is homogeneous of degree $d_i$, and ${\textrm{E}}[p_i(X)^2]=1$. We let $M_\ell(p)$ be the supremum over all choices of $S_i,d_i,p_i$ as above of $$\left({\textrm{E}}\left[p(X)\prod_{i=1}^\ell p_i(X)\right]\right)^{1/2}.$$
Note that by Cauchy-Schwartz we have that $M_\ell(p)\leq {\textrm{E}}[p(X)^2]^{1/4}$. We now define a similar quantity more closely related to what is used in [@Moments].
Let $q:({\mathbb{R}}^n)^d\rightarrow {\mathbb{R}}$ by a degree-$d$ multilinear form. Let $X^i, 1\leq i \leq d$ be independent standard $n$-dimensional Gaussians. For integers $1\leq \ell \leq d$ define $M_\ell(q)$ in the following way. Consider all possible choices of: a partition of $\{1,\ldots,d\}$ into non-empty subsets $S_1,\ldots,S_\ell$, with $S_i = \{c_{i,1},\ldots,c_{i,d_i}\}$; and a set of multilinear forms $q_i$ of degree-$d_i$ with ${\textrm{E}}[q_i(X^{c_{i,1}},\ldots,X^{c_{i,d_i}})^2] \leq 1$. We define $M_\ell(q)$ to be the supremum over all such choices of $S_i$ and $q_i$ of $$\left({\textrm{E}}\left[q(X^1,\ldots,X^d)\prod_{i=1}^\ell q_i(X^{c_{i,1}},\ldots,X^{c_{i,d_i}})\right]\right)^{1/2}.$$
We now state the moment bound whose proof will take up the rest of this Section.
\[momentProp\] Let $p$ be a homogenous degree $d$ multilinear polynomial, and $X$ a family of independent standard Gaussians, and $k\geq 2$. Then $${\textrm{E}}[|p(X)|^k] = \Theta_d\left( \sum_{\ell=1}^d M_\ell(p) k^{\ell/2} \right)^k.$$
This is essentially a version of Theorem 1 of [@Moments]:
\[LatalaThm\] For $q$ a degree-$d$ multilinear form and $X^i$ independent standard $n$-dimensional Gaussians and $k$ an integer at least 2, $${\textrm{E}}[|q(X^1,\ldots,X^d)|^k] = \Theta_d\left( \sum_{\ell=1}^d M_\ell(q) k^{\ell/2} \right)^k.$$
The basic idea of the proof is the relate $M_\ell(p)$ to $M_\ell(q)$ and ${\textrm{E}}[|p|^k]$ to ${\textrm{E}}[|q|^k]$ for $q$ the symmetric multilinear form associated to a multilinear polynomial $p$.
Let $q$ be the associated symmetric multilinear form associated to $p$. We claim that for each $\ell$ that $M_\ell(p)=\Theta_d(M_\ell(q))$. Suppose that $p_1$ and $p_2$ are degree $d$ multilinear polynomials, and $q_1$ and $q_2$ the associated symmetric multilinear forms. It is easy to see (by using the standard basis of coefficients) that ${\textrm{E}}[p_1(X)p_2(X)] = d!{\textrm{E}}[q_1(X^1,\ldots,X^d)q_2(X^1,\ldots,X^d)].$ Similarly it is easy to see that if $p$ is a degree $d$ multilinear polynomial, and $p_i$ are degree $d_i$ multilinear polynomials on distinct sets of coordinates, and $q,q_i$ their associated symmetric multilinear forms we have $$\begin{aligned}
{\textrm{E}}& \left[ p(X)\prod p_i(X)\right] = \frac{1}{\prod d_i!}{\textrm{E}}\left[q(X)\prod q_i(X^{(i)}) \right].\end{aligned}$$ Where $q(X) = q(X^1,\ldots,X^d)$, and $q_i(X^{(i)}) = q_i(X^{d_1+\ldots+d_{i-1}+1},\ldots,X^{d_1+\ldots+d_{i-1}+d_i}).$ This means that $M_\ell(p) = O_d(M_\ell(q))$ since given the appropriate $S_i,d_i,p_i$ we can use the symmetrizations of the $p_i$ to get as good a bound for $M_\ell(q)$ up to a constant factor. To show the other direction we need to show that $M_\ell(q)$ is not changed by more than a constant factor if we require that the $q_i$ are supported on disjoint sets of coordinates. But we note that if you randomly assign each coordinate to a $q_i$ and take the part that only depends on those coordinates, you loose a factor of at most $d^d$ on average.
Hence we have that $${\textrm{E}}[|q(X^1,\ldots,X^d)|^k] = \Theta_d\left( \sum_{\ell=1}^d M_\ell(p) k^{\ell/2} \right)^k.$$ We just need to show that the moments of $p$ to the moments of $q$ are the same up to a factor of $\Theta_d(1)^k$. This can be shown using the main Theorem of [@tails] which in our case states that there is some constant $C_d$ depending only on $d$ so that for any such $p,q$ and $x$, $${\textrm{Pr}}(|p(X)|>x) \leq C_d{\textrm{Pr}}(|q(X^1,\ldots,X^d)|>x/C_d)$$ and $${\textrm{Pr}}(|q(X^1,\ldots,X^d)|>x) \leq C_d{\textrm{Pr}}(|p(X)|>x/C_d).$$ Our result follows from noting that for any random variable $Y$ that $${\textrm{E}}[|Y|^k] = \int_0^\infty kx^{k-1}{\textrm{Pr}}(|Y|>x)dx.$$
Structure {#structureSec}
=========
In this Section, we will prove the following structure theorem for degree-$d$ multilinear polynomials.
\[structureProp\] Let $p$ be a degree-$d$ multilinear polynomial where the sum of the squares of its coefficients is at most 1. Let $m_1\leq m_2\leq \ldots \leq m_d$ be integers. Then there exist integers $n_1,n_2,\ldots,n_d$, $n_i = O_d(m_1m_2\cdots m_{i-1})$ and non-constant, homogeneous multilinear polynomials $h_1,\ldots,h_d$, $P_{i,j}, 1\leq i\leq d, 1\leq j \leq n_i$ so that:
1. $h_i$ is degree $i$
2. If $P_{i,a_1}\cdots P_{i,a_i}$ appears as a term in $h_i(P_{i,j})$, then the sum of the degrees of the $P_{i,a_i}$ is $d$
3. The sum of the squares of the coefficients of $h_i$ is $O_d(1)$
4. The sum of the squares of the coefficients of $P_{i,j}$ is 1
5. Each variable occurs in at most one monomial in $h_i$
6. If $Y$ is a standard Gaussian and $k\leq m_i$ then ${\textrm{E}}[|P_{i,j}(Y)|^k] = O_d(\sqrt{k})^k$.
7. $p(Y) = \sum_{i=1}^d h_i(P_{i,1}(Y),P_{i,2}(Y),\ldots,P_{i,n_i}(Y)).$
This will allow us to write $p$ in terms of other polynomials each with smaller moments. The basic idea of the proof follows from a proper interpretation of Proposition \[momentProp\]. Essentially Proposition \[momentProp\] says that the higher moments of $p$ will be small unless $p$ has some significant component consisting of a product of polynomials $P_1,\ldots,P_\ell$ of lower degree. The basic idea is that if such polynomials exist, we can split off these $P_i$ as new polynomials in our decomposition, leaving $p-P_1\cdots P_\ell$ with smaller size than $p$. We repeatedly apply this procedure to $p$ and all of the other polynomials that show up in our decomposition. Since each step decreases the size of the polynomial being decomposed, and produces only new polynomials of smaller degree, this process will eventually terminate. Beyond these ideas, the proof consists largely of bookkeeping to ensure that we have the correct number of $P$’s and that they have an appropriate number of small moments.
We first prove our statement for homogeneous, multilinear polynomials $p$. We reduce the general case to this one by writing $p$ as a sum of its homogeneous parts and decomposing each of them.
We would like to simply use the decomposition $P_{1,1}=p$ and $h_1$ is the identity, but the moments of $p$ may be too large. On the other hand, we know by Proposition \[momentProp\] that this can only be the case if $p$ has large correlation with some product of smaller degree polynomials $P_1\cdots P_k$. So if $c={\textrm{E}}[p\cdot P_1\cdots P_k]$, we can write $p'=p-c P_1\cdots P_k$. Now either $p'$ has small moments or we can break off another product of polynomials. This process must eventually terminate because when we replaced $p$ by $p'$ we decreased the expectation of its square by $c^2$. We will then apply this technique recursively to each of the $P_i$.
We define a dot product on the space of multilinear polynomials ${\left\langle P,Q\right\rangle} = {\textrm{E}}[P(Y)Q(Y)]$ where $Y$ is a standard Gaussian. Note that the square of the corresponding norm is just $|P|^2$ equals the sum of the squares of the coefficients of $P$.
We begin by letting $q=p$. We note that by Proposition \[momentProp\] that the $k^{th}$ moment of $q$ for $k\leq m_1$ is $O_d(\sqrt k)^k$ unless for some $2\leq \ell \leq d$ we have that $M_\ell(q) \geq m_1^{m_1/2}/m_1^{\ell/2}$, or equivalently, unless there exist polynomials $P_1,\ldots,P_\ell$ of norm 1, so that $c={\left\langle q,P_1\cdots P_\ell\right\rangle} \geq m_1^{(1-\ell)/2}$. If this is the case, we replace $q$ by $q'=q-cP_1\cdots P_\ell$. Note that $|q'|^2 = |q|^2-c^2$. We repeat this process with $q'$ until finally we are left with a polynomial $q$ so that for all $k\leq m_1$ the $k^{th}$ moment of $q$ is $O_d(\sqrt k)^k$ (this process must terminate since at each step we decrease $|q|^2$ by at least $m_1^{1-d}$). We now can write $p$ as $q$ plus a sum of $c_i$ times products of lower degree polynomials. It should be noted that the sum of the squares of the $c_i$ is at most 1. Letting $P_{1,1}=q$ and $h_1$ be the identity, we can now write $$p(Y) = \sum_{i=1}^d h_i(P_{i,1}(Y),P_{i,2}(Y),\ldots,P_{i,n_i}(Y)).$$ Where $|h_i|=O_d(1)$, $|P_{i,j}| \leq 1$, $n_i = O_d(m_1^{i-1})$, and for $k\leq m_1$, the $k^{th}$ moment of $P_{1,j}$ is $O_d(\sqrt k)^k.$ Unfortunately, the moments of the other $P$’s might be too large. We show by induction on $s$ that we have such a decomposition where all of the appropriate moments of the $P_{i,j}$ for $i\leq s$ are bounded and so that $n_i = O_d(m_1m_2\cdots m_{i-1})$ for all $i$.
We have already proved the $s=1$ case. To prove the general case, we first write $p$ as $\sum_{i=1}^d h_i(P_{i,1}(Y),P_{i,2}(Y),\ldots,P_{i,n_i}(Y))$ using the induction hypothesis. This satisfies all of our criteria except that the $P_{s,j}$ might have moments which are too large. We fix this by rewriting each of the $P_{s,j}$ using the same method we originally used to rewrite $p$, only guaranteing that the first $m_s$ moments are small. This will make it so that our new $P_{s,j}$ have appropriately bounded moments, but may introduce new terms in the $h_t$ for $t>s$ (if some term shows up in multiple monomials, define several $P_{i,t}$ that are equal). We need to make sure that we did not introduce too many new terms and that the sum of the squares of the coefficients is not too large.
To show the latter note that our original procedure at most doubled the sum of the squares of the coefficients. Therefore applying this to each $P_i$ in a term $cP_1\cdots P_s$ will increase the sum of the squares of the coefficients by a factor of at most $2^s$. Hence since the sum of the squares of the coefficients was $O_d(1)$ before, it still is afterwards.
Finally we need to show that our new decomposition did not introduce too many new terms. It is not hard to see that for each $P_{s,i}$ we need to introduce $O_d(m_s^{t-s})$ new $P_{t,j}$ terms. Therefore the total number of such new terms is $O(m_s^{t-s}n_s) = O_d(m_1m_2\cdots m_{t-1})$.
Finally we note that our induction terminates at $s=d$. This is because the $P_{d,j}$ must be linear polynomials of bounded norm, and therefore automatically satisfy the necessary moment bounds. This completes our inductive step and proves the Proposition.
FT-Mollification {#FTMSec}
================
We let $F$ be a degree-$d$ polynomial threshold function $F={\textrm{sgn}}(p)$, where $p$ is a degree $d$ multilinear polynomial in $n$ variables whose sum of squares of coefficients equals 1. We pick $m_1,\ldots,m_d$ (their exact sizes will be determined later). For later convenience, we assume the $m_i$ are all even. We then have a decomposition of $F$ given by Proposition \[structureProp\] as $$\begin{aligned}
F(X) & = {\textrm{sgn}}\left(\sum_{i=1}^d h_i(P_{i,1}(X),\ldots,P_{i,n_i}(X)) \right)\\ & = f(P_1(X),P_2(X),\ldots,P_d(X)) \\ & = f(P(X))\end{aligned}$$ where $P_i(X)$ is the vector-valued polynomial $(P_{i,1}(X),\ldots,P_{i,n_i}(X))$, $P$ is the vector of all of them, and $f$ is the function $f(P_1,\ldots,P_d) = {\textrm{sgn}}(\sum h_i(P_i))$. Furthermore, we have that for $k\leq d m_i$ the $k^{th}$ moment of any coordinate of any coordinate of $P_i$ is $O_d(\sqrt{k})^k$. We also have that $h_i$ is a degree $i$ multilinear polynomial the sum of the squares of whose coefficients is at most 1.
Our basic strategy now will involve approximating $f$ by a smooth function $\tilde f$, and letting $\tilde F(X) = \tilde f(P(X))$. We will then proceed to prove $$\label{ApproxEqn}
{\textrm{E}}[F(Y)] \approx_\epsilon {\textrm{E}}[\tilde F(Y)] \approx_\epsilon {\textrm{E}}[\tilde F(X)] \approx_\epsilon {\textrm{E}}[F(X)].$$ We will produce $\tilde f$ from $f$ using the technique of mollification. Namely we will have $\tilde f = f * \rho$ for an appropriately chosen smooth function $\rho$. However, we will need this $\rho$ to have several other properties so we will go into some depth here to construct it.
\[rhoLem\] Given an integer $n\geq 0$ and a constant $C$, there is a function $\rho_C:{\mathbb{R}}^n\rightarrow {\mathbb{R}}$ so that
1. $\rho_C \geq 0$.
2. $\int_{{\mathbb{R}}^n} \rho_C(x)dx = 1$.
3. For any unit vector $v\in {\mathbb{R}}^n$, and any non-negative integer $k$, $\int_{{\mathbb{R}}^n} |D_v^k \rho_C(x)|dx \leq C^k$, where $D_v^k$ is the $k^{th}$ directional derivative in the direction $v$.
4. For $D>0$, $\int_{|x|>D} |\rho(x)|dx = O\left(\left(\frac{n}{CD}\right)^2\right)$.
We prove this for $C=2$ and we note that we can obtain other values of $C$ by setting $\rho_C(x) = (C/2)^n \rho_2(Cx/2).$ We begin by defining $$B(\xi) = \begin{cases} 1 - |\xi|^2 \ \ & \textrm{if} \ |\xi|\leq 1\\ 0 \ \ & \textrm{else} \end{cases}$$ We then define $$\rho_2(x) = \rho(x) = \frac{|\hat{B}(x)|^2}{|B|_2^2}.$$ Where $\hat B$ denotes the Fourier transform of $B$. Clearly $\rho$ is non-negative. Also clearly $$\int_{{\mathbb{R}}^n} \rho(x)dx = \frac{|\hat B|_2^2}{|B|_2^2} = 1$$ by the Plancherel Theorem.
For the third property we note that $$D_v^k \rho = \frac{1}{|B|_2^2} \sum_{i=0}^k \binom{k}{i} D_v^i(\hat B) D_v^{k-i} \overline{(\hat B)}.$$
Letting $\xi$ be the dual vector corresponding to $v$ we have that $$\begin{aligned}
\left| D_v^k \rho\right|_1 & \leq \frac{1}{|B|_2^2}\sum_{i=0}^k \binom ki \left|D_v^i(\hat B) D_v^{k-i} \overline{(\hat B)}\right|_1\\
& \leq \frac{1}{|B|_2^2}\sum_{i=0}^k \binom ki \left|D_v^i(\hat B)\right|_2\left|D_v^{k-i}(\hat B)\right|_2\\
& \leq \frac{1}{|B|_2^2}\sum_{i=0}^k \binom ki \left|\xi^i B\right|_2\left|\xi^{k-i} B \right|_2\\
& \leq \frac{1}{|B|_2^2}\sum_{i=0}^k \binom ki |B|_2^2\\
& = \sum_{i=0}^k \binom ki \\
&= 2^k.\end{aligned}$$ For the last property we note that it is enough to prove that $$\int_{{\mathbb{R}}^n} |x|^2\rho(x) dx = O(n^2).$$ We have that $$\begin{aligned}
\int_{{\mathbb{R}}^n} |x|^2\rho(x) dx & = \frac{1}{|B|_2^2} \sum_{i=1}^n |x_i\hat B|_2^2\\
& = \frac{\sum_{i=1}^n\left| \frac{\partial B}{\partial \xi_i}\right|_2^2}{|B|_2^2}.\end{aligned}$$ Now $ \frac{\partial B}{\partial \xi_i} $ is $2\xi_i$ on the unit ball and 0 outside. Hence the sum of the squares of these is $2|\xi|^2$ on $|\xi|<1$ and 0 outside. Hence since both numerator and denominator above are integrals of spherically symmetric functions, their ratio is equal to $$\begin{aligned}
\int_{{\mathbb{R}}^n} |x|^2\rho(x) dx & = \frac{2\int_0^1 r^{n+1} dr}{\int_0^1 r^{n-1}(1-r^2)^2 dr}.\end{aligned}$$ Using integration by parts, the denominator is $$\begin{aligned}
\int_0^1 r^{n-1}(1-r^2)^2 dr & = \frac{4}{n}\int_0^1 r^{n+1}(1-r^2)dr \\ & = \frac{16}{n(n+2)}\int_0^1 r^{n+3}dr \\ & = \frac{16}{n(n+2)(n+4)}.\end{aligned}$$ Hence $$\int_{{\mathbb{R}}^n} |x|^2\rho(x) dx = \frac{n(n+4)}{8} = O(n^2).$$
We are now prepared to define $\tilde f$. We pick constants $C_1,\ldots,C_d$ (to be determined later). We let $$\label{rhoDefEqn}\rho(P_1,\ldots,P_d) = \rho_{C_1}(P_1)\cdot \rho_{C_2}(P_2)\cdots \rho_{C_d}(P_d).$$ Above the $\rho_{C_i}$ is defined on ${\mathbb{R}}^{n_i}$. We let $\tilde f$ be the convolution $\tilde f = f * \rho$.
Taylor Error {#TaylorSec}
============
In this Section, we prove the middle approximation of Equation \[ApproxEqn\] for appropriately large $k$. The basic idea will be to approximate $\tilde f$ by its Taylor series, $T$. $T(P(X))$ will be a polynomial of degree at most $k$ and hence ${\textrm{E}}[T(P(Y))]={\textrm{E}}[T(P(X))].$ Furthermore, we will bound the Taylor error by some polynomial $R$ and show that ${\textrm{E}}[R(P(Y))]={\textrm{E}}[R(P(X))]$ is $O(\epsilon)$ for appropriate choices of $m_i,C_i$. In particular, we let $T$ be the polynomial consisting of all of the terms of the Taylor expansion of $\tilde f$ whose total degree in the $P_i$ coordinates is less than $m_i$ for all $i$. Note that a polynomial of this form is about the best we can do since we only have control over the size of moments up to the $m_i^{th}$ moment on the $i^{th}$ block of coordinates. Our error bound will be the following
\[taylorErrorProp\] $$|T(P) - \tilde f(P)| \leq \prod_{i=1}^d \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right)-1.$$
First we prove a Lemma dealing with Taylor error for a single batch of coordinates,
\[basicTaylorLem\] If $g$ is a multivariate function, $\tilde g= g * \rho_C$ and $T$ is the polynomial consisting of all terms in the Taylor expansion of $\tilde g$ is degree less than $m$, then $$|g(x)-T(x)| \leq \frac{|g|_\infty C^m |x|^m}{m!}.$$
Let $v$ be the unit vector in the direction of $x$. Let $L$ be the line through $0$ and $x$. We note that the restriction of $T$ to $L$ is the same as the first $m-1$ terms of the Taylor series for $\tilde g|_L$. Using standard error bounds for Taylor polynomials we find that $$|g(x)-T(x)| \leq \frac{|D_v^m \tilde g|_\infty |x|^m}{m!}.$$ But $$\begin{aligned}
|D_v^m \tilde g|_\infty & = |g * D_v^m \rho_C |_\infty\\
& \leq |g|_\infty |D_v^m \rho_C|_1\\
& \leq |g|_\infty C^m.\end{aligned}$$ Plugging this in yields our result.
The basic idea of the proof will be to repeatedly apply Lemma \[basicTaylorLem\] to one batch of coordinates at a time. We begin by defining some operators on the space of bounded functions on ${\mathbb{R}}^{n_1}\times{\mathbb{R}}^{n_2}\times\cdots\times{\mathbb{R}}^{n_d}$. For such $g$, define $g^{\tilde i}$ to be the convolution of $g$ with $\rho_{C_i}$ along the $i^{th}$ set of coordinates. Define $g^{T_i}$ to be the Taylor polynomial in the $i^{th}$ set of variables of $g^{\tilde i}$ obtained by taking all terms of total degree less than $m_i$. Note that for $i\neq j$ the operations $\tilde i$ and $T_i$ commute with the operations $\tilde j$ and $T_j$ since they operate on disjoint sets of coordinates. Note that $\tilde f = f^{\tilde 1 \tilde 2\cdots \tilde d}$ and $T=f^{T_1 T_2 \cdots T_d}$. For $1\leq i \leq d$ let $f_i = f^{\tilde 1 \tilde 2\cdots \tilde i}$ and $T_i = f^{T_1 T_2 \cdots T_i}$.
We prove by induction on $s$ that $$|T_s(P)-f_s(P)| \leq \prod_{i=1}^s \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right)-1.$$ As a base case, we note that the $s=0$ case of this is trivial.
Assume that $$|T_s(P)-f_s(P)| \leq \prod_{i=1}^s \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right)-1.$$ We have that $$\begin{aligned}
|T_{s+1}&(P)- f_{s+1}(P)|\\ & \leq |T_s^{T_{s+1}}(P)-T_s^{\tilde{s+1}}(P)| + |T_s^{\tilde{s+1}}(P)-f_{s}^{\tilde{s+1}}(P)|.\end{aligned}$$ Note that $$T_s^{\tilde{s+1}}(P)-f_{s}^{\tilde{s+1}}(P) = \left(T_s - f_s\right)^{\tilde{s+1}}(P).$$ Therefore since $\tilde{s+1}$ involves only convolution with a function of $L^1$ norm 1 we have that $$|T_s^{\tilde{s+1}}(P)-f_{s}^{\tilde{s+1}}(P)| \leq |T_s(P)-f_s(P)|_{\infty,s+1}$$ where the subscript denotes the $L^\infty$ norm over just the $s+1^{st}$ set of coordinates. By the inductive hypothesis, this is at most $$\prod_{i=1}^s \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right)-1.$$
On the other hand, applying Lemma \[basicTaylorLem\] we have that $$|T_s^{T_{s+1}}(P)-T_s^{\tilde{s+1}}(P)| \leq \frac{C_{s+1}^{m_{s+1}}|P_{s+1}|^{m_{s+1}}}{m_{s+1}!}\left|T_s \right|_{\infty,s+1}.$$ By the inductive hypothesis, $$\begin{aligned}
\left|T_s \right|_{\infty,s+1} & \leq |f_s|_{\infty,s+1} + |T_s-f_s|_{\infty,s+1} \\ & \leq \prod_{i=1}^s \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right).\end{aligned}$$
Combining the above bounds, we find that $$\begin{aligned}
|T_{s+1}-f_{s+1}| \leq & \prod_{i=1}^s \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right) - 1 \\ & + \frac{C_{s+1}^{m_{s+1}}|P_{s+1}|^{m_{s+1}}}{m_{s+1}!}\prod_{i=1}^s \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right)\\
= & \prod_{i=1}^{s+1} \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right)-1.\end{aligned}$$
We can now prove the desired approximation result
\[polyapproxErrorProp\] If $\tilde F,P,T$ as above with $m_i= \Omega_d(n_i C_i^2)$, $m_i \geq \log(2^d/\epsilon)$ for all $i$, and if $k\geq dm_i$ for all $i$, then for $X$ and $Y$ are $k$-independent families of standard Gaussians, $${\textrm{E}}[\tilde F(Y)] \approx_\epsilon {\textrm{E}}[\tilde F(X)].$$
We note that since $T\circ P$ is a polynomial of degree at most $k$ we have that ${\textrm{E}}[T(P(X))] = {\textrm{E}}[T(P(Y))].$ Hence, it suffices to show that $${\textrm{E}}[|F-T|(P(X))],{\textrm{E}}[|F-T|(P(Y))] = O(\epsilon).$$ We will show this only for $X$ as $Y$ is analogous. By Proposition \[taylorErrorProp\] we have that $|F-T|$ is bounded by $$\prod_{i=1}^d \left(1+\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right)-1.$$ This is a sum over non-empty subsets $S\subseteq\{1,2,\ldots,d\}$ of $$\prod_{i\in S} \left(\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right).$$ Since there are only $2^d-1$ such $S$, it is enough to show that each term individually has expectation $O(\epsilon/2^d)$. On the other hand, we have by AM-GM that each term is at most $$\frac{1}{|S|}\sum_{i\in S} \left(\frac{C_i^{m_i}|P_i|^{m_i}}{m_i!} \right)^{|S|}.$$ Now the expectation of $|P_i|^{m_i|S|}$ is at most $n_i^{m_i|S|}$ times the average of the $m_i|S|^{th}$ moments of the coordinates of $P_i$. These by assumption are $O_d(\sqrt{m_i|S|})^{m_i|S|}$. There are $n_i$ coordinates so the moment of $|P_i|$ is at most $O_d(\sqrt{n_im_i|S|})^{m_i|S|}$. Hence the error is at most $$\begin{aligned}
O(2^d)\max_{i,s}&\left\{ O_d\left( \frac{C_i\sqrt{n_i m_i s}}{m_i}\right)^{m_i s} \right\}\\
& = O(2^d)\max_{i,s}\left\{ O_d\left( \frac{C_i\sqrt{n_i}}{\sqrt{m_i}}\right)^{m_i s} \right\}\\
& \leq O(2^d) e^{-\min_i m_i} = O(\epsilon).\end{aligned}$$
Approximation Error {#AntiConSeq}
===================
In this Section, we will prove the first and third approximations in Equation \[ApproxEqn\]. We begin with the first, namely $${\textrm{E}}[F(Y)] \approx_\epsilon {\textrm{E}}[\tilde F(Y)].$$ Our basic strategy will be to bound $$|{\textrm{E}}[F(Y)] - {\textrm{E}}[\tilde F(Y)] \leq {\textrm{E}}[|F(Y)-\tilde F(Y)|].$$ In order to get a bound on this we will first show that $F-\tilde F$ is small except where $p(Y)$ is small, and then use anti-concentration results to show that this happens with small probability. This will be true because $\rho$ is small away from $0$. We begin by proving a Lemma to this effect.
\[rhoConcentrationLem\] Let $\rho$ be the function defined in Equation \[rhoDefEqn\]. Then for any $D>0$ we have that $$\int_{\substack{(x_1,\ldots,x_d)\in {\mathbb{R}}^{n_1}\times\cdots\times{\mathbb{R}}^{n_d} \\ \exists i: |x_i| > Dn_i\sqrt{d}/C_i}} |\rho(x)|dx = O(D^{-2})$$
This will hold essentially because of the concentration property held by each $\rho_{C_i}$.
We integrate over the region where $|x_i| > Dn_i\sqrt{d}/C_i$ for each $i$. This is a product over $j\neq i$ of $\int_{{\mathbb{R}}^{n_j}} \rho_{C_j}(x_j)$ times $\int_{|x|>Dn_i\sqrt{d}/C_i} |\rho(x)|dx$. By Lemma \[rhoConcentrationLem\] the former integrals are all 1, and the latter is $O(D^{-2}/d)$. Summing over all possible $i$ yields $O(D^{-2})$.
Recall that $f$ was ${\textrm{sgn}}\circ h$, where $h=\sum h_i$ given in the decomposition of $p$ from Proposition \[structureProp\]. Recall that $\tilde f = f * \rho$. We want to bound the error in approximating $f$ by $\tilde f$. The following, is a direct consequence of Lemma \[rhoConcentrationLem\].
\[smoothErrorBoundLem\] Suppose $x=(x_1,\ldots,x_d)\in {\mathbb{R}}^{n_1}\times\cdots\times{\mathbb{R}}^{n_d}$. Suppose also that for some $D>0$ and for all $y=(y_1,\ldots,y_d)\in{\mathbb{R}}^{n_1}\times\cdots\times{\mathbb{R}}^{n_d}$ so that $|x_i-y_i|\leq Dn_i\sqrt{d}/C_i$ that $h(x)$ and $h(y)$ have the same sign, then $$|f(x)-\tilde f(x)| = O\left(\min\{ 1 , D^{-2} \}\right).$$
To show that the error is $O(1)$, we note that since $\rho\geq 0$ and $\int \rho(x)dx=1$ that $\tilde f(x) = (f * \rho)(x) \in [\inf(f),\sup(f)] \subseteq [-1,1]$. Therefore $|f-\tilde f| \leq |f| + |\tilde f| \leq 2.$
For the latter, we note that $\tilde f(x) = \int_y f(y)\rho(x-y) dy$. We note that since the total integral of $\rho$ is 1 that $$f(x) - \tilde f (x) = \int_y (f(x)-f(y))\rho(x-y) dy.$$ We note that by assumption unless $|x_i-y_i| > Dn_i\sqrt{d}/C_i$ for some $i$ that the integrand is 0. But outside of this, the integrand is at most $2\rho(x-y)$. By Lemma \[rhoConcentrationLem\] the total integral of this is $O(D^{-2})$.
We now know that $f$ is near $\tilde f$ at points $x$ not near the boundary between the $+1$ and $-1$ regions. Since we cannot directly control the size of these regions, we want to relate this to the region where $|h(x)|$ is small. This should work since unless $x$ is very large, $h$ will have derivatives that aren’t too big. In particular, we prove the following.
\[near0boundLem\] Let $x\in {\mathbb{R}}^n$. Suppose that we have $B_i\geq 0$ so that $|P_{i,j}(x)| \leq B_i$ for all $i,j$. We have that $|F(x) - \tilde F(x)|$ is at most the minimum of $O(1)$ and $$\begin{aligned}
&O_d\left(\max\left\{\left(\frac{|p(x)|}{\sum_{i=1}^d n_i^2 B_i^{i-1}/C_i} \right)^{-2} ,\left(\frac{B_iC_i}{n_i}\right)^{-2} \right\} \right).\end{aligned}$$
The bound of $O(1)$ follows immediately from Lemma \[smoothErrorBoundLem\]. For the other bound, let $$D = \min\left\{\frac{|p(x)|}{d2^d\sum_{i=1}^d n_i^2 B_i^{i-1}/C_i},\min_i\left\{\frac{B_iC_i}{n_i\sqrt{d}}\right\} \right\}.$$ By Lemma \[smoothErrorBoundLem\], it suffices to show that for any $Q=(Q_1,\ldots,Q_n)\in{\mathbb{R}}^{n_1}\times\cdots\times{\mathbb{R}}^{n_d}$ so that $|Q_i-P_i(x)|\leq Dn_i\sqrt{d}/C_i$ that $h(P(x))=p(x)$ and $h(Q)$ have the same sign. To do this, we write $h=h_1+\cdots+h_d$ and we note that $$|h(P(x))-h(Q)| \leq \sum_{i=1}^d |P_i(x)-Q_i||h_i'(z)|.$$ Where $h_i'(z)$ is the directional derivative of $h_i$ in the direction from $P_i(x)$ to $Q_i$, and $z$ is some point along this line. First, note that $|Q_i-P_i(x)|\leq B_i$. Therefore, each coordinate of $z$ is at most $2B_i$. Note that $h_i$ is a sum of at most $n_i$ monomials of degree $i$ with coefficients at most 1. The derivative of each monomial at $z$ is at most $\sqrt{d}2^d B_i^{i-1}$. Therefore, $|h_i'(z)|\leq \sqrt{d}2^d n_i B_i^{i-1}$. Therefore, $$\begin{aligned}
|h(P(x))-h(Q)| & \leq \sum_{i=1}^d |P_i(x)-Q_i||h_i'(z)|\\
& \leq \sum_{i=1}^d (D n_i \sqrt{d} / C_i )(\sqrt{d}2^d n_i B_i^{i-1})\\
& \leq D \sum_{i=1}^d d2^d n_i^2 B_i^{i-1}/C_i\\
& \leq |h(P(x))|.\end{aligned}$$ Therefore $h(P(x))$ and $h(Q)$ have the same sign, so our bound follows by Lemma \[smoothErrorBoundLem\].
We take this bound on the approximation error and prove the following Lemma on the error of expectations.
\[expectationErrorLem\] Let $Z$ be a random variable valued in ${\mathbb{R}}^n$. Let $B_i>1$ be real numbers. Let $M=\sum_{i=1}^d n_i^2 B_i^{i-1}/C_i$. Then $$\begin{aligned}
& |{\textrm{E}}[F(Z)] - {\textrm{E}}[\tilde F(Z)]| =\\
& O_d( {\textrm{Pr}}(\exists i,j:|P_{i,j}(Z)| > B_i) + M + {\textrm{Pr}}(|p(Z)|\leq \sqrt{M})).\end{aligned}$$ Furthermore, $$\begin{aligned}
{\textrm{E}}[F(Z)] \leq & {\textrm{E}}[\tilde F(Z)] \\ & + O_d( {\textrm{Pr}}(\exists i,j :|P_{i,j}(Z)| > B_i) + M) \\ & + 2{\textrm{Pr}}(-\sqrt{M}<p(Z)<0),\end{aligned}$$ and $$\begin{aligned}
{\textrm{E}}[F(Z)] \geq & {\textrm{E}}[\tilde F(Z)] \\ & + O_d( {\textrm{Pr}}(\exists i,j:|P_{i,j}(Z)| > B_i) + M)\\ & - 2{\textrm{Pr}}(0<p(Z)<\sqrt{M}).\end{aligned}$$
We note that $|F(Z) - \tilde F(Z)| = O(1)$. Also note that $\frac{1}{M} \leq \frac{B_iC_i}{n_i}$ for all $i$. The first inequality follows by noting that Lemma \[near0boundLem\] implies that unless $|P_{i,j}(Z)|>B_i$ for some $i,j$ that the following hold:
1. If $|p(z)|< \sqrt{M}$, $|F(Z) - \tilde F(Z)| \leq 2$.
2. If $|p(z)| \geq \sqrt{M}$, $|F(Z) - \tilde F(Z)| = O_d(M)$.
The other two inequalities follow from noting that if $p(z)<0$, then $F(Z)\leq\tilde F(Z)$ and if $p(Z)>0$ then $F(Z)\geq \tilde F(Z)$.
We are almost ready to prove the first of our approximation results, but we first need a theorem on the anticoncentration of Gaussian polynomials. In particular a consequence of [@anticoncentration] Theorem 8 is:
\[anticoncentrationTheorem\] Let $p$ be a degree $d$ polynomial, and $Y$ a standard Gaussian. Suppose that ${\textrm{E}}[p(Y)^2]=1$. Then, for $\epsilon>0$, $${\textrm{Pr}}(|p(Y)|<\epsilon) = O(d \epsilon^{1/d}).$$
We are now prepared to prove our approximation result.
\[yanitconcentrationErrorProp\] Let $p,F,\tilde F,h,m_i,n_i,C_i$ be as above and let $\epsilon>0$. Let $B_i = \Omega_d (\sqrt{\log(n_i/\epsilon)})$ be some real numbers. Suppose that $m_i > B_i^2$ and that $C_i = \Omega_d (n_i^2 B_i^{i-1} \epsilon^{-2d})$ for all $i$. Then, if the implied constants for the bounds on $B_i$ and $C_i$ are large enough, $$|{\textrm{E}}[F(Y)] - {\textrm{E}}[\tilde F(Y)]| = O(\epsilon).$$
We bound the error using Lemma \[expectationErrorLem\]. We note that the probability that $|P_{i,j}(Y)|\geq B_i$ can be bounded by looking at the $\log(d n_i/\epsilon)=k^{th}$ moment, yielding a probability of $\frac{O_d(\sqrt{k})^k}{B_i^k} \leq e^{-k} = \frac{\epsilon}{dn_i}$. Taking a union bound over all $j$ gives a probability of $\frac{\epsilon}{d}$. Taking a union bound over $i$ yields a probability of at most $\epsilon$.
Next we note that $$M = \sum_{i=1}^d n_i^2 B_i^{i-1}/ C_i = O_d(\epsilon^{2d}).$$ Hence if our constants were chosen to be large enough, by Theorem \[anticoncentrationTheorem\] $${\textrm{Pr}}(|p(Y)|<\sqrt{M}) = O(\epsilon).$$ This proves our result.
If we could prove Proposition \[yanitconcentrationErrorProp\] for $X$ instead of $Y$, we would be done. Unfortunately, Theorem \[anticoncentrationTheorem\] does not immediately apply for families that are merely $k$-independent. Fortunately, we can work around this to prove Proposition \[mainProp\]. In particular, we will use the inequality versions of Lemma \[expectationErrorLem\] to obtain upper and lower bounds on ${\textrm{E}}[F(X)]$ in terms of ${\textrm{E}}[{\textrm{sgn}}(p(Y)+c)]$, and make use of anticoncentration for $p(Y)$.
Let $B_i = \Omega_d(\sqrt{\log(1/\epsilon)})$ with sufficiently large constants. Define $m_i$ and $C_i$ so that $C_i = \Omega_d\left(\left(\prod_{j=1}^{i-1}m_j\right)^2 B_i^{i-1} \epsilon^{-2d}\right)$ and $m_i \geq \Omega_d\left(\left(\prod_{j=1}^{i-1}m_j\right)C_i^2\right),\log(2^d/\epsilon),B_i^2$, all with sufficiently large constants. Note that this is achievable by setting $C_i = \Omega_d\left( \epsilon^{-7^i d}\right)$, $m_i = \Omega_d\left(\epsilon^{-3\cdot7^i d} \right)$. Let $k=d\max_i m_i$. Note $k$ can be as small as $O_d(\epsilon^{-4d\cdot 7^d})$. Using these parameters, define $n_i,h_i,P_{i,j},f,\tilde f,\tilde F$ as described above. Note that since $n_i = O_d\left(\prod_{j=1}^{i-1} m_i\right)$ that $C_i = \Omega_d(n_i^2 B_i^{i-1} \epsilon^{-2d})$ and $m_i = \Omega_d (n_i C_i^2)$. Therefore, for $Y$ a family of independent standard Gaussians and $X$ a family of $k$-independent standard Gaussians, Propositions \[taylorErrorProp\] and \[yanitconcentrationErrorProp\] imply that $${\textrm{E}}[F(Y)] \approx_\epsilon {\textrm{E}}[\tilde F(Y)] \approx_\epsilon {\textrm{E}}[\tilde F(X)].$$ We note that the $M$ in Lemma \[expectationErrorLem\] is $O_d(\epsilon^{2d})$ with sufficiently small constant. Therefore, by Lemma \[expectationErrorLem\] $|{\textrm{E}}[F(X)] - {\textrm{E}}[\tilde F(X)]|$ is at most $$\begin{aligned}
O(\epsilon) + 2{\textrm{Pr}}(|p(X)|<O_d(\epsilon^{d})) + {\textrm{Pr}}(\exists i,j:|P_{i,j}(X)| > B_i ).\end{aligned}$$ We note that by looking at the $\log(d n_i/\epsilon)$ moments of the $P_{i,j}$ that the last probability is $O(\epsilon)$. Therefore, combining this with the above we get that $${\textrm{E}}[F(X)] \geq {\textrm{E}}[F(Y)]+O(\epsilon) - 2{\textrm{Pr}}(0<p(X)<O_d(\epsilon^d)),$$ and $${\textrm{E}}[F(X)] \leq {\textrm{E}}[F(Y)]+O(\epsilon) + 2{\textrm{Pr}}(-O_d(\epsilon^d)<p(X)<0).$$ But this implies that $${\textrm{E}}[{\textrm{sgn}}(p(X)-O_d(\epsilon^d))] \leq {\textrm{E}}[F(Y)] +O(\epsilon),$$ and $${\textrm{E}}[{\textrm{sgn}}(p(X)+O_d(\epsilon^d))] \geq {\textrm{E}}[F(Y)] +O(\epsilon).$$ On the other hand, applying to above to the polynomials $p\pm O_d(\epsilon^d)$, $$\begin{aligned}
{\textrm{E}}[{\textrm{sgn}}(p(Y)-O_d(\epsilon^d))] & +O(\epsilon) \leq {\textrm{E}}[F(X)] \\ & \leq {\textrm{E}}[{\textrm{sgn}}(p(Y)+O_d(\epsilon^d))]+O(\epsilon).\end{aligned}$$ But we have that $${\textrm{E}}[{\textrm{sgn}}(p(Y)-O_d(\epsilon^d))]\leq {\textrm{E}}[F(Y)] \leq {\textrm{E}}[{\textrm{sgn}}(p(Y)+O_d(\epsilon^d))].$$ Furthermore, ${\textrm{sgn}}(p(Y)-O_d(\epsilon^d))$ and ${\textrm{sgn}}(p(Y)+O_d(\epsilon^d))$ differ by at most 2, and only when $|p(Y)| = O_d(\epsilon^d)$. By Theorem \[anticoncentrationTheorem\], this happens with probability $O_d(\epsilon)$. Therefore, we have that all of the expectations above are within $O_d(\epsilon)$ of ${\textrm{E}}[F(Y)]$, and hence ${\textrm{E}}[F(X)] = {\textrm{E}}[F(Y)] + O_d(\epsilon)$. Decreasing the value of $\epsilon$ by a factor depending only on $d$ (and increasing $k$ by a corresponding factor) yields our result.
General Polynomials {#ReductionSec}
===================
We have proved our Theorem for multilinear polynomials, but would like to extend it to general polynomials. Our basic idea will be to show that a general polynomial is approximated by a multilinear polynomial in perhaps more variables.
\[multilinearLem\] Let $p$ be a degree $d$ polynomial and $\delta>0$. Then there exists a multilinear degree $d$ polynomial $p_\delta$ (in perhaps a greater number of variables) so that for every $k$-independent family of random Gaussians $X$, there is a (correlated) $k$-independent family of random Gaussians $\tilde X$ so that $${\textrm{Pr}}(|p(X)-p_\delta(\tilde X)|>\delta) < \delta.$$
We will pick some large integer $N$ (how large we will say later). If $X=(X_1,\ldots,X_n)$, we let $\tilde X = (X_{i,j}), 1\leq i\leq n, 1\leq j\leq N$. For fixed $i$ we let the collection of $X_{i,j}$ be the standard collection of $N$ standard Gaussians subject to the condition that $X_i = \frac{1}{\sqrt{N}}\sum_{j=1}^N X_{i,j}$. Equivalently, $X_{i,j} = \frac{1}{\sqrt{N}}X_i + Y_{i,j}$ where the $Y_{i,j}$ are Gaussians with variance $1-1/N$ and covariance $-1/N$ with each other.
$\tilde X$ is $k$-independent because given any $i_1,\ldots,i_k$, $j_1,\ldots,j_k$ we can obtain the $X_{i_\ell,j_\ell}$ by first picking the $X_{i_\ell}$ randomly and independently, and picking the $Y_{i_\ell,j_\ell}$ independently of those. But we note that this yields the same distribution we would get by setting all of the $X_{i_\ell,k}$ to be random independent Gaussians, and letting $X_i=\frac{1}{\sqrt{N}}\sum_{j=1}^N X_{i,j}$.
We now need to construct $p_\delta$ with the appropriate property. The idea will be to replace each term $X_i^k$ in each monomial in $p$ with some degree $k$ polynomial in the $X_{i,j}$. This will yield a multilinear degree $d$ polynomial in $\tilde X$. We will want this new polynomial to be within $\delta'$ of $X_i^k$ with probability $1-\delta'$ for $\delta'$ some small positive number depending on $p$ and $\delta$. This will be enough since if $\delta'<\delta/(2dn)$ the approximation will hold for all $i,k$ with probability at least $1-\delta/2$. Furthermore with probability $1-\delta/2$, each of the $|X_i|$ will be at most $O(\log(n/\delta))$. Therefore if this holds and each of the replacement polynomials is off by at most $\delta'$, then the value of the full polynomial will be off by at most $O(\log^d(n/\delta)\delta')$ times the sum of the coefficients of $p$. Hence if we can achieve this for $\delta'$ small enough we are done.
Hence, we have reduced our problem to the case of $p(X) = p(X_1) = X_1^d$. For simplicity of notation, we use $X$ instead of $X_1$ and $X_j$ instead of $X_{1,j}$. We note that $$X^d = N^{-d/2}\left(\sum_{i=1}^N X_i\right)^d.$$ Unfortunately, this is not a multilinear polynomial in the $X_i$. Fortunately, it almost is. Expanding it out and grouping terms based on the multiset of exponents occurring in them we find that $$X^d = N^{-d/2}\sum_{\substack{a_1\leq \ldots \leq a_k \\ \sum a_i=d}} \binom{d}{a_1,a_2,\ldots,a_k} \sum_{\mathcal{S}} \prod_{j=1}^k X_{i_j}^{a_j}.$$ Where $S$ is the set of $i_1,\ldots,i_k\in\{1,\ldots,N\}$ distinct so that $i_j < i_{j+1}$ if $a_j = a_{j+1}$. Letting $b_\ell$ be the number of $a_i$ that are equal to $\ell$ we find that this is $$N^{-d/2}\sum_{\substack{a_1\leq \ldots \leq a_k \\ \sum a_i=d}} \binom{d}{a_1,a_2,\ldots,a_k} \prod_\ell \frac{1}{b_\ell!}\sum_{\substack{i_1,\ldots,i_k\in [N] \\ i_j \ \textrm{ distinct}}} \prod_{j=1}^k X_{i_j}^{a_j}.$$ Or rewriting slightly, this is $$\sum_{\substack{a_1\leq \ldots \leq a_k \\ \sum a_i=d}} \binom{d}{a_1,a_2,\ldots,a_k} \prod_\ell \frac{1}{b_\ell!}\sum_{\substack{i_1,\ldots,i_k\in [N] \\ i_j \ \textrm{ distinct}}} \prod_{j=1}^k \left(\frac{X_{i_j}}{\sqrt{N}}\right)^{a_j}.$$ Now, with probability $1-\delta$, $\left|\sum_i \frac{X_i}{\sqrt{N}} \right| = O(\log(1/\delta))$. Furthermore with probability tending to 1 as $N$ goes to infinity, $\left(\sum_i \left(\frac{X_i}{\sqrt{N}}\right)^2 \right) = 1 + O(\delta / \log^d(1/\delta)),$ and $\left(\sum_i \left(\frac{X_i}{\sqrt{N}}\right)^a \right) = O(\delta / \log^d(1/\delta))$ for each $3\leq a \leq d$. If all of these events hold, then each term in the above with some $a_j>2$ will be $O(\delta)$, and any terms with some $a_j=2$ will be within $O(\delta)$ of $$\sum_{\substack{i_1,\ldots,i_k'\in \{1,\ldots,N\} \\ i_j \ \textrm{ distinct}}} \prod_{j=1}^{k'} \frac{X_{i_j}}{\sqrt{N}}$$ where $k'$ is the largest $j$ so that $a_j=1$. This gives a multilinear polynomial, that with probability $1-\delta$ is within $O_d(\delta)$ of $p(X)$. Perhaps decreasing $\delta$ to deal with the constant in the $O_d$ yields our result.
We can now prove Theorem \[mainTheorem\].
Let $p$ be a normalized degree $d$ polynomial. Let $k$ be as required by Proposition \[mainProp\]. Let $Y$ be a family of independent standard Gaussians and $X$ a $k$-independent family of standard Gaussians. Fix $\delta=(\epsilon/d)^d$. Let $p_\delta,\tilde X, \tilde Y$ be as given by Lemma \[multilinearLem\]. We need to show that ${\textrm{Pr}}(p(X)>0) = {\textrm{Pr}}(p(Y)>0)+O(\epsilon).$ By construction of $p_\delta$, $${\textrm{Pr}}(p(X)>0) \geq {\textrm{Pr}}(p_\delta(\tilde X)>\delta) - \delta.$$ Applying Proposition \[mainProp\] to the multilinear polynomial $p_\delta-\delta$, this is at least $${\textrm{Pr}}(p_\delta(\tilde Y)>\delta) + O(\epsilon).$$ Since $\tilde Y$ is $\ell$-independent for all $\ell$ (since $Y$ is), it is actually an independent family of Gaussians. Therefore by Theorem \[anticoncentrationTheorem\], ${\textrm{Pr}}(|p(Y)|<\delta) = O(d\delta^{1/d}) = O(\epsilon)$. Hence $${\textrm{Pr}}(p(X)>0) \geq {\textrm{Pr}}(p_\delta(\tilde Y) > -\delta) + O(\epsilon).$$ Noting that with probability $1-\delta$ that $p_\delta(\tilde Y)$ is at most $\delta$ less than $p(Y)$, this is at least $${\textrm{Pr}}(p(Y)>0) + O(\epsilon).$$ So $${\textrm{Pr}}(p(X)>0) \geq {\textrm{Pr}}(p(Y)>0) + O(\epsilon).$$ Similarly, $${\textrm{Pr}}(p(X)<0) \geq {\textrm{Pr}}(p(Y)<0) + O(\epsilon).$$ Combining these we clearly have $${\textrm{Pr}}(p(X)>0) = {\textrm{Pr}}(p(Y)>0) + O(\epsilon)$$ as desired.
Fooling PTFs of Bernoulli Random Variables
==========================================
Theorem \[mainTheorem\] should also hold when $X$ is a $k$-independent family of Bernoulli random variables and $Y$ is a fully independent family of Bernoulli random variables. The proof is essentially the same as in the Gaussian case with a few minor changes that need to be made. In particular, the following steps do not carry over immediately:
1. The reduction from the case of a general polynomial to that of a multilinear polynomial
2. Theorem \[anticoncentrationTheorem\] does not hold for Bernoulli random variables
3. Theorem \[LatalaThm\] is not stated for the Bernoulli case
The first of these problems is even easier to deal with in the Bernoulli case than in the Gaussian case. This is because any degree-$d$ polynomial is equal to some degree-$d$ multilinear polynomial on the hypercube.
The second of these problems can be dealt with by fairly standard means. In particular, the Invariance Principle of [@MOO] implies that for sufficiently regular polynomials, $p$, that $p(X)$ is anticoncentrated even for $X$ a Bernoulli random variable. We are still left with the problem of reducing ourselves to the case of a regular polynomial. This would be done using a regularity Lemma similar to that proven in [@regularity], showing that an arbitrary polynomial threshold function can be written as a decision tree on a small number of coordinates such that most of the leaves are approximated by regular polynomial threshold functions. Given a slight modification of this result telling us that these “approximations” hold even on $k$-independent inputs would allow us to reduce to the case of a regular polynomial after determining the values of $O_d(\epsilon^{-O(d)})$ coordinates.
The last of these concerns is apparently more significant, but can be dealt with by proving that Theorem \[LatalaThm\] does hold for polynomials of Bernoullis. In particular, one can show that a higher moment of a polynomial with respect to the Bernoulli distribution can be bounded in terms of the corresponding moment with respect to the Gaussian distribution. In particular, we show that:
Let $p$ be a homogeneous degree-$d$ multilinear polynomial and $k\geq 1$. Let $X$ be a Bernoulli random variable and $Y$ a Gaussian random variable. Then $${\textrm{E}}[|p(X)|^k] = O(1)^{dk}{\textrm{E}}[|p(Y)|^k].$$
Let $\sigma=(\sigma_1,\ldots,\sigma_n)$ be an $n$-dimensional Bernoulli random variable and $G=(g_1,\ldots,g_n)$ an $n$-dimensional Gaussian random variable independent of of $\sigma$. Note that $\sigma_i |g_i|$ is distributed as a Gaussian. Therefore we have that $${\textrm{E}}[|p(G)|^k] = {\textrm{E}}[|p(\sigma_1|g_1|,\ldots,\sigma_n|g_n|)|^k] = {\textrm{E}}_G[ {\textrm{E}}_\sigma[|p(\sigma_1|g_1|,\ldots,\sigma_n|g_n|)|^k]].$$ By the convexity of the $L^k$ norm this is at least $${\textrm{E}}_\sigma\left[\left|{\textrm{E}}_G[p(\sigma_1|g_1|,\ldots,\sigma_n|g_n|)]\right|^k\right].$$ On the other hand, we have that $${\textrm{E}}_G[p(\sigma_1|g_1|,\ldots,\sigma_n|g_n|)] = \sqrt{\frac{2}{\pi}}^d p(\sigma).$$ Therefore we have that $${\textrm{E}}[|p(G)|^k] \geq \sqrt{\frac{2}{\pi}}^{dk} {\textrm{E}}_\sigma[|p(\sigma)|^k].$$ As desired.
Conclusion {#kIndepConc}
==========
The bounds on $k$ presented in this paper are far from tight. At the very least the argument in Lemma \[expectationErrorLem\] could be strengthened by considering a larger range of cases of $|p(x)|$ rather than just whether or not it is larger than $\sqrt{M}$. At very least, this would give us bounds on $k$ of the form $O_d(\epsilon^{-x^d})$ for some $x$ less than 7. I suspect that the correct value of $k$ is actually $O(d^2 \epsilon^{-2})$, and in fact such large $k$ will actually be required for $p(x)=\prod_{i=1}^d(\sum_{j=1}^{k} x_{i,j})$. On the other hand, this bound is at the moment somewhat beyond our means. It would be nice at least to see if a bound of the form $k=O_d(\epsilon^{-\textrm{poly}(d)})$ can be proven. The main contribution of this work is prove that there is some sufficient $k$ that depends on only $d$ and $\epsilon$.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was done with the support of an NSF graduate fellowship.
[\[99\]]{}
Ido Ben-Eliezer, Shachar Lovett and Ariel Yadin *Polynomial Threshold Functions: Structure, Approximation and Pseudorandomness*, Manuscript, available at <http://arxiv.org/PS_cache/arxiv/pdf/0911/0911.3473v3.pdf>.
A. Carbery, and J. Wright *Distributional and $L^q$ norm inequalities for polynomials over convex bodies in ${\mathbb{R}}^n$* Mathematical Research Letters, Vol. 8(3) (2001), pp. 233–248.
I. Diakonikolas, P. Gopalan, R. Jaiswal, R. Servedio and E. Viola, *Bounded Independence Fools Halfspaces* SIAM Journal on Computing, Vol. 39(8), 2010, pp. 3441-3462.
Ilias Diakonikolas, Daniel M. Kane, and Jelani Nelson, *Bounded Independence Fools Degree-2 Threshold Functions*, Foundations of Computer Science (FOCS), 2010.
Ilias Diakonikolas, Rocco Servedio and Li-Yang Tan and Andrew Wan *A Regularity Lemma, and Low-Weight Approximators, for Low-Degree Polynomial Threshold Functions* 25th Conference on Computational Complexity (CCC) (2010).
Rafal Latala *Estimates of Moments of Tails of Gaussian Choases*, The Annals of Probability, Vol. 34 (2006), no. 6, pp. 2315-2331.
E. Mossel, R. O’Donnell, and K. Oleszkiewicz *Noise stability of functions with low influences: invariance and optimality* Proceedings of the 46th Symposium on Foundations of Computer Science (FOCS), pages 21–30, 2005.
De La Pena, V. H., and Montgomery-Smith, S. *Bounds for the tail probabilities of $U$-statistics and quadratic forms* Bulletin of the American Mathematical Society, Vol. 31 (1994), pp. 223-227.
Raghu Meka and David Zuckerman *Pseudorandom generators for polynomial threshold functions* in Proceedings of the 42nd ACM Symposium on Theory Of Computing (STOC), 2010.
Elchanan Mossel, Ryan O’Donnell and Krzysztof Oleszkiewicz *Noise stability of functions with low influences: invariance and optimality* Short version in Proceedings of 46th Annual IEEE Symposium on the Foundations of Computer Science (FOCS), full manuscript available at <http://www.cs.cmu.edu/~odonnell/papers/invariance.pdf>.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper develops a novel compositional and abstraction-based approach to synthesize edit functions for opacity enforcement in modular discrete event systems. Edit functions alter the output of the system by erasing or inserting events in order to obfuscate the outside intruder, whose goal is to infer the secrets of the system from its observation. We synthesize edit functions to solve the opacity enforcement problem in a modular setting, which significantly reduces the computational complexity compared with the monolithic approach. Two abstraction methods called opaque observation equivalence and opaque bisimulation are first employed to abstract the individual components of the modular system and their observers. Subsequently, we propose a method to transform the synthesis of edit functions to the calculation of modular supremal nonblocking supervisors. We show that the edit functions synthesized in this manner correctly solve the opacity enforcement problem.'
author:
- 'Sahar Mohajerani, Yiding Ji and Stéphane Lafortune [^1] [^2]'
title: ' **Compositional and Abstraction-Based Approach for Synthesis of Edit Functions for Opacity Enforcement**'
---
= \[draw, diamond, aspect=2, text width=4.2em, text centered, node distance=1cm, inner sep=0pt\] = \[draw, diamond, dashed,aspect=2, text width=4.2em, text centered, node distance=1cm, inner sep=0pt\] = \[draw, diamond, dotted,aspect=2, text width=4.2em, text centered, node distance=1cm, inner sep=0pt\] = \[draw=none,fill=none\]
= \[circle, scale=0.3, node distance=4cm, inner sep=0pt,draw,fill\] = \[circle, scale=1, node distance=4cm, inner sep=0pt,draw,fill\] = \[rectangle, draw, text centered, node distance=1cm, minimum height=2em\] = \[rectangle, draw, dashed,text centered, node distance=1cm, minimum height=2em\] = \[rectangle, draw, dotted,text centered, node distance=1cm, minimum height=2em\]
= \[rectangle, draw, text centered, node distance=.5cm, minimum height=2em\] = \[draw, -latex’\]
= \[draw, ellipse, anchor=west, node distance=1cm, minimum height=0.25em\] = \[draw, ellipse, dashed,anchor=west, node distance=1cm, minimum height=0.25em\] = \[draw, ellipse, dotted,anchor=west, node distance=1cm, minimum height=0.25em\]
= \[draw, diamond, aspect=2, text centered, node distance=1.5cm, inner sep=0pt\] = \[draw, diamond, aspect=2, text centered, node distance=1cm, inner sep=0pt\] = \[draw, ellipse, anchor=west, node distance=2.3cm, minimum height=0.25em\] = \[dash pattern=on 3pt off 3pt\] = \[thin, decoration=[markings,mark=at position 1 with ]{}, double distance=1pt, shorten >= 4.5pt, preaction = [decorate]{}, postaction = [draw,line width=0.4pt, white,shorten >= .5pt]{}\] = \[thin, white,line width=1.4pt, shorten >= 4.5pt\]
INTRODUCTION
============
Opacity characterizes whether the integrity of the secrets of a system can be preserved from the inference of an outside intruder, potentially with malicious purposes. The intruder is modeled as a passive observer with knowledge of the system’s structure. A system is called opaque if the intruder is unable to infer the system’s secrets from its observation.
Starting with [@bryans2005modelling; @bryans2008opacity] in the computer science literature, opacity has been extensively studied, especially in the field of discrete event systems (DES), under multiple frameworks.
For finite state automaton models, various notions of opacity have been studied, e.g., language-based opacity [@lin2011opacity], current-state opacity [@saboori2007notions], initial-state opacity [@saboori2013verification], K-step opacity [@yin2017new] and infinite-step opacity [@saboori2012verification]. Opacity has also been discussed in some other system models, like infinite state systems [@chedor2015diagnosis], modular systems [@masopust2019complexity] and Petri nets [@tong2017verification; @tong2017decidability]. Opacity under a special observer called Orwellian observer is discussed in [@mullins2014opacity] and opacity under powerful attackers is studied in [@helouet2018opacity]. A more recent work [@yin2018verification] investigates opacity for networked supervisory control systems. Furthermore, some works investigate opacity in stochastic settings, e.g., [@keroglou2017probabilistic; @berard2015probabilistic; @chen2017quantification; @wu2018privacy]. Specifically, [@yin2019infinite] presents a novel approach to tackle infinite-step and K-step opacity in stochastic DES. The survey paper [@jacob2016overview] summarizes some recent results on opacity in DES.
When opacity does not hold, it is natural to study its enforcement [@falcone2015enforcement]. One popular approach is supervisory control [@dubreil2010supervisory; @darondeau2015enforcing; @takai2008formula; @saboori2012opacity; @yin2016uniform], where some behaviors of the system are disabled before they reveal the secrets. Another widely-applied method is sensor activation [@cassez2012synthesis; @zhang2015maximum; @yin2019general], where the observability of events is dynamically changed.
Recently, a new enforcement method called insertion function was proposed in [@wu2014synthesis], which inserts fictitious events into the output of the system to obfuscate the intruder. The authors of [@ji2018enforcement] extended the method to study opacity enforcement under the assumption that the intruder may or may not know the implementation of insertion functions, while [@ji2019enforcing] discussed opacity enforcement by insertion functions under quantitative constraints. As a following work, [@wu2018synthesis] investigates a more general method called edit functions, which manipulate the output of the system by either inserting or erasing events. Then [@ji2017edit; @ji2019synthesis] considers the case when the edit function’s implementation is known to the intruder. As a summary and extension, [@ji2019synthesis] characterizes opacity enforcement by edit functions as a three-player game and proposes a novel information structure called three-player observer (TPO) to embed edit functions. A special TPO called the All Edit Structure (AES) is also introduced in [@ji2019synthesis] to characterize the edit constraints.
In this work, we elaborate the method in [@ji2019synthesis] to study opacity enforcement in a modular setting. Our motivation is as follows. To generate a three-player observer, the observer of the system needs to be calculated, which is potentially costly in computation. Furthermore, modern engineering systems usually contain multiple components that are synchronized and subject to malicious inference. In this sense, if we are to apply edit functions to enforce opacity, heavy computation is involved both from determining individual systems and synchronizing them, which may be potentially cumbersome.
To alleviate this issue, this paper applies a compositional and abstraction-based method to reduce the size of the modular system before calculating the All Edit Structure. *Bisimulation* and *observation equivalence* [@milner1989communication] are well-known methods to abstract the state space of an automaton, while they do not preserve opacity properties in general. As a variant, [@zhang2018opacity] proposes several innovative concepts termed opacity-preserving (bi)simulation relations to reduce the state space of the system in opacity verification. A compositional visible bisimulation equivalence method is discussed in [@noori2018compositional] for abstraction-based opacity verification.
For abstraction, we introduce *opaque observation equivalence* and *opaque bisimulation*, which consider the secrecy status of states when merging them. In our framework, each individual system is abstracted using opaque observation equivalence. After that, the observer is calculated. Since abstraction reduces the size of the state space, the computational complexity of calculating the observer is lowered potentially. Next, opaque bisimulation is employed to the observer of each abstracted individual system, resulting in the smallest possible automaton for future discussion.
We further leverage some results from supervisor reduction and modular supervisory control theory to reduce the complexity of supervisor synthesis. There is a rich literature on both topics, see, e.g., [@su2004supervisor; @su2010model; @schmidt2012efficient; @malik2013compositional; @mohajerani2014framework]. The main idea is to convert the construction of the monolithic All Edit Structure to a modular supervisory control problem. Specifically, we first transfer each individual three-player observer (without considering edit constraints) to its automaton form and view the set of interacting automata as the “plant" to be controlled. Then we put the edit constraint as the specification, also in an automaton form. Afterwards, we perform modular supervisory control to synthesize a least restrictive and nonblocking modular supervisor. It is shown that all the traces accepted by the supervisor represent valid edit decisions contained in the monolithic AES. Compared with the conventional monolithic approach for supervisor synthesis [@cassandras2009introduction], our compositional approach is more efficient in computation.
The presentation of this work is organized as follows. Section \[sec:preli\] gives a brief background introduction about the system model, supervisory control theory and edit functions. The general idea of the paper is presented in \[sec:generalIdea\] as a flow chart. Section \[sec:AES\] explains the abstraction methods and synchronization of three-player observers. Next, Section \[sec:sup\] transforms the calculation of the monolithic All Edit Structure to the calculation of a modular supervisor. Finally, some concluding remarks are given in Section \[sec:conclusion\].
A preliminary and partial version of this work appears in [@ji2018efficient]. The current work improves [@ji2018efficient] in the sense that [@ji2018efficient] only considers abstraction methods to synthesize edit functions in a monolithic setting, while this work also takes synchronous composition into consideration and the edit functions are synthesized by a modular approach.
MODELING FORMALISM AND BACKGROUND {#sec:preli}
=================================
Events, Automata and their Composition
--------------------------------------
In this work, we consider discrete event systems modeled as deterministic or nondeterministic automata.
A (nondeterministic) finite-state automaton is a tuple $G= \langle \Sigma, Q, \rightarrow, Q^0\rangle$, where $\Sigma_{}$ is a finite set of events, $Q$ is a finite set of states, $\intrans \subseteq Q \times \Sigma_{} \times Q$ is the *state transition relation*, and $Q^0 \subseteq Q$ is the set of *initial states*. $G$ is *deterministic* if $|Q^0| = 1$ and if $x
\trans[\sigma] y_1$ and $x \trans[\sigma] y_2$ always implies that $y_1 = y_2$.
When state marking is considered, the above definition is extended to $G=\langle \Sigma, Q, \trans,Q^0, Q^m\rangle$, where $Q^m\subseteq Q$ is the set of *marked states*. In this paper, we identify marked states using gray shading in the figures.
We assume that the system is partially observed, thus the concepts of *observable* and *unobservable* events are introduced. Since the exact identity of unobservable events is irrelevant in our later discussion of opacity, they are uniformly represented by a special event $\tau$. The event $\tau$ is never included in the alphabet $\ACT$, unless explicitly mentioned. For this reason, $\ACT_\tau=\ACT\cup\{\tau\}$ is used to represent the whole set of observable and unobservable events. Hereafter, nondeterministic automata may contain transitions labeled by $\tau$, while *deterministic* automata *never* contain $\tau$ transitions. Moreover, $P_\tau:\ACT_\tau^*\to \ACT^*$ is the *projection* that removes from strings in $\ACT_\tau^*$ all the $\tau$ events.
When automata are brought together to interact, lock-step synchronization in the style of [@hoare1978communicating] is used.
\[def:synch\] Let $G_1 =\langle\ACT_{1}, Q_1, \trans_1, Q^0_1, Q^m_1\rangle$ and $G_2 = \langle\ACT_{2}, Q_2, \trans_2, Q^0_2, Q^m_2\rangle$ be two nondeterministic automata. The *synchronous composition* of $G_1$ and $G_2$ is defined as $$\begin{split}
G_1 \sync G_2:= \left\langle \Sigma_{1}\cup \Sigma_{2}, Q_1\times Q_2,
\intrans,
Q^0_1 \times Q^0_2,
Q_1^m \times Q_2^m
\right\rangle
\end{split}$$ where $$\begin{array}{@{}r@{\quad}l@{}}
(x_1,x_2) \trans[\sigma] (y_1,y_2) &
\mbox{if } \sigma \in ({\ACT_1} \cap {\ACT_2}),\\
& x_1 \trans[\sigma]_1 y_1,\ \text{and}\
x_2 \trans[\sigma]_2 y_2 \, ; \\
(x_1,x_2) \trans[\sigma] (y_1,x_2) &
\mbox{if } \sigma \in (\ACT_1 \setminus \ACT_2)\
\cup\{\tau\}\ \\ & \text{and}\
x_1 \trans[\sigma]_1 y_1 \, ; \\
(x_1,x_2) \trans[\sigma] (x_1,y_2) &
\mbox{if } \sigma \in (\ACT_2 \setminus \ACT_1)\
\cup\{\tau\}\ \\&
\text{and}\ x_2 \trans[\sigma]_2 y_2 \, .
\end{array}$$
Importantly, synchronous composition only imposes lock-step synchronization on common events from $\ACT_1$ and $\ACT_2$.
The transition relation of an automaton $G$ is written in infix notation $x \trans[\sigma] y$, and it is extended to strings in $\Sigma_\tau^*$ by letting $x \trans[\varepsilon]
x$ for all $x\in Q$, and $x\stackrel{t\sigma}{\rightarrow}z$ if $x\stackrel{t}{\rightarrow}y$ and $y\stackrel{\sigma}{\rightarrow}z$ for some $y \in Q$. Furthermore, $x \trans[t]$ means that $x \trans[t] y$ for some $y \in Q$, and $x \trans y$ means that $x \trans[t] y$ for some $t \in
\ACT_\tau^*$. These notations also apply to state sets, where $X \trans[t]Y$ for $X, Y
\subseteq Q$ means that $x \trans[t]y$ for some $x \in X$ and $y\in Y$, and to automata, where $G \trans[t]$ means that $Q^0 \trans[t]$ ($t$ is defined in $G$) and $G \trans[t]x$ means $Q^0 \trans[t]x$.
For brevity, $p\ttrans[s]q$ for $s\in\ACT^*$ represents the existence of a string $t\in\ACT_{\tau}^*$ such that $P_\tau(t)=s$ and $p\trans[t]q$. Thus, $q\trans[u]p$ for $u\in\ACT_\tau^*$ means a path containing exactly the events in $u$, while $q\ttrans[u]p$ for $u\in\ACT^*$ means existence of a path between $p$ and $q$ with an arbitrary number of $\tau$ events between the observable events in $u$. Similarly, $p\ttrans[\tau] q$ means the existence of a string $t\in \{\tau\}^*$ such that $p\trans[t]q$.
The *language* of an automaton $G$ is defined as $\LANG(G) = \{\, s \in \ACTstar \mid G \ttrans[s] \,\}$ and the language generated by $G$ from $q\in Q$ is $\LANG(G,q)=\{s\in\ACT^*\ |\ q\ttrans[s]\}$, thus we do not include event $\tau$ in the language of an automaton. Moreover, we also introduce projections $P_i$ for $i=1,2$, which are $P_i:(\ACT_{1}\cup\ACT_{2})^*\to \ACT^*_i$ for $i=1,2$.
For a nondeterministic automaton $G= \langle {\Sigma_\tau}, Q, \rightarrow, Q^0\rangle$, the set of *unobservably reached states* of $B\in 2^Q$, is $UR(B)=\bigcup\{C\subseteq Q\ |\ B\ttrans[\tau] C \}$. Its *observer* $det(G)=\langle \ACT, X_{obs},\trans_{obs},X_{obs}^0\rangle$ is a deterministic automaton, where $X_{obs}^0=UR(Q^0)$ and $X_{obs}\subseteq 2^Q$, and $X\trans[\sigma]_{obs}Y$, where $X, Y\in X_{obs} $, if and only if $Y=\bigcup\{UR(y)\ | \ x\trans[\sigma] y\ \textnormal{for some}\ x\in X \ \textnormal{and}\ y\in Q \}$. By convention, only reachable states from $X^0_{obs}$ under $\trans_{obs}$ are considered in this paper. We also refer to the observer as the *(current-state) estimator* of the system while an observer state is referred to as *(current-state) estimate*.
A common automaton operation is the *quotient* modulo, which is an equivalence relation on sets of states.
Let $Z$ be a set. A relation $\insim \subseteq Z\times Z$ is called an *equivalence relation* on $Z$ if it is reflexive, symmetric, and transitive. Given an equivalence relation on $Z$, the *equivalence class* of $z \in Z$ is $[z] = \{\, z' \in Z \mid z \sim z' \,\}$, and $\tilde{Z} = \{\, [z]
\mid z \in Z \,\}$ is the set of all equivalence classes modulo $\sim$.
\[def:quotient\] Let $G = \langle \Sigma, Q, \rightarrow, Q^0\rangle$ be an automaton and let $\insim \subseteq Q \times Q$ be an equivalence relation. The *quotient automaton* of $G$ modulo $\sim$ is $
\tilde{G} = \langle \ACT, \tilde{Q}, \intrans\modsim, \tilde{Q}^0\rangle$, where $\intrans\modsim =\{\, ([x],\sigma,[y]) \mid x \trans[\sigma] y
\,\}$ and $\tilde{Q}^0 =\{\, [x^0] \mid x^0 \in Q^0 \,\}$.
In order to compare automata structurally, we say that an automaton is a *subautomaton* of another automaton if all states and transitions in the first automaton are contained in the second one. Formally, we have the following definition:
Let $G_1=\left\langle \Sigma_{\tau},Q_1,\rightarrow_1,Q^0_1,Q^m_1\right\rangle$ and $G_2=\left\langle\Sigma_{\tau},Q_2,\rightarrow_2, Q^0_2,Q^m_2\right\rangle$ be two automata. $G_1$ is a *subautomaton* of $G_2$, denoted by $G_1 \sqsubseteq G_2$, if $Q_1 \subseteq Q_2$, $\mathord{\rightarrow}_1 \subseteq \mathord{\rightarrow}_2$, $Q^0_1 \subseteq Q^0_2$, and $Q^m_1 \subseteq Q^m_2$.
Supervisory Control Theory {#sec:sct}
--------------------------
Considering plant $G$ and specification $K$, *supervisory control theory* provides a method to synthesize a supervisor to restrict the behavior of the plant such that the given specification is always fulfilled. The supervisor $S$ is a function defined from the language of the system $G$ to the set of events, formally, $S: \mathcal L(G)\rightarrow 2^{\Sigma}$. We also partition the set of events as *uncontrollable events* and *controllable events*, i.e., $\Sigma=\Sigma_{uc}\cup \Sigma_c$, where uncontrollable events cannot be disabled by the supervisor. In the figures the uncontrollable events are marked by an exclamation mark (!). The readers may refer to [@cassandras2009introduction] for the main results of monolithic supervisory control under full observation. Here we focus on concepts and definitions relevant to the present paper and the synthesis procedure in this paper is done on deterministic automata. Two requirements for the supervisor are *controllability* and *nonblockingness*, where controllability captures *safety* in the presence of uncontrollable events and nonblockingness focuses on *liveness* of the system.
\[defNondetControllability\] [@cassandras2009introduction] Let $G = \left\langle \Sigma,Q_G,\rightarrow_G,Q^0_G,Q^m_G\right\rangle$ and $K = \left\langle \Sigma,Q_K,\rightarrow_K,Q^0_K,Q^m_K\right\rangle$ be two deterministic automata such that $K \sqsubseteq G$. $K$ is *controllable* w.r.t. $G$ if, for all states $x \in Q_K$ and $y \in Q_G$ and for every uncontrollable event $\upsilon \in \Sigma_{uc}$ such that $x \stackrel{\upsilon}{\rightarrow}_G y$, it also holds that $x \stackrel{\upsilon}{\rightarrow}_K y$.
\[defNonblocking\] [@cassandras2009introduction] Let $G$ be a deterministic automaton. A state $x$ is called *reachable* in $G$ if $G \rightarrow x$, and *coreachable* in $G$ if $x \rightarrow Q^m$. The automaton $G$ is called reachable or coreachable if every state in $G$ has this property. $G$ is called *nonblocking* if every reachable state is coreachable.
The upper bound of controllable and nonblocking subautomata is again controllable and nonblocking. This implies the existence of a least restrictive subautomaton of the original system, which is achieved by the maximally-permissive and nonblocking supervisor.
Let $G$ be an automaton, the supremal controllable and nonblocking subautomaton of $G$ is called the supremal supervisor, denoted by $\supCN(G)$ where for all controllable and nonblocking automata $K$ w.r.t. $G$, $K\sqsubseteq \supCN(G)$.
Synthesis of $\supCN(G)$ is done by iteratively removing blocking and uncontrollable states, until a fixed point is reached, and restricting the final automaton to the remaining states and their associated transitions, for more details please see [@cassandras2009introduction; @flordal2007compositional; @WonKai:19].
In this paper, we assume that the modular system has a set of interacting components $\{G_1,\ldots, G_n\}$, and there is also a set of supervisors in a modular structure, i.e., $\SYSS=\{S_1,\ldots,S_n\}$. Here supervisor $S_i$ is responsible for controlling $G_i$. The set of modular supervisors may be synchronized as $\bigsync^n_{i=1} S_i$.
Opacity and Edit Functions {#sec:edit-fn}
--------------------------
In this work, we suppose system $G$ has certain secret information which is characterized by the set of states. Thus the state space is partitioned into two disjoint subsets: $Q=Q^S\cup Q^{NS}$ where $Q^S$ is the set of *secret states* capturing the secrets of the system, while $Q^{NS}$ is the set of *non-secret states*. When the system $\SYSG$ is modular, $\SYSG=\{G_1,\ldots, G_2\}$, the set of secret states of the system, $Q_S$, is $Q^S= \{(x_1,\ldots,x_n)\ | \exists x_i \in Q_i^S \ \}$.
Suppose there is an external intruder modeled as the observer of the system, which intends to infer the secrets of the system from its observation. Then a system is called *opaque* if the intruder is unable to determine unambiguously if the system has entered a secret state or not. Different notions of opacity have been introduced in literature and we focus on current-state opacity in this work.
\[def:currentStateOp\] A nondeterministic automaton $G$ with a set of secret states $Q^S$ is *current-state opaque* w.r.t. $Q^S$ if $(\forall s\in\LANG(G,q^0) \ \colon\ Q^0\ttrans[s]Q^S)$ then $[\ Q^0\ttrans[s]Q^{NS}]$.
The system is current-state opaque if for any string reaching a secret state there is string with the same sequence of observable events reaching a non-secret state. It is known that current-state opacity can be verified by building the standard observer automaton.
\[prop:currentStateOpEstimator\] Let $G=\auttuple{\ACT_{\tau}}$ be a nondeterministic automaton with set of secret states $Q^S$. Let $det(G) = \langle \ACT,X_{obs},\trans_{obs},X_{obs}^0\rangle$ be the current-state estimator of $G$. Then $G$ is current-state opaque w.r.t. $Q^S$ if and only if $[det(G)\trans[s]X \ \text{implies that}\ X\not\subseteq Q^S]$.
If all states violating current-state opacity are removed from the observer $det(G)$, then the accessible part of the remaining structure is called the *desired observer*, denoted by $det_d(G)=\langle \ACT,X_{obsd},\trans_{obsd},X_{obsd}^0\rangle$. The language generated by the desired observer is referred to as the *safe language*, $L_{\textit{safe}}=\LANG(det_d(G))$. Accordingly, we also define the *unsafe language*, $L_{\textit{unsafe}}=\LANG(G)\setminus L_{\textit{safe}}$.
If a system is not current state opaque then an interface based approach called *edit function* [@ji2019synthesis; @wu2018synthesis] may be applied to enforce it. An edit function may insert events into the output of the system or erase events from the output of the system. It is assumed that the intruder fails to distinguish between an inserted event and its genuine counterpart. Let $\ACT^r=\{\sigma\trans\epsilon\colon \sigma\in \ACT\}$ be the set of “event erasure” events.
\[def:editFunction\] A *deterministic edit function* is defined as $f_e: \ACT^* \times \ACT \rightarrow \ACT^*$. Given $s \in \LANG(G)$, $\sigma\in \ACT$, $$\begin{aligned}
f_e(s, \sigma) =\begin{cases}
s_I\sigma & ~\mbox{if}~s_I~\mbox{is inserted before}~\sigma\\
\epsilon & ~\mbox{nothing is inserted and}~\sigma~\mbox{is erased}\\
s_I&~\mbox{if}~s_I~\mbox{is inserted and}~\sigma~\mbox{is erased}
\end{cases}
\end{aligned}$$
With an abuse of notation, we also define a string-based edit function $\hat{f}_e$ recursively as: $\hat{f}_e(\epsilon)=\epsilon$, $\hat{f}_e(s\sigma)=\hat{f}_e(s)f_e(s,\sigma)$ for $s\in \Sigma^*$ and $\sigma\in \Sigma$. In the sequel, to ease the notational burden, we will drop the “ $\hat{}$ ” in $\hat{f}_e$ and it will be clear from the argument(s) of $f_e$ which function we are referring to (incremental single-event one or string-based one).
Two notions termed *public safety* and *private safety* were defined in [@ji2019synthesis] to characterize the behavior of edit functions. In this paper, we consider private safety alone under the assumption that the intruder does not know about the implementation of an edit function.
\[def:private\] Given $G$ and its observer $det(G)$, an edit function $f_e$ is privately safe if $\forall s \in \LANG (det(G))$, $f_e(s) \in L_{\textit{safe}}$.
Recently a three-player game structure called *three-player observer (TPO)* w.r.t. the system was defined in [@ji2019synthesis] to embed edit functions. For the sake of completeness, we recall this definition (more details are available in [@ji2019synthesis]).
\[def:tripartite\] Given a system $G$ with its observer $det(G)$ and desired observer $det_d(G)$, let $I\subseteq X_{obsd}\times X_{obs}$ be the set of information states. A three-player observer w.r.t. $G$ is a tuple of the form $T=(Q_Y, Q_Z, Q_W, \ACT, \ACT^{r}, \Theta, \trans_{yz}, \trans_{zz}, \trans_{zw}, \trans_{wy}, y_0)$, where:
- $Q_Y \subseteq I$ is the set of $Y$ states.
- $Q_Z \subseteq I\times \ACT$ is the set of $Z$ states. Let $I(z)$, $E(z)$ denote the information state component and observable event component of a $Z$ state $z$ respectively, so that $z =(I(z), E(z))$.
- $Q_W \subseteq I\times (\ACT \cup \ACT^{r})$ is the set of $W$-states. Let $I(w)$, $A(w)$ denote the information state component and action component of a $W$ state $w$ respectively, so that $w=(I(w), A(w))$.
- $\ACT$ is the set of observable events.
- $\ACT^{r}$ is the set of event-erasure events.
- $\Theta \subseteq \ACT\cup\{\epsilon\} \cup\ACT^r$ is the set of edit decisions at $Z$ states.
1. \[im:yz\] $\trans_{yz}: Q_Y \times \ACT \times Q_Z$ is the transition function from $Y$ states to $Z$ states. For $y=(x_{d}, x_{f})\in Q_Y$, $e_o\in \ACT $, we have: $
y\trans[ e_o]_{yz}z\Rightarrow [x_{f}\trans[ e_o]_{obs}]\wedge[I(z)=y]\wedge [E(z)=e_o].
$
2. \[im:zz\] $\trans_{zz}: Q_Z \times \Theta \times Q_Z$ is the transition function from $Z$ states to $Z$ states. For $z=((x_{d}, x_{f}), e_o) \in Q_Z$, $\theta\in \Theta$, we have: $
z\trans[\theta]_{zz}z'\Rightarrow [\theta \in \ACT]\wedge[I(z')=(x'_{d}, x_{f})]
\wedge[x_d\trans[\theta]_{det_d}x'_{d}]\wedge [E(z')=e_o].
$
3. \[im:zw1\] $\trans_{zw1}: Q_Z \times \Theta \times Q_W$ is the $\epsilon$ insertion transition function from $Z$ states to $W$ states. For $z=((x_{d}, x_{f}), e_o)\in Q_Z$, $\theta\in \Theta$ we have: $
z\trans[ \theta]_{zw1}w\Rightarrow [\theta=\epsilon]\wedge [I(w)=I(z)]\wedge [A(w)=e_o]
\wedge[x_{d}\trans[ e_o]_{det_d}]\wedge[x_{f}\trans[ e_o]_{obs}].
$
4. \[im:zw2\] $\trans_{zw2}: Q_Z \times \Theta \times Q_W$ is the event erasure transition function from $Z$ states to $W$ states. For $z=((x_{d}, x_{f}), e_o)\in Q_Z$, $\theta \in \Theta$, we have: $
z\trans[\theta]_{zw2}w \Rightarrow [\theta=e_o\rightarrow\epsilon]\wedge [I(w)=I(z)]
\wedge [A(w)=e_o\rightarrow\epsilon]\wedge[x_{f}\trans[e_o]_{obs}].
$
5. \[im:wy1\] $\trans_{wy1}:Q_W\times \ACT\times Q_Y$ is the transition function from $W$ states whose action component is in $\ACT$ to $Y$ states. For $w=((x_{d}, x_{f}), e_o)\in Q_W$, we have: $
w\trans[e_o]_{wy1}y\Rightarrow [y=(x'_{d}, x'_{f})]\wedge[x_{d}\trans[ e_o]_{det_d} x'_d]
\wedge[x_{f}\trans[e_o]_{obs}x'_f].
$
6. \[im:wy2\] $\trans_{wy2}: Q_W\times \ACT\times Q_Y$ is the transition function from $W$ states whose action component is in $\ACT^r$ to $Y$ states. For $w=((x_{d}, x_{f}), e_o\rightarrow \epsilon)\in Q_W$, we have: $
w\trans[ e_o]_{wy2}y\Rightarrow [y=(x_{d}, x'_{f})]\wedge[x_{f}\trans[e_o]_{obs}x'_f].
$
- $y_0=(x_{obsd,0}, x_{obs,0}) \in Q_Y$ is the initial state of $T$, where $x_{obsd,0}$ and $x_{obs,0}$ are initial states of $det_d(G)$ and $det(G)$, respectively.
In general, a three-player observer characterizes a game between a dummy player, the edit function and the environment (system). The state space of a TPO is partitioned as: $Q_Y$ states ($Y$ states) where the dummy player plays; $Q_Z$ states ($Z$ states) where the edit function plays; $Q_W$ states ($W$ states) where the environment plays. A $Y$ state contains the intruder’s estimate (left component) as well as the system’s true state estimate (right component). A $\rightarrow_{yz}$ transition is defined out of a $Y$ state, indicating that an observable event may occur and thus is received by the edit function. Then the TPO transits to a $Z$ state and the turn of the game is passed to the edit function. Notice that the observable event does not really occur and this dummy player is only introduced to help determine the decisions of edit functions.
At a $Z$ state, the edit function may choose to insert certain events (including $\epsilon$) or erase its last observed event. If a non-$\epsilon$ event is inserted, a $\rightarrow_{zz}$ transition leads the TPO to another $Z$ state, which means the edit function still has the turn to insert more events until it decides to stop insertion by inserting $\epsilon$ or by erasing the last observed event. There may be multiple transitions defined out of a $Z$ state, i.e., multiple edit decisions; we write $\Theta(z)$ to denote the set of edit decisions defined at $z \in Q_Z$ in a TPO.
If the edit function inserts nothing (respectively erases the event it receives from the dummy player), then a $\rightarrow_{zw1}$ (respectively ($\rightarrow_{zw2}$)) transition is defined and the TPO is at a $W$ state. Then the environment plays by letting the observable event executed from its preceding $Y$ state occur. Correspondingly, there are also two types of $\rightarrow_{wy}$ transitions, where $\rightarrow_{wy1}$ indicates that the executed observable event will be observed by the intruder while $\rightarrow_{wy2}$ indicates that the executed observable event will not be observed by the intruder since it has been erased by the edit function.
When the three players take turns to play, the components of each player’s states also get updated. From \[def:tripartite\], a $\rightarrow_{yz}$ transition does not change the state estimates for the intruder or the system since the player at $Y$ states is dummy and the observable events from $Y$ states do not really occur. With a $\rightarrow_{zz}$ transition, only $x_d$ is updated since $x_d$ is the estimate of the intruder and event insertion only alters the observation of the intruder. For $\rightarrow_{zw}$ transitions, we only require the observable event to be defined at $x_d$ or $x_f$. Finally, a $\rightarrow_{wy1}$ transition updates both $x_d$ and $x_f$ while a $\rightarrow_{wy1}$ transition only updates $x_f$ as the intruder does not observe the erased event. To characterize the information flow in a TPO, the notion of *run* is defined in [@ji2019synthesis].
\[def:run\] In a three-player observer $T$, a run is defined as: $\omega=y_0\xrightarrow{e_{0}}z^1_0\xrightarrow{\theta^1_{0}}z^2_0\xrightarrow{\theta^2_{0}}\cdots\xrightarrow{\theta^{m_0-1}_{0}}z^{m_0}_0\xrightarrow{\theta^{m_0}_{0}}w_0 \xrightarrow{e_0}y_1\xrightarrow{e_{1}}z^1_1\xrightarrow{\theta^1_{1}}z^2_1\xrightarrow{\theta^2_{1}}\cdots z^{m_1}_1\xrightarrow{\theta^{m_1}_1} w_1\xrightarrow{e_1}y_2\cdots\xrightarrow{e_{n}}z^1_n\xrightarrow{\theta^1_{n}}\cdots z^{m_n}_n\xrightarrow{\theta^{m_n}_n}w_n\xrightarrow{e_n}y_{n+1}$, where $y_0$ is the initial state of $T$, $e_i\in \ACT$, $\theta^{j}_i\in \Theta(z^j_i)$, $\forall 0\leq i\leq n$, $1\leq j\leq m_i$ and $n\in \mathbb{N}$, $m_i\in \mathbb N^+$.
We let $\Omega_T$ be the set of all runs in a TPO $T$. For simplicity, similar notations as for automata are defined for three-player observers and thus $T\trans[\omega]x$ denotes the existence of a run in a three-player observer. We also review the concepts of *string generated by a run* and *edit projection* defined in [@ji2019synthesis].
\[def:string-run\] Given a run $\omega$ as in Definition \[def:run\], the string generated by $\omega$ is defined as: $l(\omega)=\theta^1_{0}\theta^2_{0}\cdots\theta^{m_0-1}_{0}\theta^{m_0}_{0}e_0\theta^1_{1}\cdots \theta^{m_1}_1e_1\cdots e_{n-1}\theta^1_{n}\cdots\theta^{m_n}_{n}e_n$, where $\forall i\leq n$, $\theta^{m_i}_{i}e_i=\epsilon$ if $\theta^{m_i}_{i}=e_i\rightarrow\epsilon$.
\[def:editproj\] Given TPO $T$ and run $\omega_T$ as in \[def:run\], the edit projection $P_e: \Omega\rightarrow \mathcal L(G)$ is defined such that $P_e(\omega_T)=e_0e_1\cdots e_n$.
In a TPO, $y\in Q_Y$ is a *terminating state* if $\not\exists e_o\in \ACT$, $y\trans[e_o]$. And $w\in Q_W$ is a *deadlocking* state if $\not\exists e_o\in \ACT$, s.t. $w\trans[e_o]y$. Also $z\in Q_Z$ is a *deadlocking* state if $\not\exists \theta\in \Theta$, s.t. $z\trans[\theta]z'$ or $z \trans[\theta]w$. We call a TPO $T$ *complete* [@ji2019synthesis] if there are no deadlocking $W$ or $Z$ states in $T$ and $\forall s\in \mathcal L(G)$, $\exists \omega\in \Omega_T$, s.t. $P_e(\omega)=s$.
\[def:tobsembed\] Given a TPO $T$, a deterministic edit function $f_e$ is embedded in $T$ if $\forall s\in \mathcal L(G)$, $\exists \omega\in \Omega_T$, s.t. $P_e(\omega)=s$ and $l(\omega)=f_e(s)$.
Next, we construct the *largest three-player observer* in the sense that all the other three-player observers are subautomata of it. Such a notion is well defined by considering *all* admissible transitions at *every* state of the TPO, according to the respective conditions in \[def:tripartite\].
Edit functions are designed to erase genuine events or insert fictitious ones to mislead the intruder. In theory, it is possible to design an edit function that erases all the events of the system, although this is not desirable. To avoid this situation usually the user provides some constraints on the edit functions. The constraint that is considered in this paper is to limit the number of consecutive erasures.
\[def:feasible-dec\] The edit constraint, denoted by $\Phi$, requires that the edit function should not make $n+1$ consecutive erasures where $n\in \mathbb{N}$.
Finally, we define the *All Edit Structure (AES)* [@ji2019synthesis] by considering the edit constraint. A synthesis procedure was also presented in [@ji2019synthesis] to construct the AES. Notice that the following definition is slightly different from the AES in the preliminary version of this work [@ji2018efficient] since edit constraints are not considered in [@ji2018efficient].
\[def:aes\] Given system $G$, observer $det(G)$ and desired estimator $det_d(G)$, the All Edit Structure is defined to be the largest complete three-player observer w.r.t. $G$, which satisfies the edit constraint.
From results in [@ji2019synthesis], private safety is achievable when the AES is not empty by construction. Hereafter, we assume that the AES is non-empty in the following discussion; if it is empty, then opacity cannot be enforced by the mechanism of edit functions. It was also proven in [@ji2019synthesis] that *all* privately safe edit functions satisfying edit constraints are embedded in the AES. Formally speaking, the following result holds.
\[thm:privateaes\] Given a system $G$ and its corresponding AES under edit constraint $\Phi$, an edit function $f_e$ is privately safe and satisfies $\Phi$ if and only if $f_e \in AES$.
We end this section by briefly reviewing the pruning process discussed in [@ji2019synthesis] to construct the AES. The presence of edit constraints may preclude some undesired states from the AES, thus leaving some states without outgoing transitions, i.e.,“deadlock" $Z$ or $W$ states. Those states reflect the inability of the edit function to issue a valid edit decision (for insertion or erasure) while still maintaining opacity for all possible future behaviors, thus should be removed in the pruning process. Moreover, $Y$ states that have transitions to a deadlock $Z$ state need to be pruned as well, since $Y$-states are the states where the system issues an output event and the edit function is not allowed to prevent their occurrence.
The construction of the AES may also be interpreted as the calculation of a supervisor where the “plant” is the largest three-player observer in terms of subautomaton, including all potentially feasible edit decisions without considering edit constraints. The $Y$ states are considered as marked states. The events labeling transitions from $Y$ states to $Z$ states and from $W$ states to $Y$ states are considered as uncontrollable, while the events labeling transitions from $Z$ states to $Z$ states and $Z$ states to $W$ states are viewed as controllable. We also define the proper specification by considering edit constraints, deleting states that violate them, and taking the trim of the resulting structure. The goal is to calculate the least restrictive, controllable and nonblocking supervisor based on the plant and this specification. Similar processes of pruning game structures akin to TPOs were discussed in prior work, e.g., [@wu2014synthesis; @ji2018enforcement; @ji2019synthesis]. We will leverage this approach in the following discussion, but in the framework of *modular supervisory control*.
COMPOSITIONAL ABSTRACTION-BASED METHODOLOGY {#sec:generalIdea}
===========================================
=\[draw=black,thick, fill=gray!20\] = \[rectangle, draw, text centered, rounded corners, minimum height=2em\] = \[draw=black,thick, rectangle, draw, text centered, minimum height=2em\] = \[draw, -latex’\] = \[draw, rounded rectangle ,fill=gray!20, node distance=1.7cm, minimum height=2em\] = \[draw=black,thick, rectangle,minimum width=8.5cm\]
![The steps of Algorithm CA-AES.[]{data-label="fig:composAlg"}](algorithm.pdf)
This section presents our novel compositional and abstraction-based methodology for synthesizing modular form edit functions based on individual three-player observers after abstracting the original system. For simplicity, we call this methodology the CA-AES (Composition Abstraction-All Edit Structure) Algorithm hereafter. The input of the algorithm is a set of nondetermistic automata, $\SYSG=\{G_1, \ldots, G_n\}$ and the output is a modular representation of edit functions, which is called *Modular Edit Structure*. The algorithm is summarized in Figure \[fig:composAlg\] and its steps are as follows. We will explain how to interpret the modular representation of edit functions later.
1. The algorithm first abstracts each individual automaton, $G_i$, using *opaque observation equivalence*. This results in $\tilde{G}_i$, which has fewer states and transitions compared to the original automaton.
2. Next, we abstract the observer of $\tilde{G}_i$, i.e., $det(\tilde{G}_i)$, by *opaque bisimulation* and *bisimulation*, resulting in two abstracted deterministic automata $H_{i,ob}$ and $H_{i,b}$.
3. Then we calculate the abstracted desired observer of $G_i$ from $H_{i,ob}$, which is denoted by $H_{i,obd}$.
4. Afterward, the largest (abstracted) three-player observer of each individual component $G_i$ is calculated from the abstracted observer $H_{i,b}$ and the abstracted desired observer $H_{i,obd}$, and it is denoted by $TPO_i$.
5. The final step is to calculate a modular nonblocking and controllable supervisor, then obtain a set of modular edit functions. This is done by transforming the largest three-player observers and the edit constraint to a set of automata, i.e., $G_i^T$ and $K$, respectively. This modular approach is in contrast to calculating monolithic edit functions embedded in the monolithic AES [@ji2019synthesis].
Specifically, in step (iv), each abstracted three-player observer w.r.t. the corresponding individual system together with the constraint $\Phi$ are transformed to a set of interacting automata. Then in step (v), a modular supremal controllable and nonblocking supervisor is calculated, thereby fulfilling the edit constraint in the composed structure. Consequently, the Modular Edit Structure is itself a modular supervisor. Regarding step (v), it is possible to leverage existing efficient algorithms on modular supervisory control to calculate the Modular Edit Structure.
In the monolithic approach of calculating the AES, individual systems $G_1$ through $G_n$ are synchronized first and then the observer of the synchronized system is built. Since the computational complexity of calculating the observer is exponential, synchronizing individual components before building the observer significantly increases the complexity, which may be $2^{\prod_{i=1}^n|Q_i|}$ in the worst case where $|Q_i|$ is the cardinality of $Q_i$. Moreover, constructing the All Edit Structure is polynomial in terms of the state space of the observer, which may be potentially intractable when we deal with the synchronized system. In contrast, our compositional and abstraction-based approach reduces computational cost considerably both from abstracting individual systems and conducting computation in a modular way. However, as will be demonstrated later, some edit decisions may be omitted in the Modular Edit Structure output by Algorithm CA-AES.
The presented approach relies heavily on the use of three-player observers. We present an example to better understand the structure of such observers.
Consider the nondeterministic automaton $G_1$ with secret states set $Q_1^S=\{q_3\}$, shown in \[fig:autOFtob\]. To generate the three-player observer of $G_1$, first the observer of $G_1$ needs to be built, which is shown as $det(G_1)$ in \[fig:autOFtob\]. To generate the desired observer the state $\{q_3\}\subseteq Q_1^S$ needs to be removed. The desired observer $det_d(G_1)$ is shown in \[fig:autOFtob\]. Then we follow the procedures in [@ji2019synthesis] to build the TPO w.r.t. $det(G_1)$ in \[fig:autOFtob\] (labeled as $T'_1$). As is discussed, the game on the TPO is initiated from $Y$-state $(q_0, q_0)$ where the dummy player executes the observable event $\gamma$ (the only event defined at $q_0$ in $det(G_1)$). Then the edit function takes the turn to play at the $Z$ state $(q_0, q_0, \gamma)$ where it has two choices: insert nothing or erase $\gamma$. If $\gamma$ is erased, then the $W$ state $(q_0, q_0, \gamma\rightarrow \epsilon)$ is reached where the environment plays by executing $\gamma$. Then the turn is passed back to the dummy player and the rest of the structure is interpreted similarly.
The compositional abstraction-based approach is explained in more details in the following sections. First, in Section \[sec:AES\], we discuss abstractions at the component level and synchronization of individual three-player observers, formalizing steps (i)-(iv) of Algorithm CA-AES. Then, in Section \[sec:sup\], we discuss the last step of Algorithm CA-AES.
SYNCHRONIZATION AND ABSTRACTION OPERATIONS {#sec:AES}
==========================================
This section presents results on abstraction and composition that support steps (i)-(iv) of Algorithm CA-CAS. First, \[sec:obsEq\] describes the methods to abstract nondeterministic automata and their observers. Next, \[sec:syncTOB\] describes the process of transforming every individual three-player observer to an automaton form and shows that the automaton representation is a substructure (in the sense of subgraph) of the largest monolithic three-player observer.
Opaque observation equivalence {#sec:obsEq}
------------------------------
The first strategy used in Algorithm CA-AES to alleviate state space explosion is abstraction of system components. This subsection contains a collection of abstraction methods that can be used to abstract nondeterministc automata and their observers such that the abstracted observers and the desired observers are bisimilar to their original counterparts. The abstraction methods are based on bisimulation and observation equivalence, which are computationally efficient and can be calculated in polynomial-time [@fernandez1990implementation]. We will prove in Theorem \[thm:trisamerun\] that if we build the largest three-player observer based on the abstracted observer and the desired observer, we obtain the same runs, consequently the same edit functions as we do from the largest three-player observer based on the original observer and desired observer.
Bisimulation is a widely-used notion of abstraction that merges states with the same future behavior.
\[def:bisimgeneral\][@milner1989communication] Let $G =\langle \Sigma_{\tau}, Q, \rightarrow, Q^0\rangle$ be a nondeterministic automaton. An equivalence relation $\approx\ \subseteq Q \times Q$ is called a *bisimulation* on $G$, if the following holds for all $x_1,x_2 \in Q$ such that $x_1 \approx x_2$: if $x_1 \trans[\sigma] y_1$ for some $\sigma \in \ACT_{\tau}$, then there exists $y_2 \in Q$ such that $x_2 \trans[\sigma] y_2$, and $y_1\approx y_2$.
Bisimulation seeks to merge states with the same outgoing transitions to equivalent states *including unobservable events*, i.e., $\tau$ events. If the unobservable events are disregarded, a more general abstraction method called *weak bisimulation* or *observation equivalence* naturally comes [@milner1989communication].
\[def:obseqgeneral\] Let $G =\langle \Sigma_{\tau}, Q, \rightarrow, Q^0\rangle$ be a nondeterministic automaton. An equivalence relation $\sim\ \subseteq Q \times Q$ is called an *observation equivalence* on $G$, if the following holds for all $x_1,x_2 \in Q$ such that $x_1 \sim x_2$: if $x_1 \ttrans[s] y_1$ for some $s \in \ACTstar$, then there exists $y_2 \in Q$ such that $x_2 \ttrans[s] y_2$, and $y_1\sim y_2$.
In order to use observation equivalence for abstraction in the opacity setting, the set of secret states needs to be taken into account. In the following discussion, a restricted version of observation equivalence called *opaque observation equivalence* is employed. This notion was first defined in [@mohajerani2018transform] in the context of verifying opacity.
\[def:obseq\] Let $G =\langle \Sigma_{\tau}, Q, \rightarrow, Q^0\rangle$ be a nondeterministic automaton with set of secret states $Q^S\subseteq Q$ and set of non-secret states $Q^{NS}=Q\setminus Q^S$. An equivalence relation $\inobseq_o \subseteq Q \times Q$ is called an *opaque observation equivalence* on $G$ with respects to $Q^S$, if the following holds for all $x_1,x_2 \in Q$ such that $x_1\obseq_o x_2$:
1. if $x_1 \ttrans[s] y_1$ for some $s \in \ACTstar$, then there exists $y_2 \in Q$ such that $x_2 \ttrans[s] y_2$, and $y_1\obseq_o y_2$,
2. $x_1\in Q^S$ if and only if $x_2\in Q^S$.
We also wish to use bisimulation to abstract the observer of a nondeterministic system. Besides opaque observation equivalence, *opaque bisimulation* is also defined.
\[def:opaqueBisim\] Let $G =\langle \Sigma_{\tau}, Q, \rightarrow, Q^0\rangle$ be a nondeterministic automaton with set of secret states $Q^S\subseteq Q$ and set of non-secret states $Q^{NS}=Q\setminus Q^S$. Let $det(G)=\langle \ACT, X_{obs}, \trans_{obs},X_{obs}^0\rangle$ be the observer of $G$. An equivalence relation $\approx_{o} \subseteq X_{obs} \times X_{obs}$ is called an *opaque bisimulation equivalence* on $det(G)$ with respects to $Q^S$, if the following holds for all $X_1,X_2 \in X_{obs}$ such that $X_1\approx_{o} X_2$:
1. if $X_1 \trans[s] Y_1$ for some $s \in \ACTstar$, then there exists $Y_2 \in X_{obs}$ such that $X_2 \trans[s] Y_2$, and $Y_1\approx_{o} Y_2$,
2. $X_1\subseteq Q^S$ if and only if $X_2\subseteq Q^S$.
The first step of Algorithm CA-AES is to abstract the system using opaque observation equivalence. It has been shown in [@rutten1998automata] that if two automata are bisimilar, then their observers are also bisimilar. In this paper this result is extended such that abstracting a nondeterministic automaton using opaque observation equivalence results in an observer and a desired observer which are bisimilar to the observer and the desired observer of the original system, respectively.
\[propos:obsBisim\] Let $G =\langle \Sigma_{\tau}, Q, \rightarrow, Q^0\rangle$ be a nondeterministic automaton with set of secret states $Q^S\subseteq Q$ and set of non-secret states $Q^{NS}=Q\setminus Q^S$. Let $\sim_0$ be an opaque observation equivalence on $G$ resulting in $\tilde{G}$ and let $\approx$ be a bisimulation. Let $det_d(G)$ and $det_d(\tilde{G})$ be the desired observer of $G$ and $\tilde{G}$. Then $det(G)\approx det(\tilde{G})$ and $det_d(G)\approx det_d(\tilde{G})$.
*Proof:* First we prove that $det(G)\approx det(\tilde{G})$. To prove $det(G)\approx det(\tilde{G})$ it is enough to show that $det(G)\trans[s]X$ if and only if $det(\tilde{G})\trans[s]\tilde{X}$, which implies language equivalence between $det(G)$ and $det(\tilde{G})$ since $det(G)$ and $ det(\tilde{G})$ are deterministic. This can be shown by induction. Moreover, in the induction we also show that $x\in X$ if and only if there exist $[x']\in \tilde{X}$ such that $x\in [x']$. This is used for the second part of the proof, where we show $det_d(G)\approx det_d(\tilde{G})$.
It is shown by induction on $n\geq 0$ that $X^0\trans[\sigma_1]X^1\trans[\sigma_2]\ldots\trans[\sigma_n]X^n$ in $det(G)$ if and only if $\tilde{X}^0\trans[\sigma_1]\tilde{X}^1\trans[\sigma_2]\ldots\trans[\sigma_n]\tilde{X}^n$ in $det(\tilde{G})$ such that $x\in X^j$ if and only if $[x']\in \tilde{X}^k$, where $x\in [x']$, for $1\leq j\leq n$.
*Base case:* $n=0$. Let $X^0$ be the initial state of $det(G)$ and $\tilde{X}^0$ be the initial state of $det(\tilde{G})$. It is shown that $x\in X^0$ if and only if there exists $[x']\in \tilde{X}^0$ such that $x\in [x']$.
First, let $x\in X^0$. Then based on $UR(x^0)$, it follows that there exists $x^0\in Q^0$ such that $x^0\ttrans[\tau]x$ in $G$. Since $G\sim_o \tilde{G}$ then based on \[def:obseq\], there exists $[x'^0]\in\tilde{X}^0$ such that $[x'^0]\ttrans[\tau] [x']$ in $\tilde{G}$ such that $x^0\in [x'^0]$ and $x\in [x']$. Then based on $UR(x^0)$ it follows that $[x']\in \tilde{X}^0$.
Now let $[x']\in \tilde{X}^0$. Then based on $UR(x^0)$, it follows that there exists $[x'^0]\in Q^0$ such that $[x'^0]\ttrans[\tau][x']$ in $\tilde{G}$. Since $G\sim_o \tilde{G}$ then based on \[def:obseq\], there exists $x^0\in X^0$ such that $x^0\ttrans[\tau] x$ in $G$ such that $x^0\in [x'^0]$ and $x\in [x']$. Then based on $UR(x^0)$ it follows that $x\in X^0$.
*Inductive step*: Assume the claim holds for some $n\geq 0$, i.e, $X^0\trans[\sigma_1\sigma_2\ldots\sigma_n]X^n=X$ in $det(G)$ if and only if $\tilde{X}=\tilde{X}^0\trans[\sigma_1\sigma_2\ldots\sigma_n]\tilde{X}^n=\tilde{X}$ in $det(\tilde{G})$, such that $x\in X^k$ if and only if there exists $[x']\in \tilde{X}^k$ such that $x\in [x']$ for all $0\leq k< n$[^3]. It must be shown that $X=X^n\trans[\sigma_{n+1}]Y$ in $det(G)$ if and only if $\tilde{X}=\tilde{X}^n\trans[\sigma_{n+1}]\tilde{Y}$ in $det(\tilde{G})$ such that $x\in X$ if and only if there exists $[x']\in \tilde{X}$ such that $x\in [x']$.
First, let $X=X^n\trans[\sigma_{n+1}]Y$ in $det(G)$ and let $x\in X$. Then based on $UR(x)$ it holds that $x=x^1\ttrans[\tau]\cdots\ttrans[\tau]x^r\trans[\sigma_{n+1}]y$ in $G$, where $x^j\in X$ for all $1\leq j\leq r$ and $y\in Y$. Since $G\sim_o \tilde{G}$ it holds that $[x']=[x'^1]\ttrans[\tau]\cdots\ttrans[\tau][x'^r]\trans[\sigma_{n+1}][y']$ in $\tilde{G}$ such that $x^j\in [x'^j]$ for all $1\leq j\leq r$ and $y\in [y']$. Based on $UR(x)$ and inductive assumption it holds that $det(\tilde{G})\trans[\sigma_1\sigma_2\ldots\sigma_n] \tilde{X}^n=\tilde{X}\trans[\sigma_{n+1}]\tilde{Y}$ and $[x']\in \tilde{X}$.
Now let $\tilde{X}=\tilde{X}^n\trans[\sigma_{n+1}]\tilde{Y}$ in $det(\tilde{G})$ and let $[x]\in \tilde{X}$. Then based on $UR(x)$ it holds $[x]=[x^1]\ttrans[\tau]\cdots\ttrans[\tau][x^r]\trans[\sigma_{n+1}][y]$ in $\tilde{G}$, where $[x^i]\in \tilde{X}$ for all $1\leq i\leq r$ and $[y]\in \tilde{Y}$. Since $G\sim_o \tilde{G}$ it holds that $x'=x'^1\ttrans[\tau]\cdots\ttrans[\tau]x'^r\trans[\sigma_{n+1}]y'$ in $\tilde{G}$ such that $x'^i\in [x^i]$ for all $1\leq i\leq r$ and $y'\in [y]$. Based on $UR(x)$ and inductive assumption it holds that $det(G)\trans[\sigma_1\sigma_2\ldots\sigma_n]X^n=X\trans[\sigma_{n+1}]Y$ such that $x'\in X$.
Now we need to show that $det_d(G)\approx det_d(\tilde{G})$. It was proven above that $det(G)\approx det(\tilde{G})$, which means $det(G)\trans[s]X$ if and only if $det(\tilde{G})\trans[s]\tilde{X}$ and $x\in X$ if and only if $[x']\in \tilde{X}$, where $x\in [x']$. Therefore, it is enough to show that $X\not\in X_{obsd}$ if and only if $\tilde{X}\not\in \tilde{X}_{obsd}$.
First assume $X\subseteq Q^S$, which means for all $x\in X$ it holds that $x\in Q^S$ and $X\not\in X_{obsd}$. Since for all $x\in X$ it holds that there exist $[x']\in \tilde{X}$ such that $x\in [x']$ then based on \[def:obseq\] it holds that $[x']\in \tilde{Q}^S$. Thus, it can be concluded that for all $[x']\in \tilde{X}$ it holds that $[x']\in \tilde{Q}^S$. This means that $\tilde{X}\subseteq \tilde{Q}^S$ and consequently $\tilde{X}\not\in \tilde{X}_{obsd}$.
Now assume $\tilde{X}\subseteq \tilde{Q}^S$, which means for all $[x']\in \tilde{X}$ it holds that $[x']\in \tilde{Q}^S$ and $\tilde{X}\not\in \tilde{X}_{obsd}$. If $[x']\in \tilde{Q}^S$ then for all $x\in [x']$ it holds that $x\in Q^S$. Moreover, it was shown above that $[x']\in \tilde{X}$ if and only if $x\in X$, where $x\in [x']$. Thus, from $\tilde{X}\subseteq \tilde{Q}^S$ it follows that $X\subseteq Q^S$, which means that $X\not\in X_{obsd}$.
Thus, it can be concluded that $det_d(G)\approx det_d(\tilde{G})$.[$\blacksquare$]{}
Opaque observation equivalence seeks to merge states of a nondeterministc automaton, which are “equivalent", before constructing the observer. After calculating the observer, it is possible to further abstract the observer using opaque bisimulation. This guarantees that the smallest abstracted observer generates the same language as the original observer. In the following, Proposition \[pro:bisidetd\] shows that if opaque bisimulation is used to abstract the observer, then the abstracted desired observer is also bisimilar to the original desired observer.
\[pro:bisidetd\] Let $G =\langle \Sigma_{\tau}, Q, \rightarrow, Q^0\rangle$ be a nondeterministic automaton with set of secret states $Q^S\subseteq Q$ and set of non-secret states $Q^{NS}=Q\setminus Q^S$. Let $\approx_o$ be an opaque bisimulation on $det(G)$ resulting in $\widetilde{det(G)}$. Let $det_d(G)$ and $H_{d}$ be the desired observers of $det(G)$ and $\widetilde{det(G)}$, respectively. Then $det_d(G)\approx H_{d}$, where $\approx$ is a bisimulation relation.
*Proof:* Since $det(G)\approx_o \widetilde{det(G)}$ based on \[def:opaqueBisim\] it holds that $det(G)\trans[s]X$ if and only if $\widetilde{det(G)}\trans[s][X']$ and $X\in [X']$. Thus, it is enough show that $X\not\in X_{obs, det_d(G)}$ if and only if $[X']\not\in X_{obs,H_d}$, where $X\in [X']$.
First assume $X\subseteq Q^S$, so $X\not\in X_{obs,det_d(G)}$. Then since $X\in [X']$ based on \[def:opaqueBisim\] it holds that for all $X'\in[X']$, $X'\subseteq Q^S$. This means $[X']\subseteq Q^S$ and consequently $[X']\not\in X_{obs,H_d}$.
Then assume $[X']\subseteq Q^S$, so $[X']\not\in X_{obs,H_d}$. Since $X\in [X']$ based on \[def:opaqueBisim\], $X \subseteq Q^S$ holds, i.e., $X\not\in X_{obs,det_d(G)}$. [$\blacksquare$]{}
We now present the main results of this subsection.
\[thm:trisamerun\]
*Proof:* From Propositions \[propos:obsBisim\] and \[pro:bisidetd\] it holds that $det(G)\approx H_b$ and $det_d(G)\approx H_{obd}$. Thus, we need to show that a transition is defined in $T$ if and only if the same transition is defined in $T'$. It is shown by induction on $n\geq 0$ that $y^0\trans[\omega]q_n$ in $T$ if and only if $\tilde{y}^0\trans[\omega]\tilde{q}_n$ in $T'$.
*Base case:* $(\Rightarrow)$ First assume $y^0\trans[e_o]z^0$ in $T$, where $y^0=(X^0_{obsd},X^0_{obs})$. Based on \[def:tripartite\] it holds that $X^0_{obs}\trans[e_o]$ in $det(G)$, $I(z^0)=y^0$ and $E(z^0)=e_o$. From $X^0_{obs}\trans[e_o]$ in $det(G)$ and since $det(G)\approx {H_{b}}$ it holds that $\tilde{X}^0_{obs}\trans[e_o]$ in ${H_{b}}$. Thus, $\tilde{y}^0=(\tilde{X}^0_{obsd},\tilde{X}^0_{obs})$ and $\tilde{X}^0_{obs}\trans[e_o]$ in ${H_{b}}$, $I(\tilde{z}^0)=\tilde{y}^0$ and $E(\tilde{z}^0)=e_o$. This means $\tilde{y}^0\trans[e_o]\tilde{z}^0$ in $T'$.
$(\Leftarrow)$ Now assume $\tilde{y}^0\trans[e_o]\tilde{z}^0$ in $T'$, where $\tilde{y}^0=(\tilde{X}^0_{obsd},\tilde{X}^0_{obs})$. The same argument as $(\Rightarrow)$ holds.
*Inductive step:* Assume the claim holds for some $n\geq 0$, i.e, $n\geq 0$ that $y^0\trans[\omega]q_n$ in $T$ if and only if $\tilde{y}^0\trans[\omega]\tilde{q}_n$ in $T'$.
$(\Rightarrow)$ It must be shown that if $q_n\trans[\sigma]p_n$ in $T$ then $\tilde{q}_n\trans[\sigma]\tilde{p}_n$ in $T'$. There are six possibilities:
- $q_n=y$ is a $Y$ state and $p_n=z$ is a $Z$ state, \[def:tripartite\] \[im:yz\]. Let $y=(x_d,x_{f})$ and $\sigma\in \ACT$. Based on inductive assumption there exists $\tilde{y}=(\tilde{x}_{d},\tilde{x}_{f})$ and $\tilde{y}^0\trans[\omega]\tilde{y}$ in $T'$. From $y\trans[\sigma]z$ in $T$ and based on \[def:tripartite\] it holds that $x_{f}\trans[\sigma]$ in $det(G)$ and $I(z)=y$ and $E(z)=\sigma$. Since $det(G)\approx {H_{b}}$ it holds that there exists $\tilde{x}_{f}\trans[\sigma]$ in ${H_{b}}$, where $\tilde{x}_{f}\approx x_f$. This means $\tilde{y}\trans[\sigma]\tilde{z}$ in $T'$, where $\tilde{y}=(\tilde{x}_{d},\tilde{x}_{f})$ and $I(\tilde{z})=(\tilde{x}_{d},\tilde{x}_{f})$ and $E(\tilde{z})=\sigma$.
- $q_n=z$ is a $Z$ state and $p_n=z'$ is a $Z$ state, \[def:tripartite\] \[im:zz\]. Let $z=((x_{d},x_{f}),e_o)$ and $\sigma\in\Theta$. Then based on \[def:tripartite\] it holds that $\sigma\in \ACT$ and $I(z')= (x'_{d},x_{f})$ and $x_d\trans[\sigma]x'_d$ in $det_d(G)$ and $E(z')=e_o$. Since $det_d(G)\approx {H_{obd}}$ and from $x_d\trans[\sigma]x'_d$ in $det_d(G)$ it follows that $\tilde{x}_{d}\trans[\sigma]\tilde{x}'_{d}$ in ${H_{obd}}$, where $\tilde{x}_{d}\approx x_d$ and $\tilde{x}'_{d}\approx x'_d$, and based on the inductive assumption it follows that there exists $\tilde{z}=((\tilde{x}_{d},\tilde{x}_{f}), e_o)$ such that $\tilde{y}^0\trans[\omega]\tilde{z}$ in $T'$. Thus, $\tilde{z}\trans[\sigma]\tilde{z}'$ in $T'$, where $\tilde{z}=((\tilde{x}_{d},\tilde{x}_{f}), e_o)$ and $\sigma\in \ACT$ and $I(\tilde{z}')= (\tilde{x}'_{d},\tilde{x}_{f})$ and $\tilde{x}_d\trans[\sigma]\tilde{x}'_d$ in ${H_{obd}}$ and $E(\tilde{z}')=e_o$.
- $\tilde{q}_n\trans[\sigma]\tilde{p}_n$ is the $\epsilon$ insertion function and $q_n=z$ is a $Z$ state and $p_n=w$ is a $W$ state, \[def:tripartite\] \[im:zw1\]. Let $z=((x_{d},x_{f}),e_o)$ and $\sigma\in\Theta$. Then based on \[def:tripartite\] it holds that $\sigma=\epsilon$ and $I(w)= I(z)$ and $A(w)=e_o$ and $x_d\trans[e_o]$ and $x_f\trans[e_o]$ in $det_d(G)$ and $det(G)$, respectively. Since $det(G)\approx {H_{b}}$ and $det_d(G)\approx {H_{obd}}$, from $x_f\trans[e_o]$ in $det(G)$ it follows that $\tilde{x}_f\trans[e_o]$ in ${H_{b}}$ and from $x_d\trans[e_o]$ in $det_d(G)$ it follows that $\tilde{x}_d\trans[e_o]$ in ${H_{obd}}$. Moreover, based on the inductive assumption it holds that there exists $\tilde{z}=((\tilde{x}_{d},\tilde{x}_{f}), e_o)$ such that $\tilde{y}^0\trans[\omega]\tilde{z}$ in $T'$. Thus, $\tilde{z}\trans[\sigma]\tilde{w}$ in $T'$, where based on \[def:tripartite\] it holds that $\sigma=\epsilon$ and $I(\tilde{w})= I(\tilde{z})$ and $A(\tilde{w})=e_o$ and $\tilde{x}_d\trans[e_o]$ and $\tilde{x}_f\trans[e_o]$ in ${H_{obd}}$ and ${H_{b}}$, respectively.
- $\tilde{q}_n\trans[\sigma]\tilde{p}_n$ is the event erasure transition function and $q_n=z$ is a $Z$ state and $p_n=w$ is a $W$ state, \[def:tripartite\] \[im:zw2\]. Let $z=((x_{d},x_{f}),e_o)$ and $\sigma\in\Theta$. Then based on \[def:tripartite\] it holds that $\sigma=e_o\trans[]\epsilon$ and $I(w)= I(z)$ and $A(w)=e_o\trans[]\epsilon$ and $x_f\trans[e_o]$ in $det(G)$. Since $det(G)\approx {H_{b}}$ from $x_f\trans[e_o]$ in $det(G)$ it follows that $\tilde{x}_f\trans[e_o]$ in ${H_{b}}$. Moreover, based on the inductive assumption it follows that there exists $\tilde{z}=((\tilde{x}_{d},\tilde{x}_{f}), e_o)$ such that $\tilde{y}^0\trans[\omega]\tilde{z}$ in $T'$. Thus, $\tilde{z}\trans[\sigma]\tilde{w}$ in $T'$, where based on \[def:tripartite\] it holds that $\sigma=e_o\trans[]\epsilon$ and $I(\tilde{w})= I(\tilde{z})$ and $A(\tilde{w})=e_o\trans[]\epsilon$ and $\tilde{x}_f\trans[e_o]$ in ${H_{b}}$.
- $q_n=w$ is a $W$ state and $p_n=y$ is a $Y$ state and $\sigma\in \ACT$, \[def:tripartite\] \[im:wy1\]. Let $w=((x_{d},x_{f}),e_o)$. Then based on \[def:tripartite\] it holds that $y=(x'_d,x'_f)$ and $x_f\trans[e_o]x'_f$ in $det(G)$ and $x_d\trans[e_o]x'_d$ in $det_d(G)$. Since $det(G)\approx {H_{b}}$ and $det_d(G)\approx {H_{obd}}$, from $x_f\trans[e_o]x'_f$ in $det(G)$ it follows that $\tilde{x}_f\trans[e_o]\tilde{x}'_f$ in ${H_{b}}$ and from $x_d\trans[e_o]x'_d$ in $det_d(G)$ it follows that $\tilde{x}_d\trans[e_o]\tilde{x}'_d$ in ${H_{obd}}$. Moreover, based on the inductive assumption it holds that there exists $\tilde{w}=((\tilde{x}_{d},\tilde{x}_{f}), e_o)$ such that $\tilde{y}^0\trans[\omega]\tilde{w}$ in $T'$. Thus, $\tilde{w}\trans[\sigma]\tilde{y}$ in $T'$, where based on \[def:tripartite\] it holds that $\tilde{y} = (\tilde{x}'_d,\tilde{x}'_f)$ and $\tilde{x}_f\trans[e_o]\tilde{x}'_f$ in ${H_{b}}$ and $\tilde{x}_d\trans[e_o]\tilde{x}'_d$ in ${H_{obd}}$.
- $q_n=w$ is a $W$ state and $p_n=y$ is a $Y$ state and $\sigma\in \ACT$, \[def:tripartite\] \[im:wy2\]. Let $w=((x_{d},x_{f}),e_o\trans[]\epsilon)$. Then based on \[def:tripartite\] it holds that $y=(x_d,x'_f)$ and $x_f\trans[e_o]x'_f$ in $det(G)$. Since $det(G)\approx {H_{b}}$ and from $x_f\trans[e_o]x'_f$ in $det(G)$ it holds that $\tilde{x}_f\trans[e_o]\tilde{x}'_f$ in ${H_{b}}$. Moreover, based on the inductive assumption it holds that there exists $\tilde{w}=((\tilde{x}_{d},\tilde{x}_{f}), e_o\trans[]\epsilon)$ such that $\tilde{y}^0\trans[\omega]\tilde{w}$ in $T'$. Thus, $\tilde{w}\trans[\sigma]\tilde{y}$ in $T'$, where based on \[def:tripartite\] it holds that $\tilde{y} = (\tilde{x}_d,\tilde{x}'_f)$ and $\tilde{x}_f\trans[e_o]\tilde{x}'_f$ in ${H_{b}}$.
$(\Leftarrow)$ It must be shown that if $\tilde{q}_n\trans[\sigma]\tilde{p}_n$ in $T'$ then $q_n\trans[\sigma]p_n$ in $T$. The same argument as $(\Rightarrow)$ holds. [$\blacksquare$]{}
Theorem \[thm:trisamerun\] proves that the largest three-player observer obtained from the abstracted system (using opaque observation equivalence and opaque bisimulation) has the same set of runs with that obtained from the original system. This result is essential for the correctness of Algorithm CA-AES.
The abstractions in the worst case scenario fail to merge any states. However, as pointed out in the paper the complexity of the abstraction methods is polynomial, while the complexity of calculating the observer is exponential in the number of states. Thus, if the abstraction results in merging even few states, it can potentially reduce the complexity of calculating the observer significantly. Therefore, it is worth applying the abstraction algorithm before calculating the observers.
![System $\SYSG=\{G_1,G_2\}$ and its abstraction $\{\tilde{G}_1,\tilde{G}_2\}$. The figure also shows the largest three-player observers $T'_1$ and $T'_2$ of the abstracted components and their automata transformations, denoted by $G^T_1$ and $G^T_2$. The uncontrollable events are marked by (!).[]{data-label="fig:autOFtob"}](system.pdf){width="0.99\columnwidth"}
\[ex:abs\] Consider the nondeterministic system $\SYSG=\{G_1,G_2\}$, shown in \[fig:autOFtob\], with secret states sets $Q_1^S=\{q_3\}$ and $Q_2^S=\{s_3\}$ where all the events are observable except event $\tau$. In $G_1$ states $q_1$ and $q_2$ are opaque observation equivalent as they both have the same secrecy status and equivalent states can be reached from both, $q_1\trans[\alpha]q_3$ and $q_2\trans[\alpha]q_3$, and $q_1\trans[\tau]q_2$ and $q_2\trans[\epsilon]q_2$. Merging $q_1$ and $q_2$ results in the abstracted automaton $\tilde{G}_1$ shown in \[fig:autOFtob\]. Moreover, states $s_1$ and $s_2$ are also opaque observation equivalent and merging them results in automaton $\tilde{G}_2$ shown \[fig:autOFtob\]. After abstracting the automata, the system becomes a deterministic system. Moreover, the observers as of $\tilde{G}_1$ and $\tilde{G}_2$ are bisimilar to $det(G_1)$ and $det(G_2)$, respectively. The same is also true for the desired observer of $\tilde{G}_1$ and $\tilde{G}_2$. \[fig:autOFtob\] shows the largest three-player observers of $\tilde{G}_1$ and $\tilde{G}_2$, respectively.
Synchronous composition of TPOs {#sec:syncTOB}
-------------------------------
The second strategy used in Algorithm CA-AES to reduce computation complexity is synchronous composition of individual systems. In this work, the main advantage of our compositional approach is to build the largest three-player observer of each component individually, instead of synchronizing individual components and then building the largest monolithic three-player observer. Before synchronization, we first transfer each individual TPO to an automaton using \[def:TOBtoAutone\]. Next, the individual automata are transformed to a set of interacting automata based on \[def:TOBtoAut\]. It is shown in \[thm:TOBtoAutSync\] that the set of modular three-player observers form a subsystem of their monolithic counterpart, in the sense that some runs are omitted after synchronization. Before \[thm:TOBtoAutSync\], Lemmas \[propos:syncDet\] and \[lem:detd\] establish that synchronization of individual observers (respectively, desired observers) is isomorphic to the observer (respectively, desired observers) of the synchronized system.
\[def:TOBtoAutone\] Let $T=\langle Q_{Y},Q_{Z},Q_{W},\ACT,\ACT^{\epsilon}, \Theta, \trans_{yz},\trans_{zz},\trans_{zw},\trans_{wy},y_0\rangle$ be a three-player observer. Automaton $M^T=\langle \ACT_{M^T},Q,\trans,Q^0,Q^m\rangle$ is the monolithic transformed deterministic automaton of $T$ where,
1. $\ACT_{M^T}=\ACT\cup[\bigcup\limits_{p\in Q_Z}{\Theta}_{E(p)}]\cup[\bigcup\limits_{p\in Q_W }{\ACT }_{A(p),w}] $
2. $Q =Q_{Y}\cup Q_{Z}\cup Q_{W}$,
3. $\trans =\{(p,\alpha,q)\mid p\in Q_y\land p\trans[\alpha]_{yz} q\}\bigcup\{(p,\sigma,q)\mid p\in Q_Z\land\sigma=\alpha_{E(p)} \land p\trans[\alpha]_{zz, zw_1, zw_2} q\}\bigcup\{(p,\sigma,q)\mid p\in Q_W\land\sigma=\alpha_{A(p),w}\land p\trans[\alpha]_{wy_1, wy_2} q\}$
4. $Q^0= y_0$,
5. $Q^m= Q_Y$.
The events labeling outgoing transitions mapped from original $Y$ states in $T$, i.e., $\{(p,\alpha,q)\mid p\in Q_y\land p\trans[\alpha]_{yz} q\}$ and outgoing transitions mapped from original $W$ states in $T$, $\{(p,\sigma,q)\mid p\in Q_W\land\sigma=\alpha_{A(p),w}\land p\trans[\alpha]_{wy_1, wy_2} q\}$ are considered as uncontrollable while the other events are controllable.
In \[def:TOBtoAutone\], $\Sigma_\alpha$ represents that $\alpha$ is added to all the events of $\ACT$. To transform a three-player observer to an automaton, each state of the three-player observer is considered as an automaton state, $Q_i=Q^i_{Y}\cup Q^i_{Z}\cup Q^i_{W}$ in \[def:TOBtoAutone\]. Moreover, the information about the states needs to be considered to distinguish some transitions in the transformed automaton and to have a correct synchronization of three-player observers, since the information about state types ($Y$, $W$, $Z$) is lost in the transformation. To this end, in the transformed automaton, the events labeling the transitions from $z$ to $z$ states, from $z$ to $w$ states and from $w$ to $y$ states, need to have the information of the predecessor states reflected in them. Thus, the events in the transformed automaton have the observable event components of $Z$ states, $\bigcup\limits_{p\in Q_Z^i}{\Theta^i}_{E(p)}$, and the action components of $W$ states, $\bigcup\limits_{p\in Q_W^i}{\ACT^i}_{A(p),w}$. The initial state of the transformed automaton is the initial state of the three-player observer and the marked states are the original $Y$ states.
\[ex:TOBtoAutMonolithic\] Consider the two three-player observers $T'_1$ and $T'_2$ shown in \[fig:autOFtob\]. To transform the three-player observers $T_1$ and $T_2$ to their monolithic automata, renaming $\rho:\{\alpha, \beta, \gamma, \epsilon_\gamma, {\gamma\to\epsilon}_{\gamma}, \gamma_{\gamma\to\epsilon,w}, \gamma_{\epsilon,w}, \gamma_{\alpha}, \epsilon_\beta, {\beta\to\epsilon}_{\beta}, \beta_{\beta\to\epsilon,w}$, $\beta_{\epsilon,w}, \beta_{\alpha}, {\alpha\to\epsilon}_{\alpha}, \alpha_{\alpha\to\epsilon,w}\}\to \{\alpha, \beta, \gamma, \epsilon, \alpha\to\epsilon, \beta\to \epsilon, \gamma\to\epsilon\}$ is introduced. Next, all $Y$ states in $T'_1$ and $T'_2$ are considered as marked. \[fig:autOFtob\] shows $M_1^T$ and $M_2^T$, which are the monolithic transformed automata from $T'_1$ and $T'_2$, respectively. All the events that come from transitions defined out of $Y$ states and $W$ states in $T'_1$ and $T'_2$ are considered as uncontrollable in $G_1^T$ and $G_2^T$, thus $\alpha$, $\beta$ and $\gamma$ are uncontrollable.
This paper describes the compositional approach for modular systems. In order to have a correct interaction between the transformed automata, the transformation of the three-player observers needs to be done in the modular setting.
\[def:TOBtoAut\] Let $\mathcal{T}=\{T_1,\ldots,T_n\}$ be a three-player observer system such that $T_i=\langle Q^i_{Y},Q^i_{Z},Q^i_{W},\ACT^i,\ACT_i^{\epsilon}, \Theta^i, \trans^i_{yz},\trans^i_{zz},\trans^i_{zw},\trans^i_{wy},y^i_0\rangle$. Let $M^T_i=\langle \ACT_{M^T}^i,Q_M^i,\trans_M^i,{Q_M^0}^i,{Q_M^m}^i\rangle$ be the monolithic transformed deterministic automaton of $T_i$, based on \[def:TOBtoAutone\]. The transformed automaton system of $\mathcal{T}$ is ${\mathcal{G}}^T=\{G_1^T,\cdots,G_n^T\}$, where $G^T_i=\langle \ACT_{G^T}^i,Q_i,\trans^i,Q_i^0,Q_i^m\rangle$ and
1. $\ACT_{G^T_i}=\ACT_{M^T}^i\cup[\bigcup\limits_{\alpha\in (\ACT^j\setminus\ACT^i);j\neq i}((\ACT^i\cap \ACT^j)_{\alpha}\cup(\ACT^i\cap \ACT^j)_{\alpha,w}\cup (\ACT^i\cap \ACT^j)_{\alpha\to\epsilon,w})] $
2. $Q_i=Q_M^i$,
3. $\trans_M^i\bigcup
\{(p,\alpha,p)\mid p\in Q_y^i\ \textit{and}\ \alpha\in\bigcup\limits_{j\neq i} (\ACT^j\setminus\ACT^i)\}$
4. $Q_i^0= {Q_M^0}^i$,
5. $Q_i^m= {Q_M^m}^i$.
Since some shared events in the transformed automaton become local after incorporating the extra state information, they need to be added in the alphabet of the transformed automaton, $\bigcup\limits_{\alpha\in (\ACT^j\setminus\ACT^i);j\neq i}((\ACT^i\cap \ACT^j)_{\alpha}\cup(\ACT^i\cap \ACT^j)_{\alpha,w}\cup (\ACT^i\cap \ACT^j)_{\alpha\to\epsilon,w})$ in \[def:TOBtoAut\]. Moreover, the events not defined from $Y$ states of certain TPO $T$ but defined from $Y$ states of some other TPO $T'$ are added as self-loops at the corresponding states in the transformed automaton of $T$, $\{(p,\alpha,p)\mid p\in Q_y^i\ \textit{and}\ \alpha\in\bigcup\limits_{j\neq i} (\ACT^j\setminus\ACT^i)\}$. To create a map between the events of a TPO and its transformed automaton, *renaming* of events is necessary. Note, when the transformation of a single automaton is considered \[def:TOBtoAutone\] and \[def:TOBtoAut\] produce the same results. Thus, in the following wherever a transformed automaton is discussed we refer to \[def:TOBtoAut\]. Renaming $\rho$ simply removes the extra information from the events of the transformed automaton and maps them back to the original events in the TPO. To be more specific, $\rho$ is a map such that $\rho(\alpha_{\sigma})=\alpha$ and $\rho(\alpha)=\alpha$. Table \[table:events\] shows how the events in a transformed automaton are linked to the original events of a TPO, while the third column shows how renaming works. Specifically, in the case where events label $\rightarrow_{yz}$ transitions, renaming does not change events names.
0.4em 0.8pt 0=
0
\[ex:TOBtoAut\]
Consider the abstracted system $\tilde{\SYSG}=\{\tilde{G}_1,\tilde{G}_2\}$, shown in \[fig:autOFtob\]. The sets of secret states are $\tilde{Q}^S_1=\{q_3\}$, $\tilde{Q}^S_2=\{s_3\}$ where all the events are observable. $T'_1$ and $T'_2$ are the largest three-player observers of $\tilde{G}_1$, $\tilde{G}_2$, respectively. In Example \[ex:TOBtoAutMonolithic\] the monolithic transformed automata of $T'_1$ and $T_2'$ were generated. The three-player observer system $\{T'_1, T'_2\}$ is transformed to automata system $\mathcal{G}=\{G^T_1,G^T_2\}$, shown \[fig:autOFtob\], by adding self-loops at the marked states. Event $\beta$ is not in the alphabet of $T'_1$ so it appears as a self-loop at all marked states in $G^T_1$, which correspond to $Y$ states in $T'_1$. Similarly, $\gamma$ is added as a self-loop at marked states in $G^T_2$ since $\gamma$ is not in the alphabet of $T'_2$.
In the following, Theorem \[thm:TOBtoAutSync\] proves that if the synchronization of transformed individual three-player observers contains a transition, then the largest monolithic three-player observer w.r.t. the synchronized system also contains an equivalent transition. However, the inverse is not necessarily true as there are some behaviors in the monolithic three-player observer that are omitted in the modular structure. Before Theorem \[thm:TOBtoAutSync\], Lemma \[propos:syncDet\] [@SeaSilvaSch:12] and Lemma \[lem:detd\] establish that the modular observer and desired observer are isomorphic to their monolithic counterparts.
\[propos:syncDet\][@SeaSilvaSch:12] Let $G_1 = \langle\Sigma_1, Q_1, \rightarrow_1, Q^0_1 \rangle$ and $G_2 =\langle\Sigma_2, Q_2, \rightarrow_2, Q^0_2 \rangle$ be two nondeterministic automata. Then $det(G_1\sync G_2)$ is isomorphic to $det(G_1)\sync det(G_2)$.
\[lem:detd\] Let $G_1 = \langle Q_1, \Sigma_1, \rightarrow_1, Q^0_1\rangle$ and $G_2 = \langle Q_2, \Sigma_2, \rightarrow_2, Q^0_2\rangle$ be two nondeterministic automata with sets of secret states $Q_1^S$ and $Q_2^S$, respectively. Then $det_d(G_1)\sync det_d(G_2)$ is isomorphic to $det_d(G_1\sync G_2)$.
*Proof:* From $det(G_1\sync G_2)$ is isomorphic to $det(G_1)\sync det(G_2)$ it follows that $det(G_1\sync G_2)\trans[s]X$ if and only if $det(G_1)\sync det(G_2)\trans[s](X_1,X_2)$ and $(x_1,x_2)\in X$ if and only if $(x_1,x_2)\in X_1\times X_2$. Now we need to show that $X\not\in X_{obsd}$ if and only if $(X_1,X_2)\not\in X_{1,obsd}\times X_{2,obsd}$.
First assume $X\not\in X_{obsd}$, which means $X\subseteq Q^S$. This further means that for all $(x_1,x_2)\in X$, either $x_1\in Q^S_1$ or $x_2\in Q^S_2$, which implies either $X_1\not\in X_{1,obsd}$ or $X_2\not\in X_{2,obsd}$. Thus, $(X_1,X_2)\not\in X_{1,obsd}\times X_{2,obsd}$.
Now assume $(X_1,X_2)\not\in X_{1,obsd}\times X_{2,obsd}$. This means either $X_1\not\in X_{1,obsd}$ or $X_2\not\in X_{2,obsd}$, which implies either $X_1\subset Q^S_1$ or $X_2\subset Q^S_2$. Hence for all $(x_1,x_2)\in (X_1,X_2)=X$ either $x_1\in Q^S_1$ or $x_2\in Q^S_2$, which implies $X\subseteq Q^S$. Thus, $X\not\in X_{obsd}$.[$\blacksquare$]{}
\[thm:TOBtoAutSync\]
*Proof:* We need to show that a transition is defined in $G^T_1\sync G^T_2$ if the equivalent transition is defined in $T$. It is shown by induction on $n\geq 0$ that $(y_1^0,y_2^0)\trans[s](q_1^n,q_2^n)$ in $G^T_1\sync G^T_2$ implies $y_T^0\trans[\rho(s)]q_T$ in $T$.
Let $G^T_1\sync G^T_2\trans[s](q_1,q_2)$.
*Base case:* $n=0$. Let $(y_1^0,y_2^0)$ be the initial state of $G_1\sync G_2$, i.e., $y_1^0$ (respectively $y^0_2$) is the initial state of $T_1$ (respectively $T_2$). From \[def:tripartite\], $y_1^0=(X^0_{1,obsd},X^0_{1,obs})$ and $y_2^0=(X^0_{2,obsd},X^0_{2,obs})$, where $X^0_{i,obs}$ and $X^0_{i,obsd}$ are the initial state of $det(G_i)$ and $det_d(G_i)$ for $i\in \{1,2\}$, respectively. From Lemmas \[propos:syncDet\] and \[lem:detd\], $det(G_1\sync G_2)$ and $det_d(G_1\sync G_2)$ are isomorphic to $det(G_1)\sync det(G_2)$ and $det_d(G_1)\sync det_d(G_2)$, which implies $(X^0_{1,obs},X^0_{2,obs})$ is the initial state of $det(G_1)\sync det(G_2)$ and $det_d(G_1)\sync det_d(G_2)$. Thus $y_T^0=((X^0_{1,obsd},X^0_{2,obsd}),(X^0_{1,obs},X^0_{2,obs}))$ is the initial state of $T$.
*Inductive step:* Assume the claim holds for some $0\leq n$, which means if $(y_1^0,y_2^0)=(q_1^0,q_2^0)\trans[\sigma_0\ldots\sigma_{n-1}](q_1^n,q_2^n)$ in $G^T_1\sync G^T_2$ then $y^0\trans[\rho(\sigma_0\ldots\sigma_{n-1})]q^n$ in $T$. Now we need to show that if $(q_1,q_2)=(q_1^n,q_2^n)\trans[\sigma_n](p_1,p_2)$ in $G^T_1\sync G^T_2$ then $q=q^n\trans[\rho(\sigma_n)]p$ in $T$. From $(q_1,q_2)=(q_1^n,q_2^n)\trans[\sigma_n](p_1,p_2)$ in $G^T_1\sync G^T_2$ and based on \[def:synch\] it holds that $q_i\trans[\sigma_n]p_i$ in $G^T_i$ for $i\in \{1,2\}$, which means $q_i\trans[\sigma_n]p_i$ in $T_i$ for $i\in \{1,2\}$. Consider the following four cases for all the possible transitions:
- if $\rho(\sigma_n)=\sigma_n$ then based on \[def:TOBtoAut\] it holds that $q_i\trans[\sigma_n]p_i$ is a $yz$ transition in the original $T_i$ such that $E(p_i)=\sigma_n$ and $I(p_i)=q_i$ if $\sigma_n\in\ACT_i$ and $q_i=p_i$ otherwise for $i\in \{1,2\}$. Let $q_i=(x_{i,d},x_{i,f})$ for $i\in \{1,2\}$. Based on \[def:tripartite\] this means $x_{i,f}\trans[\sigma_n]$ in $det(G_i)$ if $\sigma_n\in\ACT_i$. Moreover, based on the inductive assumption there exists $y=(x_{d},x_{f})$ such that $y_T^0\trans[\omega]y$ in $T$, which implies $det(G_1\sync G_2)\trans[P_e(\omega)]x_f$. Since based on Lemma \[propos:syncDet\] $det(G_1\sync G_2)$ and $det(G_1)\sync det(G_2)$ are isomorphic it holds $det(G_1)\sync det(G_2)\trans[P_e(\omega)]$ and string $P_e(\omega)$ reaches $(x_{1,f},x_{2,f})$ in $det(G_1)\sync det(G_2)$. Thus, based on $x_{i,f}\trans[\sigma_n]$ in $det(G_i)$ it can be deduced that $(x_{1,f},x_{2,f})\trans[\sigma_n]$ in $det(G_1\sync G_2)$, it also implies $(x_{1,f},x_{2,f})\trans[\sigma_n]$ in $det(G_1\sync G_2)$ by Lemma \[propos:syncDet\]. This means $q_T^n=q_T\trans[\sigma_n]p_T$ is a $yz$ transition in $T$.
- if $\rho(\sigma_n)= \alpha$ and $\sigma_n=\alpha_{e_o}$. Then based on \[def:TOBtoAut\] there are three possibilities for $q_i\trans[\rho(\sigma_n)]p_i$ in $T_i$: it is a $zz$ transition or a $zw_1$ transition or a $zw_2$ transition and $e_o=E(q_i)$. Now consider the following cases:
- ${\alpha}_{e_o}\in \Sigma_{G_1^T}\setminus \Sigma_{G_2^T}$. This means $q_1\trans[{\alpha}_{e_o}]p_1$ in $G_1^T$, which implies $q_1\trans[{\sigma_n}]p_1$ in $T_1$. From ${\sigma_n}_{e_o}\not\in \Sigma_{G_2^T}$ and \[def:TOBtoAut\] it follows that $\sigma_n\not\in \Sigma_{2}$, which implies $q_2=p_2$. Consider the following three cases: If $q_1\trans[\sigma_n]p_1$ in $T_1$ is a $zz$ transition, then based on \[def:tripartite\] it holds that $x_{1,d}\trans[\sigma_n]x'_{1,d}$ in $det_d(G_1)$. If $q_1\trans[\sigma_n]p_1$ in $T_1$ is a $zw1$ transition, then based on \[def:tripartite\] it holds that $x_{1,d}\trans[\sigma_n]$ in $det_d(G_1)$ and $x_{1,f}\trans[\sigma_n]$ in $det(G_1)$. If $q_1\trans[\sigma_n]p_1$ in $T_1$ is a $zw2$ transition, then based on \[def:tripartite\] it holds that $x_{1,f}\trans[\sigma_n]$ in $det(G_1)$. In all the three cases based on Lemmas \[propos:syncDet\] and \[lem:detd\], where they show $det_d(G_1\sync G_2)$ and $ det_d(G_1)\sync det_d(G_2)$ are isomorphic, it holds that $(x_{1,d},x_{2,d})\trans[\sigma_n](x'_{1,d},x_{2,d})$ in $det_d(G_1\sync G_2)$ and $(x_{1,f},x_{2,f})\trans[\sigma_n]$ in $det(G_1\sync G_2)$. These mean $q^n=q\trans[\sigma_n]p$ in $T$ is a $zz$ transition if $q_1\trans[\sigma_n]p_1$ in $T_1$ is a $zz$ transition, $q^n=q\trans[\sigma_n]p$ is a $zw1$ transition in $T$ if $q_1\trans[\sigma_n]p_1$ is a $zw1$ transition in $T_1$ and $q^n_T=q_T\trans[\sigma_n]p_T$ is a $zw2$ transition in $T$ if $q_1\trans[\sigma_n]p_1$ is a $zw2$ transition in $T_1$.
- ${\sigma_n}_{e_o}\in \Sigma_{G_1^T}\cap \Sigma_{G_2^T}$. This means $q_i\trans[{\sigma_n}_{e_o}]p_i$ in $G_i^T$ for $i\in \{1,2\}$, which implies $q_i\trans[{\sigma_n}]p_i$ in $T_i$ for $i\in \{1,2\}$. Again there are three cases: If $q_i\trans[\sigma_n]p_i$ in $T_i$ for $i=1,2$ is a $zz$ transition. Then based on \[def:tripartite\] it holds that $x_{i,d}\trans[\sigma_n]x'_{i,d}$ in $det_d(G_i)$ for $i=1,2$. If $q_i\trans[\sigma_n]p_i$ in $T_i$ is a $zw1$ transition for $i\in \{1,2\}$. Then based on \[def:tripartite\] it holds that $x_{i,d}\trans[\sigma_n]$ in $det_d(G_i)$ and $x_{i,f}\trans[\sigma_n]$ in $det(G_i)$ for $i=1,2$. If $q_i\trans[\sigma_n]p_i$ in $T_i$ is a $zw2$ transition for $i=1,2$. Then based on \[def:tripartite\] it holds that $x_{i,f}\trans[\sigma_n]$ in $det(G_i)$ for $i=1,2$. In all the three cases based on Lemmas \[propos:syncDet\] and \[lem:detd\] where they show $det_d(G_1\sync G_2)$ and $ det_d(G_1)\sync det_d(G_2)$ are isomorphic, it holds that $(x_{1,d},x_{2,d})\trans[\sigma_n](x'_{1,d},x'_{2,d})$ in $det_d(G_1\sync G_2)$ and $(x_{1,f},x_{2,f})\trans[\sigma_n]$ in $det(G_1\sync G_2)$. These mean $q^n_T=q_T\trans[\sigma_n]p_T$ in $T$ is a $zz$ transition if $q_i\trans[\sigma_n]p_i$ in $T_i$ is a $zz$ transition, $q^n_T=q_T\trans[\sigma_n]p_T$ is a $zw1$ transition in $T$ if $q_i\trans[\sigma_n]p_i$ is a $zw1$ transition in $T_i$ and $q^n_T=q_T\trans[\sigma_n]p_T$ is a $zw2$ transition in $T$ if $q_i\trans[\sigma_n]p_i$ is a $zw2$ transition in $T_i$ for $i\in \{1,2\}$.
- ${\sigma_n}_{e_o}\in \Sigma_{G_2^T}\setminus \Sigma_{G_1^T}$. The same argument as case 1.
- if $\rho^{-1}(\sigma_n)= {\sigma_n}_{e_o,w}$ then based on \[def:TOBtoAut\] there are again three cases:
- ${\sigma_n}_{e_o,w}\in \ACT_{G^T_1}\setminus\ACT_{G^T_2}$. This means $q_1\trans[{\sigma_n}_{e_o,w}]p_1$ in $G^T_1$. From ${\sigma_n}_{e_o,w}\not\in \Sigma_{G_2^T}$ and \[def:TOBtoAut\] it follows that $\sigma_n\not\in \Sigma_{2}$, which implies $q_2=p_2$. From $\rho^{-1}(\sigma_n)= {\sigma_n}_{e_o,w}$ it holds that $p_1\trans[\sigma_n]q_1$ in $T_1$ is a $wy1$ transition. This means $x_{1,d}\trans[\sigma_n]x'_{1,d}$ in $det_d(G_1)$ and $x_{1,f}\trans[\sigma_n]x'_{1,f}$ in $det(G_1)$. Based on Lemmas \[propos:syncDet\] and \[lem:detd\], where they show $det_d(G_1\sync G_2)$ and $ det_d(G_1)\sync det_d(G_2)$ are isomorphic, it holds that $(x_{1,d},x_{2,d})\trans[\sigma_n](x'_{1,d},x_{2,d})$ in $det_d(G_1\sync G_2)$ and $(x_{1,f},x_{2,f})\trans[\sigma_n](x'_{1,f},x_{2,f})$ in $det(G_1\sync G_2)$. This means $q^n_T=q_T\trans[\sigma_n]p_T$ is a $wy1$ transition in $T$.
- ${\sigma_n}_{e_o,w}
\in \ACT_{G^T_1}\cap\ACT_{G^T_2}$. Then it follows that $q_i\trans[{\sigma_n}_{e_o,w}]p_i$ in $G^T_i$ for $i\in \{1,2\}$. This means $q_i\trans[\sigma]p_i$ in $T_i$ is a $wy1$ transition, which implies $x_{i,d}\trans[\sigma_n]x'_{i,d}$ in $det_d(G_i)$ and $x_{i,f}\trans[\sigma_n]x'_{i,f}$ in $det(G_i)$ for $i\in \{1,2\}$. Based on Lemmas \[propos:syncDet\] and \[lem:detd\], where they show $det_d(G_1\sync G_2)$ and $ det_d(G_1)\sync det_d(G_2)$ are isomorphic, it holds that $(x_{1,d},x_{2,d})\trans[\sigma_n](x'_{1,d},x'_{2,d})$ in $det_d(G_1\sync G_2)$ and $(x_{1,f},x_{2,f})\trans[\sigma_n](x'_{1,f},x'_{2,f})$ in $det(G_1\sync G_2)$. This means $q^n_T=q_T\trans[\sigma_n]p_T$ is a $wy1$ transition in $T$.
- ${\sigma_n}_{e_o,w}\in \ACT_{G^T_2}\setminus\ACT_{G^T_1}$. The same argument as case 1.
- if $\rho^{-1}(\sigma_n)= {\sigma_n}_{e_o\to\epsilon,w}$ then the same argument as above, $\rho^{-1}(\sigma_n)= {\sigma_n}_{e_o,w}$, holds. [$\blacksquare$]{}
\[thm:TOBtoAutSync\] shows that synchronization of individual transformed three-player observers is a subsystem of the largest monolithic three-player observer. Specifically, if there is a string $s$ in $G^T_1\sync G^T_2$, then there always exists a corresponding path $\rho(s)$ in $T$. The synchronized automaton form TPOs may not always be equal to the monolithic TPO since some $\rightarrow_{zz}$ transitions may not appear in the synchronized system. This happens when a state in the synchronization of TPOs is a combination of an original $Z$ and an original $Y$ state from individual TPOs, while the observable event component of the original $Z$ state is a local event. However, as there is no difference between local and shared events in the monolithic approach of obtaining TPOs, the largest monolithic TPO contains all possible transitions of edit decisions. The proof of \[thm:TOBtoAutSync\] also illustrates that for every state in the largest monolithic TPO, there exists a corresponding state in the synchronized individual TPOs in automaton form.
(ini)[$T$]{}; (yAA) [$(A,A)$]{}; (zAAc) [$(A,A),\gamma$]{}; (wAAec) [$(A,A),$ $\gamma\to\epsilon$]{}; (wAAe) [$(A,A),$ $\gamma$]{}; (yAC) [$(A,C)$]{}; (yCC) [$(C,C)$]{}; (zCCb) [$(C,C),\beta$]{}; (zACb) [$(A,C),\beta$]{}; (wCCe) [$(C,C),$ $\beta$]{}; (wCCbe) [$(C,C),$ $\beta\to\epsilon$]{}; (wACe) [$(A,C),$ $\beta$]{}; (wACbe) [$(A,C),$ $\beta\to\epsilon$]{};
(zAAb) [$(A,A),\beta$]{}; (wAAeb) [$(A,A),$ $\beta\to\epsilon$]{}; (wAAe2) [$(A,A),$ $\beta$]{}; (yAB) [$(A,B)$]{}; (yBB) [$(B,B)$]{}; (zBBc) [$(B,B),\gamma$]{}; (zABc) [$(A,B),\gamma$]{}; (wBBe) [$(B,B),$ $\gamma$]{}; (wBBce) [$(B,B),$ $\gamma\to\epsilon$]{}; (wABe) [$(A,B),$ $\gamma$]{}; (wABce) [$(A,B),$ $\gamma\to\epsilon$]{};
(yDD) [$(D,D)$]{}; (yBD) [$(B,D)$]{}; (yCD) [$(C,D)$]{}; (yAD) [$(A,D)$]{};
(zDDa) [$(D,D),\alpha$]{}; (zBDa) [$(B,D),\alpha$]{}; (zCDa) [$(C,D),\alpha$]{}; (zADa) [$(A,D),\alpha$]{};
(wDDae) [$(D,D),$ $\alpha\to\epsilon$]{}; (wBDae) [$(B,D),$ $\alpha\to\epsilon$]{}; (wCDae) [$(C,D),$ $\alpha\to\epsilon$]{}; (wADae) [$(A,D),$ $\alpha\to\epsilon$]{};
(yDE) [$(D,E)$]{}; (yBE) [$(B,E)$]{}; (yCE) [$(C,E)$]{}; (yAE) [$(A,E)$]{};
(ini) to node\[above\](yAA); (yAA) to node\[above,xshift=-3pt\][$\gamma$]{}(zAAc); (yAA) to (zAAc); (yAA) to node\[above\](zAAc); (yAA) to (zAAc); (yAA) to node\[above,xshift=3pt\][$\beta$]{}(zAAb); (yAA) to (zAAb); (yAA) to node\[above\](zAAb); (yAA) to (zAAb); (zAAc) to node\[right,xshift=3pt,yshift=5pt\][$\gamma\to\epsilon$]{}(wAAec); (zAAc) to node\[left,xshift=-3pt,yshift=5pt\][$\epsilon$]{}(wAAe); (zAAb) to node\[right,xshift=3pt,yshift=5pt\][$\beta\to\epsilon$]{}(wAAeb); (zAAb) to node\[left,xshift=-3pt,yshift=5pt\][$\epsilon$]{}(wAAe2); (wAAec) to node\[right\][$\gamma$]{}(yAC); (wAAe) to node\[left\][$\gamma$]{}(yCC); (wAAeb) to node\[right\][$\beta$]{}(yAB); (wAAe2) to node\[left\][$\beta$]{}(yBB);
(yBB) to node\[left,xshift=-3pt\][$\gamma$]{}(zBBc); (yBB) to (zBBc); (yBB) to node\[left\](zBBc); (yBB) to (zBBc); (yAB) to node\[right,xshift=3pt\][$\gamma$]{}(zABc); (yAB) to (zABc); (yAB) to node\[left\](zABc); (yAB) to (zABc); (yCC) to node\[left,xshift=-3pt\][$\beta$]{}(zCCb); (yCC) to (zCCb); (yCC) to node\[left\](zCCb); (yCC) to (zCCb); (yAC) to node\[right,xshift=3pt\][$\beta$]{}(zACb); (yAC) to (zACb); (yAC) to node\[left\](zACb); (yAC) to (zACb);
(zABc) to node\[above\][$\beta$]{}(zBBc); (zACb) to node\[above\][$\gamma$]{}(zCCb);
(zBBc) to node\[left,xshift=-3pt,yshift=3pt\][$\epsilon$]{}(wBBe); (zBBc) to node\[right,xshift=3pt,yshift=3pt\][$\gamma\to\epsilon$]{}(wBBce); (zABc) to node\[left,xshift=-3pt,yshift=3pt\][$\epsilon$]{}(wABe); (zABc) to node\[right,xshift=3pt,yshift=3pt\][$\gamma\to\epsilon$]{}(wABce); (zCCb) to node\[left,xshift=-3pt,yshift=3pt\][$\epsilon$]{}(wCCe); (zCCb) to node\[right,xshift=3pt,yshift=3pt\][$\beta\to\epsilon$]{}(wCCbe); (zACb) to node\[left,xshift=-3pt,yshift=3pt\][$\epsilon$]{}(wACe); (zACb) to node\[right,xshift=3pt,yshift=3pt\][$\beta\to\epsilon$]{}(wACbe);
(wBBe) to node\[left,xshift=-29pt,yshift=14pt\][$\gamma$]{}(yDD); (wBBce) to node\[left,xshift=-29pt,yshift=14pt\][$\gamma$]{}(yBD); (wABe) to node\[left,xshift=-30pt,yshift=15pt\][$\gamma$]{}(yCD); (wABce) to node\[left,xshift=-31pt,yshift=19pt\][$\gamma$]{}(yAD); (wCCe) to node\[right,xshift=29pt,yshift=17pt\][$\beta$]{}(yDD); (wCCbe) to node\[right,xshift=-1pt,yshift=12pt\][$\beta$]{}(yCD); (wACe) to node\[right,xshift=40pt,yshift=13pt\][$\beta$]{}(yBD); (wACbe) to node\[right,xshift=20pt,yshift=10pt\][$\beta$]{}(yAD);
(yDD) to node\[left,xshift=-3pt, yshift=4pt\][$\alpha$]{}(zDDa); (yDD) to (zDDa); (yDD) to node\[left\](zDDa); (yDD) to (zDDa); (yBD) to node\[left,xshift=-3pt, yshift=4pt\][$\alpha$]{}(zBDa); (yBD) to (zBDa); (yBD) to node\[left\](zBDa); (yBD) to (zBDa); (yCD) to node\[right,xshift=3pt, yshift=4pt\][$\alpha$]{}(zCDa); (yCD) to (zCDa); (yCD) to node\[right\](zCDa); (yCD) to (zCDa); (yAD) to node\[right,xshift=3pt, yshift=4pt\][$\alpha$]{}(zADa);
(zDDa) to node\[left,xshift=-3pt\][$\alpha\to\epsilon$]{}(wDDae); (zBDa) to node\[left,xshift=-3pt\][$\alpha\to\epsilon$]{}(wBDae); (zCDa) to node\[right,xshift=3pt\][$\alpha\to\epsilon$]{}(wCDae); (zADa) to node\[right,xshift=3pt\][$\alpha\to\epsilon$]{}(wADae);
(wDDae) to node\[left,xshift=-3pt\][$\alpha$]{}(yDE); (wBDae) to node\[left,xshift=-3pt\][$\alpha$]{}(yBE); (wCDae) to node\[right,xshift=3pt\][$\alpha$]{}(yCE); (wADae) to node\[right,xshift=3pt\][$\alpha$]{}(yAE);
(zADa) to node\[left,xshift=-3pt\][$\gamma$]{}(zCDa); (zADa) to node\[left,xshift=-27pt,yshift=-6pt\][$\beta$]{}(zBDa); (zBDa) to node\[right,xshift=-3pt, yshift=-4pt\][$\gamma$]{}(zDDa); (zCDa) to node\[right,xshift=5pt, yshift=4pt\][$\beta$]{}(zDDa);
Although the statement of Theorem \[thm:TOBtoAutSync\] concentrates on the case of two individual systems, the result can be generalized to more than two individual systems. [$\blacksquare$]{}
Notice that Theorem \[thm:TOBtoAutSync\] illustrates that some transitions are “missing” in the synchronized automaton compared with the largest monolithic three-player observer $T$. It further implies that more transitions will be missing if we synchronize more individual TPOs (in automaton form). Actually, we may locate those missing transitions and add them back to the synchronized automaton $\sync^{n}_{i=1}G^T_i$. Specifically, consider states $(q_1, q_2, \cdots, q_n)$ and $(q'_1, q'_2, \cdots, q'_n)$ in $\sync^{n}_{i=1}G^T_i$ such that $q_i, q'_i\in Q^i_Y\cup Q^i_Z$ for all $i$, i.e., every component in those states is either a $Y$ state or a $Z$ state from an individual transformed automaton. Then we add transition $(q_1, q_2, \cdots, q_n)\xrightarrow{\sigma}(q'_1, q'_2, \cdots, q'_n)$ if there exists a set of indexes $\mathbb I\in 2^{\{1, 2, \cdots n\}}$, such that for all $i\in \mathbb I$, $q_i=(x_{i,d}, x_{i, f}), q'_i=(x'_{i, d}, x_{i,f})\in Q^i_Y$ and $x_{i,d}\xrightarrow{\sigma}x'_{i,d}$ in $det_d(G_i)$; while for all $i\notin \mathbb I$, $q_i=q'_i$. Intuitively, the added transition implies that event $\sigma$ may be inserted in the largest monolithic TPO w.r.t. $\sync^{n}_{i=1}G_i$. However, due to the fact that there are no transitions defined from a $Y$ state to another $Y$ state in TPOs, those transitions are missing in $\sync^{n}_{i=1}G^T_i$, which implies the synchronized system $\sync^{n}_{i=1}G^T_i$ may only contain a subset of edit decisions in the largest monolithic TPO.
However, the above mentioned operation may not be preferred in practice since it involves explicitly synchronizing individual TPOs in their automaton form. This is usually not feasible in modular approaches and should be avoided in our Algorithm CA-AES as well. [$\blacksquare$]{}
Finally, the results of this section are formally recapped in Theorem \[pro:trisamerunAfterAbs\], which illustrates that the synchronization of transformed (automaton form) three-player observers w.r.t. individual abstracted systems contain a subset of the transitions of the largest monolithic three-player observer. The proof follows directly from Theorem \[thm:TOBtoAutSync\] and Theorem \[thm:trisamerun\].
\[pro:trisamerunAfterAbs\] Let $G_1$ and $G_2$ be two nondeterministic automaton with sets of secret states $Q_i^S\subseteq Q_i$ and sets of non-secret states $Q_i^{NS}=Q\setminus Q_i^S$ for $i=1,2$. Let $det(G_i)$ and $det_d(G_i)$ be the observer and the desired observer of $G_i$, respectively. Let $\sim_o$ be an opaque observation equivalence on $G_i$ such that $\tilde{G}_i\sim_o G_i$ for $i=1,2$. Let $H_{i,ob}\approx_o det(\tilde{G_i})$ and $H_{i,b}\approx det(\tilde{G_i})$ for $i\in \{1,2\}$, where $\approx_o$ and $\approx$ are opaque bisimulation and bisimulation, respectively. Let $T$ be the largest three-player observer w.r.t. $G_1\sync G_2$ and let $T'_i=\langle Q^i_{Y},Q^i_{Z},Q^i_{W},\ACT_i,\ACT_i^{\epsilon},
\Theta_i, \trans^i_{yz},\trans^i_{zz},\trans^i_{zw},\trans^i_{wy},y^i_0\rangle$ be the largest three-player observer w.r.t. $det_d(H_{i,ob})$ and $H_{i,b}$ for $i\in \{1,2\}$. Let $G^T_i=\langle \ACT_{G^T_i},Q_i,\trans^i,Q_i^0,Q_i^m\rangle$ for $i=1,2$ be the transformed automata of $T'_i$ and let $\rho:(\ACT_{G^T_1}\cup\ACT_{G_2^T})\to (\ACT_1\cup \ACT_1^{\epsilon}\cup
\Theta_1)\cup (\ACT_2\cup \ACT_2^{\epsilon}\cup \Theta_2)$ be a renaming. We have $[G^T_1\sync G^T_2\trans[s](q_1,q_2)]\ttrans[\tau] [T\trans[\rho(s)]q]$.
\[ex:syncTOB\] Consider the system $\SYSG=\{G_1,G_2\}$ shown in \[fig:autOFtob\]. In the first step of the compositional approach the system is abstracted by applying opaque observation equivalence, see Example \[ex:abs\]. The abstracted system $\tilde{\SYSG}=\{\tilde{G}_1,\tilde{G}_2\}$ is shown in \[fig:autOFtob\]. Next, the three player observer of individual components are built. As explained in Example \[ex:TOBtoAut\] the three-player observers of $\tilde{G}_1$ and $\tilde{G}_2$ are $T'_1$ and $T'_2$, respectively, shown in \[fig:autOFtob\]. Moreover, \[fig:autOFtob\] also shows $G^T_1$ and $G_2^T$, the transformed automata of $T'_1$ and $T'_2$, respectively. The largest monolithic three-player observer w.r.t $G_1\sync G_2$ is denoted by $T$ and is shown in \[fig:syncTOB\]. In this particular example, it can be verified that the original and abstracted three-player observers are identical (this need not be true in general); therefore, $T$ in \[fig:syncTOB\] also represents the largest three-player observer w.r.t. $\tilde{G_1}\sync \tilde{G_2}$. In $T$, we have states: $A=\{(q_0,s_0)\}$, $B=\{(q_0,s_1), (q_0,s_2)\}$, $C=\{(q_1,s_0), (q_2,s_0)\}$, $D=\{(q_1,s_1), (q_2,s_1), (q_2,s_2), (q_1,s_2)\}$ and $E=\{(q_3,s_3)\}$.
After synchronizing $G^T_1$ with $G^T_2$, we find that there are some transitions in \[fig:syncTOB\], which do not correspond to any transition in $G^T_1\sync G^T_2$. For example, no transition in $G^T_1\sync G^T_2$ corresponds to the $zz$ transition of $\beta$ from state $(A, B, \gamma)$ to $(B, B, \gamma)$ in \[fig:syncTOB\].
FROM ALL EDIT STRUCTURE CALCULATION TO SUPERVISOR SYNTHESIS {#sec:sup}
===========================================================
So far we have shown that in our compositional and abstraction-based approach, individual components can be abstracted and each largest three-player observer w.r.t. an abstracted component can be calculated individually. Then we transfer those three-player observers (TPOs) to their automaton forms. After that, we have also shown in Theorem \[pro:trisamerunAfterAbs\] that the synchronization of the transformed three-player observers results in a subsystem of the largest monolithic three-player observer up to the renaming of events.
Recall that the All Edit Structure (AES) is obtained after pruning deadlocking states from the largest TPO. Here the modular structure of the transformed TPO is kept and the calculation of a “Modular Edit Structure” can be done by mapping this problem to a modular nonblocking supervisory control problem under full observation.
As was discussed at the end of Section \[sec:edit-fn\], we pursue an approach to convert the pruning process (from the largest TPO to the AES) to a supervisory control problem. In this setting, the plant is a collection of automata transformed from individual largest TPOs obtained at the end of step (iv) of Algorithm CA-AES. The specification is the automaton form of the edit constraint. The constraint of having up to $n+1$ consecutive erasures can be modeled by a specification automaton with $n$ states where transitions are labeled by the decision events and all states are marked except the last state, which is a blocking state. After $n$ consecutive event erasures, the next transition of event erasure $\alpha\to \epsilon_\alpha$ leads the specification forward to a blocking state. If the next event is a non-erasure event, it leads the specification back to the initial state, thus resetting the sequence of erasures. Since we have a *modular* representation of the plant, we are able to leverage computationally efficient compositional techniques for modular nonblocking supervisory control problems.
\[def:spec\]
Let $\mathcal{T}=\{T_1,\ldots,T_n\}$ be a three-player observer system where $T_i=\langle Q^i_{Y},Q^i_{Z},Q^i_{W},\ACT^i,\ACT_i^{\epsilon}, \Theta^i, \trans^i_{yz},\trans^i_{zz},\trans^i_{zw},\trans^i_{wy},y^i_0\rangle$ and let $\Phi$ be the edit constraint on $\mathcal{T}$ such that there are not $n+1$ consecutive event erasures. Then $K=\langle \ACT_{K}, Q_K, \trans_K, Q_K^0, Q^K_m\rangle$ is the automaton form of $\Phi$ where,
- $Q_K=\{x_1,\ldots,x_n\}$
- $\trans_K=\bigcup_{1\leq i\leq n-1}\{(x_i,\alpha\to\epsilon_{\alpha},x_{i+1})\mid
p\trans[\alpha\to\epsilon]_{zw2}q\ \textnormal{and} \ E(p)=\alpha\}\cup
\bigcup_{1\leq i\leq n-2}\{(x_{i+1},\epsilon_\alpha,x_{1})\mid
p\trans[\epsilon]_{zw1}q\ \textnormal{and}\ E(p)=\alpha\}\cup
\bigcup_{1\leq i\leq n-2}\{(x_{i+1},\alpha_\sigma,x_{1})\mid
p\trans[\alpha]_{zz}q\ \textnormal{and}\ E(p)=\sigma\}$ $\cup
\{(x_{1},\epsilon_\alpha,x_{1})\mid
p\trans[\epsilon]_{zw1}q\ \textnormal{and}\ E(p)=\alpha\}\cup
\{(x_{1},\alpha_\sigma,x_{1})\mid
p\trans[\alpha]_{zz}q\ \textnormal{and}\ E(p)=\sigma\}$,
- $Q_K^0=x_1$
- $Q_m^K=\{x_1,\ldots,x_{n-1}\}$
![Automaton $K$ is the automaton form of the constraint $\Phi$ in Example \[ex:spec\], $S$ and $SP$ ares the supervisor and the selected path in Example \[ex:final\]](specSup.pdf){width="0.85\columnwidth"}
.\[fig:sup\]
\[ex:spec\] Consider the transformed system $\SYSG^T=\{G_1^T,G_2^T\}$ shown in \[fig:autOFtob\]. Assume the constraint $\Phi$ only allows one erasure. The specification automaton of this constraint is shown in \[fig:sup\] as $K$. Automaton $K$ has three states. As there is no constraint on event insertion, the events related with event insertion just form self-loops at the initial state of $K$. On the other hand, by executing $\alpha\to\epsilon_\alpha$, $\beta\to\epsilon_\beta$ or $\gamma\to\epsilon_\gamma$ the specification transits from $x_1$ to $x_2$. Next, at $x_2$ if the edit decision is to certain events, then the system goes back to the initial state $x_1$, thus allowing more event erasures since there are no consecutive erasures. However, if another event erasure occurs from $x_2$, the system goes to the blocking state $x_3$.
The following theorem establishes that the three-player observer of the system under constraint $\Phi$ and the transformed system synchronized with the specification $K$ have the same runs up to a renaming of the events.
\[thm:spec\] Let $G=\langle \Sigma, Q, \rightarrow, Q^0 \rangle$ be a nondeterministic automaton with the set of secret states $Q^S\subseteq Q$. Let $T'=\langle Q'_{Y},Q'_{Z},Q'_{W},\ACT,\ACT^{\epsilon},
\Theta, \trans'_{yz},\trans'_{zz},\trans'_{zw},\trans'_{wy},y'_0\rangle$ be the largest three-player observer of $G$ under the edit constraint $\Phi$ which prohibits $n+1$ consecutive event erasures. Let $T=\langle Q_{Y},Q_{Z},Q_{W},\ACT,\ACT^{\epsilon},
\Theta, \trans_{yz},\trans_{zz},\trans_{zw},\trans_{wy},y_0\rangle$ be the largest three-player observer w.r.t. $G$ when no constraint is considered and let $G^T=\langle\Sigma_{G^T}, Q, \rightarrow, Q^0 \rangle$ be the automaton transformation of $T$. Let $K$ be the specification automaton of $\Phi$. Then $G^T\sync K\trans[s]$ if and only if $T'\trans[\rho(s)]$.
*Proof:* Clearly $T'\sqsubseteq T$ and $\ACT_K\subseteq \ACT_{G^T}$. First let $G^T\sync K\trans[s](q_G,q_K)\trans[\sigma](p_G,p_K)$, let $P_K:\ACT_{G^T}\to\ACT_K$ and let $P_E:\ACT_{G^T}\to \bigcup_{\sigma\in\ACT}\{\sigma\to\epsilon_\sigma\in\ACT_K\}$ be a map that removes all the events except event erasures from $\ACT_K$. From $G^T\sync K\trans[s](q_G,q_K)\trans[\sigma](p_G,p_K)$ and $\ACT_K\subseteq \ACT_{G^T}$ and \[def:synch\], it holds that $G^T\trans[s]q_G\trans[\sigma]p_G$ and $K\trans[P_K(s)]q_K\trans[P_K(\sigma)]p_K$. Now consider three cases:
- $q_K=x_i$, $p_K=x_{i+1}$ and $p_K, q_K\in Q_m^K$. Then $|P_E(s)|<n$. This implies $\sigma$ is an erasure event but there are not $n$ consecutive erasures in $s$, so $T'\trans[\rho(s)]q_T\trans[\rho(\sigma)]p_T$.
- $q_K=x_i$, $p_K=x_{i+1}$, $q_K\in Q_m^K$ but $p_K\not\in Q_m^K$. Then $|P_E(s)|=n$ and $K\trans[P_K(s)]q_K\trans[\sigma]p_K$, which implies $G^T\sync K\trans[s\sigma](p_G,p_K)$ and $(p_G,p_K)$ is a blocking state. This further indicates there are $n$ consecutive erasures in $s$,so $T'\trans[\rho(s)]q_T\not\trans[]$.
- $q_K=x_i$, $p_K=x_{1}$ and $p_K, q_K\in Q_m^K$. This implies $\sigma$ is a non-erasure event, so $T'\trans[\rho(s)]q_T\trans[\rho(\sigma)]p_T$.
Now assume $T'\trans[\rho(s)]q_T$. This means $T\trans[\rho(s)]q_T$, which implies $G^T\trans[s]q_T$. Consider two cases. 1. $|P_E(s)|<n$. Then $K\trans[P_K(s)]q_K$, which implies $G^T\sync K\trans[s](q_G,q_K)$. 2. $|P_E(s)|=n$. This means $T'\trans[\rho(s)]q_T\not\trans[]$ and $K\trans[P_K(s)]q_K\not\trans[]$, which implies $G^T\sync K\trans[s](q_G,q_K)\not\trans$.[$\blacksquare$]{}
The following theorem shows that equivalent states are removed in solving the supervisory control problem to obtain an automaton satisfying the edit constraint and in the pruning process to obtain the AES from the largest TPO.
\[thm:supervisor\]
*Proof:* The pruning process to obtain the AES and the supervisor synthesis procedure are both iterative, which remove states at each iteration. We will show by induction that at each iteration a state is removed in $G^T$ by the supervisory control synthesis procedure if and only if the corresponding state is removed by the pruning process from $T$. *Base case:* Clearly $G^T$ and $T$ have the same transition relation.
*Inductive step:* Assume the claim holds for some $n>0$. We let $X^n_G$ be the state space of $G^T$ at the $n$-th iteration of supervisor synthesis process and $X^n_T$ be the state set of $T$ at the $n$-th iteration of pruning process. Then $S\trans[\rho^{-1}(s)]q \Leftrightarrow AES\trans[\rho(s)]q$ holds for all $q\in X^n_G$ ($q\in X^n_T$). Now we need to show that $X^{n+1}_G$ and $X^{n+1}_T$ are also equal. Assume $G^T\trans[s]q$, which implies $T\trans[\rho(s)]q$ based on \[def:TOBtoAut\].
$(\ttrans)$ First we show that if $q\not\in X^{n+1}_T$, which means $q$ is removed by the pruning process at the $n$-th iteration, so $q\not\in X^{n+1}_G$. Then $q\not\in X^{n+1}_T$ if it is a deadlock state or it is a $Y$ state and there exists $e_o\in \ACT$ such that $q\trans[e_o]z'$, where $z'$ is a deadlock $Z$ state. If $q$ is a deadlock state then based on \[def:tripartite\] it holds that $q$ is either a $W$ state or a $Z$ state. Thus, consider the following three case:
- $q$ is a $W$ state. Then $\not\exists e_o\in\ACT$ such that $q\trans[e_o]$. Then based on \[def:TOBtoAut\] it holds that either $q\not\trans[{e_o}_{e_o}]$ or $q\not\trans[{e_o}_{e_o\to\epsilon}]$ in $G^T$ either, which means $q\not\trans$. Thus, $q$ is a blocking state in $G^T$ and $q\not\in \Theta^{nonb}(X^n_G)$.
- $q$ is a $Z$ state. If $q$ is a deadlock state then $\not\exists\theta\in \Theta$ such that $q\trans[\theta]z'$ or $q\trans[\theta]w$ in $T$. Then based on \[def:TOBtoAut\] it holds that either $q\trans[\theta_{E(q)}]z'$ or $q\trans[\theta_{E(q)}]w$ does not exist in $G^T$, which means $q\not\trans[]$. Thus, $q$ is a blocking state in $G^T$ and $q\not\in \Theta^{nonb}(X^n_G)$.
- $q$ is a $Y$ state and $q\trans[e_o]z'$, where $z'$ is a deadlock $Z$ state. Then based on \[def:TOBtoAut\] it holds that $q\trans[e_o]z'$ in $G^T$ and $e_o\in\ACT_u$. As it was shown above if $z'$ is a deadlock state in $T$ then $z'$ is also a deadlock state in $G^T$, thus removed by the supervisor synthesis procedure. If $z'$ is removed, i.e, $z'\not\in X_G^n$, then $q\not\in\Theta^{cont}(X^n_G)$.
Thus, if $q$ is removed in the pruning of $T$, then $q$ is also removed from $G^T$ in supervisor synthesis.
$(\Leftarrow)$ Now we show that if $q\not\in\Theta^{nonb}(X^n_G)\cap\Theta^{cont}(X^n_G)$, then $q\not\in X^{n+1}_T$ and $q$ needs to be removed from $T$ by the pruning process. If $q\not\in\Theta^{nonb}(X^n_G)\cap\Theta^{cont}(X^n_G)$, there are two cases:
- $q\not\in\Theta^{nonb}(X^n_G)$. Then it holds that $q$ is a blocking state, which means $q\trans[\alpha]$ does not exists in $G^T$. There are three possibilities for $q$, it can be a $Y$, $Z$ or $W$ state. If $q$ is a $Y$ state in $T$ then $q\in Q^m$ in $G$, which means $q\in \hat{\Theta}^{nonb}$, which contradicts the assumption. Thus, $q$ can only be a $W$ or $Z$ state in $T$. In both cases from $q\not\trans[\alpha]$ in $G^T$ and \[def:TOBtoAut\], it holds that $\not\exists\rho(\alpha)\in \Theta$ such that $q\trans[\rho(\theta)]$ in $T$. This means $q$ is a deadlock state in $T$ and $q\not\in X^{n+1}_T$
- $q\not\in\Theta^{cont}(X^n_G)$. This means that $q\trans[u]z$ in $G^T$ such that $u\in \ACT_u$ and $z\not\in X^n_G$. Based on \[def:TOBtoAut\], this means that $q$ is a $Y$ state in $T$ and $z$ is a $Z$ state. It was shown above that if $z\not\in X^n_G$ then $z\not\in X^n_T$, which means $z$ is a deadlock $Z$ state. Thus $q$ will be removed by pruning process, which means $q\not\in X^{n+1}_T$.
Thus, if $q$ is removed by synthesis from $G^T$ then $q$ is also removed from $T$ by pruning.[$\blacksquare$]{}
Theorem \[thm:supervisor\] proves that when it comes to imposing the edit constraint, the pruning process from the largest TPO to the AES removes equivalent states with the synthesis procedure of a supremal supervisor. Hence no information is lost when we apply the supervisory control approach to enforce the edit constraints and obtain edit functions. This result is essential to show that the transformation of the TPO to an equivalent automaton and the constraint $\Phi$ to specification $K$ is correctly done in \[def:TOBtoAutone\] and \[def:spec\], respectively. The next step is to consider the modular representation of the system. In that case, we will use the transformation in \[def:TOBtoAut\], which results in a set of automata transformations of the individual three-player observers, with necessary self-loops to capture the synchronization among the components. Finally, we combine the results about abstraction and decomposition, which results in Theorem \[thm:complete\].
\[thm:complete\] Let $\SYSG=\{G_1,\ldots,G_n\}$ be a modular nondeterministic system with sets of secret states $Q_i^S$. Let $AES$ be the All Edit Structure of $\SYSG$ under constraint $\Phi$. Let $det(G_i)$ and $det_d(G_i)$ be the observer and the desired observer of $G_i$, respectively. Let $\sim_o$ be an opaque observation equivalence on $G_i$ such that $\tilde{G_i}\sim_o G_i$ for $i=1,\cdots,n$. Let $H_{i,ob}\approx_o det(\tilde{G_i})$ and $H_{i,b}\approx
det(\tilde{G_i})$ for $i=1,\cdots,n$, where $\approx_o$ and $\approx$ are opaque bisimulation and bisimulation, respectively. Let $T'_i$ be the largest three-player observer of $det_d(H_{i,ob})$ and $H_{i,b}$ for $i=1,\cdots,n$ with the event set $\ACT_{i,T}$. Let $G_i^T$ be the transformed automaton of $T'_i$ and $K$ be the automaton specification. Let $P_e:\Omega\to\LANG(\SYSG)$ be an edit projection and $l(\omega)$ be a string generated by run ( \[def:editproj\] and \[def:string-run\]). Let $\SYSS$ be the least restrictive controllable and nonblocking supervisor calculated from $\{G^T_{1},\ldots,G_{n}^T, K\}$ and let $\rho:(\ACT_{G_1^T}\cup\cdots\cup\ACT_{G_n^T})
\to (\ACT_{1,T}\cup\cdots\cup\ACT_{n,T})$ be the renaming map. Then \[$\forall s\in \LANG(\SYSG)$, $\exists t\in\LANG(\SYSS)$: $P_e(\rho(t))=s$\]$\Rightarrow$ \[$l(\rho(t))=f_e(s)$ where $f_e\in AES$\].
*Proof:* The proof follows directly from Theorems \[thm:supervisor\] and \[thm:spec\], in combination with \[pro:trisamerunAfterAbs\].[$\blacksquare$]{}
\[thm:complete\] essentially shows the proof for all the steps shown in \[fig:composAlg\]. The theorem shows that Algorithm CA-AES correctly synthesizes edit functions for opacity enforcement in a modular form, therefore, the algorithm is sound. It also reveals that the problem of calculating the modular representation of the All Edit Structure can be transformed to synthesizing modular supervisors. The advantage of such a transformation is that we may leverage various existing approaches for calculating a modular supremal nonblocking supervisor in the literature; see, e.g., [@feng2008supervisory; @schmidt2012efficient; @mohajerani2014framework; @mohajerani2017compositional]. Therefore, we can obtain a modular representation of the All Edit Structure, which is noticeably efficient to compute. Then we may synchronize individual components in the Modular Edit Structure, which results in a subsystem of the monolithic AES. However, as was pointed out in Section \[sec:AES\], some edit decisions are omitted after the synchronization. In practice, it is usually desired to retain the modular structure and extract an edit function from it, much in the same way as a set of modular supervisors control a plant. The extraction process is explained next.
Each step of extracting a valid edit decision is described in \[fig:selectingFe\]. Here the edit function is an interface between the system’s output and the outside environment. Assume that the system outputs event $\gamma$, then the edit function makes an edit decision for that event and the edited string will be output to the external observers.
Specifically, this process contains the following steps. $(1)$ when $\gamma$ is received by the edit function, all the components of the Modular Edit Structure are in states that correspond to $Y$ states of the All Edit Structure. $(2)$ At these states, event $\gamma$ is executed and states of all the components in the Modular Edit Structure are updated simultaneously. After the execution of $\gamma$, each component of the Modular Edit Structure is at a state that corresponds to a $Z$ state of the All Edit Structure. $(3)$ Then assume there are multiple transitions defined out of such a current state, we need to select one common transition which corresponds to a specific edit decision and can be viewed as making a control decision from the current state. Note that as the selected transition needs to be accepted by all the components, thus it may happen spontaneously in all the components of the system. The solution of the modular supervisory control problem guarantees the existence of such a common transition out of the current $Z$ states.
Algorithm CA-AES returns the edited string $\rho(\sigma_0\ldots\sigma_k)$ for event $\gamma$ when every component of the Modular Edit Structure reaches a new state corresponding to a $Y$ state of the AES. At that point, the Modular Edit Structure is ready to process the next event output by the system, and the above steps repeat. Meanwhile, the algorithm keeps track of the states of the Modular Edit Structure as its components evolve. Based on \[thm:supervisor\], the edited string $\rho(\sigma_0\ldots\sigma_k)$ is accepted by the monolithic AES. This finally confirms that the extracted edit decision from the Modular Edit Structure corresponds a valid edit decision in the monolithic AES. The above process is illustrated in the following example.
\[ex:final\] Consider the nondeterministic system $\SYSG=\{G_1,G_2\}$ shown in \[fig:autOFtob\]. As it was shown in Example \[ex:abs\], the system can be abstracted using opaque observation equivalence. After abstraction, the system becomes deterministic, which means there is no need to calculate the observers of $G_1$ and $G_2$. The largest three-player observers of $\tilde{G_1}$ and $\tilde{G_2}$ are $T'_1$ and $T'_2$, respectively, shown in \[fig:autOFtob\]. Next, the three-player observers are transformed to automata $G_1^T$ and $G_2^T$ shown in \[fig:autOFtob\], as explained in Example \[ex:TOBtoAut\].
Assume the user adds an edit constraint such that only one consecutive erasure is allowed as in Example \[ex:spec\]. The specification automaton of this constraint is $K$, shown in \[fig:sup\]. Due to this constraint, the $Y$ states $(A,D)$ and $(B,E)$ are considered undesired states in $T$, shown in \[fig:syncTOB\] and they should not be reached when we synthesize edit functions. Since $(A, D)$ is not allowed, its successor states $(A, D, \alpha)$, $(A, D, \alpha\rightarrow\epsilon)$ and $(A, E)$ become unreachable from the initial state $(A, A)$. Those three states together with $(A,D)$ and $(B,E)$ are drawn in dashed lines in \[fig:syncTOB\] and are to be removed in the next step. Furthermore, states $(B, D, \alpha \rightarrow \epsilon)$, $(A, C, \beta\rightarrow \epsilon)$ and $(A, B, \gamma\rightarrow \epsilon)$ become deadlocking after $(A,D)$ and $(B,E)$ are removed. They are drawn in dotted lines in \[fig:syncTOB\] and are also to be removed.
Following the compositional approach with supervisor reduction presented in [@su2004supervisor; @mohajerani2017compositional], we calculate a least restrictive and nonblocking supervisor for the transformed automaton, which is shown in \[fig:sup\] as automaton $S$. All the paths accepted by this supervisor represent valid edit decisions. Consider the accepted path $SP$ shown in \[fig:sup\], it corresponds to an edit function’s decisions for string $\gamma\beta\alpha$ such that $f_e(\gamma\beta\alpha)=\gamma(\gamma\to\epsilon)\gamma\beta\epsilon\beta\alpha\gamma(\alpha\to\epsilon)\alpha$, which is shown by thick lines in $T$, \[fig:syncTOB\]. As is seen, $\rho(SP)=f_e(\gamma\beta\alpha)$. Specifically, when event $\gamma$ is output by the system, $SP$ returns $\gamma({\gamma\to\epsilon}_\gamma)(
\gamma_{\gamma\to\epsilon,w})$, which means erasing the $\gamma$. Next, event $\beta$ is output by the system and it is unchanged according to $SP$. Finally, $\alpha$ is output by the system and $SP$ returns $\alpha(\gamma_\alpha)(\alpha\to\epsilon_\alpha)(\alpha_{\alpha\to\epsilon,w})$, which means erasing $\alpha$ and inserting $\gamma$. Similarly, we may consider other paths accepted by the supervisor in \[fig:sup\] to track edit decisions on other strings, then we have a complete picture of how an edit function works.
From the result in Theorem \[thm:TOBtoAutSync\], our modular algorithm CA-AES results in fewer edit decisions compared with the monolithic approach in [@ji2019synthesis], due to the synchronization process in Section \[sec:AES\]. This indicates that our method may not be complete in the sense that even if Algorithm CA-AES does not return any modular form edit functions, the monolithic approach may still return valid edit functions. This may be viewed as the tradeoff of reducing computational complexity by the modular method. [$\blacksquare$]{}
CONCLUSION {#sec:conclusion}
==========
This paper investigated a compositional and abstraction-based approach to synthesize edit functions for opacity enforcement in a modular setting, given a set of individual systems. The edit functions modify the system’s output by inserting and erasing events, under the constraint of limited number of event erasures. The Three-Player Observer (TPO) and All Edit Structure (AES) proposed in our prior work were employed here; these discrete structures embed edit functions and reflect the constraints. The monolithic approach first synchronizes all individual systems, then calculates the monolithic AES to obtain edit functions. In contrast, the compositional approach first exploits the modular structure and builds individual TPOs. Then, it incorporates the edit constraint and calculates the Modular Edit Structure in a nonblocking modular supervisory control manner to obtain edit functions. In addition, we also applied abstraction methods to reduce the state space of the system before opacity enforcement. We showed that the abstraction processes preserve opacity. Combining system composition and abstraction, we proposed an efficient approach to enforce opacity for complex systems containing multiple components.
[10]{}
B. B[é]{}rard, K. Chatterjee, and N. Sznajder. Probabilistic opacity for [M]{}arkov decision processes. , 115(1):52–59, 2015.
J. W. Bryans, M. Koutny, L. Mazar[é]{}, and P. Y. A. Ryan. Opacity generalised to transition systems. , 7(6):421–435, 2008.
J. W. Bryans, M. Koutny, and P.Y.A. Ryan. Modelling opacity using [P]{}etri nets. , 121:101–115, 2005.
C. G. Cassandras and S. Lafortune. . Springer, 2008.
F. Cassez, J. Dubreil, and H. Marchand. Synthesis of opaque systems with static and dynamic masks. , 40(1):88–115, 2012.
S. Ch[é]{}dor, C. Morvan, S. Pinchinat, and H. Marchand. Diagnosis and opacity problems for infinite state systems modeled by recursive tile systems. , 25(1-2):271–294, 2015.
J. Chen, M. Ibrahim, and R. Kumar. Quantification of secrecy in partially observed stochastic discrete event systems. , 14(1):185–195, 2017.
P. Darondeau, H. Marchand, and L. Ricker. Enforcing opacity of regular predicates on modal transition systems. , 25(1-2):251–270, 2015.
J. Dubreil, P. Darondeau, and H. Marchand. Supervisory control for opacity. , 55(5):1089–1100, 2010.
Y. Falcone and H. Marchand. Enforcement and validation (at runtime) of various notions of opacity. , 25(4):531–570, 2015.
L. Feng and W. M. Wonham. Supervisory control architecture for discrete-event systems. , 53(6):1449–1461, 2008.
J.-C. Fernandez. An implementation of an efficient algorithm for bisimulation equivalence. , 13(2-3):219–236, 1990.
H. Flordal, R. Malik, M. Fabian, and K. [Å]{}kesson. Compositional synthesis of maximally permissive supervisors using supervision equivalence. , 17(4):475–504, 2007.
L. H[é]{}lou[ë]{}t, H. Marchand, and L. Ricker. Opacity with powerful attackers. In [*Proceedings of the 14th IFAC International Workshop on Discrete Event Systems*]{}, pages 475–482, 2018.
C. A. R. Hoare. Communicating sequential processes. , 21(8):666–677, 1978.
R. Jacob, J.-J. Lesage, and J.-M. Faure. Overview of discrete event systems opacity: Models, validation, and quantification. , 2016.
Y. Ji and S. Lafortune. Enforcing opacity by publicly known edit functions. In [*Proceedings of the 56th IEEE Conference on Decision and Control*]{}, pages 4866–4871, 2017.
Y. Ji, Y.-C. Wu, and S. Lafortune. Enforcement of opacity by public and private insertion functions. , 93:369––378, 2018.
Y. Ji, Yin. X, and S. Lafortune. Enforcing opacity by insertion functions under multiple energy constraints. , accepted, 2019.
Y. Ji, X. Yin, and S. Lafortune. Opacity enforcement using nondeterministic publicly-known edit functions. , DOI: 10.1109/TAC.2019.2897553, to appear, 2019.
C. Keroglou and C. N. Hadjicostis. Probabilistic system opacity in discrete event systems. , 28(2):289–314, 2018.
F. Lin. Opacity of discrete event systems and its applications. , 47(3):496–503, 2011.
R. Malik and R. Leduc. Compositional nonblocking verification using generalized nonblocking abstractions. , 58(8):1891–1903, 2013.
T. Masopust and X. Yin. Complexity of detectability, opacity and [A]{}-diagnosability for modular discrete event systems. , 101:290–295, 2019.
R. Milner. , volume 84. Prentice Hall New York, 1989.
S. Mohajerani, Y. Ji, and S. Lafortune. Efficient synthesis of edit functions for opacity enforcement using bisimulation-based abstractions. In [*Proceedings of the 57th IEEE Conference on Decision and Control*]{}, pages 3573–3578, 2018.
S. Mohajerani and S. Lafortune. Transforming opacity verification to nonblocking verification in modular systems. , 2018, under review.
S. Mohajerani, R. Malik, and M. Fabian. A framework for compositional synthesis of modular nonblocking supervisors. , 59(1):150–162, 2014.
S. Mohajerani, R. Malik, and M. Fabian. Compositional synthesis of supervisors in the form of state machines and state maps. , 76:277–281, 2017.
J. Mullins and M. Yeddes. Opacity with orwellian observers and intransitive non-interference. In [*Proceedings of the 12th IFAC International Workshop on Discrete Event Systems*]{}, pages 344–349, 2014.
M. Noori-Hosseini, B. Lennartson, and C. Hadjicostis. Compositional visible bisimulation abstraction applied to opacity verification. In [*Proceedings of the 14th IFAC International Workshop on Discrete Event Systems*]{}, pages 434–441, 2018.
J. JMM Rutten. Automata and coinduction (an exercise in coalgebra). In [*Intl. Conference on Concurrency Theory*]{}, pages 194–218, 1998.
A. Saboori and C. N. Hadjicostis. Verification of infinite-step opacity and complexity considerations. , 57(5):1265–1269, 2012.
A. Saboori and C.N. Hadjicostis. Notions of security and opacity in discrete event systems. In [*Proceedings of the 46th IEEE Conference on Decision and Control*]{}, pages 5056–5061. IEEE, 2007.
A. Saboori and C.N. Hadjicostis. Opacity-enforcing supervisory strategies via state estimator constructions. , 57(5):1155–1165, 2012.
A. Saboori and C.N. Hadjicostis. Verification of initial-state opacity in security applications of discrete event systems. , 246:115–132, 2013.
K. W. Schmidt and J. Eduardo R. Cury. Efficient abstractions for the supervisory control of modular discrete event systems. , 57(12):3224–3229, 2012.
C. Seatzu, M. Silva, and J. H. van Schuppen. . SPRINGER, 2012.
R. Su, J. H. van Schuppen, and J. E. Rooda. Model abstraction of nondeterministic finite-state automata in supervisor synthesis. , 55(11):2527–2541, 2010.
R. Su and W. M. Wonham. Supervisor reduction for discrete-event systems. , 14(1):31–53, 2004.
S. Takai and Y. Oka. A formula for the supremal controllable and opaque sublanguage arising in supervisory control. , 1(4):307–311, 2008.
Y. Tong, Z. Li, C. Seatzu, and A. Giua. Decidability of opacity verification problems in labeled petri net systems. , 80:48–53, 2017.
Y. Tong, Z. Li, C. Seatzu, and A. Giua. Verification of state-based opacity using [P]{}etri nets. , 62(6):2823–2837, 2017.
W. M. Wonham and K. Cai. . SPRINGER, 2019.
B. Wu and H. Lin. Privacy verification and enforcement via belief abstraction. , 2(4):815–820, 2018.
Y.-C. Wu and S. Lafortune. Synthesis of insertion functions for enforcement of opacity security properties. , 50(5):1336–1348, 2014.
Y.-C. Wu, V. Raman, B. C. Rawlings, S. Lafortune, and S. A. Seshia. Synthesis of obfuscation policies to ensure privacy and utility. , 60(1):107–131, 2018.
X. Yin and S. Lafortune. A uniform approach for synthesizing property-enforcing supervisors for partially-observed discrete-event systems. , 61(8):2140–2154, 2016.
X. Yin and S. Lafortune. A new approach for the verification of infinite-step and [K]{}-step opacity using two-way observers. , 80:162–171, 2017.
X. Yin and S. Lafortune. A general approach for optimizing dynamic sensor activations for discrete event systems. , 105:376–383, 2019.
X. Yin and S. Li. Verification of opacity in networked supervisory control systems with insecure control channels. In [*Proceedings of 57th IEEE Conference on Decision and Control*]{}, pages 4851–4856, 2018.
X. Yin, Z. Li, W. Wang, and S. Li. Infinite-step opacity and [K]{}-step opacity of stochastic discrete-event systems. , 99:266–274, 2019.
B. Zhang, S. Shu, and F. Lin. Maximum information release while ensuring opacity in discrete event systems. , 12(3):1067–1079, 2015.
K. Zhang, X. Yin, and M. Zamani. Opacity of nondeterministic transition systems: A (bi) simulation relation approach. , DOI: 10.1109/TAC.2019.2908726, to appear, 2019.
[^1]: The work of the first author was supported by the Swedish Research Council. The work of the second and third authors was supported in part by US NSF grants CNS-1421122 and CNS-1738103.
[^2]: Sahar Mohajerani, Yiding Ji and Stéphane Lafortune are with the Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, USA. [{saharm;jiyiding;[email protected]}]{}
[^3]: Since the base case of the induction is proven for $n=0$, $X^0\trans[\epsilon]$, the inductive step is considered true for $0\leq k< n$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Precise and elegant coordination of a prosthesis across many degrees of freedom is highly desired for rehabilitation of people with limb deficiency. Processing the electrical neural signals, collected from the surface of the remnant muscles of the stump, is a common way to activate certain function of the artificial limb. Based on the assumption that there are distinguishable and repeatable signal patterns among different types of muscular activation, the problem of the prosthesis control reduces to the pattern recognition. Widely accepted classical methods for pattern recognition, however, can not provide simultaneous and proportional control of the artificial limb. Here we show that quantum information processing of the neural signals allows us to overcome above difficulties suggesting a very simple scheme for myoelectric control of artificial limb with advanced functionalities.'
author:
- Michael Siomau
- Ning Jiang
title: Myoelectric Control of Artificial Limb by Quantum Information Processing
---
Information embedded within the electrical neural pulses, controlling muscle contractions in our body, can be extracted from surface or intramuscular myoelectric signals, which are typically summarized into so-called electromyogram (EMG). While acquisition of intramuscular signals could cause serious ethical concerns and may lead to infection due to its invasive procedure, for the last forty years, surface EMG signals have been the only source powered prosthesis control [@Jiang:12]. Currently, the commercial prostheses utilize simple processing of the surface electromyogram, and can provide very limited functionalities [@Parker:04]. To improve the functionality of myoelectrically controlled prosthesis, pattern classification algorithms for surface EMG have been extensively investigated in the academia. It has been shown, for instance, that with properly selected features and classifiers, one can achieve very high classification accuracy (more than 10 classes of movements with less than 5% of classification error) [@Scheme:11].
The academic success of myoelectric control based on pattern classification, however, has not translated into significant commercial and clinical impact as one would have expected. So far, none of the commercial prostheses is using pattern classification based controller. One of the main problems with the pattern classification for myoelectric control is that it leads to very unnatural (for the user) control scheme [@Jiang:12]. While natural movements are continuous and require activations of several degrees of freedom (DOF) simultaneously and proportionally, classical schemes for pattern recognition allow activation of only one class that corresponds to a particular action in one decision, i.e. sequential control. Moreover, all these classes as well as their superpositions must be previously learned. Simultaneous activation of two DOFs is thus recognized as a new class of action, but not as a combination of known actions. This means higher complexity of the classifier and consequently less robustness. In addition, this also implies that the user must spend more time in training and learning, resulting in higher rehabilitation cost and more frustration.
Recently there are attempts in addressing the issue of simultaneous and proportional control of prosthesis. Instead of taking classic pattern recognition approach, for example, multilayer perceptron neural network was used to estimate joint force [@Jiang:09] and joint angles [@Jiang:12a] of the actions from surface EMG. One of the main limitations of this multilayer perceptron approach is that it still requires the data from combined activation of multiple DOFs in the training set. Alternatively, non-negative matrix factorization has been used to estimate the joint force [@Jiang:09] and the joint angles [@Rehbaum:12]. This approach only requires data from single DOF activations for calibration, but has inferior performance in comparison to multilayer perceptron network approach. It is worth noting that both mentioned control schemes require solving optimization problems during training stage, which in turn demand significant computational power for multiple DOF. The training, moreover, needs to be performed on large data sets with typical size of $10^4 -
10^5$ samples, which need to be acquired during time-consuming sessions. Taking into account that the classical control schemes, such as multilayer perceptron, linear perceptron, linear discriminant analysis, Gaussian mixture model and hidden Markov model, do not show significant difference in classification power [@Hargrove:07], in this paper we develop a radical approach to the problem of myoelectric control based on quantum information processing.
During the last two decades, information encoding into the states of quantum systems and its processing according to the laws of quantum mechanics have been demonstrating impressive advantages over classical information processing [@Nielsen:00]. Typically these advantages are discussed in the context of computational complexity [@Galindo:02] and communication security [@Gisin:02]. Recently, however, we have shown that quantum information processing can dramatically increase the capabilities of the simplest learning machine – perceptron [@Siomau:13]. It has been demonstrated, in particular, that quantum perceptron is able to learn arbitrary logical functions, perform classification of data with an overlap and even conclude when the given features do not belong to previously seen classes – an unattainable tasks for its classical counterpart. Moreover, quantum perceptron learning rule does not involve any optimization procedure that makes the quantum information processing immune to the curse of dimensionality, which seriously limits power of classical learning models [@Kecman:01].
In this paper we adopt the concept of quantum perceptron to provide simultaneous and proportional myoelectric control of the artificial limb. While superposition principle of quantum mechanics allows us to recognize combinations of actions without their previous learning, the fact that quantum state $\ket{\psi}$ is normalized, i.e. $ \langle \psi \ket{\psi} = 1$, ensures proportional control of the prosthesis.
Let us focus on a particular case of limb deficiency – transradial amputation, or the amputation at the forearm. Our choice is mainly based on two factors. First of all, this type of limb deficiency represents a large portion of the upper-limb amputations. Second, wrist movements are complex enough and require activation of multiple DOFs. In our analysis we consider a wrist prosthesis with three DOF, namely flextion-extension, radial-ulnar deviation and pronation-supination, as illustrated in Figure \[fig-1\].
![The three DOF of the wrist to be emulated by the prosthesis.[]{data-label="fig-1"}](wrist.jpg){width="60mm"}
Let us suppose that $n$ electrodes are placed on the surface of the deficient limb. Particular choice of the location of the electrodes strongly depends on particular amputation and remnant musculature, therefore it is discussed elsewhere. There are four classical features that can be extracted from the raw signals of each electrode channel: mean absolute value, zero crossing, slope sign change and wave length [@Hudgins:93]. Thus, for each of the four features the feature space is of dimension of $n$. Let one of these features, mean absolute value for example, be encoded into the states of a (discrete) $n$-dimensional quantum system, so that the component $a_i \ket{i}$ of the state $\ket{\psi} = \sum a_i \ket{i}$ represents the mean absolute value of the signal from the $i$th electrode. Here, the amplitude of the signal $a_i$ is normalized over amplitudes from all the channels by factor $\sqrt{\sum
|a_i|^2}$ to fulfill the normalization condition $ \sum |a_i|^2 =
1$. Let the quantum state $\ket{\psi}$ be the input of a single-layer network of quantum perceptrons. In the most simple architecture that we are going to consider, the number of the perceptrons in the network is equal to the number of DOFs to be controlled, so that each quantum perceptron governs just one specific DOF. Without loss of generality, let us suppose that the perceptron $D1$ controls flextion-extension, $D2$ – radial-ulnar deviation, while $D3$ – pronation-supination. The control scheme is represented schematically in Figure \[fig-2\].
![The principal scheme of the myoelectric control of the artificial limb. The electrodes placed on the surface of the deficient or normal (as in the left picture) limb detect the electrical neural signals, which are summarized in EMG (two particular examples of EMG compound of seven electrode channels are displayed upward). These figures reproduced from [@Jiang:12a], with permission from Biomed Central. EMG features are encoded into the amplitudes of quantum states. During the learning stage, the three sets of positive operator valued measurements (POVM) operators are deduced, which control the three DOF. These operators are utilized to decide on kind of action to be performed by the prosthesis during the autonomous operation stage.[]{data-label="fig-2"}](scheme.jpg){width="80mm"}
Before free control of the artificial limb, the subject with the limb deficiency is instructed to focus on performing particular actions to learn how to control the prosthesis. Technically, the aim of the learning stage is to deduce the association between the EMG features and the joint kinematic. For example, given a corresponding command, the subject is trying to perform flextion activating DOF $D1$. EMG features, extracted from received myoelectric signals, are encoded into the quantum state $\ket{f} = \sum f_i \ket{i}$. Since efficient control of artificial limb demands not just recognition between particular classes, but also prediction of a degree of chosen action (angle or force), the learning must be repeated for different angles of flection. The quantum states, constructed from EMG features corresponding to different angles $\theta_i$ are combined into POVM operator $P_f^{D1} = \ket{\sum \alpha_i f_i}
\bra{\sum \alpha_i f_i }$, where $\alpha_i = \theta_i/ \sum_i
\theta_i$ [@comment-1].
Following the above procedure, learning extension leads to the construction of operator $P_e^{D1}$. The operators $P_f^{D1}$ and $P_e^{D1}$ summarize all training data that activate the $D1$. However, these operators do not necessarily form a legitimate POVM. The fundamental property of POVM operators is that they form a complete set, therefore, we need to define the third operator $P_0^{D1} = I - P_f^{D1} - P_e^{D1}$ to fulfill this condition. Here $I$ is the identity operator. Operator $P_0^{D1}$ thus collects all features that correspond to EMG activities, which do not lead to activation of the $D1$. In the same way as described above, the learning procedure is to be repeated for the other two DOFs $D2$ and $D3$. As the result of the learning stage we have deduced nine operators that govern the three DOF according to the classification of the selected feature (mean absolute value in our case). The learning must be repeated for the other features the same way as discussed above.
Let us now see how the trained network of quantum perceptrons responds to the EMG during subject’s autonomous control of the artificial limb. Desiring to perform an action, the subject generates neural signals which, as before, are acquired from the limb surface, expressed in EMG and encoded into a quantum state $\ket{\psi}$. Each quantum perceptron computes the expectation values $\bra{\psi} P \ket{\psi}$ for the nine operators [@comment-2]. The expectation values are interpreted as the strength of activation. For $D1$, for example, the three expectation values are computed as $f = \bra{\psi} P_f^{D1} \ket{\psi}$, $e =
\bra{\psi} P_e^{D1} \ket{\psi}$ and $0_{D1} = \bra{\psi} P_0^{D1}
\ket{\psi}$. The expectation values $f$ and $e$ contain information of how likely the given EMG correspond to flextion or extension. The expectation value $0_{D3}$ gives the probability that the flextion-extension DOF is inactive. If $f > e$, then the prostheses makes flextion on the degree $\frac{(f - e) \theta_f}{1 - {\rm Tr}
\left( P_f^{D1} P_e^{D1}\right)}$, where $\theta_f$ is the maximal flextion observed in the training data and ${\rm Tr} (...)$ stands for trace operation and gives the overlap between the operators $P_f^{D1}$ and $P_e^{D1}$. If, in contrast, $e
> f$, the prosthesis performs extension on $\frac{(e - f)
\theta_e}{1 - {\rm Tr} \left( P_f^{D1} P_e^{D1}\right)}$, where $\theta_e$ is the maximal observed extension. Here, we assumed that a linear combination of DOFs corresponds to a linear combination of corresponding features. This assumption is valid for mean absolute value, although may not be true for an arbitrary chosen feature [@Jiang:09].
As the result of the analysis of the given state $\ket{\psi}$, which encodes mean absolute value feature, the network of three quantum perceptrons returns three angles to the mechanical system of the prosthesis. These angles define the simultaneous and proportional action completely. Interestingly, there is additional control information encoded into the three expectation values $0_{D1},
0_{D2}$ and $0_{D3}$. As we have noted before, the expectation value $0_{D1}$ tells us the probability that the state $\ket{\psi}$ does not represent neither flextion nor extension. This may be interpreted as the given signal activates the other two DOFs, i.e. $0_{D1} = D2 + D3$. Such interpretation gives us three additional equations with three unknowns. Resolving this system of linear algebra equations, we can deduce additional proportionality between the activation of the three DOFs.
We tested the proposed control scheme on a set of EMG acquired from an able-bodied subject who performs wrist contractions. Eight electrodes were placed around the circumference of the forearm, with equal inter-electrode distance [@Jiang:09]. It is important to note that even for a single DOF there are two types of movements: direct action from the rest position and return action to the rest position. EMG for these actions may differ significantly, although formally correspond to the same spatial angles. In our analysis we always use direct actions both for the training and recognition of new patterns.
Our training data set collects only those movements of the subject that activate one of the three degrees of freedom. The test data set contains complex movements without any restrictions. The raw EMG that correspond to all these movements were recorded with $1024$ Hz sampling rate. Mean absolute value feature was extracted from the EMG averaging the signals over $100$ ms window. In our analysis we focused on just two DOFs D1 and D3, as they are the most desired functions not yet have commercial availability from trans-radial amputees. The purpose is to test whether our control scheme can recognize the combinations of these movements, being trained on just single DOF activations, i.e. the system can extrapolate automatically from single DOF data to their arbitrary combinations.
To test the control scheme, we initially used small sets for training consisting of just 500 data strings for each action, i.e. flextion, extension, pronation and supination. Each data string consists of the mean absolute values acquired from the eight channels and the respective joint angles of the DOF activation. The testing data contains 8216 data strings divided on 55 blocks; each block encodes a multiple DOF activation in a certain angle range. In 15 of the 55 blocks, the classification error appeared: 5 mistakes in flextion-extension, 9 – in pronation-supination and once both DOF were misclassified. To analyze the accuracy of classification, in cases when it was successful, we used so-called performance index, which is widely used as a global indicator of quality of the estimation [@Jiang:12a]. The performance index for an $k^{\rm
th}$ DOF is given by $$\label{PI}
R^2_k = 1 - \frac{\sum \left( \widetilde{\alpha_i} - \alpha_i
\right)^2}{ \sum \left( \alpha_i - \overline{\alpha_i} \right)^2} \,
,$$ where $\alpha_i$ it the joint angle of the $k^{\rm th}$ DOF, $\widetilde{\alpha_i}$ is the corresponding estimate, $\overline{\alpha_i}$ is temporal average of $\alpha_i$ and the summation is to be done over all data samples. Similarly, the global performance of the estimator is defined through the sum over all $K$ DOFs as $$\label{PIg}
R^2 = 1 - \frac{ \sum_{k=1}^K \sum \left( \widetilde{\alpha_i} -
\alpha_i \right)^2}{ \sum_{k=1}^K \sum \left( \alpha_i -
\overline{\alpha_i} \right)^2} \, .$$
The performance of flextion-extension recognition is found to be $0,834$, while the performance of pronation-supination classification is $0.224$. It is not surprising that the performance for $D3$ is much lower than for $D1$. Previous studies also reported lower performance on D3 than D1 [@Jiang:09; @Jiang:12a]. This is mainly due to the fact that muscle responsible for D3 are deep muscles. As a results, their EMG are easily masked by the EMG of superficial muscles, such as flexor muscles and extensor muscles. This masking effect is particularly pronounced during combined activations of D1 and D3. Nevertheless, the global performance of the two degrees of freedom recognition $0,715$ is comparable to the performance of classical schemes [@Jiang:09; @Jiang:12a].
It is important to check, how the efficiency of the estimation changes with the growth of the training set size, because using more data for the training is a standard way to improve performance of the model. We subsequently used large data sets for the training: 2000 samples for each action. For testing we used the same 55 blocks of unseen data. In result, in 10 blocks errors were detected: 3 for $D1$ and 7 for $D3$. Moreover, the global performance increased to $0,736$. Thus, the efficiency of classification increased with the size to the training set.
It is important to stress on the role of the overlap between operators $P_f^{D1}$ and $P_e^{D1}$ and between operators $P_p^{D3}$ and $P_s^{D3}$ in estimation of the respective DOFs. Indeed, these pairs of operators, by their construction, differ specifically in those components that are crucial for activation of the corresponding actions. It has been observed that the overlaps vary in certain range with the growth of training data approaching (on the large scale) some optimal value. As closer the overlap to the optimal values as higher the performance of estimation. In practice, closeness of the overlaps to their optimal values would indicate that further learning is not necessary.
Finally, we would like to admit that there are many nontrivial ways to further improve our control scheme. For example, it may be beneficial to include more quantum perceptrons in the network, so that several perceptrons control one DOF (direct and return actions, for example). It is also possible to construct a database of POVM operators out of the training data, so that one set of operators controls movements in a particular angle interval. A particular set of operators can be called from the database depending on overall intensity of the EMG. Also, different features may be encoded and analyzed simultaneously due to the fact that quantum amplitudes are complex numbers in general. Finally, our control scheme can be combined with classical control schemes, since its implementation does not demand any optimization.
In conclusion, we showed that quantum information processing can realize intuitive simultaneous and proportional prosthetic control over multiple DOFs, with only training on individual DOF data. We suggested a simple control scheme for the wrist prosthesis having advantages of optimization absence during the learning stage and not demanding large training data sets (size of which is proportional to the time spend by a subject in a lab and by implication his/her frustration in the procedure).
We would like to thank Dario Farina for his comments and suggestions, Johnny LG Vest-Nielsen for assistance in data acquisition and Ivan Vujaklija for art work. This work is partly supported by European Commission via the Industrial Academia Partnerships and Pathways (IAPP) under, Grant No. 251555 (AMYO); and the German Ministry for Education and Research (BMBF) via the Bernstein Focus Neurotechnology (BFNT) Göttingen, under Grant No. 01GQ0810.
[15]{}
N. Jiang, S. Dosen, K.-R. Müller and D. Farina, doi:10.1109/MSP.2012.2203480.
P. Parker, K. Englehart and B. Hugdins, *Control of Upper Limb Prosthesis*, (1st ed., Wiley-IEEE Press, 2004).
E. Scheme and K. Englehart, J. Rehab. Res. Dev. **48**, 643 (2011).
N. Jiang, K.B. Englehart and P.A. Parker, IEEE Trans. Biomed. Eng. **56**, 1070 (2009).
N. Jiang, J.L.G. Vest-Nielsen, S. Muceli and D. Farina, J. Neuroeng. Rehab. **9**, 42 (2012).
H. Rehbaum, N. Jiang and D. Farina, 34th Ann. Int. Conf. IEEE Eng. Med. Biol. Soc. San Diego, USA.
L.J. Hargrove, K. Englehart and B. Hudgins, IEEE Trans. Biomed. Eng. **54**, 847 (2007).
M.A. Nielsen and I.L. Chuang, *Quantum Computation and Quantum Information*, (Cambridge University Press, 2000).
A. Galindo and M.A. Martin-Delgado, Rev. Mod. Phys. **74**, 347 (2002).
N. Gisin, G. Ribordy, W. Tittel and H. Zbinden, Rev. Mod. Phys. **74**, 145 (2002).
M. Siomau, arXiv:1210.6626v2.
V. Kecman, *Learning and Soft Computing: Support Vector Machines, Neural Networks, and Fuzzy Logic Models*, (MIT Press, 2001).
B. Hudgins, P. Parker and R.N. Scott, IEEE Trans. Biomed. Eng. **40**, 82 (1993).
Let us suppose we are given just two feature vectors $\ket{a} = \sum
a_i \ket{i}$ and $\ket{b} = \sum b_i \ket{i}$ represented through quantum states of a $n$-dimensional quantum system and we know that the vectors belong to one class. Let us define the operator $P =
\frac{\ket{a+b} \bra{a+b}}{ {\rm Tr} \left( \ket{a+b} \bra{a+b}
\right)}$. Being a part of a POVM set, this operator ensures nonzero probability of outcome for any given linear combination of vectors $\alpha \ket{a} + \beta \ket{b}$, where $|\alpha|^2 +|\beta|^2 = 1$. Taking into account that statistics of future inputs can not be deduced, it is reasonable to accept that all superpositions of the vectors $\ket{a}$ and $\ket{b}$ can be observed with equal probability. The operator $P$ maximizes the average probability of detection of an arbitrary superposition of the two vectors; thus it is optimal POVM element by construction. This consideration generalizes straightforwardly to the case of an arbitrary number of feature vectors. If two feature vectors $\ket{a} = \sum a_i \ket{i}$ and $\ket{b} = \sum b_i \ket{i}$ represent the same motion but on differen angles $\theta_a$ and $\theta_b$ respectively, the operator $P = \ket{ \alpha_a a + \alpha_b b} \bra{\alpha_a a + \alpha_b b}$, where $\alpha_a = \theta_a / \left(\theta_a + \theta_b\right)$ and $\alpha_b = \theta_b / \left(\theta_a + \theta_b\right)$, takes into account contribution of the feature vectors to the activation of the motion. This definition of the operator $P$ is not unique. It leads, in particular, to the normalization of the all action angles to the maximal angle observed in the training data.
A single measurement with the POVM on a given feature vector implies three possible outcomes, namely the feature vector belongs to either the first or the second class, or the feature vector is essentially different from the two classes. However, the result of the single measurement is of no use for myoelectric control of artificial limb, because it would lead us to a choice between the two actions and no action. The expectation values of the operators, in contrast, are much more informative: they give us probabilistic interpretation what kind of action the given feature vector corresponds to.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Here we consider the 2D free boundary incompressible Euler equation with surface tension. We prove that the surface tension does not prevent a finite time splash or splat singularity, i.e. that the curve touches itself either in a point or along an arc. To do so, the main ingredients of the proof are a transformation to desingularize the curve and a priori energy estimates.
0.3cm *Keywords: Euler, incompressible, blow-up, water waves, splash, splat, surface tension.*
author:
- |
Angel Castro, Diego Córdoba, Charles Fefferman,\
Francisco Gancedo and Javier Gómez-Serrano\
\
*Dedicated to Peter Constantin on his 60th Birthday*
title: Finite time singularities for water waves with surface tension
---
Introduction
============
In this paper we continue the work in [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:finite-time-singularities-Euler] and [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:splash-water-waves] where we show the formation of singularities for the free boundary incompressible Euler equations. Here we prove that in two space dimensions the free boundary problem develops finite time “splash" and “splat" singularities when surface tension is taken into account (see below, in Section \[SectionInitialData\], the precise definition of the splash and splat curves).
In order to describe the evolution of a fluid with a moving domain $\Omega(t)\subset{\Bbb{R}}^2$, the 2D incompressible Euler equations are used: $$\label{Euler}
(v_t+v\cdot{\nabla}v)(x,y,t)=-{\nabla}p(x,y,t)-(0,1),\quad (x,y)\in\Omega(t)$$ with the fluid velocity $v(x,y,t)\in{\Bbb{R}}^2$ and the pressure $p(x,y,t)\in{\Bbb{R}}$. The vector $-(0,1)$ represents the external gravitational force (the acceleration due to gravity is taken equal to one for the sake of simplicity). The free boundary $$\label{Parametriza}
\partial\Omega(t)=\{z(\alpha,t)=(z_1(\alpha,t),z_2(\alpha,t)):\alpha\in{\Bbb{R}}\}$$ is smooth and convected by the velocity field $$\label{VeloFrontera}
z_t(\alpha,t)\cdot z_{\alpha}^\bot(\alpha,t)=v(z(\alpha,t),t)\cdot z_{\alpha}^\bot(\alpha,t),$$ which is assumed to be incompressible and irrotational $$\label{IncomIrro}
{\nabla}\cdot v(x,y,t)=0,\quad {\nabla}^{\bot}\cdot v(x,y,t)=0, \qquad (x,y)\in\Omega(t).$$ Here we study the relevance of considering the Laplace-Young condition for which the pressure on the interface $\partial\Omega(t)$ is proportional to its curvature, meaning that the surface tension effect is considered: $$\label{L-Y}
-p(z({\alpha},t),t)=\frac{\tau}{2} \frac{z_{{\alpha}{\alpha}}({\alpha},t)\cdot z_{\alpha}^\bot({\alpha},t)}{|z_{\alpha}({\alpha},t)|^3} \equiv \frac{\tau}{2}K.$$ Above $\tau>0$ is the surface tension coefficient.
The results in this paper can be shown for three different scenarios:
1. $\Omega(t)$ a compact domain: $z({\alpha},t)$ is a $2\pi$-periodic function in $\alpha$.
2. Asymptotically flat case: $z({\alpha},t)-({\alpha},0)\to 0$ as $\alpha\to\infty$.
3. $\Omega(t)$ periodic in the horizontal variable: $z({\alpha},t)-({\alpha},0)$ is a $2\pi$-periodic function in $\alpha$.
The problem to study here is the potential formation of singularities for the system (\[Euler\]-\[L-Y\]) with smooth interface and smooth velocity field with finite energy as initial data: $$\begin{aligned}
\label{VeloInicial}
\begin{split}
\Omega(0)&=\Omega_0,\quad \partial\Omega_0=\{z_0(\alpha):\alpha\in{\Bbb{R}}\},\\
v(x,y,0)&=v_0(x,y),\quad \int_{\Omega_0}|v_0(x,y)|^2dxdy< +\infty.
\end{split}\end{aligned}$$ The smooth initial curve $z_0({\alpha})$ must satisfy the arc-chord condition: $$\label{arcchord}
|z_0({\alpha})-z_0(\beta)|\geq c_{AC}|{\alpha}-\beta|,\quad \mbox{for all }{\alpha},\,\beta\in{\Bbb{R}},$$ where $c_{AC}>0$ is the arc-chord constant. The study of this quantity has been employed by other authors to prove local existence (see for example [@Wu:well-posedness-water-waves-2d], [@Wu:well-posedness-water-waves-3d]). We will quantify how our curve $z({\alpha})$ satisfies the arc-chord condition through the following quantity
$${\mathcal{F}}(z)=\frac{|\beta|}{|z({\alpha})-z({\alpha}-\beta)|},\quad {\alpha},\beta\in [-\pi,\pi].$$
Throughout the paper we will only focus on scenario 3 for the sake of simplicity. From now on, we will denote $\Omega_0\cap[-\pi,\pi]\times{\Bbb{R}}$ by $\Omega_0$ by abuse of notation (a fundamental domain in the period).
We establish the main result in the paper for the system (\[Euler\]-\[L-Y\]).
Consider $z_0({\alpha})-({\alpha},0)\in H^k({\mathbb{T}})$ for $k\geq 5$. Then there exist a family of initial data satisfying and the arc-chord condition and a time $T_s>0$ such that the interface $z({\alpha},t) \in H^{k}({\mathbb{T}})$ from the unique smooth solution of the system (\[Euler\]-\[arcchord\]) on the time interval $[0,T_s]$ touches itself at a single point (“splash" singularity) or along an arc (“splat" singularity) at time $t=T_s$.
These solutions can be extended to the periodic $3D$ setting considering scenarios invariant under translations in one coordinate direction. In [@Coutand-Shkoller:finite-time-splash], Coutand-Shkoller consider additional $3D$ splash and splat singularities. The case with small initial data was treated by Wu in the two dimensional case [@Wu:almost-global-wellposedness-2d] and the three dimensional case was studied by Wu [@Wu:global-wellposedness-3d] and Germain et al. [@Germain-Masmoudi-Shatah:global-solutions-gravity-water-waves-annals].
For other long time behaviour results see Alvarez-Lannes [@AlvarezSamaniego-Lannes:large-time-existence-water-waves], Castro et al. [@Castro-Cordoba-Fefferman-Gancedo-LopezFernandez:rayleigh-taylor-breakdown] and the references therein.
In order to prove this theorem we proceed as in [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:finite-time-singularities-Euler] and [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:splash-water-waves]. Using it is easy to declare that $v$ is harmonic in $\Omega(t)$. This fact allows us to introduce the moment $\omega({\alpha},t)$ by elementary potential theory as follows: $$v(x,y,t)=\frac{PV}{2\pi}\int_{{\Bbb{R}}}\frac{(x-z_1(\beta,t),y-z_2(\beta,t)))^{\bot}}{|(x,y)-z(\beta,t)|^2}\omega(\beta,t)d\beta,$$ where PV denotes principal value at infinity. This moment is also known in the literature as the vorticity amplitude. Then the system (\[Euler\]-\[L-Y\]) is equivalent to the following evolution equations which are only written in terms of the free boundary $z({\alpha},t)$ and the amplitude $\omega({\alpha},t)$: $$\label{em}
z_t({\alpha},t)=BR(z,\omega)({\alpha},t)+c({\alpha},t)z_{{\alpha}}({\alpha},t),$$ $$\begin{aligned}
\begin{split}\label{cEuler}
\omega_t({\alpha},t)&=-2BR_t(z,\omega)({\alpha},t)\cdot
z_{{\alpha}}({\alpha},t)-\Big(\frac{\omega^2}{4|{\partial_{\alpha}}z|^2}\Big)_{{\alpha}}({\alpha},t) +(c\omega)_{{\alpha}}({\alpha},t)\\
&\quad+2c({\alpha},t) BR_{{\alpha}}(z,\omega)({\alpha},t)\cdot z_{{\alpha}}({\alpha},t)-2
(z_2)_{\alpha}({\alpha},t)+\tau \left(\frac{z_{{\alpha}{\alpha}}({\alpha},t)\cdot z_{\alpha}^\bot({\alpha},t)}{|z_{\alpha}({\alpha},t)|^3}\right)_{{\alpha}}
\end{split}\end{aligned}$$ (for details see for example [@Cordoba-Cordoba-Gancedo:interface-water-waves-2d Section 2]). Above $BR(z,\omega)$ is the Birkhoff-Rott integral defined by $$\label{BR}
BR(z,\omega)=\frac{1}{2\pi}PV\int_{{\Bbb{R}}}\frac{(z({\alpha},t)-z(\beta,t))^{\bot}}{|z({\alpha},t)-z(\beta,t)|^2}\omega(\beta,t)d\beta,$$ and $c({\alpha},t)$ is arbitrary since the boundary is convected by the normal velocity .
Local existence in Sobolev spaces was first achieved by Wu [@Wu:well-posedness-water-waves-2d] assuming initially the arc-chord condition. For other variations and results see [@Craig:existence-theory-water-waves; @Nalimov:cauchy-poisson; @Beale-Hou-Lowengrub:growth-rates-linearized; @Yosihara:gravity-waves; @Wu:well-posedness-water-waves-3d; @Christodoulou-Lindblad:motion-free-surface; @Lindblad:well-posedness-motion; @Coutand-Shkoller:well-posedness-free-surface-incompressible; @Shatah-Zeng:geometry-priori-estimates; @Zhang-Zhang:free-boundary-3d-euler; @Lannes:well-posedness-water-waves; @Ambrose-Masmoudi:zero-surface-tension-3d-waterwaves; @Ambrose:well-posedness-vortex-sheet; @Lannes:stability-criterion; @Alazard-Burq-Zuily:water-wave-surface-tension; @Alazard-Metivier:paralinearization; @Cordoba-Cordoba-Gancedo:interface-water-waves-2d].
The strategy of the proof of the main result is to establish a local existence theorem from the initial data that has a splash or a splat singularity (notice that the equations are time reversible invariant). Since the curve self-intersects (failure of the arc-chord condition), it is not clear if the amplitude of the vorticity remains smooth and the meaning of equations (\[em\]-\[cEuler\]). In order to deal with these obstacles we use a conformal map $$P(w)=\Big(\tan\Big(\frac{w}{2}\Big)\Big)^{1/2},\quad w\in{\Bbb{C}},$$ whose intention is to keep apart the self-intersecting points taking the branch of the square root above passing through those crucial points. Here $P(z)$ will refer to a 2 dimensional vector whose components are the real and imaginary parts of $P(z_1 + iz_2)$. We also make sure that $\Omega(t)\cup\partial\Omega(t) $ do not contain any singular point of the transformation $P$. Then potential theory helps us to get the following analogous evolution equations for the new curve $${\tilde{z}}({\alpha},t)=P(z({\alpha},t))$$ and the new amplitude $\tilde{\omega}$: $$\begin{aligned}
\label{zeq}
\tilde{z}_{t}({\alpha},t) & = Q^2({\alpha},t)BR(\tilde{z},\tilde{\omega})({\alpha},t) + \tilde{c}({\alpha},t)\tilde{z}_{{\alpha}}({\alpha},t),\end{aligned}$$ $$\begin{aligned}
\label{eqomega}
\tilde{\omega}_{t}({\alpha},t) =& -2 BR_t(\tilde{z},\tilde{\omega})({\alpha},t) \cdot \tilde{z}_{{\alpha}}({\alpha},t) - |BR(\tilde{z},\tilde{\omega})|^{2} (Q^{2})_{{\alpha}}({\alpha},t) - \Big(\frac{Q^2({\alpha},t)\tilde{\omega}({\alpha},t)^2}{4|\tilde{z}_{{\alpha}}({\alpha},t)|^{2}}\Big)_{{\alpha}}\nonumber \\
& + 2\tilde{c}({\alpha},t) BR_{\alpha}(\tilde{z},\tilde{\omega}) \cdot \tilde{z}_{{\alpha}}({\alpha},t) + \left(\tilde{c}({\alpha},t)\tilde{\omega}({\alpha},t)\right)_{{\alpha}}
- 2 \left(P^{-1}_2(\tilde{z}({\alpha},t))\right)_{{\alpha}} \nonumber \\
&+\tau\left(\frac{Q^{3}}{|\tilde{z}_{{\alpha}}({\alpha},t)|^{3}}(\tilde{z}_{{\alpha}}^{T}HP_{2}^{-1} \tilde{z}_{{\alpha}} \nabla P_{1}^{-1} \cdot \tilde{z}_{{\alpha}}
- \tilde{z}_{{\alpha}}^{T}HP_{1}^{-1} \tilde{z}_{{\alpha}} \nabla P_{2}^{-1} \cdot \tilde{z}_{{\alpha}})\right)_{{\alpha}} \nonumber\\
&+\tau\left(Q\frac{\tilde{z}_{{\alpha}{\alpha}}({\alpha},t)\cdot\tilde{z}_{{\alpha}}^{\bot}({\alpha},t)}{|\tilde{z}_{{\alpha}}({\alpha},t)|^3}\right)_{{\alpha}} \end{aligned}$$ where $$Q^2({\alpha},t) = \left|\frac{dP}{dw}(P^{-1}(\tilde{z}({\alpha},t)))\right|^{2},$$
and $HP_{i}^{-1}$ denotes the Hessian matrix of $P_{i}^{-1}$, which is the $i$-th ($i = \{1,2\}$) component of the transformation $P^{-1}$.
Here, we choose $\tilde{c}({\alpha},t)$ in such a way that $|\tilde{z}_{{\alpha}}({\alpha},t)| = A(t)$. This particular choice of $\tilde{c}$ was first introduced by Hou et al. in [@Hou-Lowengrub-Shelley:removing-stiffness] and was later used by Ambrose [@Ambrose:well-posedness-vortex-sheet] and Ambrose-Masmoudi [@Ambrose-Masmoudi:zero-surface-tension-2d-waterwaves]. The choice of $\tilde{c}$ implies
$$\begin{aligned}
\tilde{c}({\alpha},t) & = \frac{{\alpha}+\pi}{2\pi}\int_{-\pi}^{\pi}(Q^{2}BR(\tilde{z},\tilde{\omega}))_\beta(\beta,t)\cdot\frac{\tilde{z}_{\beta}(\beta,t)}{|\tilde{z}_{\beta}(\beta,t)|^{2}}d\beta \\
& - \int_{-\pi}^{{\alpha}}(Q^{2}BR(\tilde{z},\tilde{\omega}))_\beta(\beta,t)\cdot\frac{\tilde{z}_{\beta}(\beta,t)}{|\tilde{z}_{\beta}(\beta,t)|^{2}}d\beta\end{aligned}$$
It is easy to check that if we take $Q \equiv 1$ in (\[zeq\]-\[eqomega\]) we recover (\[em\]-\[cEuler\]).
We also define the function $$\begin{aligned}
\label{varphi}
\tilde{\varphi}(\alpha,t)=\frac{Q^2(\alpha,t)\tilde{\omega}(\alpha,t)}{2| \tilde{z}_\alpha(\alpha,t)|}-\tilde{c}(\alpha,t)| \tilde{z}_\alpha(\alpha,t)|\end{aligned}$$ introduced by Beale et al. for the linear case [@Beale-Hou-Lowengrub:growth-rates-linearized] and by Ambrose-Masmoudi for the nonlinear one [@Ambrose-Masmoudi:zero-surface-tension-2d-waterwaves]. This function will be used to prove local existence in Sobolev spaces.
In the sections below, we show a local existence theorem based on energy estimates. Section \[SectionInitialData\] is devoted to provide the appropriate initial data for the splash and splat singularities. In Section \[SectionEnergyWRT\] we choose an energy which does not need a precise sign on the Rayleigh-Taylor function. In Section \[SectionEnergyRT\] we choose a different energy that involves the sign of the Rayleigh-Taylor function and the estimates are uniform with respect to the surface tension coefficient. These two energies are based on the ones obtained in the non-tilde domain by Ambrose ([@Ambrose:well-posedness-vortex-sheet]) and Ambrose-Masmoudi ([@Ambrose-Masmoudi:zero-surface-tension-3d-waterwaves]).
The Rayleigh-Taylor function is given by the following formula $$\begin{aligned}
\begin{split}\label{R-T}
\sigma \equiv& \left(BR_{t}(\tilde{z},\tilde{\omega}) + \frac{\tilde{\varphi}}{|\tilde{z}_{{\alpha}}|}BR_{{\alpha}}(\tilde{z},\tilde{\omega})\right) \cdot \tilde{z}_{{\alpha}}^{\perp} + \frac{\tilde{\omega}}{2|\tilde{z}_{{\alpha}}|^{2}}\left(\tilde{z}_{{\alpha}t} + \frac{\tilde{\varphi}}{|\tilde{z}_{{\alpha}}|}\tilde{z}_{{\alpha}{\alpha}}\right) \cdot \tilde{z}_{{\alpha}}^{\perp} \\
& + Q\left|BR(\tilde{z},\tilde{\omega}) + \frac{\tilde{\omega}}{2|\tilde{z}_{{\alpha}}|^{2}}\tilde{z}_{{\alpha}}\right|^{2}(\nabla Q)(\tilde{z}) \cdot \tilde{z}_{{\alpha}}^{\perp}
+ (\nabla P_{2}^{-1})(\tilde{z}) \cdot \tilde{z}_{{\alpha}}^{\perp}.
\end{split}\end{aligned}$$
All solutions that we will consider throughout the paper will have finite energy, as discussed in [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:finite-time-singularities-Euler]. The system satisfies the conservation of the mechanical energy. We define it this way: (not to be confused with the subsequent definitions of some other energies, see sections \[SectionEnergyWRT\] and \[SectionEnergyRT\]).
$$\begin{aligned}
\mathcal{E}_S(t) & = \frac{1}{2}\int_{\Omega_f(t)}|v(x,y,t)|^2dxdy + \frac{1}{2}\int_{-\pi}^{\pi}(z_2(\alpha,t))^2{\partial_{\alpha}}z_{1}(\alpha,t) d\alpha
+ \frac{\tau}{2} \int_{-\pi}^{\pi} |{\partial_{\alpha}}z({\alpha},t)|d\alpha \\
& \equiv \mathcal{E}_k(t) + \mathcal{E}_p(t) + \mathcal{E}_\tau(t),\end{aligned}$$
where $z(\alpha,t) = (z_{1}(\alpha,t), z_{2}(\alpha,t)), u(\alpha,t) = v(z(\alpha,t),t)$, and $\Omega_f(t)=\Omega(t) \cap [-\pi,\pi] \times \mathbb{R}$ is a fundamental domain in the water region in a period, then it follows that the energy is conserved.
$$\begin{aligned}
\frac{d\mathcal{E}_k(t)}{dt} & = \int_{\Omega_f(t)}v(x,y,t)(v_t(x,y,t) + v(x,y,t)\cdot \nabla v(x,y,t))dxdy \nonumber \\
& = \int_{\Omega_f(t)}v(x,y,t)(-\nabla p(x,y,t) - (0,1))dxdy \nonumber \\
& = -\int_{\Omega_f(t)}v(x,y,t)(\nabla (p(x,y,t) + y))dxdy \nonumber \\
& = - \int_{\partial (\Omega_f(t))}v(x,y,t)\cdot \overrightarrow{n} y ds
+ \int_{\partial (\Omega_f(t))}v(x,y,t)\cdot \overrightarrow{n} \frac{\tau}{2} K ds\nonumber \\
& = -\int_{-\pi}^{\pi}z_{2}(\alpha,t)u(\alpha,t)\cdot {\partial_{\alpha}}z^\bot(\alpha,t)d\alpha
+\frac{\tau}{2} \int_{-\pi}^{\pi}u(\alpha,t)\cdot {\partial_{\alpha}}z^\bot(\alpha,t) \frac{{\partial_{\alpha}}^{2} z({\alpha},t) \cdot {\partial_{\alpha}}z^{\perp}({\alpha},t)}{|{\partial_{\alpha}}z({\alpha},t)|^{3}} d\alpha\end{aligned}$$
where we have used the incompressibility of the fluid ($\nabla \cdot v = 0$) and Laplace-Young’s condition for the pressure on the interface. Next $$\begin{aligned}
\frac{d\mathcal{E}_p(t)}{dt} & = \int_{-\pi}^{\pi}z_2(\alpha,t)\partial_t z_2(\alpha,t){\partial_{\alpha}}z_{1}(\alpha,t)d\alpha + \frac{1}{2}\int_{-\pi}^{\pi}(z_{2}(\alpha,t))^2
{\partial_t}{\partial_{\alpha}}z_{1}(\alpha,t)d\alpha \nonumber \\
& = \int_{-\pi}^{\pi}z_2(\alpha,t)\partial_t z_2(\alpha,t){\partial_{\alpha}}z_{1}(\alpha,t)d\alpha - \int_{-\pi}^{\pi}z_{2}(\alpha,t){\partial_{\alpha}}z_{2}(\alpha,t)\partial_t z_1(\alpha,t)d\alpha \nonumber \\
& = \int_{-\pi}^{\pi}z_2(\alpha,t)u(\alpha,t)\cdot {\partial_{\alpha}}z^\bot(\alpha,t)d\alpha.\end{aligned}$$
$$\begin{aligned}
\frac{d\mathcal{E}_\tau(t)}{dt} & = \frac{\tau}{2}\int_{-\pi}^{\pi} \frac{{\partial_{\alpha}}z({\alpha},t) \cdot {\partial_{\alpha}}\partial_t z({\alpha},t)}{|{\partial_{\alpha}}z({\alpha},t)|}d\alpha
= -\frac{\tau}{2}\int_{-\pi}^{\pi} \frac{{\partial_{\alpha}}^{2} z({\alpha},t) \cdot \partial_t z({\alpha},t)}{|{\partial_{\alpha}}z({\alpha},t)|}d\alpha \nonumber \\
& = -\frac{\tau}{2}\int_{-\pi}^{\pi} \frac{{\partial_{\alpha}}^{2} z({\alpha},t) \cdot u({\alpha},t)}{|{\partial_{\alpha}}z({\alpha},t)|}d\alpha
= -\frac{\tau}{2}\int_{-\pi}^{\pi} \frac{{\partial_{\alpha}}^{2} z({\alpha},t) \cdot {\partial_{\alpha}}z^{\perp}({\alpha},t)}{|{\partial_{\alpha}}z({\alpha},t)|^{3}}u({\alpha},t) \cdot {\partial_{\alpha}}^{\perp} z({\alpha},t) d\alpha\end{aligned}$$
Adding all the derivatives we get the desired result.
Properties of the curvature in the tilde domain {#SectionCurvature}
===============================================
In this section we will rewrite the term corresponding to the curvature $K(z({\alpha},t))$ in the new tilde variables $\tilde{z}({\alpha},t)$.
We will proceed step by step. Let us recall that the curvature is defined by $$\begin{aligned}
K({\alpha},t) = \frac{z_{{\alpha}{\alpha}}({\alpha},t) \cdot z_{{\alpha}}^{\bot}({\alpha},t)}{|z_{{\alpha}}({\alpha},t)|^{3}}\end{aligned}$$
We begin with the term $|z_{{\alpha}}({\alpha},t)|^{3}$. We have that $$\begin{aligned}
|\tilde{z}_{{\alpha}}({\alpha},t)|^{2} = \langle {\partial_{\alpha}}P(z({\alpha},t)), {\partial_{\alpha}}P(z({\alpha},t))\rangle
= \langle \nabla P(z({\alpha},t)) \cdot z_{{\alpha}}({\alpha},t), \nabla P(z({\alpha},t)) \cdot z_{{\alpha}}({\alpha},t)\rangle\end{aligned}$$ Since $P$ and $P^{-1}$ are conformal, by the Cauchy-Riemann equations $$\begin{aligned}
\nabla P(z({\alpha},t))^{T} \nabla P(z({\alpha},t)) = Q^{2}({\alpha},t)\text{Id}_2,\end{aligned}$$ that implies that $$\begin{aligned}
|\tilde{z}_{{\alpha}}({\alpha},t)|^{3} = Q^{3}({\alpha},t)|z_{{\alpha}}({\alpha},t)|^{3}\end{aligned}$$
We move to the other term $$\begin{aligned}
\langle z_{{\alpha}{\alpha}}({\alpha},t), z_{{\alpha}}^{\bot}({\alpha},t)\rangle
& = \langle {\partial_{\alpha}}\left(\nabla P^{-1}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}}({\alpha},t)\right), (\nabla P^{-1}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}}({\alpha},t))^{\bot}\rangle \\
& = \langle \nabla P^{-1}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}{\alpha}}({\alpha},t), (\nabla P^{-1}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}}({\alpha},t))^{\bot}\rangle \\
& + \langle {\partial_{\alpha}}\left(\nabla P^{-1}(\tilde{z}({\alpha},t))\right) \cdot \tilde{z}_{{\alpha}}({\alpha},t), (\nabla P^{-1}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}}({\alpha},t))^{\bot}\rangle \equiv W + X\end{aligned}$$
Again, by the Cauchy-Riemann equations $$\begin{aligned}
W = \frac{1}{Q^2({\alpha},t)}\langle \tilde{z}_{{\alpha}{\alpha}}({\alpha},t), \tilde{z}_{{\alpha}}({\alpha},t)^{\bot}\rangle\end{aligned}$$
Developing the terms in $X$, we get $$\begin{aligned}
\left(\nabla P^{-1}(\tilde{z}({\alpha},t))\right) \cdot \tilde{z}_{{\alpha}}({\alpha},t)
= \left(
\begin{array}{c}
\tilde{z}^{T}_{{\alpha}}({\alpha},t) \cdot HP^{-1}_{1}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}}({\alpha},t) \\
\tilde{z}^{T}_{{\alpha}}({\alpha},t) \cdot HP^{-1}_{2}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}}({\alpha},t) \\
\end{array}
\right),\end{aligned}$$
where $HP^{-1}_{i}$ denotes the Hessian of the $i$-th component of $P^{-1}$ ($i = 1,2$). Hence, we can write $X$ as $$\begin{aligned}
X & =
- \tilde{z}^{T}_{{\alpha}}({\alpha},t) \cdot HP^{-1}_{1}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}}({\alpha},t) \nabla P^{-1}_{2}(\tilde{z}({\alpha},t)) \cdot \tilde{z}({\alpha},t) \\
& + \tilde{z}^{T}_{{\alpha}}({\alpha},t) \cdot HP^{-1}_{2}(\tilde{z}({\alpha},t)) \cdot \tilde{z}_{{\alpha}}({\alpha},t) \nabla P^{-1}_{1}(\tilde{z}({\alpha},t)) \cdot \tilde{z}({\alpha},t).\end{aligned}$$
This means that $$\begin{aligned}
K({\alpha},t) = Q({\alpha},t) \frac{\tilde{z}_{{\alpha}{\alpha}}({\alpha},t) \cdot \tilde{z}^{\bot}_{{\alpha}}({\alpha},t)}{|\tilde{z}({\alpha},t)|^{3}}
+ X({\alpha},t) \frac{Q({\alpha},t)^{3}}{|\tilde{z}({\alpha},t)|^{3}} \equiv Q({\alpha},t) \tilde{K}({\alpha},t) + M({\alpha},t)\end{aligned}$$
We will now try to simplify further by exploiting the Cauchy-Riemann equations. We can calculate the Hessian and the gradient terms as:
$$\begin{aligned}
P^{-1}_{1,x}(\tilde{z}) = \Re\left(\frac{4\tilde{z}}{1+\tilde{z}^{4}}\right) \equiv \Re(a) \\
P^{-1}_{1,y}(\tilde{z}) = \Re\left(\frac{4i\tilde{z}}{1+\tilde{z}^{4}}\right) \equiv -\Im(a) \\
P^{-1}_{2,x}(\tilde{z}) = \Im\left(\frac{4\tilde{z}}{1+\tilde{z}^{4}}\right) \equiv \Im(a) \\
P^{-1}_{2,y}(\tilde{z}) = \Im\left(\frac{4i\tilde{z}}{1+\tilde{z}^{4}}\right) \equiv \Re(a) \\
P^{-1}_{1,x,x}(\tilde{z}) = \Re\left(\frac{4(1-3\tilde{z}^{4})}{(1+\tilde{z}^{4})^{2}}\right) \equiv \Re(b) \\
P^{-1}_{1,x,y}(\tilde{z}) = \Re\left(\frac{4i(1-3\tilde{z}^{4})}{(1+\tilde{z}^{4})^{2}}\right) \equiv -\Im(b) \\
P^{-1}_{2,x,x}(\tilde{z}) = \Im\left(\frac{4(1-3\tilde{z}^{4})}{(1+\tilde{z}^{4})^{2}}\right) \equiv \Im(b) \\
P^{-1}_{2,x,y}(\tilde{z}) = \Im\left(\frac{4i(1-3\tilde{z}^{4})}{(1+\tilde{z}^{4})^{2}}\right) \equiv \Re(b) \\\end{aligned}$$
Therefore the Hessians are $$HP_{1}^{-1} =
\left(
\begin{array}{cc}
\Re(b) & -\Im(b) \\
-\Im(b) & -\Re(b)
\end{array}
\right), \quad
HP_{2}^{-1} =
\left(
\begin{array}{cc}
\Im(b) & \Re(b) \\
\Re(b) & -\Im(b)
\end{array}
\right),$$
Calculating further:
$$\tilde{z}_{{\alpha}}^{T} HP_{2}^{-1} \tilde{z}_{{\alpha}} = \Re(b)(2\tilde{z}_{{\alpha}}^{1}\tilde{z}_{{\alpha}}^{2}) + \Im(b)((\tilde{z}^{1}_{{\alpha}})^2 - (\tilde{z}^{2}_{{\alpha}})^2)$$
$$\tilde{z}_{{\alpha}}^{T} HP_{1}^{-1} \tilde{z}_{{\alpha}} = \Re(b)((\tilde{z}^{1}_{{\alpha}})^2 - (\tilde{z}^{2}_{{\alpha}})^2) - \Im(b)(2\tilde{z}_{{\alpha}}^{1}\tilde{z}_{{\alpha}}^{2})$$
$$\begin{aligned}
X_1 & = \Re(a)\Re(b)(2(\tilde{z}_{{\alpha}}^{1})^2\tilde{z}_{{\alpha}}^{2}) + \Re(a)\Im(b)((\tilde{z}^{1}_{{\alpha}})^2\tilde{z}_{{\alpha}}^{1} - (\tilde{z}^{2}_{{\alpha}})^2\tilde{z}_{{\alpha}}^{1}) \\
& + \Im(a)\Re(b)(-2\tilde{z}_{{\alpha}}^{1}(\tilde{z}_{{\alpha}}^{2})^2) + \Im(b)\Im(b)((\tilde{z}^{1}_{{\alpha}})^2\tilde{z}_{{\alpha}}^{2} - (\tilde{z}^{2}_{{\alpha}})^2\tilde{z}_{{\alpha}}^{2}) \\
X_2 & = \Re(b)\Re(b)((\tilde{z}^{1}_{{\alpha}})^2\tilde{z}_{{\alpha}}^{2} - (\tilde{z}^{2}_{{\alpha}})^2\tilde{z}_{{\alpha}}^{2}) + \Re(a)\Im(b)(-2\tilde{z}_{{\alpha}}^{1}(\tilde{z}_{{\alpha}}^{2})^2) \\
& + \Im(a)\Im(b)(2(\tilde{z}_{{\alpha}}^{1})^2\tilde{z}_{{\alpha}}^{2}) + \Im(a)\Re(b)((\tilde{z}^{1}_{{\alpha}})^2\tilde{z}_{{\alpha}}^{1} - (\tilde{z}^{2}_{{\alpha}})^2\tilde{z}_{{\alpha}}^{1}) \\\end{aligned}$$
This means $$\begin{aligned}
X = X_1 - X_2 & = ((\tilde{z}_{{\alpha}}^{1})^2+(\tilde{z}_{{\alpha}}^{2})^2)(\tilde{z}_{{\alpha}}^{2}(\Re(a)\Re(b)+\Im(a)\Im(b)) + \tilde{z}_{{\alpha}}^{1}(\Re(a)\Im(b)-\Im(a)\Re(b))) \\
& \equiv ((\tilde{z}_{{\alpha}}^{1})^2+(\tilde{z}_{{\alpha}}^{2})^2)\langle G(z), \tilde{z}_{{\alpha}}\rangle. \\\end{aligned}$$
We can see that
$$\begin{aligned}
-\frac{Q_{{\alpha}}}{Q^{3}} & = \frac{1}{2}{\partial_{\alpha}}\left(\frac{1}{Q^{2}}\right) = {\partial_{\alpha}}(\Re(a)^{2} + \Im(a)^2) \\
& = \Re(a)\Re(b)\tilde{z}^{1}_{{\alpha}} - \Re(a)\Im(b)\tilde{z}^{2}_{{\alpha}}
+ \Im(a)\Im(b)\tilde{z}^{1}_{{\alpha}} + \Im(a)\Re(b)\tilde{z}^{2}_{{\alpha}} \\
& = \langle G(z), \tilde{z}_{{\alpha}}^{\perp} \rangle\end{aligned}$$
by the Cauchy-Riemann equations.
If we take one derivative in space of $X$, we obtain
$$\begin{aligned}
{\partial_{\alpha}}X & = ((\tilde{z}_{{\alpha}}^{1})^2+(\tilde{z}_{{\alpha}}^{2})^2)\langle \nabla G(\tilde{z}) \cdot \tilde{z}_{{\alpha}}, \tilde{z}_{{\alpha}}\rangle + ((\tilde{z}_{{\alpha}}^{1})^2+(\tilde{z}_{{\alpha}}^{2})^2)\langle G(\tilde{z}), \tilde{z}_{{\alpha}{\alpha}}\rangle \\
& = ((\tilde{z}_{{\alpha}}^{1})^2+(\tilde{z}_{{\alpha}}^{2})^2)\langle \nabla G(\tilde{z}) \cdot \tilde{z}_{{\alpha}}, \tilde{z}_{{\alpha}}\rangle + |\tilde{z}_{{\alpha}}|^{3} \tilde{K}\langle G(\tilde{z}), \tilde{z}_{{\alpha}}^{\perp}\rangle \\
& = ((\tilde{z}_{{\alpha}}^{1})^2+(\tilde{z}_{{\alpha}}^{2})^2)\langle \nabla G(\tilde{z}) \cdot \tilde{z}_{{\alpha}}, \tilde{z}_{{\alpha}}\rangle - |\tilde{z}_{{\alpha}}|^{3} \tilde{K}\frac{Q_{{\alpha}}}{Q^{3}},\end{aligned}$$
This implies
$$\begin{aligned}
K = Q\tilde{K} - Q^{3}\frac{X}{|\tilde{z}|^{3}}
\Rightarrow K_{{\alpha}} = (Q\tilde{K})_{{\alpha}} + \frac{Q^{3}}{|\tilde{z}_{{\alpha}}|}\langle \nabla G(\tilde{z}) \cdot \tilde{z}_{{\alpha}}, \tilde{z}_{{\alpha}}\rangle - \tilde{K}Q_{{\alpha}} = (Q\tilde{K})_{{\alpha}} + M_1 + M_2\end{aligned}$$
Later, we will see that the $M_1$ is a low order term and can be absorbed by the energy.
Initial data {#SectionInitialData}
============
For initial data we are interested in considering a self-intersecting curve in one point. More precisely, we will use as initial data *splash curves* which are defined this way:
\[defsplash\] We say that $z({\alpha}) = (z_1({\alpha}),z_2({\alpha}))$ is a *splash curve* if
1. $z_{1}({\alpha}) - {\alpha}, z_2({\alpha})$ are smooth functions and $2\pi$-periodic.
2. $z({\alpha})$ satisfies the arc-chord condition at every point except at $\alpha_1$ and $\alpha_2$, with $\alpha_1 < \alpha_2$ where $z({\alpha}_1) = z({\alpha}_2)$ and $|z_{{\alpha}}({\alpha}_1)|, |z_{{\alpha}}({\alpha}_2)| > 0$. This means $z({\alpha}_1) = z({\alpha}_2)$, but if we remove either a neighborhood of ${\alpha}_1$ or a neighborhood of ${\alpha}_2$ in parameter space, then the arc-chord condition holds.
3. The curve $z(\alpha)$ separates the complex plane into two regions; a connected water region and a vacuum region (not necessarily connected). The water region contains each point $x+iy$ for which y is large negative. We choose the parametrization such that the normal vector $n=\frac{(-{\partial}_\alpha z_2(\alpha), {\partial}_\alpha z_1(\alpha))}{|{\partial}_\alpha z(\alpha)|}$ points to the vacuum region. We regard the interface to be part of the water region.
4. We can choose a branch of the function $P$ on the water region such that the curve $\tilde{z}({\alpha}) = (\tilde{z}_1({\alpha}),\tilde{z}_2({\alpha})) = P(z({\alpha}))$ satisfies:
1. $\tilde{z}_1({\alpha})$ and $\tilde{z}_2({\alpha})$ are smooth and $2\pi$-periodic.
2. $\tilde{z}$ is a closed contour.
3. $\tilde{z}$ satisfies the arc-chord condition.
We will choose the branch of the root that produces that $$\lim_{y \to -\infty}P(x+iy) = -e^{-i \pi/4}$$ independently of $x$.
5. $P(w)$ is analytic at $w$ and $\frac{dP}{dw}(w) \neq 0$ if $w$ belongs to the interior of the water region. Furthermore, $(\pm \pi, 0)$ and $(0,0)$ belong to the vacuum region.
6. $\tilde{z}({\alpha}) \neq q^l$ for $l=0,...,4$, where $$\label{points}
q^0=\left(0,0\right),\quad
q^1=\left(\frac{1}{\sqrt{2}},\frac{1}{\sqrt{2}}\right),\quad
q^2=\left(\frac{-1}{\sqrt{2}},\frac{1}{\sqrt{2}}\right),\quad
q^3=\left(\frac{-1}{\sqrt{2}}, \frac{-1}{\sqrt{2}}\right),\quad
q^4=\left(\frac{1}{\sqrt{2}}, \frac{-1}{\sqrt{2}}\right).$$
Moreover, we will define a *splat curve* as a splash curve but replacing condition (2) by the fact that the curve touches itself along an arc, instead of a point.
Let us note that in order to measure when the transformation $P$ is regular, we need to control the distance to the points $q^{l}$. In order to do so, we introduce the function
$$m(q^{l})({\alpha}, t) \equiv |\tilde{z}({\alpha},t)-q^{l}|$$
for $l = 0,\ldots, 4$.
We have performed numerical simulations, as explained in [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:splash-water-waves] with the following initial data on the non-tilde domain:
$$z^{0}_{1}({\alpha}) = {\alpha}+ \frac{1}{4}\left(-\frac{3\pi}{2} - 1.9\right)\sin({\alpha})+\frac{1}{2}\sin(2{\alpha})+\frac{1}{4}\left(\frac{\pi}{2} - 1.9\right)\sin(3{\alpha})$$ $$z^{0}_{2}({\alpha}) = \frac{1}{10}\cos({\alpha}) - \frac{3}{10}\cos(2{\alpha}) + \frac{1}{10}\cos(3{\alpha})$$
Note that $z\left(\frac{\pi}{2}\right) = z\left(-\frac{\pi}{2}\right)$ (splash). Instead of prescribing an initial condition for $\tilde{\omega}$, we prescribed the normal component of the velocity to ensure a more controlled direction of the fluid. From that we got the initial $\tilde{\omega}({\alpha},0)$ using the following relations. Let $\psi$ be such that $\nabla^{\perp}\psi = v$ and $\Psi({\alpha})$ its restriction to the interface. The initial normal velocity is then prescribed by setting $$u^{0}_{n}({\alpha})|z_{{\alpha}}({\alpha})| = \Psi_{{\alpha}}({\alpha}) = 3 \cdot \cos({\alpha}) - 3.4 \cdot \cos(2{\alpha}) + \cos(3{\alpha}) + 0.2\cos(4{\alpha}).$$
The reader may easily check that the above $z^{0}_{1}$ and $z^{0}_{2}$ yield a splash curve, i.e. the conditions in Definition \[defsplash\] are satisfied. See Figure \[PictureSplash\].
![Splash singularity. The interface self intersects in a point.[]{data-label="PictureSplash"}](ZoomNontildaRedPolished.eps)
In order to get an initial data for the splat singularity, one only needs to perturb the splash curve so that it $z_{0}^{1}({\alpha}) = 0$ on a neighbourhood of both $\alpha = \pm \frac{\pi}{2}$. The normal velocity can be the same since it has the right sign (the one that separates the curve). By continuity, the Rayleigh-Taylor function should remain positive.
For the case where the energy is independent on the surface tension coefficient (see Section \[SectionEnergyRT\]), we need the curve to satisfy the Rayleigh-Taylor condition initially. This is always the case when the surface tension coefficient is small enough. To illustrate this phenomenon, we have plotted in the next figure the Rayleigh-Taylor condition for different values of the surface tension coefficient and the initial condition described above. We can see that for small enough values of $\tau$ (0 and 0.1): the Rayleigh-Taylor condition $\sigma$ is strictly positive. For bigger values of $\tau$, the Rayleigh-Taylor condition $\sigma$ has distinct sign.
![Rayleigh-Taylor function for different values of $\tau$: $\tau = 0$ (blue), $\tau = 0.1$ (red), $\tau = 0.5$ (green), $\tau = 1$ (black)[]{data-label="PictureRTCoeff"}](RayleighTaylorBlue0Red01Green05Black1.eps)
Energy without the Rayleigh-Taylor condition {#SectionEnergyWRT}
============================================
In this section, we prove local existence in the tilde domain, where the time of existence depends on the surface tension coefficient. This theorem has the advantage that the initial data does not need to satisfy the Rayleigh-Taylor condition and it works for every $\tau > 0$.
\[localexistencetildeWRT\] Let $k \geq 3$. Let $\tilde{z}^{0}({\alpha})$ be the image of a splash curve by the map $P$ parametrized in such a way that $|{\partial_{\alpha}}\tilde{z}^{0}({\alpha})| = \frac{L}{2\pi}$, where $L$ is the length of the curve in a fundamental period, and such that $\tilde{z}_{1}^{0}({\alpha}), \tilde{z}_2^{0}({\alpha}) \in H^{k+2}({\mathbb{T}})$. Let $\tilde{\omega}({\alpha},0) \in H^{k+\frac{1}{2}}({\mathbb{T}})$. Then there exist a finite time $T > 0$, a time-varying curve $\tilde{z}({\alpha},t) \in C([0,T];H^{k+2})$, and a function $\tilde{\omega}({\alpha},t) \in C([0,T];H^{k+\frac{1}{2}})$ providing a solution of the water wave equations (\[zeq\] - \[eqomega\]).
The proof below is based in the following energy estimates:
The energy {#primeraenergia}
----------
We will define the energy for $k \geq 3$ as
$$E^2_{k}(t) = \mathcal{EE}^{2}(t) + \underbrace{2|\tilde{z}_{{\alpha}}|^{3}\int Q^{2k+1}\left({\partial_{\alpha}}^{k}(\tilde{K})\right)^{2}}_{A}
+ \underbrace{\frac{1}{\tau}\int Q^{2k+2}{\partial_{\alpha}}^{k}(\tilde{\omega})\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))}_{B}
+ \underbrace{\frac{1}{2|\tilde{z}_{{\alpha}}|\tau^{2}}\int Q^{2k+3}({\partial_{\alpha}}^{k}(\tilde{\omega}))^{2}\tilde{\omega}^{2}}_{C},$$
$$\mathcal{EE}^{2}(t) = \|\tilde{z}\|_{L^{2}}^{2} + \|\tilde{\omega}\|_{L^{2}}^{2} +\|\mathcal{F}(z)\|^2_{L^\infty}(t) +\sum_{l=0}^4\frac{1}{m(q^l)(t)},$$
where $m(q^{l})(t) = \min_{{\alpha}\in{\mathbb{T}}}q^{l}({\alpha},t)$ for $l = 0,\ldots,4$ and $\Lambda = (-\Delta)^{1/2}.$ From now on, we will denote the Hilbert transform of a function $f$ by $H(f)$, where $$\begin{aligned}
H(f)({\alpha}) = \frac{PV}{\pi}\int_{-\pi}^{\pi}\frac{f({\alpha}-\beta)}{2\tan\left(\frac{\beta}{2}\right)}d\beta.\end{aligned}$$
Recall that the operator $\Lambda$ can also be written as $\Lambda(f) = {\partial_{\alpha}}H(f)$.
The energy estimates
--------------------
The energy estimates for $\mathcal{EE}$ were proved in [@Cordoba-Cordoba-Gancedo:interface-water-waves-2d] and in [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:finite-time-singularities-Euler]. In this section we will focus on the new terms ($A$, $B$ and $C$).
### $\tilde{K}$
$$\tilde{K}_t = \text{NICE3 } + \frac{Q^{2}}{2|\tilde{z}_{{\alpha}}|^{3}}H(\tilde{\omega}_{{\alpha}{\alpha}}) + \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}(Q^{2})_{{\alpha}}H(\tilde{\omega}_{{\alpha}}),$$
where NICE3 means $$\int Q^{j}{\partial_{\alpha}}^{k}(\tilde{K}) {\partial_{\alpha}}^{k}(NICE3) \leq CE_{k}^{p}(t)$$ for some positive constants $C, p$ and any $j$.
We start writing $\tilde{K}_t$
$$\begin{aligned}
\tilde{K}_t & = \frac{-3}{|\tilde{z}_{{\alpha}}|^{5}}\tilde{z}_{{\alpha}t} \cdot \tilde{z}_{{\alpha}} \tilde{z}_{{\alpha}{\alpha}} \cdot \tilde{z}_{{\alpha}}^{\perp} + \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}\left(\tilde{z}_{{\alpha}{\alpha}t} \cdot \tilde{z}_{{\alpha}}^{\perp} + \tilde{z}_{{\alpha}{\alpha}} \cdot \tilde{z}_{{\alpha}t}^{\perp}\right) = P_0 + P_1 + P_2\end{aligned}$$
Calculating further $P_0$ we get that $$\begin{aligned}
P_0 & = \frac{-3}{|\tilde{z}_{{\alpha}}|^{5}}(Q^{2}BR + c \tilde{z}_{{\alpha}})_{{\alpha}} \cdot \tilde{z}_{{\alpha}} \tilde{z}_{{\alpha}{\alpha}} \cdot \tilde{z}_{{\alpha}}^{\perp} = \text{NICE3},\end{aligned}$$
by the estimates proved in the Appendix.
On the one hand, developing $P_2$, we obtain $$\begin{aligned}
P_2 & = \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}\left(\tilde{z}_{{\alpha}{\alpha}} \cdot \tilde{z}_{{\alpha}t}^{\perp}\right)
= -\frac{1}{|\tilde{z}_{{\alpha}}|^{3}}\left(\tilde{z}_{{\alpha}{\alpha}}^{\perp} \cdot \tilde{z}_{{\alpha}t}\right)=\\
& = -\frac{1}{|\tilde{z}_{{\alpha}}|^{3}}\left((Q^{2} BR)_{{\alpha}} + (\tilde{c}\tilde{z}_{{\alpha}})_{{\alpha}}\right) \cdot \tilde{z}_{{\alpha}{\alpha}}^{\perp} \\
& = \text{NICE3 } -\frac{1}{|\tilde{z}_{{\alpha}}|^{3}}\left(\frac{Q^{2}}{2}\frac{\tilde{z}_{{\alpha}}^{\perp}}{|\tilde{z}_{{\alpha}}|^{2}}H(\tilde{\omega}_{{\alpha}}) + \tilde{c}_{{\alpha}}\tilde{z}_{{\alpha}}\right) \cdot \tilde{z}_{{\alpha}{\alpha}}^{\perp} = \text{NICE3},\end{aligned}$$ since $\tilde{c}_{{\alpha}}$ is as regular as $\tilde{\omega}, \tilde{z}_{{\alpha}{\alpha}}$ and therefore bounded in $H^{k}$. On the other, $P_1$ gives rise to $$\begin{aligned}
P_1 & = \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}\left(\tilde{z}_{{\alpha}{\alpha}t} \cdot \tilde{z}_{{\alpha}}^{\perp}\right) = \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}\left((Q^{2} BR)_{{\alpha}{\alpha}} + (\tilde{c}\tilde{z}_{{\alpha}})_{{\alpha}{\alpha}}\right) \cdot \tilde{z}_{{\alpha}}^{\perp} = P_{1,1} + P_{1,2}\end{aligned}$$ We can further develop $P_{1,2}$ to obtain $$\begin{aligned}
P_{1,2} = \text{NICE3},\end{aligned}$$ since the terms vanish either by integrating by parts, by being a dot product between two orthogonal vectors or because $\tilde{c}_{{\alpha}} = \text{NICE3}$. We also have that $$\begin{aligned}
P_{1,1} & = \text{NICE3 } + \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}\left(2(Q^{2})_{{\alpha}} BR_{{\alpha}} + Q^{2} BR_{{\alpha}{\alpha}}\right)\cdot \tilde{z}_{{\alpha}}^{\perp} = \text{NICE3 } + P_{1,1,1} + P_{1,1,2}\end{aligned}$$ The only term in $BR_{{\alpha}}$ which is not NICE3 is when we hit with the derivative in $\tilde{\omega}$. Therefore $$\begin{aligned}
P_{1,1,1} & = \text{NICE3 } + \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}2(Q^{2})_{{\alpha}} \frac{1}{2}H(\tilde{\omega}_{{\alpha}})\end{aligned}$$ Finally, regarding $P_{1,1,2}$ and keeping in mind that hitting with all the derivatives in $z$ leads us to a term which has the factor $\tilde{z}_{{\alpha}{\alpha}{\alpha}} \cdot \tilde{z}_{{\alpha}} = - |\tilde{z}_{{\alpha}{\alpha}}|^{2}$, giving us the extra regularity we needed to integrate the term. $$\begin{aligned}
P_{1,1,2} & = \text{NICE3 } + \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{3}}\tilde{z}_{{\alpha}}^{\perp} \cdot \left(\frac{1}{2\pi}\int \frac{(\tilde{z}({\alpha})-\tilde{z}(\beta))^{\perp}}{|\tilde{z}({\alpha})-\tilde{z}(\beta)|^{2}}\tilde{\omega}_{{\alpha}{\alpha}}(\beta)d\beta\right) \\
& = \text{NICE3 } + \frac{Q^{2}}{2|\tilde{z}_{{\alpha}}|^{3}}H(\tilde{{\omega}}_{{\alpha}{\alpha}}).\end{aligned}$$ We should notice that there doesn’t appear a term proportional to $H(\tilde{{\omega}}_{{\alpha}})$ since the kernel that results from subtracting the Hilbert transform has room for two derivatives instead of one.
Adding all the previous estimates together we get the desired result.
### $\tilde{\omega}$
We first notice that $M_1$ (one of the terms in the curvature) is of the order of $z_{{\alpha}}$ and therefore it can be absorbed by the energy. Hence $${\partial_{\alpha}}K = (\tilde{K}Q)_{{\alpha}} - \tilde{K}Q_{{\alpha}} + \text{low order terms}$$
We will follow the proof done by Ambrose in [@Ambrose:well-posedness-vortex-sheet]. Taking into account the estimates for the implicit operator done in [@Cordoba-Cordoba-Gancedo:interface-water-waves-2d], we are left to see the impact of the $Q$ factor in the singular term $(\tilde{c}\tilde{\omega})_{{\alpha}}$, since the impact into the others is either trivial (the ones that come from the factor proportional to the curvature) or is zero (the rest of the terms).
$${\partial_{\alpha}}^{k}(\tilde{c}_{{\alpha}} \tilde{\omega}) = \text{NICE35 } + \frac{Q^{2}\tilde{\omega}^{2}}{2|\tilde{z}_{{\alpha}}|}H({\partial_{\alpha}}^{k}(\tilde{K})),$$
where NICE35 means $$\int Q^{j} \Lambda ({\partial_{\alpha}}^{k} (\tilde{\omega}) )\text{NICE35} \leq CE_{k}^{p}(t)$$ for some positive constants $C, p$ and any $j$.
The most singular term is when we hit all the derivatives in $\tilde{c}_{{\alpha}}$, since if we hit all of them in $\tilde{\omega}$, that term would belong to NICE35. Developing the new terms
$$\begin{aligned}
{\partial_{\alpha}}^{k}(\tilde{c}_{{\alpha}} \tilde{\omega}) & = \text{NICE35 } - \tilde{\omega} {\partial_{\alpha}}^{k}\left((Q^{2}BR)_{{\alpha}} \cdot \frac{\tilde{z}_{{\alpha}}}{|\tilde{z}_{{\alpha}}|^{2}}\right) \\
& = \text{NICE35 } - \frac{Q^{2} \tilde{\omega} }{|\tilde{z}_{{\alpha}}|^{2}} \cdot {\partial_{\alpha}}^{k}\left(\tilde{z}_{{\alpha}} \cdot \int \frac{(\tilde{z}_{{\alpha}}({\alpha}) - \tilde{z}_{{\alpha}}(\beta))^{\perp}}{|\tilde{z}({\alpha}) - \tilde{z}(\beta)|^{2}}\tilde{\omega}(\beta)d\beta\right) \\
& = \text{NICE35 } -\frac{Q^{2} \tilde{\omega}^{2}}{2|\tilde{z}_{{\alpha}}|^{4}}{\partial_{\alpha}}^{k}\left(\tilde{z}_{{\alpha}} \cdot H(\tilde{z}_{{\alpha}{\alpha}}^{\perp})\right) \\
& = \text{NICE35 } + \frac{Q^{2} \tilde{\omega}^{2}}{2|\tilde{z}_{{\alpha}}|} H\left({\partial_{\alpha}}^{k} (\tilde{K})\right). \\\end{aligned}$$
$${\partial_{\alpha}}^{k}(\tilde{c} \tilde{\omega}_{{\alpha}}) = \text{NICE35 },$$ where NICE35 means $$\int Q^{j} \Lambda ({\partial_{\alpha}}^{k} (\tilde{\omega}) )\text{NICE35} \leq CE_{k}^{p}(t)$$ for some positive constants $C, p$ and any $j$.
The most singular term is when we hit all the derivatives in $\tilde{{\omega}}_{{\alpha}}$, since if we hit all of them in $\tilde{c}$, that term would belong to NICE35. Thus, we have to estimate
$$\begin{aligned}
\int Q^{j} H {\partial_{\alpha}}^{k+1}(\tilde{\omega}) {\partial_{\alpha}}^{k+1}(\tilde{{\omega}}) \tilde{c}
& = - \int {\partial_{\alpha}}^{k+1}(\tilde{\omega}) H({\partial_{\alpha}}^{k+1}(\tilde{{\omega}}) Q^{j}\tilde{c}) \\
& = \frac{1}{2}\int {\partial_{\alpha}}^{k+1}(\tilde{\omega}) \left[H({\partial_{\alpha}}^{k+1}(\tilde{{\omega}})) \tilde{c} Q^{j} - H({\partial_{\alpha}}^{k+1}(\tilde{{\omega}}) \tilde{c} Q^{j})\right]
\leq CE_{k}^{p}(t),\end{aligned}$$
and therefore it is NICE35.
Calculations of the time derivative of the energy
-------------------------------------------------
Using the previous lemmas and propositions, we can get the following estimates for the derivative of the energy:
$$\begin{aligned}
\frac{dA}{dt} & = \text{OK } +
2\int Q^{2k+1}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k}\left(Q^{2}H(\tilde{\omega}_{{\alpha}{\alpha}}) + 4QQ_{{\alpha}}H(\tilde{\omega}_{{\alpha}})\right) \\
& = \text{OK } + 2\int 2kQ^{2k+2}Q_{{\alpha}}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k-1}\left(H(\tilde{\omega}_{{\alpha}{\alpha}})\right) \\
& + 2\int Q^{2k+3}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k}\left(H(\tilde{\omega}_{{\alpha}{\alpha}})\right) \\
& + 2\int 4Q^{2k+2}Q_{{\alpha}}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k}\left(H(\tilde{\omega}_{{\alpha}})\right) \\
& = A^{1} + A^{2} + A^{3},\end{aligned}$$
where we will say that a term is OK if it is controlled by the energy.
We should be careful while estimating $B_t$ because
$$\begin{aligned}
\frac{dB}{dt} & = \text{OK } + \frac{1}{\tau}\int Q^{2k+2}{\partial_{\alpha}}^{k}(\tilde{\omega}_t)\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))
+ \frac{1}{\tau}\int Q^{2k+2}{\partial_{\alpha}}^{k}(\tilde{\omega})\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}_t)) \\
& = \text{OK } + \frac{1}{\tau}\int Q^{2k+2}{\partial_{\alpha}}^{k}(\tilde{\omega}_t)\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))
+ \frac{1}{\tau}\int \Lambda(Q^{2k+2}{\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{\omega}_t) \\
& = \text{OK } + \frac{2}{\tau}\int Q^{2k+2}{\partial_{\alpha}}^{k}(\tilde{\omega}_t)\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))
+ \frac{1}{\tau}\int (Q^{2k+2})_{{\alpha}}H({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{\omega}_t) \\\end{aligned}$$
Hence
$$\begin{aligned}
\frac{dB}{dt} & = \text{OK } + \frac{2}{\tau}\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))\frac{Q^{2}\tilde{\omega}^{2}}{2|\tilde{z}_{{\alpha}}|}H({\partial_{\alpha}}^{k}(\tilde{K})) \\
& + 2\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}((Q\tilde{K} + \frac{Q^{3}}{|\tilde{z}_{{\alpha}}|^{3}}X)_{{\alpha}}) \\
& + \int (2k+2)Q^{2k+1}Q_{{\alpha}}H({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k+1}((\tilde{K}Q)) \\
& = \text{OK } + \frac{2}{\tau}\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))\frac{Q^{2}\tilde{\omega}^{2}}{2|\tilde{z}_{{\alpha}}|}H({\partial_{\alpha}}^{k}(\tilde{K})) \\
& + 2\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}((Q\tilde{K})_{{\alpha}})
- 2\int Q^{2k+2}Q_{{\alpha}}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K}) \\
& + \int(2k+2) Q^{2k+2}Q_{{\alpha}}H({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k+1}(\tilde{K}) \\
& = \text{OK } + B^{1} + B^{2} + B^{3} + B^{4}\end{aligned}$$
$$\frac{dC}{dt} = \text{OK } + \frac{1}{|\tilde{z}_{{\alpha}}|\tau}\int Q^{2k+4}\tilde{\omega}^{2}{\partial_{\alpha}}^{k}(\tilde{\omega}){\partial_{\alpha}}^{k+1}(\tilde{K}) =
\text{OK } + C^{1}$$
Development of the derivative in $B$
------------------------------------
We start from the development of $B^1$, $B^2$, $B^3$ and $B^{4}$. We trivially have:
$$\begin{aligned}
B^1 & = \frac{1}{\tau}\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))\frac{Q^{2}\tilde{\omega}^{2}}{|\tilde{z}_{{\alpha}}|}H({\partial_{\alpha}}^{k}(\tilde{K})) \\
B^3 & = - 2\int Q^{2k+2}Q_{{\alpha}}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K}) \\
B^{4} & = \text{OK } -\int(2k+2) Q^{2k+2}Q_{{\alpha}}H({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K})\end{aligned}$$
We now look at $B^{2}$. We can decompose it in the following way $$\begin{aligned}
B^{2} & = 2\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}(Q_{{\alpha}}\tilde{K} + Q\tilde{K}_{{\alpha}}) \\
& = \text{OK } + 2\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))(Q_{{\alpha}}{\partial_{\alpha}}^{k}(\tilde{K}) + Q{\partial_{\alpha}}^{k+1}(\tilde{K}) + kQ_{{\alpha}}{\partial_{\alpha}}^{k}(\tilde{K})) \\
& = \text{OK } + B^{2,1} + B^{2,2} + B^{2,3}\end{aligned}$$
We can write down the terms $B^{2,1}$ and $B^{2,3}$ in the form $$\begin{aligned}
B^{2,1} & = 2\int Q^{2k+2}H({\partial_{\alpha}}^{k+1}(\tilde{\omega}))Q_{{\alpha}}{\partial_{\alpha}}^{k}(\tilde{K}) \\
B^{2,3} & = 2k\int Q^{2k+2}H({\partial_{\alpha}}^{k+1}(\tilde{\omega}))Q_{{\alpha}}{\partial_{\alpha}}^{k}(\tilde{K})\end{aligned}$$
Integrating by parts in $B^{2,2}$ we establish
$$\begin{aligned}
B^{2,2} & = -2\int Q^{2k+3}\Lambda({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K}) \\
& -2(2k+3)\int Q^{2k+2}Q_{{\alpha}}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K}) \\
& = B^{2,2,1} + B^{2,2,2}\end{aligned}$$
Again, $B^{2,2,2}$ can easily be reduced to the canonical form $$\begin{aligned}
B^{2,2,2} & = -2(2k+3)\int Q^{2k+2}Q_{{\alpha}}H({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K})\end{aligned}$$
Collection of the terms
-----------------------
We will split all the uncontrolled terms into three categories: high order and low order types I and II and we will see that the sum of the terms in each category adds up to low enough order terms, denoted by OK.
### High Order
$$\begin{array}{lr}
\displaystyle 2\int Q^{2k+3}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k}\left(H(\tilde{\omega}_{{\alpha}{\alpha}})\right) & (A^{2})
\end{array}$$ $$\begin{array}{lr}
\displaystyle -2\int Q^{2k+3}\Lambda({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K}) & (B^{2,2,1})
\end{array}$$
No terms from $C$.
### Low Order Type I
$$\begin{array}{lr}
\displaystyle 2\int 2kQ^{2k+2}Q_{{\alpha}}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k-1}\left(H(\tilde{\omega}_{{\alpha}{\alpha}})\right) & (A^{1}) \\
\displaystyle 2\int 4Q^{2k+2}Q_{{\alpha}}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k}\left(H(\tilde{\omega}_{{\alpha}})\right) & (A^{3}) \\
\end{array}$$ $$\begin{array}{lr}
\displaystyle - 2\int Q^{2k+2}Q_{{\alpha}}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K}) & (B^{3}) \\
\displaystyle -\int(2k+2) Q^{2k+2}Q_{{\alpha}}H({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K}) & (B^{4}) \\
\displaystyle 2\int Q^{2k+2}H({\partial_{\alpha}}^{k+1}(\tilde{\omega}))Q_{{\alpha}}{\partial_{\alpha}}^{k}(\tilde{K})) & (B^{2,1})\\
\displaystyle 2k\int Q^{2k+2}H({\partial_{\alpha}}^{k+1}(\tilde{\omega}))Q_{{\alpha}}{\partial_{\alpha}}^{k}(\tilde{K}) & (B^{2,3}) \\
\displaystyle -2(2k+3)\int Q^{2k+2}Q_{{\alpha}}H({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\partial_{\alpha}}^{k}(\tilde{K}) & (B^{2,2,2}) \\
\end{array}$$
No terms from $C$.
### Low Order Type II
No terms from $A$.
$$\begin{array}{lr}
\displaystyle \frac{1}{\tau}\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))\frac{Q^{2}\tilde{\omega}^{2}}{|\tilde{z}_{{\alpha}}|}H({\partial_{\alpha}}^{k}(\tilde{K}))
& (B^{1})
\end{array}$$
$$\begin{array}{lr}
\displaystyle \frac{1}{|\tilde{z}_{{\alpha}}|\tau}\int Q^{2k+4}\tilde{\omega}^{2}{\partial_{\alpha}}^{k}(\tilde{\omega}){\partial_{\alpha}}^{k+1}(\tilde{K}) & (C^{1})
\end{array}$$
Regularized system
------------------
Now, let $\tilde{z}^{{\varepsilon},\delta,\mu} ({\alpha},t)$ be a solution of the following system (compare with (\[zeq\] - \[eqomega\])): $$\begin{aligned}
\label{epsdelmuzt}
\begin{split}\tilde{z}^{{\varepsilon},\delta,\mu}_t({\alpha},t)&=\phi_{\delta} * \phi_{\delta} * \left(Q^2(\tilde{z}^{{\varepsilon},\delta,\mu})BR(\tilde{z}^{{\varepsilon},\delta,\mu},\tilde{{\omega}}^{{\varepsilon},\delta,\mu})\right)({\alpha},t)+\phi_{\mu} * \left(\tilde{c}^{{\varepsilon},\delta,\mu}\left(\phi_{\mu} * {\partial_{\alpha}}\tilde{z}^{{\varepsilon},\delta,\mu}\right)\right)({\alpha},t),
\end{split}\end{aligned}$$
$$\begin{aligned}
\label{epsdelmuwt}
\tilde{\omega}&^{{\varepsilon},\delta,\mu}_{t}=\phi_{\delta} * \phi_{\delta} *\left(-2 BR_t(\tilde{z}^{{\varepsilon},\delta,\mu},\tilde{\omega}^{{\varepsilon},\delta,\mu}) \cdot \tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}}- |BR(\tilde{z}^{{\varepsilon},\delta,\mu},\tilde{\omega}^{{\varepsilon},\delta,\mu})|^{2} (Q^{2}(\tilde{z}^{{\varepsilon}\delta,\mu}))_{{\alpha}}\right. \nonumber \\
&- \Big(\frac{Q^2\tilde{(\omega^{{\varepsilon},\delta,\mu})^2}}{4|\tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}}|^{2}}\Big)_{{\alpha}}
+ 2\overline{c}^{{\varepsilon},\delta,\mu} BR_{\alpha}(\tilde{z}^{{\varepsilon},\delta,\mu},\omega^{{\varepsilon},\delta,\mu}) \cdot \tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}} + \left(\overline{c}^{{\varepsilon},\delta,\mu}\tilde{\omega}^{{\varepsilon},\delta,\mu}\right)_{{\alpha}}- 2 \left(P^{-1}_2(\tilde{z}^{{\varepsilon},\delta,\mu}({\alpha},t))\right)_{{\alpha}} \nonumber \\
&+\tau \left( \frac{Q^{3}(\tilde{z}^{{\varepsilon},\delta,\mu})}{|\tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}}({\alpha},t)|^{3}}
(\tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}})^{T}HP_{2}^{-1} \tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}}\nabla P_{1}^{-1}\cdot \tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}}
-(\tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}})^{T}HP_{1}^{-1} \tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}}\nabla P_{2}^{-1} \cdot \tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}})\right)_{{\alpha}}
\nonumber\\
&\left.+\tau\left(Q\frac{\tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}{\alpha}}\cdot (\tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}})^{\bot}}{|\tilde{z}^{{\varepsilon},\delta,\mu}_{{\alpha}}|^3}\right)_{{\alpha}}\right)- {\varepsilon}\phi_\mu*\phi_\mu*\left(\Lambda (\tilde{\omega}^{{\varepsilon},\delta,\mu})\frac{1}{Q^{2k+3}}\right)\end{aligned}$$
$\tilde{z}^{{\varepsilon},\delta,\mu}({\alpha},0)=\tilde{z}_0({\alpha})$ and $\tilde{{\omega}}^{{\varepsilon},\delta,\mu}({\alpha},0)=\tilde{{\omega}}_0({\alpha})$ for ${\varepsilon}>0$, $ \delta > 0, \mu > 0$. The functions $\phi_{\delta}$ and $\phi_{\mu}$ are even mollifiers,
$$\begin{aligned}
\tilde{c}^{{\varepsilon},\delta,\mu}({\alpha})=&\frac{{\alpha}+\pi}{2\pi}\int_{-\pi}^\pi\frac{\partial_{\beta} \tilde{z}^{{\varepsilon},\delta,\mu}({\beta}))}{|\partial_{\beta}
\tilde{z}^{{\varepsilon},\delta,\mu}({\beta})|^2}\cdot \phi_{\delta} * \phi_{\delta} * (\partial_{\beta} (Q^2(\tilde{z}^{{\varepsilon},\delta,\mu})({\beta}) BR(\tilde{z}^{{\varepsilon},\delta,\mu},\tilde{{\omega}}^{{\varepsilon},\delta,\mu}))({\beta})) d{\beta}\\
& -\int_{-\pi}^{\alpha}\frac{\partial_{\beta} \tilde{z}^{{\varepsilon},\delta,\mu}(\beta)}{|\partial_{\beta} \tilde{z}^{{\varepsilon},\delta,\mu}(\beta)|^2}\cdot\phi_{\delta} * \phi_{\delta} * (\partial_{\beta} (Q^2(\tilde{z}^{{\varepsilon},\delta,\mu})(\beta)
BR(\tilde{z}^{{\varepsilon},\delta,\mu},\tilde{{\omega}}^{{\varepsilon},\delta,\mu}))(\beta)) d\beta,\end{aligned}$$
and $$\begin{aligned}
\overline{c}^{{\varepsilon},\delta,\mu}({\alpha})=&\frac{{\alpha}+\pi}{2\pi}\int_{-\pi}^\pi\frac{\partial_{\beta} \tilde{z}^{{\varepsilon},\delta,\mu}({\beta}))}{|\partial_{\beta}
\tilde{z}^{{\varepsilon},\delta,\mu}({\beta})|^2}\cdot (\partial_{\beta} (Q^2(\tilde{z}^{{\varepsilon},\delta,\mu})({\beta}) BR(\tilde{z}^{{\varepsilon},\delta,\mu},\tilde{{\omega}}^{{\varepsilon},\delta,\mu}))({\beta})) d{\beta}\\
& -\int_{-\pi}^{\alpha}\frac{\partial_{\beta} \tilde{z}^{{\varepsilon},\delta,\mu}(\beta)}{|\partial_{\beta} \tilde{z}^{{\varepsilon},\delta,\mu}(\beta)|^2}\cdot (\partial_{\beta} (Q^2(\tilde{z}^{{\varepsilon},\delta,\mu})(\beta)
BR(\tilde{z}^{{\varepsilon},\delta,\mu},\tilde{{\omega}}^{{\varepsilon},\delta,\mu}))(\beta)) d\beta,\end{aligned}$$
The RHS of the evolution equations for $\tilde{z}^{{\varepsilon},\delta,\mu}$ and $\tilde{\omega}^{{\varepsilon},\delta,\mu}$ are Lipschitz in the spaces $H^{k+2}({\mathbb{T}})$ and $H^{k+\frac{1}{2}}({\mathbb{T}})$ since they are mollified. Therefore we can solve (\[epsdelmuzt\]-\[epsdelmuwt\]) for short time, thanks to Picard’s theorem.
Now, we can perform energy estimates to get uniform bounds in $\mu$ (we just deal with a transport term and a dissipative) and we can let $\mu$ go to zero. The energy estimates that we can get are the following:
$$\begin{aligned}
& \frac{d}{dt}\left(\|\tilde{z}^{{\varepsilon},\delta,\mu}\|^2_{H^5}
+\|{\mathcal{F}}(\tilde{z}^{{\varepsilon},\delta,\mu})\|^2_{L^\infty}+\|\tilde{{\omega}}^{{\varepsilon},\delta,\mu}\|^2_{H^{3+\frac12}}+\sum_{l=0}^{4}\frac{1}{m^{{\varepsilon},\delta,\mu}(q^l)}\right)(t)\\
&\leq
C(\delta)\left(\|\tilde{z}^{{\varepsilon},\delta,\mu}\|^2_{H^5}+\|{\mathcal{F}}(\tilde{z}^{{\varepsilon},\delta,\mu})\|^2_{L^\infty}+
\|\tilde{{\omega}}^{{\varepsilon},\delta,\mu}\|^2_{H^{3+\frac12}}+\sum_{l=0}^4\frac{1}{m^{{\varepsilon},\delta,\mu}(q^l)}\right)^j(t).\end{aligned}$$
We should note that for the new system without the $\phi_\mu$ mollifier, the length of the tangent vector $|{\partial_{\alpha}}\tilde{z}^{\delta}|$ is now constant in space and depends only on time. Next we will perform energy estimates as in the previous case by using the curvature $\tilde{K}^{\delta}$ from the curve $\tilde{z}^{\delta}$.
Similarly, we get (let us omit the superscript $\delta,\varepsilon$ in $\tilde{z}^{\delta,\varepsilon}$ and $\tilde{{\omega}}^{\delta,\varepsilon}$)
- $$\tilde{K}_t = \text{NICE3 } + \frac{Q^{2}}{2|\tilde{z}_{{\alpha}}|^{3}}{\phi_{\delta} * \phi_{\delta} *}H(\tilde{\omega}_{{\alpha}{\alpha}} )+ \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}(Q^{2})_{{\alpha}}{\phi_{\delta} * \phi_{\delta} *}H(\tilde{\omega}_{{\alpha}}),$$
- $${\partial_{\alpha}}^{k}(\overline{c}_{{\alpha}} \tilde{\omega}) = \text{NICE35 } + \frac{Q^{2}\tilde{\omega}^{2}}{2|\tilde{z}_{{\alpha}}|}H({\partial_{\alpha}}^{k}(\tilde{K})),$$
- $${\partial_{\alpha}}^{k}(\tilde{c} \tilde{\omega}_{{\alpha}}) = \text{NICE35 },$$
and the following collection of terms:
### High Order
$$\begin{array}{lr}
\displaystyle 2\int Q^{2k+3}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k}{\phi_{\delta} * \phi_{\delta} *}\left(H(\tilde{\omega}_{{\alpha}{\alpha}})\right) & (A^{2})
\end{array}$$ $$\begin{array}{lr}
\displaystyle -2\int Q^{2k+3}\Lambda({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k}(\tilde{K}) & (B^{2,2,1}) \\
\displaystyle -2\frac{{\varepsilon}}{\tau}\|{\partial_{\alpha}}^{k+1} \tilde{\omega}\|^{2}_{L^{2}} & (D)
\end{array}$$
No terms from $C$.
### Low Order Type I
$$\begin{array}{lr}
\displaystyle 2\int 2kQ^{2k+2}Q_{{\alpha}}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\partial_{\alpha}}^{k-1}{\phi_{\delta} * \phi_{\delta} *}\left(H(\tilde{\omega}_{{\alpha}{\alpha}})\right) & (A^{1}) \\
\displaystyle 2\int 4Q^{2k+2}Q_{{\alpha}}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k}\left(H(\tilde{\omega}_{{\alpha}})\right) & (A^{3}) \\
\end{array}$$ $$\begin{array}{lr}
\displaystyle - 2\int Q^{2k+2}Q_{{\alpha}}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega})){\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k}(\tilde{K}) & (B^{3}) \\
\displaystyle -\int(2k+2) Q^{2k+2}Q_{{\alpha}}H({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k}(\tilde{K}) & (B^{4}) \\
\displaystyle 2\int Q^{2k+2}H({\partial_{\alpha}}^{k+1}(\tilde{\omega}))Q_{{\alpha}}{\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k}(\tilde{K})) & (B^{2,1})\\
\displaystyle 2k\int Q^{2k+2}H({\partial_{\alpha}}^{k+1}(\tilde{\omega}))Q_{{\alpha}}{\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k}(\tilde{K}) & (B^{2,3}) \\
\displaystyle -2(2k+3)\int Q^{2k+2}Q_{{\alpha}}H({\partial_{\alpha}}^{k+1}(\tilde{\omega})){\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k}(\tilde{K}) & (B^{2,2,2}) \\
\end{array}$$
No terms from $C$.
### Low Order Type II
No terms from $A$.
$$\begin{array}{lr}
\displaystyle \frac{1}{\tau}\int Q^{2k+2}\Lambda({\partial_{\alpha}}^{k}(\tilde{\omega}))\frac{Q^{2}\tilde{\omega}^{2}}{|\tilde{z}_{{\alpha}}|}{\phi_{\delta} * \phi_{\delta} *}H({\partial_{\alpha}}^{k}(\tilde{K}))
& (B^{1})
\end{array}$$
$$\begin{array}{lr}
\displaystyle \frac{1}{|\tilde{z}_{{\alpha}}|\tau}\int Q^{2k+4}\tilde{\omega}^{2}{\partial_{\alpha}}^{k}(\tilde{\omega}){\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k+1}(\tilde{K}) & (C^{1})
\end{array}$$
We note that throughout this section we have repeatedly used the following commutator estimate for convolutions:
$$\label{commmol}
\|\phi_{\delta} * ({\partial_{\alpha}}f g) - g \phi_{\delta} * ({\partial_{\alpha}}f)\|_{L^{2}} \leq C\|{\partial_{\alpha}}g\|_{L^{\infty}} \|f\|_{L^{2}},$$
where the constant $C$ is independent of $\delta, f$ and $g$.
Also using this commutator estimate we can find all the cancelations we need in the previous collection of terms of low order type I and II to obtain a suitable energy estimate.
Regarding the high order terms, we will do the estimates in detail. We will see the need for the dissipative term since there are terms that escape for half of a derivative.
$$\begin{aligned}
A^{2} + B^{2,2,1} + D & =
2\int Q^{2k+3}\left({\partial_{\alpha}}^{k}(\tilde{K})\right){\phi_{\delta} * \phi_{\delta} *}\left(H({\partial_{\alpha}}^{k+2} \tilde{\omega})\right) \\
& -2\int Q^{2k+3}H({\partial_{\alpha}}^{k+2}(\tilde{\omega})){\phi_{\delta} * \phi_{\delta} *}{\partial_{\alpha}}^{k}(\tilde{K})
-2{\varepsilon}\|{\partial_{\alpha}}^{k+1} \tilde{\omega}\|^{2}_{L^{2}} \\
& = 2 \int {\partial_{\alpha}}^{k}(\tilde{K})\left(Q^{2k+3} {\phi_{\delta} * \phi_{\delta} *}H({\partial_{\alpha}}^{k+2} \tilde{\omega}) - {\phi_{\delta} * \phi_{\delta} *}\left(Q^{2k+3} H ({\partial_{\alpha}}^{k+2} \tilde{\omega})\right)\right)
-2{\varepsilon}\|{\partial_{\alpha}}^{k+1} \tilde{\omega}\|^{2}_{L^{2}} \\
& \leq \|{\partial_{\alpha}}^{k} \tilde{K}\|_{L^{2}}\|{\partial_{\alpha}}Q^{2k+3}\|_{L^{\infty}}\|{\partial_{\alpha}}^{k+1}\tilde{\omega}\|_{L^{2}} -2{\varepsilon}\|{\partial_{\alpha}}^{k+1} \tilde{\omega}\|^{2}_{L^{2}}
\leq C({\varepsilon})E^{p}(t),
$$
which is uniform in $\delta$. This proves that we can pass to the limit $\delta \to 0$.
Finally, by applying the a priori energy estimates to the new system (which only depend on ${\varepsilon}$) we can pass to the limit ${\varepsilon}\to 0$ since now we don’t have the previous problems and $A^{2} + B^{2,2,1} = 0$.
Energy with the Rayleigh-Taylor condition {#SectionEnergyRT}
=========================================
In this section, we prove local existence in the tilde domain, where the time of existence does not depend on the surface tension coefficient. In this theorem, we need initial data to satisfy the Rayleigh-Taylor condition as we explain in Section \[SectionInitialData\]. This Rayleigh-Taylor condition will hold in particular if the surface tension coefficient is small enough.
\[localexistencetilde\] Let $k \geq 3$. Let $\tilde{z}^{0}({\alpha})$ be the image of a splash curve by the map $P$ parametrized in such a way that $|{\partial_{\alpha}}\tilde{z}^{0}({\alpha})| = \frac{L}{2\pi}$, where $L$ is the length of the curve in a fundamental period, and such that $\tilde{z}_{1}^{0}({\alpha}), \tilde{z}_2^{0}({\alpha}) \in H^{k+2}({\mathbb{T}})$. Let $\tilde{\varphi}({\alpha},0) \in H^{k+\frac{1}{2}}({\mathbb{T}})$ be as in and let $\tilde{\omega}({\alpha},0) \in H^{k-1}({\mathbb{T}})$. Then there exist a finite time $T > 0$, a time-varying curve $\tilde{z}({\alpha},t) \in C([0,T];H^{k+2})$, and functions $\tilde{\omega}({\alpha},t) \in C([0,T];H^{k-1})$ and $\tilde{\varphi} \in C([0,T]; H^{k+\frac{1}{2}})$ providing a solution of the water wave equations (\[zeq\] - \[eqomega\]). Assume that initially, the Rayleigh-Taylor condition is strictly positive.
In order to prove this theorem we will use the solutions we have obtained in theorem \[localexistencetildeWRT\] for $\tau>0$. We will perform energy estimates on these solutions.
The energy {#the-energy}
----------
We will define the energy for $k \geq 3$ as
$$\begin{aligned}
E^2_{k}(t) & = \mathcal{EE}^{2}(t) + \underbrace{\tau \frac{|\tilde{z}_{{\alpha}}|}{2}\int Q^{2k+1}\left({\partial_{\alpha}}^{k}(\tilde{K})\right)^{2}}_{A}
+ \underbrace{\int Q^{2k-2}{\partial_{\alpha}}^{k}(\tilde{{\varphi}})\Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}}))}_{B} \\
& + \underbrace{|\tilde{z}_{{\alpha}}|^{2} \tau \int (\mathcal{C}\|\tilde{K}(t)\|_{H^{1}} + \tilde{K}) Q^{2k+1}{\partial_{\alpha}}^{k-1}(\tilde{K})\Lambda({\partial_{\alpha}}^{k-1}(\tilde{K}))}_{C}
+ \underbrace{2|\tilde{z}_{{\alpha}}| \int \mathcal{C}\|\tilde{K}(t)\|_{H^{1}} Q^{2k-2}\left({\partial_{\alpha}}^{k}(\tilde{{\varphi}})\right)^{2}}_{D} \\
& + \underbrace{|\tilde{z}_{{\alpha}}|^{2} \int \sigma Q^{2k}\left({\partial_{\alpha}}^{k-1}(\tilde{K})\right)^{2}}_{E}
+ \frac{|\tilde{z}_{{\alpha}}|^{2}}{m(Q^{2k}\sigma)(t)},\end{aligned}$$
where $m(Q^{2k}\sigma)=\min_{{\alpha}\in{\mathbb{T}}}Q^{2k}(\tilde{z}({\alpha},t))\sigma({\alpha},t)$ and $\mathcal{C}$ is a sufficiently large constant such that $C$ is strictly positive. Remember that $\tilde{\varphi}$ was introduced in Equation \[varphi\].
At this point is important to notice the following.
The following sentences hold. \[lemmaequivenergies\]
1. Let $\tilde{\varphi}\in H^{3+\frac{1}{2}}$, $\tilde{{\omega}}\in H^{2}$ and $z\in H^k$ with $k\geq 4$. Then $\tilde{{\omega}}\in H^3$.
2. Let $\tilde{\varphi}\in H^{3+\frac{1}{2}}$, $\tilde{{\omega}}\in H^{3}$ and $z\in H^k$ with $k\geq 5$. Then $\tilde{{\omega}}\in H^{3.5}$.
3. Let $\tilde{\omega}\in H^{3+\frac{1}{2}}$, and $\tilde{z}\in H^k$ with $k\geq 5$. Then $\tilde{\varphi}\in H^{3.5}$.
This lemma shows that for a fixed $\tau>0$ the energy of this section is equivalent to this one in section \[primeraenergia\]. This allows us to use this energy to extend the solutions of the theorem \[localexistencetildeWRT\] up to a time $T$ which does not depend on $\tau$ (for a small enough $\tau$).
The energy estimates
--------------------
Again, we will only focus on the new terms ($A-E$) since the estimates for the other ones were proved in [@Cordoba-Cordoba-Gancedo:interface-water-waves-2d] and in [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:finite-time-singularities-Euler].
### $\tilde{K}$
$$\begin{aligned}
\tilde{K}_t & = \text{NICE3B } + \frac{Q^{2}}{2|\tilde{z}_{{\alpha}}|^{3}}H(\tilde{\omega}_{{\alpha}{\alpha}}) + \frac{1}{|\tilde{z}_{{\alpha}}|^{3}}(Q^{2})_{{\alpha}}H(\tilde{\omega}_{{\alpha}}) \\
& = \text{NICE3B } + \frac{1}{|\tilde{z}_{{\alpha}}|^{2}}H(\tilde{{\varphi}}_{{\alpha}{\alpha}}) - \frac{1}{|\tilde{z}_{{\alpha}}|}(\tilde{K}\tilde{{\varphi}})_{{\alpha}},\end{aligned}$$
where NICE3B means $$\int Q^{j}{\partial_{\alpha}}^{k}(\tilde{K}) {\partial_{\alpha}}^{k}(NICE3B) \leq CE_{k}^{p}(t)$$ for some positive constants $C, p$ and any $j$.
The first equality follows from the proof from the last section since the energies are equivalent (see Lemma \[lemmaequivenergies\]). We now prove the second one. We begin by using the relation to get $$\begin{aligned}
\tilde{K}_t & = \text{NICE3B } + \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}}H\left(\left(\frac{\tilde{{\varphi}}}{Q^{2}}\right)_{{\alpha}{\alpha}}\right)
+ \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|}H\left(\left(\frac{\tilde{c}}{Q^{2}}\right)_{{\alpha}{\alpha}}\right) \\
& + \frac{2(Q^{2})_{{\alpha}}}{|\tilde{z}_{{\alpha}}|^{2}}H\left(\left(\frac{\tilde{{\varphi}}}{Q^{2}}\right)_{{\alpha}}\right)
+ \frac{2(Q^{2})_{{\alpha}}}{|\tilde{z}_{{\alpha}}|}H\left(\left(\frac{\tilde{c}}{Q^{2}}\right)_{{\alpha}}\right) = I + J \\\end{aligned}$$ We can easily see that $$\tilde{c}_{{\alpha}} = - \frac{\tilde{z}_{{\alpha}}}{|\tilde{z}_{{\alpha}}|^{2}} \cdot (Q^{2} BR)_{{\alpha}} = \text{NICE3B}$$ since it is at the level of $\tilde{{\omega}}_{{\alpha}}, \tilde{z}_{{\alpha}{\alpha}}$ but we gain one derivative by multiplying by the tangential direction. This proves that $$\begin{aligned}
J = \text{NICE3B } + \frac{2(Q^{2})_{{\alpha}}}{|\tilde{z}_{{\alpha}}|^{2}}H\left(\frac{\tilde{{\varphi}}_{{\alpha}}}{Q^{2}}\right).\end{aligned}$$ Looking now to $\tilde{c}_{{\alpha}{\alpha}}$ we can see that $$\begin{aligned}
\tilde{c}_{{\alpha}{\alpha}} = - \frac{\tilde{z}_{{\alpha}{\alpha}}}{|\tilde{z}_{{\alpha}}|^{2}} \cdot (Q^{2} BR)_{{\alpha}}
- \frac{\tilde{z}_{{\alpha}}}{|\tilde{z}_{{\alpha}}|^{2}} \cdot (Q^{2} BR)_{{\alpha}{\alpha}} = I_1 + I_2.\end{aligned}$$ Using the standard estimates, the only thing that causes trouble in $I_1$ is when all the derivatives hit $\tilde{\omega}$ and therefore $$\begin{aligned}
I_1 = \text{NICE3B } - K\frac{Q^{2}}{2|\tilde{z}_{{\alpha}}|}H(\tilde{{\omega}}_{{\alpha}}).\end{aligned}$$ Regarding $I_2$, again, we need all the derivatives to hit $BR$ to get the most singular terms, which are
$$\begin{aligned}
I_2 & = \text{NICE3B } - \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}} \tilde{z}_{{\alpha}} \cdot \left[\frac{2}{2\pi}\int_{-\pi}^{\pi} \frac{(\tilde{z}_{{\alpha}}({\alpha}) - \tilde{z}_{{\alpha}}(\beta))^{\perp}}{|\tilde{z}({\alpha})-\tilde{z}(\beta)|^{2}}\tilde{\omega}_{{\alpha}}({\alpha}-\beta)d\beta\right] \\
& - \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}} \tilde{z}_{{\alpha}} \cdot \left[\frac{1}{2\pi}\int_{-\pi}^{\pi} \frac{(\tilde{z}({\alpha}) - \tilde{z}(\beta))^{\perp}}{|\tilde{z}({\alpha})-\tilde{z}(\beta)|^{2}}\tilde{\omega}_{{\alpha}{\alpha}}({\alpha}-\beta)d\beta\right] \\
& - \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}} \tilde{z}_{{\alpha}} \cdot \left[\frac{1}{2\pi}\int_{-\pi}^{\pi} \frac{(\tilde{z}_{{\alpha}{\alpha}}({\alpha}) - \tilde{z}_{{\alpha}{\alpha}}(\beta))^{\perp}}{|\tilde{z}({\alpha})-\tilde{z}(\beta)|^{2}}\tilde{\omega}_{{\alpha}}({\alpha}-\beta)d\beta\right] \\
& = \text{NICE3B } +\frac{2Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}} \frac{1}{2}\frac{\tilde{z}_{{\alpha}} \cdot \tilde{z}_{{\alpha}{\alpha}}^{\perp}}{|\tilde{z}_{{\alpha}}|^{2}}H(\tilde{\omega}_{{\alpha}}) - \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}}\frac{\tilde{z}_{{\alpha}} \cdot \tilde{z}_{{\alpha}{\alpha}}^{\perp}}{|\tilde{z}_{{\alpha}}|^{2}}H(\tilde{\omega}_{{\alpha}}) - \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}}\frac{1}{2}\frac{\tilde{\omega}}{|\tilde{z}_{{\alpha}}|^{2}} \tilde{z}_{{\alpha}} \cdot H(\tilde{z}_{{\alpha}{\alpha}{\alpha}}^{\perp})\end{aligned}$$
Collecting all the terms from $I_1$ and $I_2$, we obtain $$\begin{aligned}
\frac{\tilde{c}_{{\alpha}{\alpha}}}{Q^{2}} & = \text{NICE3B } + \frac{1}{2|\tilde{z}_{{\alpha}}|}H((\tilde{K}\tilde{\omega})_{{\alpha}}) \\
& = \text{NICE3B } + \frac{1}{Q^{2}}H((\tilde{K} \tilde{{\varphi}})_{{\alpha}}). \\\end{aligned}$$
We can finally write the total contribution as
$$\begin{aligned}
\tilde{K}_t & = \text{NICE3B } + \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}}H\left(\frac{\tilde{{\varphi}}_{{\alpha}{\alpha}}}{Q^{2}}\right)
- \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}}H\left(\frac{4Q_{{\alpha}}\tilde{{\varphi}}_{{\alpha}}}{Q^{3}}\right) \\
& - \frac{1}{|\tilde{z}_{{\alpha}}|}(\tilde{K} \tilde{{\varphi}})_{{\alpha}}
+ \frac{2(Q^{2})_{{\alpha}}}{|\tilde{z}_{{\alpha}}|^{2}}H\left(\frac{\tilde{{\varphi}}_{{\alpha}}}{Q^{2}}\right) \\
& = \text{NICE3B } + \frac{Q^{2}}{|\tilde{z}_{{\alpha}}|^{2}}H\left(\frac{\tilde{{\varphi}}_{{\alpha}{\alpha}}}{Q^{2}}\right)- \frac{1}{|\tilde{z}_{{\alpha}}|}(\tilde{K} \tilde{{\varphi}})_{{\alpha}} \\
& = \text{NICE3B } + \frac{1}{|\tilde{z}_{{\alpha}}|^{2}}H\left(\tilde{{\varphi}}_{{\alpha}{\alpha}}\right)- \frac{1}{|\tilde{z}_{{\alpha}}|}(\tilde{K} \tilde{{\varphi}})_{{\alpha}}\end{aligned}$$
as we wanted to prove.
### $\tilde{\varphi}$
Throughout this section, we will use the following estimate which was proved in [@Castro-Cordoba-Fefferman-Gancedo-GomezSerrano:finite-time-singularities-Euler] for the case without surface tension. The proof is exactly the same for the case with it.
$$\begin{aligned}
\varphi_{{\alpha}t} = \text{NICE2B } + \frac{\tilde{{\varphi}} \tilde{{\varphi}}_{{\alpha}{\alpha}}}{|\tilde{z}_{{\alpha}}|} - Q^{2} \sigma \tilde{K} + \tau \frac{Q^{2}}{2|\tilde{z}_{{\alpha}}|}\left((\tilde{K}Q)_{{\alpha}} + M_{{\alpha}}\right),\end{aligned}$$
where NICE2B means $$\int Q^{j}\Lambda({\partial_{\alpha}}^{k}(\tilde{\varphi})) {\partial_{\alpha}}^{k-1}(NICE2B) \leq CE_{k}^{p}(t)$$ for some positive constants $C, p$ and any $j$.
Calculations of the time derivative of the energy
-------------------------------------------------
Using the previous lemmas and propositions, we can get the following estimates for the derivative of the energy: $$\begin{aligned}
\frac{dA}{dt} & = \text{OK } + \frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k+1}{\partial_{\alpha}}^{k}(\tilde{K}){\partial_{\alpha}}^{k}(H(\tilde{{\varphi}}_{{\alpha}{\alpha}}))
- \tau\int Q^{2k+1}{\partial_{\alpha}}^{k}(\tilde{K}){\partial_{\alpha}}^{k}((\tilde{K} \tilde{{\varphi}})_{{\alpha}}) \\
& = \text{OK } + \frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k+1}{\partial_{\alpha}}^{k}(\tilde{K}){\partial_{\alpha}}^{k}(H(\tilde{{\varphi}}_{{\alpha}{\alpha}}))
- \tau\int Q^{2k+1}{\partial_{\alpha}}^{k}(\tilde{K}){\partial_{\alpha}}^{k+1}( \tilde{{\varphi}})\tilde{K}
= \text{OK } + A^1 + A^2\end{aligned}$$ Again, we need to be careful while computing the derivative of $B$ as in Section \[SectionEnergyWRT\]. We obtain $$\begin{aligned}
\frac{dB}{dt} & = 2\int Q^{2k-2} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k-1}(\tilde{{\varphi}}_{{\alpha}t})
+ \int (Q^{2k-2})_{{\alpha}} H({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k-1}(\tilde{{\varphi}}_{{\alpha}t}) \\
& = \text{OK } -2\int Q^{2k-2} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k-1}\left(\frac{\tilde{{\varphi}} \tilde{{\varphi}}_{{\alpha}{\alpha}}}{|\tilde{z}_{{\alpha}}|}\right) \\
&-2\int Q^{2k-2} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k-1}(Q^{2} \sigma \tilde{K}) \\
&+\frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k-2} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(Q^{2} (Q\tilde{K})_{{\alpha}})\\
&-\frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k}Q_{{\alpha}} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(\tilde{K})\\
& + \int \frac{\tau}{|\tilde{z}_{{\alpha}}|}(k-1)Q^{2k+2}Q_{{\alpha}} H({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k+1}(\tilde{K}) \\
& = \text{OK } + B^{1} + B^{2} + B^{3}+B^{4} + B^{5}\end{aligned}$$
Development of the derivative of the $B$ term
---------------------------------------------
We begin noticing that $B^{1} = \text{OK}$, as it was proved in [@Cordoba-Cordoba-Gancedo:interface-water-waves-2d]. Integrating by parts in $B^{5}$, we have that $$\begin{aligned}
B^{5} & = -\int \frac{\tau}{|\tilde{z}_{{\alpha}}|}(k-1)Q^{2k+2}Q_{{\alpha}} H({\partial_{\alpha}}^{k+1}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(\tilde{K})
\end{aligned}$$
Furthermore, the only singular terms arising from $B^{2}$ are when all derivatives hit either $\tilde{K}$ or $\sigma$, this gives us $$\begin{aligned}
B^2 = \text{OK } -2\int Q^{2k} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k-1}(\sigma) \tilde{K} - 2\int Q^{2k} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k-1}(\tilde{K}) \sigma = \text{OK } + B^{2,1} + B^{2,2}.\end{aligned}$$
However, the only singular term of the Rayleigh-Taylor condition that is not in $H^{k-1}$ is the one belonging to $BR_t(\tilde{z},\tilde{\omega}) \cdot \tilde{z}_{{\alpha}}$ when the time derivative hits $\omega$, this means
$$\begin{aligned}
B^{2,1} & = \text{OK } -\tau\int Q^{2k} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) \tilde{K} H({\partial_{\alpha}}^{k}(\tilde{K}Q)) \\
& = -\tau\int Q^{2k+1} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) \tilde{K} H({\partial_{\alpha}}^{k}(\tilde{K})) \\\end{aligned}$$
Finally, developing $B^{3}$ we obtain $$\begin{aligned}
B^{3} & = \frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k-2} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(Q^{3}\tilde{K}_{{\alpha}}) \\
& + \frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k-2} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(Q^{2}Q_{{\alpha}} \tilde{K}) \\
& = B^{3,1} + B^{3,2}\end{aligned}$$
Modulo lower order terms we can see that $$\begin{aligned}
B^{3,2} = \text{OK } + \frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k}Q_{{\alpha}} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(\tilde{K})\end{aligned}$$
We can continue splitting $B^{3,1}$ into $$\begin{aligned}
B^{3,1}& = \text{OK } + \frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k+1} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k+1}(\tilde{K})
+ \frac{\tau}{|\tilde{z}_{{\alpha}}|}\int 3kQ^{2k}Q_{{\alpha}} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(\tilde{K}) \\
& = \text{OK }-\frac{\tau}{|\tilde{z}_{{\alpha}}|}\int Q^{2k+1} \Lambda({\partial_{\alpha}}^{k+1}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(\tilde{K})
+ \frac{\tau}{|\tilde{z}_{{\alpha}}|}\int (k-1)Q^{2k}Q_{{\alpha}} \Lambda({\partial_{\alpha}}^{k}(\tilde{{\varphi}})) {\partial_{\alpha}}^{k}(\tilde{K}) \\
& = \text{OK } + B^{3,1,1} + B^{3,1,2}\end{aligned}$$ where in the last equality we have performed an integration by parts. We can observe that
$$B^{3,2} + B^{4} = B^{3,1,2} + B^{5} = 0, \quad B^{3,1,1} + A^{1} = 0$$
We will now see that $B^{2,2}$ cancels with the term arising from the derivative of $E$. Taking into account the previous lemmas
$$\frac{dE}{dt} = 2\int \sigma Q^{2k}\left({\partial_{\alpha}}^{k-1}(\tilde{K})\right)H({\partial_{\alpha}}^{k+1}(\tilde{{\varphi}})) = \text{OK } -B^{2,2}$$
Finally, we will see that the contributions from the time derivatives of $C$ and $D$ cancel $B^{2,1}$ and $A^{2}$. We start by noticing that, modulo lower order terms $A^{2} = B^{2,1}$. Furthermore
$$\begin{aligned}
\frac{dC}{dt} & = \text{OK } + 2\tau \int (\mathcal{C}\|\tilde{K}(t)\|_{H^{1}} + \tilde{K}) Q^{2k+1}H({\partial_{\alpha}}^{k+1}(\tilde{{\varphi}}))\Lambda({\partial_{\alpha}}^{k-1}(\tilde{K})) \\
\frac{dD}{dt} & = \text{OK } + 2\tau \int \mathcal{C}\|\tilde{K}(t)\|_{H^{1}} Q^{2k+1}\left({\partial_{\alpha}}^{k}(\tilde{{\varphi}})\right) {\partial_{\alpha}}^{k+1}(\tilde{K}),\end{aligned}$$
which, by integration by parts results in
$$\begin{aligned}
\frac{dC}{dt} + \frac{dD}{dt} + A^{2} + B^{2,1} = \text{OK}.\end{aligned}$$
Adding all the contributions, we can bound the derivative in time of the energy by a power of the energy.
Helpful estimates for the Birkhoff-Rott operator
================================================
In this Appendix we will prove some of the estimates used throughout the paper for the sake of clarity to the reader.
We begin with a classical decomposition of the Birkhoff-Rott operator. We should notice that we can write it in the following ways. On one hand:
$$\begin{aligned}
BR(\tilde{z},\tilde{\omega})({\alpha})&=\frac{1}{2\pi}\int_{-\pi}^\pi \left(\frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}-\frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2\tan(\beta/2)}\right)\tilde{\omega}({\alpha}-\beta)d\beta \\
& +\frac{1}{2\pi}\int_{-\pi}^\pi \left(\frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2\tan(\beta/2)}\right)\tilde{\omega}({\alpha}-\beta)d\beta \\
& = \frac{1}{2\pi}\int_{-\pi}^\pi \left(\frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}-\frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2\tan(\beta/2)}\right)\tilde{\omega}({\alpha}-\beta)d\beta+\frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2}H(\tilde{\omega})({\alpha}) \\
& = \frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2}H(\tilde{\omega})({\alpha}) + \text{l.o.t}(\tilde{\omega}).\end{aligned}$$
On the other hand:
$$\begin{aligned}
BR(\tilde{z},\tilde{\omega})({\alpha})&=\frac{1}{2\pi}\int_{-\pi}^\pi \frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}
(\tilde{\omega}({\alpha}-\beta)-\tilde{\omega}({\alpha}))d\beta
+ \frac{\tilde{\omega}({\alpha})}{2\pi}\int_{-\pi}^\pi \frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}d\beta\\
&=\frac{1}{2\pi}\int_{-\pi}^\pi \frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}
(\tilde{\omega}({\alpha}-\beta)-\tilde{\omega}({\alpha}))d\beta \\
& + \frac{\tilde{\omega}({\alpha})}{2\pi}\int_{-\pi}^\pi (\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot\left(\frac{1}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2} - \frac{1}{4|\tilde{z}_{{\alpha}}({\alpha})|^{2}\sin^{2}\left(\frac{\beta}{2}\right)}\right)d\beta \\
&+ \frac{\tilde{\omega}({\alpha})}{2|\tilde{z}_{{\alpha}}({\alpha})|^{2}} \Lambda(\tilde{z}^{\perp}({\alpha}))\\
& = \frac{\tilde{\omega}({\alpha})}{2|\tilde{z}_{{\alpha}}({\alpha})|^{2}} \Lambda(\tilde{z}^{\perp}({\alpha}))+ \text{l.o.t}(\tilde{z}).\end{aligned}$$
See [@Ambrose-Masmoudi:zero-surface-tension-2d-waterwaves], [@Cordoba-Cordoba-Gancedo:interface-water-waves-2d] for more details concerning the lower order terms.
We will now prove energy estimates for the Birkhoff-Rott integral, showing that it is as regular as ${\partial_{\alpha}}\tilde{z}$. The proof is taken from [@Cordoba-Cordoba-Gancedo:interface-water-waves-2d Section 6].
The following estimate holds $$\begin{aligned}
\label{nsibr}
\|BR(\tilde{z},\tilde{\omega})\|_{H^k}\leq
C(\|{\mathcal{F}}(\tilde{z})\|^2_{L^\infty}+\|\tilde{z}\|^2_{H^{k+1}}+\|\tilde{\omega}\|^2_{H^{k}})^j,\end{aligned}$$ for $k\geq 2$, where $C$ and $j$ are constants independent of $\tilde{z}$ and $\tilde{\omega}$.
Using this estimate for $k=2$ we find easily that $$\begin{aligned}
\label{nliibr}
\|{\partial_{\alpha}}BR(\tilde{z},\tilde{\omega})\|_{L^\infty}\leq
C(\|{\mathcal{F}}(\tilde{z})\|^2_{L^\infty}+\|\tilde{z}\|^2_{H^{3}}+\|\tilde{\omega}\|^2_{H^{2}})^j.\end{aligned}$$
We shall present the proof for $k=2$. Let us write $$\begin{aligned}
\begin{split}
BR(\tilde{z},\tilde{\omega})({\alpha},t)&=\frac{1}{2\pi}\int_{-\pi}^\pi C_1({\alpha},\beta)\tilde{\omega}({\alpha}-\beta)d\beta
+\frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2}H(\tilde{\omega})({\alpha})
\end{split}\end{aligned}$$ where $C_1$ is given by
$$\label{fb}
C_1({\alpha},\beta)=\frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}-\frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2\tan(\beta/2)},$$
We shall show that $\|C_1\|_{L^\infty}\leq C\|{\mathcal{F}}(\tilde{z})\|^2_{L^{\infty}}\|\tilde{z}\|^2_{C^2}$. To do so we split $C_1=D_1+D_2+D_3$ where $$D_1=\frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)-{\partial_{\alpha}}\tilde{z}({\alpha})\beta)^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2},\quad D_2={\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})[\frac{\beta}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}-\frac{1}{|{\partial_{\alpha}}\tilde{z}({\alpha})|^2\beta}],$$ and $$D_3=\frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{|{\partial_{\alpha}}\tilde{z}({\alpha})|^2}[\frac1\beta-\frac{1}{2\tan(\beta/2)}].$$ The inequality $$\label{ncll}|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)-{\partial_{\alpha}}\tilde{z}({\alpha})\beta|\leq \|\tilde{z}\|_{C^2}|\beta|^2$$ yields easily $|D_1|\leq \|\tilde{z}\|_{C^2}\|{\mathcal{F}}(\tilde{z})\|^2_{L^{\infty}}$.
Then we can rewrite $D_2$ as follows: $$D_2={\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})[\frac{({\partial_{\alpha}}\tilde{z}({\alpha})\beta-(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)))\cdot ({\partial_{\alpha}}\tilde{z}({\alpha})\beta+(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)))}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2\beta}],$$ and, in particular, we have $$|D_2|\leq \frac{|{\partial_{\alpha}}\tilde{z}({\alpha})\beta-(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))|(|{\partial_{\alpha}}\tilde{z}({\alpha})\beta|+|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|)}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2|{\partial_{\alpha}}\tilde{z}({\alpha})||\beta|}.$$ Using we find that $|D_2|\leq 2
\|\tilde{z}\|_{C^2}\|{\mathcal{F}}(\tilde{z})\|^2_{L^{\infty}}$.
Next let us observe that since $\beta \in [-\pi,\pi]$ gives $|D_3|\leq
C\|{\mathcal{F}}(\tilde{z})\|_{L^{\infty}}$.
The boundedness of the term $C_1$ in $L^\infty$ gives us easily $$\label{brl2}
\|BR(\tilde{z},\tilde{\omega})\|_{L^2}\leq C\|{\mathcal{F}}(\tilde{z})\|^2_{L^\infty}\|\tilde{z}\|^2_{C^2}\|\tilde{\omega}\|_{L^2}.$$ In ${\partial_{\alpha}}^2BR(\tilde{z},\tilde{\omega})$, the most singular terms are given by $$P_1({\alpha})=\frac{1}{2\pi}PV\int_{-\pi}^\pi{\partial_{\alpha}}^2\tilde{\omega}({\alpha}-\beta)\frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^{\bot}}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}d\beta,$$ $$P_2({\alpha})=\frac{1}{2\pi}PV\int_{-\pi}^\pi\tilde{\omega}({\alpha}-\beta)\frac{({\partial_{\alpha}}^2\tilde{z}({\alpha})-{\partial_{\alpha}}^2\tilde{z}({\alpha}-\beta))^{\bot}}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}d\beta,$$ $$P_3({\alpha})=-\frac{1}{\pi}PV\int_{-\pi}^\pi\tilde{\omega}({\alpha}-\beta)\frac{(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^4}\big(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))\cdot({\partial_{\alpha}}^2 \tilde{z}({\alpha})-{\partial_{\alpha}}^2 \tilde{z}({\alpha}-\beta))\big)d\beta.$$ Again we have the expression $$\begin{aligned}
\begin{split}
P_1({\alpha})&=\frac{1}{2\pi}\int_{-\pi}^\pi C_1({\alpha},\beta){\partial_{\alpha}}^2\tilde{\omega}({\alpha}-\beta)d\beta
+\frac{{\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{2|{\partial_{\alpha}}\tilde{z}({\alpha})|^2}H({\partial_{\alpha}}^2\tilde{\omega})({\alpha})d{\alpha},
\end{split}\end{aligned}$$ giving us $$\begin{aligned}
\label{q1nl22dbr}
\begin{split}
|P_1({\alpha})|&\leq C\|{\mathcal{F}}(\tilde{z})\|^j_{L^\infty}\|\tilde{z}\|^j_{C^2}(\|{\partial_{\alpha}}^2\tilde{\omega}\|_{L^2}+|H({\partial_{\alpha}}^2\tilde{\omega})(\alpha)|).
\end{split}\end{aligned}$$ Next let us write $P_2=Q_1+Q_2+Q_3$ where $$Q_1({\alpha})=\frac{1}{2\pi}\int_{-\pi}^\pi(\tilde{\omega}({\alpha}-\beta)-\tilde{\omega}({\alpha}))\frac{({\partial_{\alpha}}^2 \tilde{z}({\alpha})-{\partial_{\alpha}}^2 \tilde{z}({\alpha}-\beta))^{\bot}}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}d\beta,$$ $$Q_2({\alpha})=\frac{\tilde{\omega}({\alpha})}{2\pi}\int_{-\pi}^\pi({\partial_{\alpha}}^2\tilde{z}({\alpha})-{\partial_{\alpha}}^2\tilde{z}({\alpha}-\beta))^{\bot}\big(\frac{1}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}-\frac{1}{|{\partial_{\alpha}}\tilde{z}({\alpha})|^2|\beta|^2}\big)d\beta,$$ $$Q_3({\alpha})=\frac{1}{2\pi}\frac{\tilde{\omega}({\alpha})}{|{\partial_{\alpha}}\tilde{z}({\alpha})|^2}\int_{-\pi}^\pi\!\!({\partial_{\alpha}}^2\tilde{z}({\alpha})\!-\!{\partial_{\alpha}}^2\tilde{z}({\alpha}\!-\!\beta))^{\bot}\big(\frac{1}{|\beta|^2}\!-\!\frac{1}{4\sin^2(\beta/2)}\big)d\beta\!+\!\frac{1}{2}\frac{\tilde{\omega}({\alpha})}{|{\partial_{\alpha}}\tilde{z}({\alpha})|^2}{\Lambda}({\partial_{\alpha}}^2\tilde{z^{\bot}})({\alpha}),$$ where ${\Lambda}={\partial_{\alpha}}H$.
Using that $$|{\partial_{\alpha}}^2\tilde{z}({\alpha})-{\partial_{\alpha}}^2\tilde{z}({\alpha}-\beta)|\leq|\beta|^\delta\|\tilde{z}\|_{C^{2,\delta}},$$ we get $|Q_1({\alpha})|+|Q_2({\alpha})|\leq
\|\tilde{\omega}\|_{C^1}\|{\mathcal{F}}(\tilde{z})\|^{j}\|\tilde{z}\|^j_{C^{2,\delta}},$ while for $Q_3$ we have $$\begin{aligned}
|Q_3({\alpha})|&\leq C\|\tilde{\omega}\|_{L^\infty}\|{\mathcal{F}}(\tilde{z})\|_{L^\infty}(\|\tilde{z}\|_{C^2}+|{\Lambda}({\partial_{\alpha}}^2\tilde{z}^{\bot})({\alpha})|),\end{aligned}$$ that is $$\label{q2nl22dbr}
|P_2({\alpha})|\leq (1+|{\Lambda}({\partial_{\alpha}}^2\tilde{z}^{\bot})({\alpha})|)\|\tilde{\omega}\|_{C^1}\|{\mathcal{F}}(\tilde{z})\|^{j}\|\tilde{z}\|^j_{C^{2,\delta}}.$$ Let us now consider $P_3=Q_4+Q_5+Q_6+Q_7+Q_8+Q_9$, where $$\begin{aligned}
Q_4=\frac{-1}{\pi}\int_{-\pi}^\pi(\tilde{\omega}({\alpha}\!-\!\beta)\!-\!\tilde{\omega}({\alpha}))\frac{(\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta))^\bot}{|\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta)|^4}\big((\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta))\!\cdot\!({\partial_{\alpha}}^2 \tilde{z}({\alpha})\!-\!{\partial_{\alpha}}^2 \tilde{z}({\alpha}\!-\!\beta))\big)d\beta,\end{aligned}$$ $$Q_5=-\frac{\tilde{\omega}({\alpha})}{\pi}\int_{-\pi}^\pi\frac{(\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta)\!-\!{\partial_{\alpha}}\tilde{z}({\alpha})\beta)^{\bot}}{|\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta)|^4}\big((\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta))\cdot({\partial_{\alpha}}^2 \tilde{z}({\alpha})\!-\!{\partial_{\alpha}}^2 \tilde{z}({\alpha}\!-\!\beta))\big)d\beta,$$ $$Q_6=-\frac{\tilde{\omega}({\alpha}){\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{\pi}\int_{-\pi}^\pi\frac{\beta(\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta)\!-\!{\partial_{\alpha}}\tilde{z}({\alpha})\beta)\cdot({\partial_{\alpha}}^2 \tilde{z}({\alpha})\!-\!{\partial_{\alpha}}^2 \tilde{z}({\alpha}\!-\!\beta))}{|\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta)|^4}d\beta,$$ $$Q_7=-\frac{\tilde{\omega}({\alpha}){\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{\pi}{\partial_{\alpha}}\tilde{z}({\alpha})\cdot\!\!\int_{-\pi}^\pi\beta^2({\partial_{\alpha}}^2 \tilde{z}({\alpha})-{\partial_{\alpha}}^2 \tilde{z}({\alpha}-\beta))\big(\frac{1}{|\tilde{z}({\alpha})\!-\!\tilde{z}({\alpha}\!-\!\beta)|^4}-\frac{1}{|{\partial_{\alpha}}\tilde{z}({\alpha})|^4|\beta|^4}\big)d\beta,$$ $$Q_8=-\frac{\tilde{\omega}({\alpha}){\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{\pi|{\partial_{\alpha}}\tilde{z}({\alpha})|^4}{\partial_{\alpha}}\tilde{z}({\alpha})\cdot\!\!\int_{-\pi}^\pi({\partial_{\alpha}}^2 \tilde{z}({\alpha})-{\partial_{\alpha}}^2 \tilde{z}({\alpha}-\beta))\big(\frac{1}{|\beta|^2}-\frac{1}{4\sin^2(\beta/2)}\big)d\beta,$$ and $$Q_9=-\frac{\tilde{\omega}({\alpha}){\partial^{\bot}_{\alpha}}\tilde{z}({\alpha})}{|{\partial_{\alpha}}\tilde{z}({\alpha})|^4}{\partial_{\alpha}}\tilde{z}({\alpha})\cdot {\Lambda}({\partial_{\alpha}}^2 \tilde{z}({\alpha})).$$ Proceeding as before we get $$|P_3({\alpha})|\leq C(1+|{\Lambda}({\partial_{\alpha}}^2\tilde{z})({\alpha})|)\|\tilde{\omega}\|_{C^1}\|{\mathcal{F}}(\tilde{z})\|^j_{L^\infty}\|\tilde{z}\|^j_{C^{2,\delta}},$$ which together with and gives us the estimate $$|(P_1+P_2+P_3)({\alpha})|\leq C(1+|{\Lambda}({\partial_{\alpha}}^2\tilde{z})({\alpha})|+|H({\partial_{\alpha}}^2\tilde{\omega})(\alpha)|)\|\tilde{\omega}\|_{C^1}(\|{\mathcal{F}}(\tilde{z})\|^j_{L^\infty}+\|\tilde{z}\|^j_{H^3}).$$ For the rest of the terms in ${\partial_{\alpha}}^2 BR(\tilde{z},\tilde{\omega})$ we obtain analogous estimates allowing us to conclude the equality $$\label{enl22dbr}
\|{\partial_{\alpha}}^2 BR(\tilde{z},\tilde{\omega})\|_{L^2}\leq C(1+\|{\partial_{\alpha}}^3\tilde{z}\|_{L^2}+\|{\partial_{\alpha}}^2\tilde{\omega}\|_{L^2}) \|\tilde{\omega}\|_{C^1}\|{\mathcal{F}}(\tilde{z})\|^j_{L^\infty}\|\tilde{z}\|^j_{C^{2,\delta}}.$$ Finally the Sobolev inequalities yield for $k=2$.
The following estimate will also be helpful $$\begin{aligned}
\label{nsibr}
\|{\partial_{\alpha}}BR(\tilde{z},\tilde{\omega}) \cdot {\partial_{\alpha}}\tilde{z}\|_{H^k}\leq
C(\|{\mathcal{F}}(\tilde{z})\|^2_{L^\infty}+\|\tilde{z}\|^2_{H^{k+2}}+\|\tilde{\omega}\|^2_{H^{k}})^j,\end{aligned}$$ for $k\geq 2$, where $C$ and $j$ are constants independent of $\tilde{z}$ and $\tilde{\omega}$.
In ${\partial_{\alpha}}BR(\tilde{z},\tilde{\omega}) \cdot {\partial_{\alpha}}\tilde{z}$, the most singular terms are given by $$R_1({\alpha})=\frac{1}{2\pi}PV\int_{-\pi}^\pi{\partial_{\alpha}}\tilde{\omega}({\alpha}-\beta)\frac{{\partial_{\alpha}}\tilde{z}({\alpha}) \cdot (\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^{\bot}}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}d\beta,$$ $$R_2({\alpha})=\frac{1}{2\pi}PV\int_{-\pi}^\pi\tilde{\omega}({\alpha}-\beta)\frac{{\partial_{\alpha}}\tilde{z}({\alpha}) \cdot ({\partial_{\alpha}}\tilde{z}({\alpha})-{\partial_{\alpha}}\tilde{z}({\alpha}-\beta))^{\bot}}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2}d\beta,$$ $$R_3({\alpha})=-\frac{1}{\pi}PV\int_{-\pi}^\pi\tilde{\omega}({\alpha}-\beta)\frac{{\partial_{\alpha}}\tilde{z}({\alpha}) \cdot (\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^4}\big(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))\cdot({\partial_{\alpha}}\tilde{z}({\alpha})-{\partial_{\alpha}}\tilde{z}({\alpha}-\beta))\big)d\beta.$$
$R_2$ can be estimated in the same way as $P_2$. Regarding $R_1$, one can write it as $$R_1({\alpha})=\frac{1}{2\pi}PV\int_{-\pi}^\pi{\partial_{\alpha}}\tilde{\omega}({\alpha}-\beta)\left[\frac{{\partial_{\alpha}}\tilde{z}({\alpha}) \cdot (\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^{\bot}}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^2} - \frac{{\partial_{\alpha}}\tilde{z}_{{\alpha}} \cdot {\partial_{\alpha}}\tilde{z}_{{\alpha}}^{\bot}}{|\tilde{z}_{{\alpha}}({\alpha})|^{2}}\right]d\beta.$$
Now, since ${\partial_{\alpha}}\tilde{\omega}({\alpha}-\beta) = -\partial_{\beta} \tilde{\omega}({\alpha}-\beta)$, one can integrate by parts and bound the resulting kernel (which has order -1) giving $$|R_1| \leq C(\|{\mathcal{F}}(\tilde{z})\|^2_{L^\infty}+\|\tilde{z}\|^2_{H^{k+2}}+\|\tilde{\omega}\|^2_{H^{k}})^j.$$
Finally, $R_3$ can be written in the form
$$\begin{aligned}
R_3({\alpha})=-\frac{1}{\pi}PV\int_{-\pi}^\pi\tilde{\omega}({\alpha}-\beta)&\left[\frac{{\partial_{\alpha}}\tilde{z}({\alpha}) \cdot (\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))^\bot}{|\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta)|^4}\big(\tilde{z}({\alpha})-\tilde{z}({\alpha}-\beta))\cdot({\partial_{\alpha}}\tilde{z}({\alpha})-{\partial_{\alpha}}\tilde{z}({\alpha}-\beta))\big)\right. \\
&\left.- \frac{{\partial_{\alpha}}\tilde{z}({\alpha}) \cdot {\partial_{\alpha}}\tilde{z}({\alpha})^{\bot}}{\beta|\tilde{z}_{{\alpha}}({\alpha})|^{4}}{\partial_{\alpha}}\tilde{z}({\alpha}) \cdot \partial_{{\alpha}}^{2} \tilde{z}({\alpha})\right]d\beta,\end{aligned}$$
and bound $R_3$ by the kernel (which has order 0) in $L^{\infty}$ norm and $\omega$ in $L^{2}$ norm. This completes the proof.
Then, the following corollary is immediate
$$\begin{aligned}
\label{nsibr}
\|\tilde{c}_{{\alpha}}\|_{H^k}\leq
C(\|{\mathcal{F}}(\tilde{z})\|^2_{L^\infty}+\|\tilde{z}\|^2_{H^{k+2}}+\|\tilde{\omega}\|^2_{H^{k}})^j,\end{aligned}$$
for $k\geq 2$, where $C$ and $j$ are constants independent of $\tilde{z}$ and $\tilde{\omega}$.
[**Acknowledgements**]{} {#acknowledgements .unnumbered}
------------------------
AC, DC, FG and JGS were partially supported by the grant [MTM2011-26696]{} of the MCINN (Spain) and the grant StG-203138CDSIF of the ERC. FG acknowledges support from the Ramón y Cajal program. CF was partially supported by NSF grant DMS-0901040. We are grateful for the support of the Fundación General del CSIC.
[99]{}
T. Alazard, N. Burq, and C. Zuily. On the water-wave equations with surface tension. *Duke Math. J.*, 158(3):413–499, 2011.
T. Alazard and G. Métivier. Paralinearization of the Dirichlet to Neumann operator, and regularity of three-dimensional water waves. *Comm. Partial Differential Equations*, 34(10-12):1632-1704, 2009.
B. Alvarez-Samaniego and D. Lannes. Large time existence for 3D water-waves and asymptotics. *Invent. Math.*, 171(3):485-541, 2008.
D. M. Ambrose. Well-posedness of vortex sheets with surface tension. *SIAM J. Math. Anal.*, 35(1):211–244 (electronic), 2003.
D. M. Ambrose and N. Masmoudi. The zero surface tension limit of two-dimensional water waves. *Comm. Pure Appl. Math.*, 58(10):1287–1315, 2005.
D. M. Ambrose and N. Masmoudi. The zero surface tension limit of three-dimensional water waves. *Indiana Univ. Math. J.*, 58(2):479-521, 2009.
J. T. Beale, T. Y. Hou, and J. S. Lowengrub. Growth rates for the linearized motion of fluid interfaces away from equilibrium. *Comm. Pure Appl. Math.*, 46(9):1269-1301, 1993.
A. Castro, D. Córdoba, C. Fefferman, F. Gancedo, and J. Gómez-Serrano. Finite time singularities for the free boundary incompressible Euler equations. *Arxiv preprint arXiv:1112.2170*, 2011.
A. Castro, D. Córdoba, C. Fefferman, F. Gancedo, and J. Gómez-Serrano. Splash singularity for water waves. *Proceedings of the National Academy of Sciences*, 109(3):733-738, 2012.
A. Castro, D. Córdoba, C. Fefferman, F. Gancedo, and M. López-Fernández. Rayleigh-Taylor breakdown for the Muskat problem with applications to water waves. *Ann. of Math. (2)*, 175(2):909–948, 2012.
D. Christodoulou and H. Lindblad. On the motion of the free surface of a liquid. *Comm. Pure Appl. Math.*, 53(12):1536-1602, 2000.
A. Córdoba, D. Córdoba, and F. Gancedo. Interface evolution: water waves in 2-D. *Adv. Math.*, 223(1):120-173, 2010.
D. Coutand and S. Shkoller. Well-posedness of the free-surface incompressible Euler equations with or without surface tension. *J. Amer. Math. Soc.*, 20(3):829-930, 2007.
D. Coutand and S. Shkoller. On the finite-time splash and splat singularities for the 3-D free-surface Euler equations. *Arxiv preprint arXiv:1201.4919*, 2012.
W. Craig. An existence theory for water waves and the Boussinesq and Korteweg-de Vries scaling limits. *Comm. Partial Differential Equations*, 10(8):787-1003, 1985.
P. Germain, N. Masmoudi, and J. Shatah. Global solutions for the gravity water waves equation in dimension 3. *Ann. of Math.(2)*, 175(2):691-754, 2012.
T. Y. Hou, J. S. Lowengrub, and M. J. Shelley. Removing the stiffness from interfacial flows with surface tension. *J. Comput. Phys.*, 114(2):312-338,1994.
D. Lannes. Well-posedness of the water-waves equations. *J. Amer. Math. Soc.*, 18(3):605- 654, 2005.
D. Lannes. A stability criterion for two-fluid interfaces and applications. *Arxiv preprint arXiv:1005.4565*, 2010.
H. Lindblad. Well-posedness for the motion of an incompressible liquid with free surface boundary. *Ann. of Math. (2)*, 162(1):109-194, 2005.
V. I. Nalimov. The Cauchy-Poisson problem. *Dinamika Splo$\breve{s}$n. Sredy*, (Vyp. 18 Dinamika Zidkost. so Svobod. Granicami):104-210, 1974.
J. Shatah and C. Zeng. Geometry and a priori estimates for free boundary problems of the Euler equation. *Comm. Pure Appl. Math.*, 61(5):698-744, 2008.
S. Wu. Well-posedness in Sobolev spaces of the full water wave problem in 2-D. *Invent. Math.*, 130(1):39-72, 1997.
S. Wu. Well-posedness in Sobolev spaces of the full water wave problem in 3-D. *J. Amer. Math. Soc.*, 12(2):445-495, 1999.
S. Wu. Almost global wellposedness of the 2-D full water wave problem. *Invent. Math.*, 177(1):45-135, 2009.
S. Wu. Global wellposedness of the 3-D full water wave problem. *Invent. Math.*, 184(1):125-220, 2011.
H. Yosihara. Gravity waves on the free surface of an incompressible perfect fluid of finite depth. *Publ. Res. Inst. Math. Sci.*, 18(1):49-96, 1982.
P. Zhang and Z. Zhang. On the free boundary problem of three-dimensional incompress- ible Euler equations. *Comm. Pure Appl. Math.*, 61(7):877-940, 2008.
----------------------------------------------------- -----------------------------------------------------
**Angel Castro**
[Département de Mathématiques et Applications]{}
[École Normale Supérieure]{}
[45, Rue d’Ulm, 75005 Paris]{}
[Email: [email protected]]{}
**Diego Córdoba** **Charles Fefferman**
[Instituto de Ciencias Matemáticas]{} [Department of Mathematics]{}
[Consejo Superior de Investigaciones Científicas]{} [Princeton University]{}
[C/ Nicolás Cabrera, 13-15]{} [1102 Fine Hall, Washington Rd, ]{}
[Campus Cantoblanco UAM, 28049 Madrid]{} [Princeton, NJ 08544, USA]{}
[Email: [email protected]]{} [Email: [email protected]]{}
**Francisco Gancedo** **Javier Gómez-Serrano**
[Departamento de Análisis Matemático]{} [Instituto de Ciencias Matemáticas]{}
[Universidad de Sevilla]{} [Consejo Superior de Investigaciones Científicas]{}
[C/ Tarfia, s/n ]{} [C/ Nicolás Cabrera, 13-15]{}
[Campus Reina Mercedes, 41012 Sevilla]{} [Campus Cantoblanco UAM, 28049 Madrid]{}
[Email: [email protected]]{} [Email: [email protected]]{}
----------------------------------------------------- -----------------------------------------------------
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The study of the spatially resolved Star Formation Rate-Mass ([$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}) relation gives important insights on how galaxies assemble at different spatial scales. Here we present the analysis of the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}of 40 local cluster galaxies undergoing ram pressure stripping drawn from the GAs Stripping Phenomena in galaxies (GASP) sample. Considering their integrated properties, these galaxies [ show]{} a SFR enhancement with respect to undisturbed galaxies of similar stellar mass; we now exploit spatially resolved data to investigate the origin and location of the excess. Even on $~\sim1$kpc scales, stripping galaxies present a systematic enhancement of [$\rm \Sigma_{SFR}$]{}($\sim 0.35$ dex at [$\rm \Sigma_\ast$]{}=$\rm 10^{8}M_\odot \,kpc^{-2}$) at any given [$\rm \Sigma_\ast$]{}compared to their undisturbed counterparts. The excess is independent on the degree of stripping and of the amount of star formation in the tails and it is visible at all galactocentric distances within the disks, suggesting that the star formation is most likely induced by compression waves from ram pressure. Such excess is larger for less massive galaxies and decreases with increasing mass. As stripping galaxies are characterised by ionised gas beyond the stellar disk, we also investigate the properties of 411 star forming clumps found in the galaxy tails. At any given stellar mass density, these clumps are systematically forming stars at a higher rate than in the disk, but differences are reconciled when we just consider the mass formed in the last few 10$^8$yr ago, suggesting that on these timescales the local mode of star formation is similar in the tails and in the disks.'
author:
- Benedetta Vulcani
- 'Bianca M. Poggianti'
- Stephanie Tonnesen
- 'Sean L. McGee'
- Alessia Moretti
- Jacopo Fritz
- Marco Gullieuszik
- 'Yara L. Jaffé'
- Andrea Franchetto
- Neven Tomičić
- Matilde Mingozzi
- Daniela Bettoni
- Anna Wolter
bibliography:
- 'references.bib'
title: 'GASP XXX. The spatially resolved SFR-Mass relation in stripping galaxies in the local universe'
---
,
Introduction
============
The existing correlation between a [ galaxy’s]{} stellar mass ([$\rm M_\ast$]{}) and its ongoing Star Formation Rate (SFR) is one of the most widely studied relations in modern astrophysics [see @Speagle2014 for a compilation]. Specifically, it relates the stars that have been formed throughout the entire galaxy life to the ongoing SFR, allowing us to investigate the process of star formation and thus galaxy evolution as a whole. Overall, among star-forming galaxies, higher stellar mass systems undergo more intense star formation activity than lower mass systems [e.g. @Noeske2007b]. The existence of such relation [ and especially its low dispersion (0.2-0.3 dex at all redshifts)]{} points to a scenario where galaxies form through secular processes rather than stochastic merger-driven star-forming episodes. Their evolution throughout cosmological time and across environments must therefore be regulated by the same universal laws [e.g., @Noeske2007a; @Noeske2007b; @Bouche2010; @Daddi2010; @Genzel2010; @Tacconi2010; @Dave2012; @Dayal2013; @Dekel2013; @Lilly2013; @Feldmann2015; @Tacchella2016].
[ Only galaxies with SFRs well above the main sequence [@Rodighiero2011; @Rodighiero2015; @Silverman2015] contrast this picture as they can be interpreted as evidence of starbursts triggered by mergers or external inflows; however, the recent observational evidence on the young age of these systems [e.g., @daCunha2015; @Ma2015] points toward an alternative interpretation in line with the in situ scenario.]{}
The MS has been studied for the first time by [@Brinchmann2004] for local galaxies and later confirmed for high-redshift galaxies by several works [e.g., @Salim2007; @Noeske2007b; @Elbaz2007; @Daddi2007; @Speagle2014; @Schreiber2015; @Kurczynski2016; @Santini2017; @Tacchella2016; @Pearson2018; @Popesso2019; @Morselli2019]. These studies are based on integrated quantities, therefore consider galaxies as a whole, not always distinguishing among the morphological components or excluding regions hosting Active Galactic Nuclei (AGN), if any. In addition, many observations cover only partially the optical extent of the galaxies and are thus subject to aperture effects.
To overcome these issues, in recent years, efforts have been devoted to analyze the [ SFR-[$\rm M_\ast$]{}]{} relation on smaller scales, using spatially resolved data, therefore characterizing the so called “local” relation, in contrast to the “global” one based on integrated properties.
[ The comparison between the local and global relation can shed light on the interplay between different galaxy scales, i.e. on the physical processes connecting local parameters of star formation and feedback to the global star formation in galaxies [e.g., @Semenov2018]. It can also help to determine the minimum scale at which the mechanism that drives the star formation activity with respect to the stellar mass could be universal.]{}
While some studies attempted to characterise the small scales using photometric data [@Abdurrouf2017; @Abdurrouf2018; @Morselli2018; @Hemmati2020], the great step forward for this kind of analysis has been possible thanks to the advent of large integral field spectroscopic (IFS) surveys [e.g. @Sanchez2012; @Bundy2015; @Bryant2015]. All studies report the existence of a correlation even at smaller scales (down to the sizes of molecular clouds), thus implying that the star formation process is regulated by physical processes that act on sub-galactic scales [@RosalesOrtega2012; @Sanchez2013; @CanoDiaz2016; @Hsieh2017; @Lin2017; @Pan2018; @Liu2018; @Medling2018; @Hall2018; @Erroz2019; @CanoDiaz2019; @Vulcani2019b; @Bluck2020; @Enia2020; @Morselli2020]. Nonetheless, the slope, intercept and scatter of the relation vary significantly among different works. These discrepancies are most likely due to the different sample selection, star formation indicator, dust correction, and fitting procedure adopted by the various authors.
Moreover, some authors highlight that the spatially resolved relation varies dramatically from galaxy to galaxy [e.g., @Hall2018; @Vulcani2019b] and that some specific galaxy populations can deviate from the general relation [e.g. @CanoDiaz2019; @Medling2018]. Investigating which specific populations do not follow the general trends, both on local and global scales, can give useful insight on their evolution.
[@CanoDiaz2019; @Medling2018] have found that galaxies of different [ global]{} morphology occupy distinct loci on the [ spatially resolved]{} MS: the earlier the morphological type, the lower is on average the [ spatially resolved SFR ([$\rm \Sigma_{SFR}$]{})]{}, even for galaxies of similar [ spatially resolved [$\rm M_\odot$]{}([$\rm \Sigma_\ast$]{})]{}. Similarly, also on global scales the SFR-mass relation depends on morphology, with late-type galaxies having systematically higher SFR values than early types [e.g., @Calvi2018MorphologyUniverse].
[@Ellison2020] have shown that variations in Star Formation Efficiency [SFE = [$\rm \Sigma_{SFR}$]{}/$\Sigma_{\rm H2}$, e.g., @Genzel2015] are responsible for variations in [$\rm \Sigma_{SFR}$]{}on kpc-scales, therefore galaxies that highly deviate from the fit of the relation have very high SFE. Similar results have been obtained also using global quantities: [@Genzel2015; @Silverman2015; @Silverman2018; @Tacconi2018] showed that there is a correlation between the position of a galaxy relative to the MS ($\Delta$SFR) and its total SFE. [ [@Saintonge2012; @Saintonge2016] have highlighted also a dependence on the global cold gas reservoirs, with in addition systematic variations in the molecular-to-atomic ratio. However, [@Ellison2020] have shown that on local scales the dependence on gas fraction is only secondary to the SFE and weaker.]{}
Also galaxies in the densest environments have been shown to deviate from the general field population, on global scales. Both at $z=0$ and up to $z\sim1$, cluster galaxies can be as star forming as field galaxies, but a population of galaxies with a suppressed SFR at any given mass has been detected [@Vulcani2010ComparingEnvironments; @Paccagnella2016; @Guglielmo2019; @Old2020]. On local scales, only [@Vulcani2019b] have compared the local MS of galaxies in clusters and field, but the way their sample was assembled (i.e. morphologically undisturbed star-forming galaxies on the global MS) prevented them from investigating eventual differences in the large population of morphologically disturbed galaxies in clusters.
These results suggest that the local star formation (at scales $>\sim$1 kpc$^2$) is established by some universal process, but it is modulated partially by global properties, such as the morphology of the galaxy or the gas fraction.
The local MS has been shown also to drive the global one [see also @Hsieh2017; @CanoDiaz2016], most likely through the existence of the size-mass relation: on local scales the mean [$\rm \Sigma_{SFR}$]{}and [$\rm \Sigma_\ast$]{}values for all galaxies are quite similar, regardless of the galaxy size, while on global scales more extended galaxies are also more massive and more star-forming [@Vulcani2019b].
In this context, another population worth investigating are the cluster galaxies that are currently losing their gas via ram-pressure stripping (RPS) due to their motion through the intracluster medium [ICM; @Gunn1972], before being fully quenched [@Vulcani2020]. The most spectacular examples of galaxies losing gas are the so-called jellyfish galaxies. They are at the peak of the stripping and show tails with ionized gas and bright blue knots downstream of the disks, indicating substantial SF in their tails, and asymmetric disks of young stars [e.g., @Cortese2007; @Smith2010; @Fumagalli2014; @Fossati2016; @Consolandi2017; @Poggianti2017; @Moretti2018; @Gullieuszik2017; @Bellhouse2017; @Boselli2018].
Observationally, [ it has been shown that]{} RPS generally enhances the star formation before quenching it (@Crowl2006 [@Merluzzi2013; @Kenney2014], but see @Crowl2008 for a different interpretation). [@Vulcani2018_L] showed that stripping galaxies lay above the SFR-[$\rm M_\ast$]{}relation of undisturbed galaxies, indicating that star formation is boosted in the disks during stripping. Additional star formation takes place in the tails (see also @Fumagalli2014 [@Poggianti2017; @Roman2019; @Cramer2018; @Boselli2018], but @Boselli2016 for different results). This observed enhancement is linked to the higher molecular gas reservoir these galaxies have. Moretti et al. (submitted) have indeed shown that galaxies at peak stripping are very efficient in converting HI into H$_2$.
Simulations overall support the observational results (@Kronberger2008 [@Kapferer2009], but @TonnesenBryan2012 did not find a significant star formation enhancement), even though they are not always concordant on the portion of the galaxy which shows the enhancement. [@Kronberger2008] found that even though new stars are mainly formed in the central parts of the disk, a significant fraction forms also in the wake of the galaxy, while [@Kapferer2009] found a shift in the star formation from the disk to the wake, with a net SFR suppression in the disk. [@Roediger2014] showed that star formation enhancements take place only in regions of sufficiently low initial interstellar medium pressure, which will be stripped soon afterward. [@Troncoso2020] divided the galaxy with a plane perpendicular to the galaxy velocity direction and found an enhancement in the half of the galaxy approaching the cluster center. [@Bekki2014; @Steinhauser2016] found that the enhancement depends strongly on the satellite mass, orbit, and inclination angle.
In this context, a great step forward on the characterisation and interpretation of stripping galaxies has been possible thanks to the GAs Stripping Phenomena in galaxies with MUSE (GASP[^1]) project, an ESO Large Programme granted 120hr of observing time with the integral field spectrograph MUSE that was completed in 2018. GASP allows us to study galaxies in the local universe in various stages of RPS in clusters [@Jaffe2018] and provides us with the unique possibility of looking for trends and performing comparisons in a homogeneous sample, reducing possible biases. A complete description of the survey can be found in @Poggianti2017.
In this paper, we make use of the GASP sample to investigate for the first time the spatially resolved SFR-[$\rm M_\ast$]{}relation of [ galaxies currently being stripped by ram pressure - called from now on “stripping galaxies” for brevity]{}, with the aim of understanding the origin of the global enhancement observed in @Vulcani2018_L. The first part of the paper will focus only on the galaxy disks, excluding the contribution of the galaxy tails. We will therefore compare stripping galaxies to the control sample studied in @Vulcani2019b, to investigate what drives the observed enhancement and localise where such enhancement is. In the second part of the paper we will instead only focus on the tails of the stripping galaxies and characterise the star forming properties of the clumps detected in the galaxy wakes, complementing the characterisation of the clumps presented in @Poggianti2019.
The paper is divided in the following sections. Sections \[sec:data\] presents the data sample and analysis, and describes the identification and characterisation of the clumps (Section \[sec:clumps\]). Section \[sec:results\] includes the results: Section \[sec:results\_disk\] focuses on the comparison between the stripping and control sample disk [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relations, Section \[sec:results\_clumps\] includes the study of the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relations for the clumps. In Section \[sec:disc\] we discuss the results and conclude.
We adopt a [@Chabrier2003] initial mass function (IMF) in the mass range 0.1-100 M$_{\odot}$. The cosmological constants assumed are $\Omega_m=0.3$, $\Omega_{\Lambda}=0.7$ and H$_0=70$ km s$^{-1}$ Mpc$^{-1}$.
 
Data sample and data analysis {#sec:data}
=============================
Data sample
-----------
All the observations used in this paper have been obtained in the context of the GASP project. The survey targeted 114 galaxies at redshift $0.04<z<0.1$, spanning a wide range of galaxy stellar masses ($10^9<$[$\rm M_\ast$]{}/[$\rm M_\odot$]{}$<10^{11.5}$) and located in different environments (galaxy clusters, groups, filaments and isolated). GASP includes both galaxies selected as stripping candidates and undisturbed galaxies.
The sample of galaxies analysed in this paper is drawn from the GASP cluster sample and has been presented in @Vulcani2018_L. Briefly, it includes galaxies with signs of mild, moderate, and extreme stripping, as well as truncated disks. Uncertain cases, as well interacting galaxies [ identified on the basis of stellar tails and/or companions in the same field of view]{} were disregarded. We further exclude from this sample JO149 and JO95 since for these galaxies effective radii could not be determined (see below). The total stripping sample includes 40 galaxies. We refer to Table 1 of @Vulcani2018_L for the list of the objects, along with redshifts, coordinates, integrated stellar masses and star formation rates.
When needed, we will also use the control sample of galaxies presented in @Vulcani2018_L [ and already exploited in @Vulcani2019b]{}. This sample includes cluster+field galaxies that are undisturbed and do not show any clear sign of environmental effects (ram pressure stripping, tidal interaction, mergers, gas accretion, or other interactions) on their spatially resolved star formation distribution. [ Similarly to what done in @Vulcani2019b, we exclude from the sample JO93 and P19482 that, after a careful inspection of their [H$\alpha$]{}maps, turned out to be in an initial phase of stripping. The final sample includes ]{} 30 galaxies, 16 of which are cluster members and 14 field galaxies. Table 2 of @Vulcani2018_L presents the galaxies included in the control sample. Note that in @Vulcani2019b we did not find any difference between undisturbed galaxies in clusters and in the field. The result was somehow expected, as those cluster members most likely just entered their cluster from the field and have had no time yet to feel cluster specific processes.
In @Vulcani2019b we already compared our results to literature results [e.g. @CanoDiaz2016; @Hsieh2017], highlighting how different observational strategies, along with sample selection, analyzing method, fitting recipe and spatial resolution, play an important role in the determination of the parameters that better describe the relations. In what follows we will therefore only use our own control sample, which is treated in the exactly same way as our primary sample.
The left panel of Figure \[fig:mass\] shows the total stellar mass distribution of the galaxies entering the sample, compared to that of the control sample.[^2]
Data analysis {#sec:analysis}
-------------
A complete description of the survey strategy, observations, data reduction and analysis procedure is presented in @Poggianti2017.
Briefly, data were reduced with the most recent available version of the MUSE pipeline[^3] and datacubes were averaged filtered in the spatial direction with a 5$\times$5 pixel kernel, corresponding to our worst seeing conditions of 1$^{\prime\prime}$ = 0.7-1.3 kpc at the redshifts of the GASP galaxies. All the forthcoming results are therefore valid on a scale of $\sim1$ kpc.
We corrected the reduced datacube for extinction due to our Galaxy and subtracted the stellar-only component of each spectrum derived with our spectrophotometric code [[sinopsis]{}]{}[@Fritz2017]. [[sinopsis]{}]{}also provides [ stellar masses]{} for each MUSE spaxel.
Emission line fluxes and errors, along with the underlying continuum, were derived using the IDL software [[kubeviz]{}]{}[@Fossati2016]. We consider as reliable only spaxels with S/N([H$\alpha$]{})$>$5. [H$\alpha$]{}luminosities corrected both for stellar absorption and for dust extinction were used to compute SFRs, adopting the [@Kennicutt1998a]’s relation: $\rm SFR (M_{\odot}
\, yr^{-1}) = 4.6 \times 10^{-42} L_{\rm H\alpha} (erg \, s^{-1})$. The extinction was estimated from the Balmer decrement assuming a value $\rm H\alpha/H\beta = 2.86$ and the [@Cardelli1989] extinction law. The MUSE data reach a surface brightness detection limit of $\rm V\sim 27 \, mag \, arcsec^{-2}$ and $\rm H\alpha \sim 10^{-17.6} \, erg \, s^{-1}\, cm^{-2} \, arcsec^{-2}$ at the 3$\sigma$ confidence level [@Poggianti2017], which translates into a [$\rm \Sigma_{SFR}$]{}limit of $\rm \sim 7 \times 10^{-5} \, M_\odot \, yr^{-1} \, kpc^{-2}$.
We employed the standard diagnostic diagram \[OIII\]5007/$\rm H\beta$ vs \[OI\]6300/$\rm H\alpha$ to separate the regions powered by star formation from regions powered by AGN or LINER emission.[^4] Only spaxels with a S/N$>$3 in all emission lines involved are considered. We adopted the division lines by @Kauffmann2003. For the majority of the galaxies most of the [H$\alpha$]{}is powered by photoionization (plots not shown, see e.g. Fig. 2 in @Poggianti2019), even though 11 galaxies host an AGN in their center (see M. Radovich et al. in prep.). To compute SFRs, we considered only the spaxels whose ionised flux is powered by star formation.
For each spaxel in each galaxy we also computed the galactocentric radius fixing the centre of the galaxy to the peak of the stellar mass map. The radius is then expressed in units of $r_e$, which is computed on I-band images by measuring the radius of an ellipse including half of the total light of the galaxy [@Franchetto2020]. We remind the reader that our observations cover the entire optical extension of the galaxy, up to several effective radii, so our data are not affected by aperture loss. All quantities were corrected for the effect of inclination. The inclination distribution of the sample is shown in the right panel of Fig.\[fig:mass\]. Galaxies with $i > 70^{\circ}$ will not be excluded from the analysis, even though their results must be taken with caution, so they will be highlighted in the following plots.
In the following analysis, we will consider separately spaxels within and outside galaxy disks. We use the definition of galaxy boundaries developed by @Gullieuszik2020. [ Briefly, for each galaxy, the galaxy boundaries were estimated by inspecting the stellar isophote corresponding to a surface brightness 1$\sigma$ above the average sky background level. The stellar isophotes were derived by using the continuum map obtained by the KUBEVIZ model of the [H$\alpha$]{}+\[N II\] lineset. As for stripping galaxies the isophote does not have elliptical symmetry, mainly because of the emission from stars born in the stripped tail, @Gullieuszik2020 fit an ellipse to the undisturbed side of the isophote and used the same ellipse to replace the isophote on the disturbed side. The resulting contour defines a mask that we used to discriminate the galaxy main body and the ram-pressure stripped tail.]{} Everything inside of the isophote represents the galaxy disk, the rest constitutes the galaxy tail.
By definition, control sample galaxies have negligible [H$\alpha$]{}flux (therefore SFR) in the tails. Therefore, comparisons between the stripping and the control sample will be performed using only the spaxels belonging to the galaxy disks.
### Identification and Characterization of [H$\alpha$]{}clumps {#sec:clumps}
For the stripping sample, we will also investigate the properties of [H$\alpha$]{}clumps detected outside the galaxy disks. The clumps have [H$\alpha$]{}surface brightness typically between $\rm 10^{-16.5}-10^{-15} \, erg \, s^{-1} \, cm^{-2} \, arcsec^{-2}$. @Poggianti2017 describes in detail how these clumps are identified. Briefly, these are defined by searching the local minima of the laplace + median filtered [H$\alpha$]{}MUSE image.[^5] The boundaries of these clumps (i.e. their radius, having assumed circular symmetry) are estimated considering outgoing shells until the average counts reach a threshold value that defines the underlying diffuse emission.
SFRs of the clumps have been computed in the same way as for the single spaxels, only for the clumps whose main ionisation mechanism is photoionisation by young stars, always according to the \[OIII\]5007/$\rm H\beta$ vs \[OI\]6300/$\rm H\alpha$ diagrams. @Poggianti2019 showed the BPT diagrams for the clumps in the tails for 16/40 galaxies in the sample.
Following @Poggianti2019, stellar mass estimates of clumps in the tails have been obtained running [[sinopsis]{}]{}with an upper limit to the age of the stellar populations ($2\times 10^8$ yr). This choice avoids having very low levels of unrealistically old stars in the tails, whose light contribution is insignificant, but whose integrated stellar mass can result in overestimating the stellar mass. In this way, we are more likely to get a fair value of the total stellar mass. Note that @Poggianti2019 tested that these stellar mass values do not change significantly varying the upper age limit between a few $10^7$ and $10^9$ yr, therefore the measurement is stable.
[$\rm \Sigma_{SFR}$]{}and [$\rm \Sigma_\ast$]{}are obtained by dividing the SFR and M$_\ast$ obtained frm the integrated spectrum of the clumps by the area of the clumps, obtained assuming circular symmetry. Note that as both values are divided by the same amount, even in the cases our sizes could be overestimated due to the seeing which is always about 1$^{\prime\prime}$ (see @Poggianti2019) the correlation in maintained. In addition, the correlation is valid as long as the tracers of star formation and the stellar mass have the same spatial distribution.
Our final sample includes 411 [H$\alpha$]{}clumps. These ones have been selected for being found outside the stellar disk, are powered by SFR and have S/N$>$3 in all the lines involved in the BPT.
-------------------- --------------------------- --------------------------- ----------------------------------------------------------------------------------------------------------
stripping (S) -14.79$^{+0.04}_{-0.05}$ 1.641$^{+0.006}_{-0.006}$ \[fig:SFR\_Mass\_all\], \[fig:sfr\_mass\_dist\_mass\], \[fig:blobs\_sfrd\]
control -15.28$^{+0.06}_{-0.06}$ 1.659$^{+0.008}_{-0.008}$ \[fig:SFR\_Mass\_all\], \[fig:SFR\_Mass\_SFR\], \[fig:SFR\_Mass\_Jstage\], \[fig:sfr\_mass\_dist\_mass\]
S SFR$_{out}<$0.02 -13.72$^{+0.08}_{-0.07}$ 1.502$^{+0.009}_{-0.01}$ \[fig:SFR\_Mass\_SFR\]
S SFR$_{out}>$0.02 -15.10$^{+0.06}_{-0.06}$ 1.681$^{+0.007}_{-0.007}$ \[fig:SFR\_Mass\_SFR\]
S Jstage=0.5 -15.2$^{+0.1}_{-0.1}$ 1.71$^{+0.02}_{-0.02}$ \[fig:SFR\_Mass\_Jstage\]
S Jstage=1 -14.00$^{+0.07}_{-0.07}$ 1.544$^{+0.01}_{-0.009}$ \[fig:SFR\_Mass\_Jstage\]
S Jstage=2 -15.93$^{+0.07}_{-0.07}$ 1.780$^{+0.009}_{-0.01}$ \[fig:SFR\_Mass\_Jstage\]
S Jstage=3 -18.0$^{+0.2}_{-0.3}$ 1.97$^{+0.03}_{-0.03}$ \[fig:SFR\_Mass\_Jstage\]
JO10 -17.9$^{+ 0.2}_{- 0.2}$ 1.89$^{+0.02}_{-0.02}$ \[fig:sfr\_mass\_dist\_mass\]
JO112 -19.2$^{+0.4}_{-0.4}$ 2.28$^{+ 0.05}_{-0.05}$ \[fig:sfr\_mass\_dist\_mass\]
JO113 -13.4$^{+0.3}_{-0.3}$ 1.54$^{+ 0.05}_{- 0.04}$ \[fig:sfr\_mass\_dist\_mass\]
JO135 -12.9$^{+ 0.2}_{- 0.2}$ 1.34$^{+0.03}_{-0.02}$ \[fig:sfr\_mass\_dist\_mass\]
JO138 -14.1$^{+0.4}_{-0.4}$ 1.56$^{+ 0.06}_{-0.06}$ \[fig:sfr\_mass\_dist\_mass\]
JO13 -18.6$^{+0.3}_{- 0.3}$ 2.20$^{+0.04}_{-0.04}$ \[fig:sfr\_mass\_dist\_mass\]
JO141 -13.4$^{+ 0.1}_{-0.2}$ 1.40$^{+ 0.02}_{- 0.02}$ \[fig:sfr\_mass\_dist\_mass\]
JO144 -15.0$^{+0.2}_{-0.2}$ 1.62$^{+0.03}_{-0.03}$ \[fig:sfr\_mass\_dist\_mass\]
JO147 -12.2$^{+0.1}_{-0.1}$ 1.23$^{+ 0.01}_{-0.01}$ \[fig:sfr\_mass\_dist\_mass\]
JO159 -16.5$^{+0.5}_{-0.6}$ 1.87$^{+0.07}_{-0.07}$ \[fig:sfr\_mass\_dist\_mass\]
JO160 -16.4$^{+0.2}_{-0.3}$ 1.87$^{+0.03}_{-0.03}$ \[fig:sfr\_mass\_dist\_mass\]
JO162 -11.1$^{+ 0.2}_{-0.3}$ 1.21$^{+0.04}_{-0.04}$ \[fig:sfr\_mass\_dist\_mass\]
JO175 -11.78$^{+0.089}_{-0.09}$ 1.21$^{+0.01}_{-0.01}$ \[fig:sfr\_mass\_dist\_mass\]
JO181 -21$^{+1}_{-2}$ 2.7$^{+0.3}_{-0.2}$ \[fig:sfr\_mass\_dist\_mass\]
JO194 -19.6$^{+0.3}_{-0.3}$ 2.21$^{+0.04}_{-0.04}$ \[fig:sfr\_mass\_dist\_mass\]
JO197 -16.4$^{+0.2}_{-0.2}$ 1.82$^{+0.03}_{-0.03}$ \[fig:sfr\_mass\_dist\_mass\]
JO200 -11.9$^{+0.1}_{-0.1}$ 1.21$^{+0.01}_{-0.01}$ \[fig:sfr\_mass\_dist\_mass\]
JO201 -13.8$^{+0.2}_{-0.2}$ 1.54$^{+0.02}_{-0.02}$ \[fig:sfr\_mass\_dist\_mass\]
JO204 -11.2$^{+0.1}_{-0.1}$ 1.15$^{+0.02}_{-0.02}$ \[fig:sfr\_mass\_dist\_mass\]
JO206 -21.1$^{+0.6}_{-0.6}$ 2.48$^{+0.08}_{-0.07}$ \[fig:sfr\_mass\_dist\_mass\]
JO23 -26.1$^{+0.8}_{-0.9}$ 3.1$^{+0.1}_{-0.1}$ \[fig:sfr\_mass\_dist\_mass\]
JO27 -12.3$^{+0.2}_{-0.2}$ 1.37$^{+0.04}_{-0.03}$ \[fig:sfr\_mass\_dist\_mass\]
JO28 -19$^{+1}_{-1}$ 2.2$^{+0.2}_{-0.1}$ \[fig:sfr\_mass\_dist\_mass\]
JO36 -28$^{+1}_{-1}$ 3.3$^{+0.1}_{-0.1}$ \[fig:sfr\_mass\_dist\_mass\]
JO47 -21.3$^{+0.8}_{-0.8}$ 2.6$^{+0.1}_{-0.1}$ \[fig:sfr\_mass\_dist\_mass\]
JO49 -13.5$^{+0.2}_{-0.2}$ 1.41$^{+0.03}_{-0.03}$ \[fig:sfr\_mass\_dist\_mass\]
JO60 -11.2$^{+0.1}_{-0.1}$ 1.23$^{+0.02}_{-0.02}$ \[fig:sfr\_mass\_dist\_mass\]
JO69 -16.9$^{+0.4}_{-0.4}$ 2.00$^{+0.06}_{-0.05}$ \[fig:sfr\_mass\_dist\_mass\]
JO70 -10.11$^{+0.09}_{-0.09}$ 1.04$^{+0.01}_{-0.01}$ \[fig:sfr\_mass\_dist\_mass\]
JO85 -13.6$^{+0.1}_{-0.1}$ 1.51$^{+0.02}_{-0.02}$ \[fig:sfr\_mass\_dist\_mass\]
JO93 -14.2$^{+0.2}_{-0.2}$ 1.54$^{+0.02}_{-0.02}$ \[fig:sfr\_mass\_dist\_mass\]
JW100 -11.9$^{+0.2}_{-0.2}$ 1.19$^{+0.03}_{-0.03}$ \[fig:sfr\_mass\_dist\_mass\]
JW108 -25.5$^{+0.5}_{-0.5}$ 2.88$^{+0.06}_{-0.06}$ \[fig:sfr\_mass\_dist\_mass\]
JW10 -24$^{+1}_{-2}$ 3.0$^{+0.2}_{-0.2}$ \[fig:sfr\_mass\_dist\_mass\]
JW115 -13.7$^{+0.8}_{-0.9}$ 1.545$^{+0.1}_{-0.1}$ \[fig:sfr\_mass\_dist\_mass\]
JW29 -24.2$^{+0.9}_{-1}$ 3.1$^{+0.1}_{-0.1}$ \[fig:sfr\_mass\_dist\_mass\]
JW56 -11.9$^{+0.6}_{-0.7}$ 1.30$^{+0.1}_{-0.09}$ \[fig:sfr\_mass\_dist\_mass\]
clumps out -13.0$^{+0.5}_{-0.5}$ 1.63$^{+0.08}_{-0.07}$ \[fig:blobs\_sfr\]
global stripping -10$^{+2}_{-5}$ 0.9$^{+0.5}_{-0.2}$ \[fig:blobs\_sfr\]
-------------------- --------------------------- --------------------------- ----------------------------------------------------------------------------------------------------------
: Least-squares regression parameters for the different samples. Note that JO171, J1079 and JW39 are not listed because the fit is not meaningful. The Figure where the relation is used is also listed. Details on the fitting method are given in the text. \[tab:param\]
![Spatially resolved SFR-[$\rm M_\ast$]{}([$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}) relation for all spaxels in the galaxy disks of all stripping galaxies (blue). Superimposed in a blue scale are contours representing the 15th, 35th, 65th, 85th and 98th percentiles. Superimposed in red[ - to yellow]{} scale as shaded areas are contours representing the same percentiles for the control sample discussed in @Vulcani2019b. Thick blue and dashed red lines show the fit to the relation, for the stripping and control sample, respectively. Transparent lines show samples from the posterior, indicating the scatter in the fit. The magenta dotted line represents the effective threshold in spatially resolved specific SFR entailed by the adopted cuts in S/N and corresponds to $\rm 10^{-11.2}\, yr^{-1} \, kpc^{-2}$. Galaxies in the stripping sample have a systematically higher [$\rm \Sigma_{SFR}$]{}at any given [$\rm \Sigma_\ast$]{}than galaxies in the control sample. \[fig:SFR\_Mass\_all\] ](cfr_jelly_cs_cont_allspaxels_in_MCMC_inv_v2.png)
Results {#sec:results}
=======
Disk [$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}relation for stripping and control sample galaxies {#sec:results_disk}
----------------------------------------------------------------------------------------------------
### The disk [$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}relation of all galaxies
Figure \[fig:SFR\_Mass\_all\] shows the spatially resolved SFR-[$\rm M_\ast$]{}([$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}) relation considering all galaxies in the stripping sample, using the 139727 spaxels whose emission is dominated by star formation in the galaxy disks. Here we consider together both the clumps and the diffuse emission. The effective threshold in spatially resolved specific SFR entailed by the adopted cuts in S/N corresponds to $\rm 10^{-11.2}\, yr^{-1} \, kpc^{-2}$ [@Vulcani2019b] and it is also shown. A correlation between the two quantities is immediately visible, with spaxels with higher [$\rm \Sigma_\ast$]{}typically having higher values of [$\rm \Sigma_{SFR}$]{}. The correlation spans more than four orders of magnitude in both [$\rm \Sigma_\ast$]{}and [$\rm \Sigma_{SFR}$]{}. Spaxels with [$\rm \Sigma_\ast$]{}$\rm > 10^{9} M_\odot kpc^{-2}$ form a quite thin [$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}relation, which seems shifted towards lower [$\rm \Sigma_{SFR}$]{}values than that of the whole population. We will investigate later on who is responsible for such trend. Overall, the scatter of the relation is $\sim0.4$ dex. We measured this value subdividing the sample in 10 [$\rm \Sigma_\ast$]{}bins and computing the standard deviation of [$\rm \Sigma_{SFR}$]{}in each bin separately. We then took the mean value of the standard deviations.
As comparison, Figure \[fig:SFR\_Mass\_all\] also shows as density contours the relation for the 92020 disk spaxels of the galaxies belonging to the GASP control sample from @Vulcani2019b. As discussed in that paper, the scatter of this sample is lower, being $\sim0.3$ dex. Overall, the datapoints of the stripping sample extend both at high and low [$\rm \Sigma_\ast$]{}with respect to the control sample. In addition, at any given [$\rm \Sigma_\ast$]{}points of stripping galaxies have systematically higher [$\rm \Sigma_{SFR}$]{}than control sample galaxies. In the stripping sample, the most external contour, including 98th of the total population, systematically extends towards higher [$\rm \Sigma_{SFR}$]{}values than that of the control sample. In contrast, the lower edge of the same contour is very similar for the two samples. This is very similar to the effective threshold in [$\rm \Sigma_{SFR}$]{}we adopt and could be driven by that detection limit. To further support the results on a statistical ground, we perform a linear regression fitting using the python module Pystan, a package for Bayesian inference. The parameters describing the fit, along with errors are tabulated in Tab.\[tab:param\], where all the fits discussed in the rest of the paper can also be found. Slopes are compatible within $2\sigma$, while intercepts are different at more than $3\sigma$ level. For reference, at $\rm \log(\Sigma_\ast [M_\sun/kpc^{-2}])=8$ the $\rm \Delta(\log(\Sigma_{SFR} [M_\sun/yr/kpc^{-2}]))$ between the two fits is $\sim 0.35$ dex.
To understand these differences, we can analyse which galaxies contribute to the different portions of the graph.
Figure \[fig:violin\_mass\] shows the distribution of [$\rm \Sigma_\ast$]{}as a function of the galaxy stellar mass. Data distributions are shown in terms of violin plots, which give the probability density of the data at different values, smoothed by a kernel density estimator. [ Unlike bar graphs with means and error bars, violin plots contain all data points. The shape of the violin displays frequencies of values: the thicker part of the violin shape means that the values in that y-axis section of the violin have higher frequency, and the thinner part implies lower frequency. Violin plots also highlight the maximum extension of the data, and the presence of different peaks, their position and relative amplitude. The maximum width of each violin is set the same for all galaxies, for display purposes.]{}
The Figure clearly shows that the [$\rm \Sigma_\ast$]{}range depends on $M_\ast$. Spaxels in low mass galaxies do not reach high [$\rm \Sigma_\ast$]{}values: galaxies with $M_\ast <10^{9.75} M_\odot$ have [$\rm \Sigma_\ast$]{}always lower than 10$\rm ^8 M_\odot kpc^{-2}$. In contrast in massive galaxies [$\rm \Sigma_\ast$]{}can reach and even exceed 10$\rm ^9 M_\odot kpc^{-2}$. [ At $M_\ast >10^{10.4} M_\odot$, stripping and control sample galaxies have a different behaviour: there are no control sample galaxies with [$\rm \Sigma_\ast$]{}$>$10$\rm ^{9.25} M_\odot kpc^{-2}$, while there are 6 stripping galaxies (15% of the sample) in the same [$\rm \Sigma_\ast$]{}regime. This is only partially due to the different mass distributions of the two samples (Fig.\[fig:mass\]): limiting the comparison to [$\rm M_\ast$]{}$<10^{11}$ [$\rm M_\odot$]{}we still have 5 stripping galaxies with [$\rm \Sigma_\ast$]{}$>$10$\rm ^{9.25} M_\odot kpc^{-2}$.]{}
It is very important to compare samples at given stellar mass. Figure \[fig:delta\] compares them in bins of both stellar mass and galactocentric distance, to better localise also the spatial position of the star formation enhancement. Three stellar mass bins (defined by these boundaries: $\log($[$\rm M_\odot$]{}\[$M_\odot$\]) = \[8.7, 9.7, 10.6, 11.5\]) and four galactocentric distance bins (defined by these boundaries: r/r$_e$ = \[0, 0.5, 1, 1.5, 5\]) are considered, for a total of 12 independent bins.[^6] In each bin, we compute the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relation of the control sample galaxies in that bin and then we measure for both samples the difference between the measured [$\rm \Sigma_{SFR}$]{}and the [$\rm \Sigma_{SFR}$]{}expected from the control sample fit, given the measured [$\rm \Sigma_\ast$]{}. Figure \[fig:delta\] shows the distribution of such differences. Distributions of the two samples are clearly different. In most of the cases, the stripping sample distribution is shifted towards higher values. [ In each stellar mass and galactocentric distance bins,]{} the K-S test is able to state with very high confidence level ($p-value<<$0.01) that distributions are always drawn from different parent samples. Also median values are different in most of the cases (except for intermediate distances in the highest mass bins), when errors on medians (=1.235$\times \sigma/\sqrt{N}$ with $\sigma$ standard deviation of the distribution and $N$ number of points) are considered. These errors though are very tiny, given the high number of data points in each distribution. Figure \[fig:delta\] shows instead the standard deviation of the distributions, which are indeed quite broad. [ Some differences in the median values between stripping and control galaxies are evident: in the galaxy central regions (r/r$_e<$1) such difference is larger (= stripping sample has enhanced [$\rm \Sigma_{SFR}$]{}with respect to the control sample) among the least massive galaxies and decreases with increasing stellar mass. In the external regions, such difference does not hold anymore. Fixing the stellar mass bin, very central (r/r$_e<$0.5) and external regions (r/r$_e>$1.5) have a larger enhancement than intermediate regions. We note, however, that in the most massive bins there are only three control sample galaxies, so comparisons might not be meaningful.]{}[ This Figure overall suggests that in stripping galaxies the [$\rm \Sigma_{SFR}$]{}is enhanced with respect of the control sample, both fixing the stellar mass and the galactocentric distance. ]{}
In addition to stellar mass, there are also other galaxy characteristics that can influence the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relation. As discussed in @Jaffe2018 [@Vulcani2018_L], and B. M. Poggianti et al. (in prep.), galaxies in the stripping sample can be categorised based on the stage of the stripping (mild, moderate and extreme, and truncated disks, see examples in Fig. 2 of @Jaffe2018). Gullieuszik et al. (2020) also studied galaxies as a function of the amount of the SFR in the tails. We can therefore inspect the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relation of galaxies of these different categories, to determine whether the offset is determined by one of these groups. Figure \[fig:SFR\_Mass\_SFR\] focuses on stripping galaxies with total SFR in the tails $<0.02$ [$\rm M_\odot$]{}yr$^{-1}$ (left) and $>0.02$ [$\rm M_\odot$]{}yr$^{-1}$ (right), separately. Six galaxies have low level of SFR in the tails (JW29, JO138, JO23, JO197, JW108, JO10), while all the rest have $SFR_{out}>0.02$ [$\rm M_\odot$]{}yr$^{-1}$. Comparing these samples to the control sample, it appears evident that both galaxies with low and high SFR in the tails are characterized by [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relations shifted high compared to that of the control sample galaxies. Similar conclusions are reached if instead of a cut in absolute value of SFR$_{out}$ we adopt a cut in SFR$_{out}$/SFR$_{tot}$=20%. Linear regression fits are statistically different when comparing [ both the sample with low and high SFR in the tails and the control sample (Tab. \[tab:param\])]{}. Similarly, Figure \[fig:SFR\_Mass\_Jstage\] shows that the [$\rm \Sigma_{SFR}$]{}enhancement is present in stripping galaxies showing any degree of stripping. Galaxies with both mild, moderate and extreme stripping (Jstage = 0.5, 1, 2) show a shift of the contours towards higher [$\rm \Sigma_{SFR}$]{}values at any given mass, compared to the control sample contours. Galaxies with Jstage =0.5 seem to extend less toward high [$\rm \Sigma_{SFR}$]{}at high [$\rm \Sigma_\ast$]{}values than the other stripping galaxies, but this might be due to fact that there are no very massive galaxies among Jstage =0.5 galaxies. The slopes of the fits are always statistically different from that of the control sample. Galaxies with Jstage =2 seem to be the main responsible for the strip at high [$\rm \Sigma_{SFR}$]{}and [$\rm \Sigma_\ast$]{}values. In contrast, Jstage=3 - which are the truncated disks - show a quite different behaviour. These galaxies are characterised by extremely narrow relations. In their very central regions the correlation between [$\rm \Sigma_{SFR}$]{}and [$\rm \Sigma_\ast$]{}is very tight. In this case the fitting parameters [ describing the data of Jstage=3 galaxies]{} agree within the uncertainties [ with those of the control sample]{}, but the different data distribution is outstanding.
These results show that most of the differences between the stripping and control samples are not simply due to galaxies with very long tails and with high level of SFR in the tails, that is galaxies at the peak of the stripping, but to galaxies in all the stripping stages.
![Distribution of the fraction of star forming spaxels in galaxy disks in the stripping (blue) and control (orange) sample. Overall, the fraction of star forming spaxels within the galaxy main body is smaller in galaxies in the stripping sample. \[fig:frac\]](jelly_sfing_fraction_pergal_v2.png)
This result, along with the global enhancement observed in @Vulcani2018_L, is even more significant if we consider that in stripping galaxies the portion of the disks which is actually star forming is much smaller than in the control sample galaxies. Fig.\[fig:frac\] shows the distribution of the portion of galaxy disk powered by either star formation or LINER/AGN, assuming some star formation is present also in the LINER/AGN dominated regions, for the two samples and highlights how stripping sample galaxies have overall a lower fraction of star forming spaxels than the control sample galaxies. Excluding the spaxels powered mainly by AGN would only increase the differences.
### The galaxy-by-galaxy [$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}relation
While taking into account the properties of the galaxies together allows us to study the general trends and analyse the galaxy population as a whole, it does not allow us to understand if all galaxies follow similar relations or if each galaxy is characterized by a different slope, intercept, and scatter. Figure \[fig:sfr\_mass\_dist\_mass\] presents the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relation for each galaxy separately, distinguishing among spaxels at different galactocentric distances. Galaxies with high inclination ($i>70^{\circ}$) are indicated with an asterisk. Galaxies hosting an AGN are surrounded by a red square. It appears evident that, even though overall in most cases a correlation does exist, each object spans a distinct locus on the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}plane: some galaxies show quite elongated sequences, some others are characterised by a cloud rather than a sequence. This is similar to what we found in @Vulcani2019b for the control sample. Overall trends with distance are detected, with spaxels in the cores having higher values of [$\rm \Sigma_\ast$]{}and [$\rm \Sigma_{SFR}$]{}. Few cases deviates from such trend (e.g. JO36, JO27, JW29, JO147). These are most likely due to the high inclination of the galaxies that mixes spaxels at different distances and entails high levels of dust extinction.
Some galaxies have all spaxels above the total fit of the relation (e.g. JO113), while most of them have spaxels both above and below the line. Typically, especially for massive galaxies, galaxy cores are always below the fit. As seen in Fig.\[fig:SFR\_Mass\_all\], spaxels with [$\rm \Sigma_\ast$]{}$\rm > 10^{9} M_\odot kpc^{-2}$ have typically very thin [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relations, probably indicating an homogeneity of the star forming properties in the galaxy cores. In @Vulcani2019b we showed that masking the spaxels most likely located in the galaxy bulge - whose size has been obtained applying a fitting on the I-band images (see A. Franchetto et al. in prep.)- did not affect the results, showing how the suppression [ of the [$\rm \Sigma_{SFR}$]{}]{} extends beyond the galaxy bulge. As also highlighted in Fig.\[fig:SFR\_Mass\_Jstage\], Jstage=3 (JO10, JO23, JO36, JW108) follow very thin relations. All spaxels of JO10 are well below the fit of the relation, the other
truncated disks cross the relation, even though they all show a quite suppressed [$\rm \Sigma_{SFR}$]{}given their [$\rm \Sigma_\ast$]{}. Note that they were not outliers in the global SFR-mass relation [@Vulcani2018_L].
The presence of AGN seems not to influence the trends, but it is important to note that almost all massive galaxies in the sample host an AGN, therefore it is not possible to disentangle the two effects.
Fitting the relation to each galaxy separately (Table \[tab:param\]), when the fit is meaningful,[^7] the slope of the relation is generally different than that of the total fit, highlighting the large galaxy-by-galaxy variation.
To better relate the [$\rm \Sigma_{SFR}$]{}and [$\rm \Sigma_\ast$]{}distribution of each galaxy to its global properties, we compute again the difference between the measured [$\rm \Sigma_{SFR}$]{}and the [$\rm \Sigma_{SFR}$]{}expected from the total control sample fit, given the measured [$\rm \Sigma_\ast$]{}. We show the distribution of the differences in Fig. \[fig:violin\], using the violin plots, and we sort galaxies for increasing total stellar mass (top) and for increasing $\Delta$(SFR) (bottom). Following @Vulcani2018_L, $\Delta$(SFR) [ is the difference between the SFR of each galaxy and the value derived from the control sample fit given the galaxy mass]{}. These violin plots show also the median and the interquartile ranges. Overall, considering both samples together, the median $\Delta$([$\rm \Sigma_{SFR}$]{}) increases with increasing $\Delta$(SFR) (Pearson’s correlation coefficient= 0.52, 2-tailed p-value = 3.0$\cdot10^{-6}$) and decreases with increasing stellar mass (Pearson’s correlation coefficient= -0.62, 2-tailed p-value = 1.6$\cdot10^{-8}$). Stripping galaxies not only populate the most massive end of the mass distribution (see Fig. \[fig:mass\]), but also the highest end of the $\Delta$(SFR) distribution. The maximum value of $\Delta$(SFR) in the control sample is 0.35 dex, in the stripping sample 0.7 dex. Binning galaxies according to their stellar mass, we find that at all masses galaxies in the stripping sample have a systematically higher median $\Delta$([$\rm \Sigma_{SFR}$]{}) than their control sample counterparts. Nonetheless, the difference decreases with increasing stellar mass. In contrast, binning galaxies according to their $\Delta$(SFR), no strong differences are found between the median values of the two samples, at any $\Delta$(SFR).
These results highlight a link between local and global properties of the galaxies. $\Delta$[$\rm \Sigma_{SFR}$]{}is influenced by stellar mass and most likely has an effect on the $\Delta$(SFR) measured on global scales.
 
[$\rm \Sigma_{SFR}$]{}and [$\rm \Sigma_\ast$]{}properties of the [H$\alpha$]{}clumps outside the galaxy disks in stripped galaxies {#sec:results_clumps}
-----------------------------------------------------------------------------------------------------------------------------------
![[H$\alpha$]{}maps for four galaxies in the stripping sample, shown as an example. Magenta circles show the [H$\alpha$]{}clumps outside the stellar disk (i.e. in the tail). \[fig:Ha\_map\]](Ha_map_jelly_4.png)
![[$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}relation for all spaxels in the disks of all stripping galaxies (blue). Superimposed are shown the values of the [H$\alpha$]{}clumps detected in the tails (magenta). Clumps outside the galaxy bodies do not follow a clear [$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}. They have systematically higher [$\rm \Sigma_{SFR}$]{}values at any given [$\rm \Sigma_\ast$]{}than spaxels in the galaxy disks and do not lie on the extrapolation of the disk relation. \[fig:blobs\_sfrd\]](cfr_spaxels_in_blobs_sfrd_MCMC_inv.png)
Differently from control sample galaxies, stripping galaxies (except for truncated disks) are characterized by the existence of material outside the galaxy disk, that is the galaxy tail. In this section we therefore focus only on the stripping sample and study the properties of the [H$\alpha$]{}clumps detected in the tails, identified following the procedure described in Sec.\[sec:analysis\]. As an example, Figure \[fig:Ha\_map\] shows the [H$\alpha$]{}maps of four galaxies of the stripping sample.
@Poggianti2019 characterised the properties of the clumps, considering only 16 galaxies. They found that the star forming clumps are dynamically quite cold, have a median [H$\alpha$]{}velocity dispersion $\sigma= 27$ [$\rm km \, s^{-1}$]{}, a median [H$\alpha$]{}luminosity L([H$\alpha$]{})$=4\times 10^{38}\, erg \, s^{-1}$, a median SFR=0.003 [$\rm M_\odot$]{}yr$^{-1}$ and [$\rm M_\ast$]{}$=3\times10^6$ [$\rm M_\odot$]{}. They characterised the tail clumps scaling relations (M$_{gas}$-[$\rm M_\odot$]{}, L([H$\alpha$]{})-$\sigma$, SFR-M$_{gas}$), but they did not focus specifically on the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relation, as we do here.
First of all, we note that the number of star forming clumps in the tail varies from galaxy to galaxy and seems not to be strictly related to the galaxy stellar mass. Ten galaxies have no star forming clumps outside the stellar disk (JO10, JO112, JO13, JO138, JO197, JO23, JW108, JW115, JW29, JW56).
Figure \[fig:blobs\_sfrd\] shows the [$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}relation for the 411 [H$\alpha$]{}clumps in the tails, overlaid to the relation of the galaxy disks (from Fig. \[fig:SFR\_Mass\_all\]). The [$\rm \Sigma_\ast$]{}of the clumps spans the range $10^5-10^7$ [$\rm M_\odot$]{}kpc$^{-2}$, the [$\rm \Sigma_{SFR}$]{}spans the range $10^{-4}-10^{-2}$ [$\rm M_\odot$]{}yr$^{-1}$ kpc$^{-2}$, occupying a very different locus from that of the spaxels in the galaxy disks. A cloud rather than a well defined relation is evident. A Pearson’s correlation test is not able to retrieve a significant correlation (coefficient= 0.19, 2-tailed p-value = 4.0$\cdot10^{-5}$). Compared to the extrapolation of the [$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}relation of the disk spaxels towards low mass surface densities, that of the clumps is shifted towards higher values at any given [$\rm \Sigma_\ast$]{}. We stress, however, that while the y-axis of the plots are comparable, the [$\rm \Sigma_\ast$]{}values of the clumps in the tails and of the spaxels in the disks have a different meaning therefore a fair comparison is not possible. [$\rm \Sigma_\ast$]{}for the tail clumps represent the stellar mass formed only during the ongoing star formation episode, therefore the “true clump mass”. In contrast, in the galaxy disks, the masses are “projected stellar masses” inflated by the underlying old stellar populations.
![[$\rm \Sigma_\ast$]{}- distance relation for all the star forming clumps in the tails of the stripping sample. Distance is in unit of $r_e$. The black line shows the linear fit. Clumps farther away from the galaxy disks [ have systematically lower [$\rm \Sigma_\ast$]{}values]{}. \[fig:blobs\_dist\_m\]](blobs_dist_m.png)
Overall, clumps can be found as far as 80 kpc from the galaxy center, with a median value of 25 kpc. Note that given that the clumps are extraplanar, when measuring their distance we simply compute the euclidean distance from the galaxy center, without considering the inclination to correct for projection effects. Clumps farther away from the galaxy disks are systematically less dense in mass: Fig.\[fig:blobs\_dist\_m\] shows an anticorrelation between the clump distance and [$\rm \Sigma_\ast$]{}. This trend is supported by the Pearson’s correlation test (correlation coefficient= -0.3, 2-tailed p-value = 1.0$\cdot10^{-9}$). In contrast, there seems not to be a clear correlation with [$\rm \Sigma_{SFR}$]{}(plot not shown).
![Global SFR-mass relation for the stripping sample (grey) and the [H$\alpha$]{}clumps found outside the disk (magenta). Blue points represent the SFR-mass relation of the galaxies when only the mass formed in the last $2\times10^8$ yr is considered (see text for details). They grey line is the fit to the global relation, the purple line is the fit representing both the magenta and blue points. Transparent lines show samples from the posterior, indicating the scatter in the fit. The relation for the clumps is shifted towards lower mass and SFR values and is much steeper when we compare the clumps and the global values for the stripping sample; it has instead the same slope when we consider only the amount of stellar mass produced only in the most recent epochs for the galaxies, suggesting that the local mode of star formation is very similar in the galaxy disks and tails.\[fig:blobs\_sfr\]](jelly_cfr_total_blobs_mass2e8_MCMC_inv_allfit.png)
As for the clumps we also have integrated values, we can compare their SFR-[$\rm M_\ast$]{}relation to that of stripping galaxies, computed taking into account only the SFRs and masses measured within the galaxy disks. Fig. \[fig:blobs\_sfr\] shows that the relation for the clumps is not only shifted towards lower mass and SFR values, but it is also much steeper. This result might suggest that the clumps are not simply a smaller scale of the galaxies. However, in this comparison, stellar masses are not computed in exactly the same way. To overcome this issue, for the galaxy integrated values, we can compute the stellar mass in the same way we did for the clumps, i.e. excluding the contribution of the stellar populations older than $2\times10^8$ yr. In this way we can inspect the amount of stellar mass produced only in the most recent epochs. This is shown in Fig.\[fig:blobs\_sfr\] when we plot using the blue symbols only the mass formed in the last $2\times10^8$ yr: the global relation is now simply an extension of that traced by the clumps, suggesting that the local mode of star formation is very similar in the galaxy disks and tails.
discussion and conclusions {#sec:disc}
==========================
The analysis of the spatially resolved Star Formation Rate- Mass relation can help the understanding of how galaxies assemble at different spatial scales. Comparing the relation of galaxies located in different environments can also shed light on the role of environmental processes in enhancing on suppressing the star formation and eventually in the galaxy quenching.
In this paper we have investigated the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}of 40 local cluster galaxies selected for showing signs of the effects of ram pressure stripping. We have also contrasted the results with those obtained inspecting a sample of 30 undisturbed galaxies, presented in @Vulcani2019b. The 70 galaxies are drawn from the GAs Stripping Phenomena in galaxies (GASP) sample [@Poggianti2017], and data have been analysed in a homogeneous way, therefore no systematic affects the results.
In @Vulcani2018_L we compared the integrated properties of these two samples, and found that stripping galaxies occupy the upper envelope of the undisturbed sample SFR-[$\rm M_\ast$]{}relation, showing a systematic enhancement of the SFR at any given mass. The star formation enhancement occurs both in the disk and in the tails. In this paper we aimed at further investigating and spatially localising such SFR enhancement.
The first result of the paper is presented in Figure \[fig:SFR\_Mass\_all\], which showed that even on $~\sim1$kpc scales, stripping galaxies present a systematic enhancement of [$\rm \Sigma_{SFR}$]{}at any given [$\rm \Sigma_\ast$]{}compared to their undisturbed counterparts. This excess is as large as $\sim 0.35$ dex at [$\rm \Sigma_\ast$]{}=$\rm 10^{8}M_\odot \,kpc^{-2}$. This result is only partially driven by the different mass distribution between the two samples. The excess is overall independent on the degree of stripping (except for the truncated disks) and of the amount of star formation in the tails, but is larger for less massive galaxies and decreases with increasing mass.
Interestingly, analyzing the ALMA data of a subset of the GASP galaxies, @Moretti2020 have shown that galaxies undergoing ram pressure stripping have a much larger H$_2$ reservoir - which is the fuel for star formation - than normal galaxies, suggesting that this physical process causes the conversion of large amounts of HI into the molecular phase in the disk. This result can also explain the excess in SFR that we observe.
The presence of AGNs seems not to affect the results. In addition, as the greatest differences between stripping and control sample galaxies is observed at low masses, where galaxies in our sample do not host AGNs, results can not even be driven by a contamination of AGN spaxels not correctly identified by the BPT diagram.
In a simple model of a galaxy, gas is in pressure equilibrium that is set by the gravitational potential. Thus in order for ram pressure to have any effect on this gas, it must be larger than the disk gas pressure. This is set by the restoring force in galaxies, and often used for determining whether gas can be removed from a disk (@Jaffe2018 [@Gullieuszik2020]. By the same argument, gas cannot be compressed unless ram pressure is stronger than the gravitationally-set pressure. Therefore, one would expect that because compression is relatively stronger than the gas pressure in the outskirts of galaxies, we should see a SFR enhancement in the outskirts of ram pressure stripped galaxies. Indeed, this is seen in the simulations of [@Roediger2014].
However, Fig. \[fig:delta\] suggests that the boost in the SFR surface density happens both in the inner and outer regions with respect to the control sample, and across a range of galaxy masses. Therefore a simple compression argument is not as easily applied. We argue that there are two probable causes for this. First, as has been found in simulations [e.g., @TonnesenBryan2012], dense gas that is not stripped from the outskirts of galaxies can lose angular momentum via shear from the ICM, and spiral towards the center. This dense gas may then undergo star formation near the galaxy center, increasing the local star formation rate. However, we may then predict that the star formation in the outskirts would decrease as dense gas migrated inward. The second cause takes into account the varying temperature in the ISM. Any compression wave from ram pressure may drive shocks in cold clouds (with lower sound speeds), inducing star formation. As this does not require cloud migration it may be the more appealing picture for how star formation surface density can increase at all galactic radii.
Overall we could not detect a clear dependence on the [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relation of the stripping stage, indicating that the [$\rm \Sigma_{SFR}$]{}enhancement appears as soon as the stripping begins and is maintained throughout the different stripping phases. Only truncated disks, representative of the final stage of stripping galaxies show a different behaviour. The fit of the relation is very similar to that of the control sample, but it is much narrower. [ [@Fritz2017] have suggested that the existence of truncated disks points to an outside-in quenching scenario [see also @Boselli2016]]{}, with the galaxy cores being the last portions of galaxies still able to produce stars. Our results though suggest that these cores are still undisturbed, and produce new stars at the same rate as undisturbed galaxies, [ similarly to what found by [@KoopmannKenney2004a; @KoopmannKenney2004b] for a sample of galaxies in the Virgo cluster.]{}
The analysis presented in this paper also highlights the existence of a large galaxy-by-galaxy variation (Fig.\[fig:sfr\_mass\_dist\_mass\]), similar to that found in @Vulcani2019b for the galaxies of the control sample. In many cases, especially for the most massive galaxies, we found that galaxy cores are always below the fit, suggesting that these regions are deprived of star formation and therefore supporting an inside-out quenching scenario according to which the suppression of the star formation occurs in the galaxy cores first and then extends to the outskirts. This behaviour is not due to the presence of a bulge [@Vulcani2019b] nor to the presence of an AGN. Indeed we consider only the spaxels powered by star formation.
A point that we did not explore here is the location of the enhancement with respect to the galaxy disk and the motion of the galaxy. In the literature, this has been done for NGC 2276, where the observed enhancement in SFR on one side of the galaxy has been explained in terms of a combination of both tidal forces and ram pressure [@Gruendl1993; @Hummel1995; @Rasmussen2006; @Wolter2015; @Tomicic2018]. Inspired by this result, [@Troncoso2020], using the EAGLE simulation, looked for the effects of the ICM on the spatially resolved star-formation activity in galaxies. They found that dividing each galaxy in two halves using the plane perpendicular to the velocity direction, differentiating the galaxy part approaching to the cluster center (the leading half), and the opposite one (the trailing half), there is an enhancement of the SFR, SFE, and interstellar medium pressure in the leading half with respect to the trailing one. Their results suggest that RP is boosting the star formation by gas compression in the leading half, and transporting the gas to the trailing half. As subdividing galaxies based on the velocity cut proposed by [@Troncoso2020] is not feasible observationally, the same authors suggest to use instead the plane that maximizes the SFR difference, showing that it is in most cases well aligned to the velocity vector. This analysis, certainly relevant for understanding the result of this paper, is indeed deferred to a future work (I. Gaspar et al., in prep.). Analyzing the spatially resolved SFR of jellyfish galaxies in their initial stripping phase, they preliminarily find a SFR enhancement on the leading side (i.e. the half galaxy closest to the cluster center) of the galaxy disk. If this preliminary finding holds up, then assuming that galaxies move toward the cluster center, this relative triggering could correspond to the RPS compression.
Finally, in the last part of the paper, we have focused on the star forming clumps detected in the tails of stripping galaxies and investigated their local and global SFR-Mass relation. These clumps can be found as far as 80 kpc from the galaxy center and we detected an anticorrelation between their distance and their [$\rm \Sigma_\ast$]{}, indicating that further away clumps typically [ have lower [$\rm \Sigma_\ast$]{}values]{}.
Investigating their [$\rm \Sigma_{SFR}$]{}-[$\rm \Sigma_\ast$]{}relation, a cloud rather than a well defined relation is evident. The [$\rm \Sigma_\ast$]{}of the clumps spans the range $10^5-10^7$ [$\rm M_\odot$]{}kpc$^{-2}$, the [$\rm \Sigma_{SFR}$]{}spans the range $10^{-4}-10^{-2}$ [$\rm M_\odot$]{}yr$^{-1}$ kpc$^{-2}$, occupying a very different locus from that of the spaxels in the galaxy disks. Compared to the extrapolation of the [$\rm \Sigma_{SFR}$]{}- [$\rm \Sigma_\ast$]{}relation of the disk spaxels towards low mass surface densities, that of the clumps is shifted towards high values at any given [$\rm \Sigma_\ast$]{}. We remind the reader, though, that a fair comparison is nor straightforward, as the [$\rm \Sigma_\ast$]{}values of the clumps in the tails represent the stellar mass formed only during the ongoing star formation episode, therefore the “true clump mass”. In contrast, in the galaxy disks, the masses are “projected stellar masses” inflated by the underlying old stellar populations.
Considering global values, the clumps SFR-Mass relation is much steeper than that of the galaxy disk and it is not simply the extrapolation of the galaxy relation. Nonetheless, if we exclude the contribution of the stellar populations older than few $10^8$ yr in the galaxy disk values, adopting the same approach used for the clumps, the global relation becomes an extension of that traced by the clumps, suggesting that the local mode of star formation is very similar in the galaxy disks and tails.
We thank the referee for their comments that helped us to improve the manuscript. Based on observations collected at the European Organisation for Astronomical Research in the Southern Hemisphere under ESO programme 196.B-0578. This project has received funding from the European Reseach Council (ERC) under the Horizon 2020 research and innovation programme (grant agreement N. 833824). We acknowledge financial contribution from the contract ASI-INAF n.2017-14-H.0, from the grant PRIN MIUR 2017 n.20173ML3WW\_001 (PI Cimatti) and from the INAF main-stream funding programme (PI Vulcani).
[^1]: <http://web.oapd.inaf.it/gasp/index.html>
[^2]: Colour images of all GASP galaxies along with [H$\alpha$]{}images can be consulted on a webpage at <http://web.oapd.inaf.it/gasp/gasp_atlas>.
[^3]: <http://www.eso.org/sci/software/pipelines/muse>
[^4]: Among the various line-ratio diagrams, the one based on the \[OI\] is the most sensitive to physical processes different from Star Formation (e.g. thermal conduction from the surrounding hot ICM, turbulence and shocks) and can therefore be considered as a conservative lower limit of the real star formation budget [@Poggianti2019]. In the appendix of @Vulcani2019b we have shown that results for the control sample are qualitatively independent on the choice of the diagnostic diagram.
[^5]: [ Note that the laplacian filtering measures the second spatial derivative of an image and is commonly defined using a negative peak, that is why we are looking for minima.]{}
[^6]: Note that results do not change if we adopt as upper limit for the mass bins $\log($[$\rm M_\ast$]{}\[$M_\odot$\]) = 11, which is the maximum mass of control sample galaxies.
[^7]: A reliable fit could not be retrieved for JO179, JO171, and JW39.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The onset of quartetting, i.e. $\alpha$-particle condensation, in symmetric nuclear matter is studied with the help of an in-medium modified four nucleon equation. It is found that at very low density quartetting wins over pairing, because of the strong binding of the $\alpha$-particles. The critical temperature can reach values up to around 6 MeV. Also the disappearance of $\alpha$-particles with increasing density, i.e. the Mott transition, is investigated. In finite nuclei the Hoyle state, that is the ${0_2}^+$ of $^{12}$C, is identified as an “$\alpha$-particle condensate” state. It is conjectured that such states also exist in heavier $n\alpha$-nuclei, like $^{16}$O, $^{20}$Ne, etc. For instance the 6-th $0^+$ state of $^{16}$O at 15.1 MeV is identified from a theoretical analysis as being a strong candidate for an $\alpha$ condensate state. Exploratory calculations are performed for the density dependence of the $\alpha$ condensate fraction at zero temperature to address the suppression of the four-particle condensate below nuclear-matter density. Possible quartet condensation in other systems is discussed briefly.\
Keywords: nuclear matter, $\alpha$-matter, superfluidity, Bose-Einstein condensation, strongly coupled systems
author:
- |
Y. Funaki$^1$, T. Yamada$^2$, H. Horiuchi$^{3,4}$, G. Röpke$^5$, P. Schuck$^{6,7}$ and A. Tohsaki$^3$\
$^1$Nishina Center for Accelerator-Based Science,\
The Institute of Physical and Chemical Research (RIKEN), Wako 351-0198, Japan\
$^2$Laboratory of Physics, Kanto Gakuin University, Yokohama 236-8501, Japan\
$^3$Research Center for Nuclear Physics (RCNP), Osaka University,\
Ibaraki, Osaka 567-0047, Japan\
$^4$ International Institute for Advanced Studies, Kizugawa 619-0225, Japan\
$^5$Institut für Physik, University of Rostock, Universitätsplatz 1,\
18051 Rostock, Germany\
$^6$Institut de Physique Nucléaire, CNRS, UMR8608, Orsay, F-91406, France\
$^7$Université Paris-Sud, Orsay, F-91505, France\
title: ' Alpha-Particle Condensation in Nuclear Systems'
---
Introduction
============
One of the most amazing phenomena in quantum many-particle systems is the formation of quantum condensates. At present, the formation of condensates is of particular interest in strongly coupled fermion systems in which the crossover from Bardeen-Cooper-Schrieffer (BCS) pairing to Bose-Einstein condensation (BEC) may be investigated. Among very different quantum systems such as the electron-hole-exciton system in excited semiconductors, atoms in traps at extremely low temperatures, etc., nuclear matter is especially well suited for the study of correlation effects in a quantum liquid.
Neutron matter, nuclear matter, but also finite nuclei are superfluid. However, at low density, nuclear matter will not cluster into pairs, i.e. deuterons but rather into $\alpha$ -particles which are much more stable. Also heavier clusters, starting with Carbon, may be of importance but are presently not considered for condensation phenomena. Therefore, one may ask the question whether there exists quartetting, i.e. $\alpha$-particle condensation, in nuclei, analogous to nuclear pairing. The only nucleus which in its ground state has a pronounced $\alpha$ -cluster structure is $^8$Be. In section 4 we will show a figure of $^8$Be in the laboratory frame and in the intrinsic deformed frame. We will see that $^8$Be is formed out of two $\alpha$-particles roughly 4 fm apart, only mildly interpenetrating one another. Actually $^8$Be is slightly unstable and the two $\alpha$’s only hold together via the Coulomb barrier. Because of the large distance of the two $\alpha$-particles, the 0$^+$ ground state of $^8$Be has, in the laboratory frame, a spherical density distribution whose average is very low: about 1/3 of ordinary saturation density $\rho_0$. $^8$Be is, therefore, a very large object with an rms radius of about 3.7 fm to be compared with the nuclear systematics of $R = r_0A^{1/3}$ = 2.44 fm. Definitely $^8$Be is a rather unusual and, in its kind, unique nucleus. One may ask the question what happens when one brings a third $\alpha$-particle alongside of $^8$Be. We know the answer: the 3-$\alpha$ system collapses to the ground state of $^{12}$C which is much denser than $^8$Be and can not accommodate with its small radius of 2.4 fm three more or less free $\alpha$-particles barely touching one another. One nevertheless may ask the question whether the dilute three $\alpha$ configuration $^8$Be-$\alpha$, or rather $\alpha$-$\alpha$-$\alpha$, may not form an isomeric or excited state of $^{12}$C. That such a state indeed exists will be one of the main subjects of our considerations. Once one accepts the idea of the existence of an $\alpha$-gas state in $^{12}$C, there is no reason why equivalent states at low density should not also exist in heavier $n\alpha$-nuclei, like $^{16}$O, $^{20}$Ne, etc. The possible existence of a loosely bound 4$\alpha$ state in $^{16}$O will be another topic of our presentation. In a mean field picture, i.e. all $\alpha$’s being ideal bosons ( in this context remember that the first excited state of an $\alpha$-particle is at $\sim$ 20 MeV, by factors higher than in all other nuclei), all $\alpha$’s will occupy the lowest $0S$-state, i.e. they will condense. This forms, of course, not a macroscopic condensate but it can be understood in the same sense as we know that nuclei are superfluid because of the presence of a finite number of Cooper pairs. On the other hand, for example during the cooling process of compact stars [@ShapT], where one predicts the presence of $\alpha$-particles [@ST], a real macroscopic phase of condensed $\alpha$’s may be formed. In the present contribution we will mainly concentrate on nuclear systems but we also can think about the possibility of quartetting in other Fermi-systems. One should, however, keep in mind that a pre-requisite for its existence is, as in nuclear physics, that there are four different types of fermions. For example to form quartets with cold atoms one could try to trap fermions in four different magnetic substates, a task which eventually seems possible [@Salo]. Also theoretical works in this direction have appeared in the mean while [@Lecheminant; @miyake]. The fact that the $\alpha$-particle condensates in nuclei do not form the ground state, may give raise to questions. In this respect, one should note that Bose condensates of cold atoms in traps also are not in their ground state which is a solid. It is a question of time scales. $\alpha$-particle condensate states in nuclei usually live four orders of magnitude longer than typical nuclear times.
In the next section we will investigate how the binding energy of various nuclear clusters change with density. In section 3 we study the critical temperature of $\alpha$-particle condensation in infinite matter via an in-medium four-nucleon equation (Thouless criterion) and in section 4 we treat $\alpha$-particle condensation in finite nuclei. In section 5 we give our results for $^{12}$C and in section 6 the ones for $^{16}$O. In section 7, we briefly discuss the question of the occupation numbers of the $\alpha$-particles in the various nuclear states with emphasis on the condensate states. In section 8 we present a simplified calculation of the condensate fraction in $\alpha$-matter. Finally in section 9 we conclude with outlook and further discussions.
Nuclear clusters in the medium
==============================
With increasing density of nuclear matter, medium modifications of single-particle states as well as of few-nucleon states become of importance. The self-energy of an $A$-particle cluster can in principle be deduced from contributions describing the single-particle self-energies as well as medium modifications of the interaction and the vertices. A guiding principle in incorporating medium effects is the construction of [*consistent*]{} (“conserving”) approximations, which treat medium corrections in the self-energy and in the interaction vertex at the same level of accuracy. This can be achieved in a systematic way using the Green functions formalism [@KKER]. At the mean-field level, we have only the Hartree-Fock self-energy $\Gamma^{\rm HF} =
\sum_2 V(12,12)_{\rm ex} f(2)$ together with the Pauli blocking factors, which modify the interaction from $V(12,1'2')$ to $V(12,1'2')[1 - f(1) - f(2)]$, with $f(1)=[1+\exp(E^{\rm HF}(1)-\mu)/T]^{-1}$. In the case of the two-nucleon system ($A=2$), the resulting effective wave equation which includes those corrections reads $$\label{two_part_bind}
\left[E^{\rm HF}(1)+E^{\rm HF}(2)-E_{2,n,P}\right] \psi_{2,n,P}(12) +
\sum_{1'2'}[1-f(1)-f(2)]\,\,V(12,1'2')
\psi_{2,n,P}(1'2')=0.$$ This [*effective wave equation*]{} describes bound states as well as scattering states. The onset of pair condensation is achieved when the binding energy $E_{d,P=0}$ coincides with $2 \mu$.
Similar equations have been derived from the Green function approach for the case $A = 3$ and $A = 4$, describing triton/helion ($^3$He) nuclei as well as $\alpha$-particles in nuclear matter. The effective wave equation contains in mean field approximation the Hartree-Fock self-energy shift of the single-particle energies as well as the Pauli blocking of the interaction. We give the effective wave equation for $A=4$, $$\begin{aligned}
\label{four_part_bind}
&& \left[E^{\rm HF}(1)+E^{\rm HF}(2)+ E^{\rm HF}(3)+E^{\rm HF}(4)
-E_{4,n,P}\right]
\psi_{4,n,P}(1234) \nonumber\\ && +
\sum_{i<j}\sum_{1'2'3'4'}[1-f(i)-f(j)]V(ij,i'j')\prod_{k\neq i,j}\delta_{k,k'}
\psi_{4,n,P}(1'2'3'4')=0.
\label{EWE}\end{aligned}$$ A similar equation is obtained for $A=3$.
The effective wave equation has been solved using separable potentials for $A=2$ by integration. For $A=3,4$ we can use a [*Faddeev approach*]{} [@Beyer]. The shifts of binding energy can also be calculated approximately via perturbation theory. In Fig. \[shifts\] we show the shift of the binding energy of the light clusters ($d, t/h$ and $\alpha$) in symmetric nuclear matter as a function of density for temperature $T$ = 10 MeV.
It is found that the cluster binding energy decreases with increasing density. Finally, at the [*Mott density*]{} $\rho_{A,n,P}^{\rm Mott}(T)$ the bound state is dissolved. The clusters are not present at higher densities, merging into the nucleonic medium. For a given cluster type characterized by $A,n$, we can also introduce the Mott momentum $P^{\rm Mott}_{A,n}(\rho,T)$ in terms of the ambient temperature $T$ and nucleon density $\rho$, such that the bound states exist only for $P \ge P^{\rm Mott}_{A,n}(\rho,T)$. We do not present an example here, but it is intuitively clear that a cluster with high c.o.m. momentum with respect to the medium is less affected by the Pauli principle than a cluster at rest.
Four-particle condensates and quartetting in nuclear matter
===========================================================
In general, it is necessary to take into account of [*all bosonic clusters*]{} to gain a complete picture of the onset of superfluidity. As is well known, the deuteron is weakly bound as compared to other nuclei. Higher $A$-clusters can arise that are more stable. In this section, we will consider the formation of $\alpha$-particles, which are of special importance because of their large binding energy per nucleon ($\sim 7$ MeV). We will not include tritons or helions, which are fermions and not so tightly bound. Moreover, we will not consider nuclei in the iron region, which have even larger binding energy per nucleon than the $\alpha$-particle and thus constitute, in principle, the dominant component at low temperatures and densities. However, the latter are complex structures of many particles and are strongly affected by the medium as the density increases for given temperature, so that they are assumed not to be of relevance in the density region considered here.
The in-medium wave equation for the four-nucleon problem has been solved using the Faddeev-Yakubovski technique, with the inclusion of Pauli blocking. The binding energy of an $\alpha$-like cluster with zero c.o.m. momentum vanishes at around $\rho_0/10$, where $\rho_0 \simeq 0.16$ nucleons/fm$^3$ denotes the saturation density of isospin-symmetric nuclear matter, see Fig. \[shifts\]. Thus, the four-body bound states make no significant contribution to the composition of the system above this density. Given the medium-modified bound-state energy $E_{4,P}$, the bound-state contribution to the EOS is $$\rho_4(\beta,\mu) = \sum_P\left[e^{\beta(E_{4,P}
- 2 \mu_p-2 \mu_n)} -1\right]^{-1}\,.$$ We will not include the contribution of the excited states or that of scattering states. Because of the large specific binding energy of the $\alpha$ particle, low-density nuclear matter is predominantly composed of $\alpha$ particles. This observation underlies the concept of $\alpha$ matter and its relevance to diverse nuclear phenomena.
As exemplified by Eq. (\[EWE\]), the effect of the medium on the properties of an $\alpha$ particle in mean-field approximation (i.e., for an uncorrelated medium) is produced by the Hartree-Fock self-energy shift and Pauli blocking. The shift of the $\alpha$-like bound state has been calculated using perturbation theory [@RMS] as well as by solution of the Faddeev-Yakubovski equation [@Beyer]. It is found that the bound states of clusters $d$, $t$, and $h$ with $A<4$ are already dissolved at a Mott density $\rho_\alpha^{\rm Mott} \approx \rho_0/10$, see Fig. \[shifts\]. Since Bose condensation only is of relevance for $d$ and $\alpha$, and the fraction of $d$, $t$ and $h$ becomes low compared with that of $\alpha$ with increasing density, we can neglect the contribution of them to an equation of state. Consequently, if we further neglect the contribution of the four-particle scattering phase shifts in the different channels, we can now construct an equation of state $\rho(T, \mu) =\rho^{\rm free}(T, \mu) +
\rho^{{\rm bound}, d}(T, \mu)
+\rho^{{\rm bound}, \alpha}(T, \mu)$ such that $\alpha$-particles determine the behavior of symmetric nuclear matter at densities below $\rho_\alpha^{\rm Mott}$ and temperatures below the binding energy per nucleon of the $\alpha$-particle. The formation of deuteron clusters alone gives an incorrect description because the deuteron binding energy is small, and the abundance of $d$-clusters is small compared with that of $\alpha$-clusters. In the low density region of the phase diagram, $\alpha$-matter emerges as an adequate model for describing the nuclear-matter equation of state.
With increasing density, the medium modifications – especially Pauli blocking – will lead to a deviation of the critical temperature $T_c(\rho)$ from that of an ideal Bose gas of $\alpha$-particles (the analogous situation holds for deuteron clusters, i.e., in the isospin-singlet channel).
Symmetric nuclear matter is characterized by the equality of the proton and neutron chemical potentials, i.e., $\mu_p=\mu_n=\mu$. Then an extended Thouless condition based on the relation for the four-body T-matrix (in principle equivalent to Eq. (\[EWE\]) at eigenvalue 4$\mu$) $$\begin{aligned}
{\rm T}_4(1234,1''2''3''4'', 4 \mu)& =& \sum_{1'2'3'4'} \Biggl\{
\frac{V(12,1'2')[1-f(1)-f(2)] }{ 4
\mu - E_1-E_2-E_3-E_4 }\delta(3,3')\delta(4,4')\nonumber\\
&& \qquad \qquad + {\rm cycl.} \Biggr\}
{\rm T}_4(1'2'3'4',1''2''3''4'', 4 \mu)\end{aligned}$$ serves to determine the onset of Bose condensation of $\alpha$-like clusters, noting that the existence of a solution of this relation signals a divergence of the four-particle correlation function. An approximate solution has been obtained by a variational approach, in which the wave function is taken as Gaussian incorporating the correct solution for the two-particle problem [@RSSN].
The results are presented in Figs. \[fig:trans\_mu\] and \[fig:trans\_dens\]. An important consequence of those is that at the lowest temperatures, Bose-Einstein condensation occurs for $\alpha$ particles rather than for deuterons. As the density increases within the low-temperature regime, the chemical potential $\mu$ first reaches $-7$ MeV, where the $\alpha$’s Bose-condense. By contrast, Bose condensation of deuterons would not occur until $\mu$ rises to $-1.1$ MeV.
The [*“quartetting”*]{} transition temperature sharply drops as the rising density approaches the critical Mott value at which the four-body bound states disappear. At that point, pair formation in the isospin-singlet deuteron-like channel comes into play, and a deuteron condensate will exist below the critical temperature for BCS pairing up to densities above the nuclear-matter saturation density $\rho_0$, as described in the previous Section. The critical density at which the $\alpha$ condensate disappears is estimated to be $\rho_0/3$. Therefore, $\alpha$-particle condensation primarily only exists in the Bose-Einstein-Condensed (BEC) phase and there does not seem to exist a phase where the quartets acquire a large extension as Cooper pairs do in the weak coupling regime. However, the variational approach of Ref. [@RSSN] on which this conclusion is based represents only a first attempt at the description of the transition from quartetting to pairing. The detailed nature of this fascinating transition remains to be clarified.
Many different questions arise in relation to the possible physical occurrence and experimental manifestations of quartetting: Can we observe the hypothetical “$\alpha$ condensate” in nature? What about thermodynamic stability? What happens with quartetting in asymmetric nuclear matter? Are more complex quantum condensates possible? What is their relevance for finite nuclei? As discussed below, the special type of microscopic quantum correlation associated with quartetting may be important in nuclei, its role in these finite inhomogeneous systems being similar to that of pairing.
Description of Alpha-Particle Condensate States in Self-Conjugate 4n Nuclei
===========================================================================
Let us discuss the possibility of quartetting in nuclei. The only nucleus having a pronounced $\alpha$-cluster structure in its ground state is $^8$Be. In Fig. \[profiles\](a), we show the result of an exact calculation of the density distribution of $^8$Be in the laboratory frame. In Fig. \[profiles\](b) we show, for comparison, the result of the same calculation in the intrinsic, deformed frame. We see a pronounced two $\alpha$-cluster structure where the two $\alpha$’s are $\sim$ 4 fm apart, giving rise to a very low average density $\rho \sim \rho_0/3$ as seen in Fig. \[profiles\](a). As already discussed in the introduction, $^8$Be is a rather unusual and unique nucleus. One may be intrigued by the question, already raised earlier, whether loosely bound $\alpha$-particle configurations may not also exist in heavier $n\alpha$-nuclei, at least in excited states, naturally close to the $n\alpha$ disintegration threshold. Since $\alpha$-particles are rather inert bosons ( first excited state at $\sim$ 20 MeV), these $\alpha$-particles then would all condense in the lowest $S$-wavefunction, very much in the same way as do bosonic atoms in magneto-optical traps [@String]. This question and exploring related issues of quartetting in finite nuclei will consume most of the rest of the present study.
In fact, we will be able to offer strong arguments that the $0_2^+$ state of $^{12}$C at 7.654 MeV is a state of $\alpha$-condensate nature. Later, also indications for the existence of an analogous state in $^{16}$O will be discussed.
First, it should be understood that the $0_2^+$ state in $^{12}$C is in fact hadronically unstable (as $^8$Be), being situated about 300 keV above the three $\alpha$-break up threshold. This state is stabilized only by the Coulomb barrier. It has a width of $8.7$ eV and a corresponding lifetime of $7.6\times 10^{-17}$ s. As well known, this state is of paramount astrophysical (and biological!) importance due to its role in the creation of $^{12}$C in stellar nucleosynthesis. Its existence was predicted in 1953 by the astrophysicist Fred Hoyle [@hoyle]. His prediction was confirmed experimentally a few years later by Willy Fowler and coworkers at Caltech [@fowler]. It is also well known that this [*Hoyle state*]{}, as it is now called, is a notoriously difficult state for any nuclear theory to explain. For example, the most modern no-core shell-model calculations predict the $0_2^+$ state in $^{12}$C to lie at around 17 MeV above the ground state – more than twice the actual value [@nocore]. This fact alone tells us that the Hoyle state must have a very unusual structure. It is easy to understand that, should it indeed have the proposed loosely bound three $\alpha$-particle structure, a shell-model type of calculation would have great difficulties in reproducing its properties.
An important development bearing on this issue took place some thirty years ago. Two Japanese physicists, M. Kamimura [@kamimura] and K. Uegaki [@uegaki], along with their collaborators, almost simultaneously reproduced the Hoyle state from a microscopic theory. They employed a twelve-nucleon wave function together with a Hamiltonian containing an effective nucleon-nucleon interaction. At that time, their work did not attract the attention it deserved; the true importance of their achievement has been appreciated only recently. The two groups started from practically the same ansatz for the $^{12}$C wave function, which has the following three $\alpha$-cluster structure: $$\langle \vec{r}_1...\vec{r}_{12}|^{12}{\rm C} \rangle =
{\cal{A}}\left[\chi({{\vec{s}}},\vec{t})\phi_1\phi_2\phi_3\right]\,.$$ In this expression, the operator ${\cal A}$ imposes antisymmetry in the nucleonic degrees of freedom and $\phi_i$, with $i=1,2,3$ for the three $\alpha$’s, is an intrinsic $\alpha$-particle wave function of prescribed Gaussian form, $$\phi(\vec{r}_1,\vec{r}_2,\vec{r}_3, \vec{r}_4) =
\exp\left\{-\left[(\vec{r}_1-\vec{r}_2)^2
+(\vec{r}_1-\vec{r}_3)^2 +\cdots \right]/{8b^2}\right\}\,,
\label{alphawf}$$ where the size parameter $b$ is adjusted to fit the rms of the free $\alpha$-particle, and $\chi({{\vec{s}}},\vec{t})$ is a yet-to-be determined three-body wave function for the c.o.m.motion of the three $\alpha$’s, their corresponding Jacobi coordinates being denoted by ${\vec{s}}$ and $\vec{t}$. The unknown function $\chi$ was determined via calculations based on the Generator Coordinate Method [@uegaki] (GCM) and the Resonating Group Method [@kamimura] (RGM) calculations using the Volkov I and Volkov II nucleon-nucleon forces, which fit $\alpha$-$\alpha$ phase shifts. The precise solution of this complicated three body problem, carried out three decades ago, was truly a pioneering achievement, with results fulfilling expectations. The position of the Hoyle state, as well as other properties including the inelastic form factor and transition probability, successfully reproduced the experimental data. Other states of $^{12}$C below and around the energy of the Hoyle state were also successfully described. Moreover, it was already recognized that the three $\alpha$’s in the Hoyle state form sort of a gas-like state. In fact, this feature had previously been noted by H. Horiuchi [@hori] prior to the appearance of Refs. [@kamimura; @uegaki], based on results from the orthogonality condition model (OCM) [@saitoh]. All three Japanese research groups concluded from their studies that the linear-chain state of three $\alpha$-particles, postulated by Morinaga many years earlier [@mori] as an interpretation of the Hoyle state, had to be rejected.
Although the evidence for interpreting the Hoyle state in terms of an $\alpha$ gas was stressed in the cited papers from the late 1970’s, two important aspects of the situation were missed at that time. First, because the three $\alpha$’s move in identical $S$-wave orbits, one is dealing with an $\alpha$-condensate state, albeit not in the macroscopic sense, and that this may be a quite general phenomenon, also in heavier self-conjugate nuclei. The second important point is that the complicated three-body wave function $\chi({\vec{s}},{\vec{t}})$ for the c.o.m. motion of the three $\alpha$’s can be replaced by a structurally and conceptually very simple microscopic three-$\alpha$ wave function of the condensate type, which has practically 100 percent overlap with the previously constructed ones [@thsr] [@cbec] (see also Ref. [@Hackenbroich]). We now describe this condensate wave function.
We start by examining the BCS wave function of ordinary fermion pairing, obtained by projecting the familiar BCS ground-state ansatz onto an $N$-particle subspace of Fock space. In the position representation, this wave function is $$\langle \vec{r}_1\cdots \vec{r}_N|{\rm BCS}\rangle = {\cal {A} }
\left[\phi(\vec{r}_1,\vec{r}_2)\phi(\vec{r}_3,\vec{r}_4)\cdots \phi(\vec{r}_{N-1}
\vec{r}_N)\right]\,, \label{eq:1}$$ where $\phi(\vec{r}_1,\vec{r}_2)$ is the Cooper-pair wave function (including spin and isospin), which is to be determined variationally through the familiar BCS equations. The condensate character of the BCS ansatz is borne out by the fact that within the antisymmetrizer $\cal A$, one has a product of $N/2$ times the same pair wave function $\phi$, with one such function for each distinct pair in the reference partition of $\{1,2,\ldots,N-1,N\}$. Formally, it is now a simple matter to generalize (\[eq:1\]) to quartet or $\alpha$-particle condensation. We write $$\langle \vec{r}_1,\ldots,\vec{r}_N|\Phi_{n\alpha}\rangle = {\cal {A} }\left[
\phi_{\alpha}(\vec{r}_1, \vec{r}_2, \vec{r}_3, \vec{r}_4)\phi_{\alpha}
(\vec{r}_5,\ldots , \vec{r}_8)
\cdots \phi_{\alpha}(\vec{r}_{N-3},\ldots, \vec {r}_N)\right]\,, \label{eq:2}$$ where $\phi_{\alpha}$ is the wave function common to all condensed $\alpha$-particles. Of course, finding the variational solution for this function is, in general, extraordinarily more complicated than finding the Cooper pair-wave function $\phi$ of Eq. (\[eq:1\]). Even so, in the present case that the $\alpha$-particle is the four-body cluster involved, and for applications to relatively light nuclei, the complexity of the problem can be reduced dramatically. This possibility stems from the fact, already known to the authors of Refs. [@kamimura; @uegaki], that, due to the BEC-character of $\alpha$-particle condensation (see above), an excellent variational ansatz for the intrinsic wave function of the $\alpha$-particle is provided \[as in Eq. (\[alphawf\])\], by a Gaussian form with only the size parameter $b$ to be determined. In addition – and here resides the essential point of our wave function – even the c.o.m. motion of the system of $\alpha$-particles can be described very well by a Gaussian wave function with, this time, a size parameter $B \gg b$ to account for the motion over the nuclear space. We therefore write $$\phi_\alpha(\vec{r}_1,\vec{r}_2,\vec{r}_3,\vec{r}_4) =
e^{{\displaystyle{-2}}{\scriptstyle\vec{R}^2}{\displaystyle{/B^2}}}
\phi(\vec{r}_1-\vec{r}_2,\vec{r}_1-\vec{r}_3,\cdots)\,, \label{eq:3}$$ where $\vec{R}= (\vec{r}_1+\vec{r}_2+\vec{r}_3+\vec{r}_4)/4$ is the c.o.m. coordinate of one $\alpha$-particle and $\phi(\vec{r}_1-\vec{r}_2,...)$ is the same intrinsic $\alpha$-particle wave function of Gaussian form as already used in Refs. [@kamimura; @uegaki] and given explicitly in Eq. (\[alphawf\]). Naturally, in Eq. (\[eq:2\]) the center of mass $\vec{X}_{\rm cm}$ of the three $\alpha$’s, i.e., of the whole nucleus, should be eliminated; this is easily achieved by replacing $\vec{R}$ by $
\vec{R}-\vec{X}_{\rm cm}$ in each of the $\alpha$ wave functions in Eq. (\[eq:2\]). The $\alpha$-particle condensate wave function specified by Eqs. (\[eq:2\]) and (\[eq:3\]), proposed in Ref. [@thsr] and henceforth called the THSR wave function, now depends on only two parameters, $B$ and $b$. The wave function (\[eq:2\]) with (\[eq:3\]) is pictorially represented in Fig. \[fig:osc\]. The expectation value of an assumed microscopic Hamiltonian $H$, $${\cal {H}}(B,b)=\frac{\langle\Phi_{n\alpha}(B,b)|H|\Phi_{n\alpha}(B,b)\rangle}
{\langle \Phi_{n\alpha}|\Phi_{n\alpha}\rangle}\,, \label{eq:4}$$ can be evaluated, and the corresponding two-dimensional energy surface can be quantized using the two parameters $B$ and $b$ as Hill-Wheeler coordinates.
![Pictorial representation of the THSR wave function for $n=3$ ($^{12}$C). The three $\alpha$-particles are trapped in the $0S$-state of a wide harmonic oscillator $(B)$ and the four nucleons of each $\alpha$ are confined in the $0s$-state of a narrow one $(b)$. All nucleons are antisymmetrised.[]{data-label="fig:osc"}](mix_para.eps)
Before presenting the results, let us discuss the THSR wave function in somewhat more detail. This innocuous-looking variational ansatz, namely Eq. (\[eq:2\]) together with Eq. (\[eq:3\]), is actually more subtle than it might at first appear. One should realize that two limits are incorporated exactly. One is obtained by choosing $B=b$, for which Eq. (\[eq:2\]) reduces to a standard Slater determinant with harmonic-oscillator single-nucleon wave functions, leaving the oscillator length $b$ as the single adjustable parameter. This holds because the right-hand-side of expression (\[eq:3\]), with $B=b$, becomes a product of four identical Gaussians, and the antisymmetrization creates all the necessary $P$, $D$, etc. harmonic oscillator wave functions automatically [@thsr]. On the other hand, when $B \gg b$, the density of $\alpha$-particles is very low, and in the limit $B \rightarrow \infty$, the average distance between $\alpha$’s is so large that the antisymmetrisation between them can be neglected, i.e., the operator $\cal {A}$ in front of Eq. (\[eq:2\]) becomes irrelevant and can be removed. In this limiting case, our wave function then describes an ideal gas of independent, condensed $\alpha$-particles – it is a pure product state of $\alpha$’s! An elucidating study on this aspect is given in Ref. [@yamada1].\
Evidently, however, in realistic cases the antisymmetrizer $\cal {A}$ cannot be neglected, and evaluation of the expectation value (\[eq:4\]) becomes a nontrivial analytical task. The Hamiltonian in Eq. (\[eq:4\]) was taken to be the one used in Ref. [@tohsaki_F1], which features an effective nucleon-nucleon force of the Gogny type, with parameters fitted to $\alpha$-$\alpha$ scattering phase shifts as available about fifteen years ago. This force also leads to very reasonable properties of ordinary nuclear matter. Our theory is therefore free of any adjustable parameters. The energy landscapes ${\cal {H}}(B,b)$ for various $n\,\alpha$ nuclei are shown in Fig. \[fig:E\_surface\] [@funaki_be; @tohsaki_nara]. We see that they all have qualitatively the same structure. From the minimum point on, with increasing $B$-parameter, developes a valley with constant $b$ which takes approximately the value of the free space $\alpha$-particle. The valley then goes over a saddle point which indicates the disintegration of the nucleus into n $\alpha$-particles. It is interesting to see that the minimum of the energy surfaces does not correspond to the Slater determinant case with $b=B$ but rather a quite substantial gain in energy due to four body correlations can be observed even for the ground state in these light nuclei.
![Contour map of the energy surface ${\cal {H}}(B,b)$ for (a) $^8$Be, (b) $^{12}$C and (c) $^{16}$O. The variables $B$ and $R_0$ are connected by the relation, $B^2=b^2+2R_0^2$. Numbers attached to the contour lines are the binding energies in units of MeV.[]{data-label="fig:E_surface"}](E_surface.eps)
It is evident that for large numbers of $\alpha$-particles the explicit antisymmetrisation demanded in our wave function of Eq. (\[eq:2\]), will encounter great difficulties. On the other hand in the case of many quartets or a macroscopic number of it in the condensate, in particular in infinite nuclear matter, one may transform our number conserving condensate wave function into a coherent state and proceed in a similar way as in the BCS case of standard pairing. Roughly this can go along the following lines. The first step is to define an $\alpha$-particle coherent state
$$|\alpha \rangle \sim \exp\Big(\frac{1}{4!}\sum_{1234}\Phi_{1234}^{(\alpha)}c_1^\dagger c_2^\dagger c_3^\dagger c_4^\dagger\Big)|{\rm vac}\rangle .$$
Next comes to mix pair creators and pair destructors: $$Q_{\alpha}^\dagger =\frac{1}{2} \sum_{12}\{ X_{12}^{\alpha}c_1^\dagger c_2^\dagger - Y_{12}^{\alpha}c_1c_2\} .$$
If we replaced the pair operators by bosons, this would correspond to a Hartree-Fock-Bogoliubov approach for bosons, with, in analogy to the fermion case, a bosonic gap equation, etc. However, here we want to keep the fermionic structure of the pairs. The amplitudes $X,Y$ shall obey the following orthonormality, respectively completeness relations $XX^\dagger - YY^\dagger = 1$. We also demand: $Q_{\alpha}|\alpha \rangle=0$, and the corresponding order parameter is a quartet expectation value: $\langle c_1^\dagger c_2^\dagger c_3^\dagger c_4^\dagger \rangle \sim XY$.
A self-consistent set of quartet equations can then be constructed: $$\left(
\begin{array}{cc}
A & \Delta_4\\
-\Delta_4 & -A\\
\end{array}
\right)
\left(
\begin{array}{cc}
X\\
Y\\
\end{array}
\right)
=E
\left(
\begin{array}{cc}X\\
Y\\
\end{array}
\right),$$ where $$\Delta_4 = \langle [c^\dagger c^\dagger ,[H,c^\dagger c^\dagger ]]\rangle \sim v_{....}\langle c_1^\dagger c_2^\dagger c_3^\dagger c_4^\dagger \rangle,$$ and $v_{....}$ stands for the matrix element of the interaction and the bosonic “gap” is then very schematically given by $$\Delta_4 \sim \sum v_{....}XY.$$
Similar equations from an analogous procedure can be obtained mixing a fermion destructor with three creators: $q_\alpha^\dagger=\sum U_{123}^\alpha c_1^\dagger c_2^\dagger c_3^\dagger -\sum V_1^\alpha c_1$. The final equation for quartetting is a very intuitive extension of the equation for the BCS pairing order parameter $\langle cc \rangle$ [@Ring_Schuck]. $$(\epsilon_1+\epsilon_2+\epsilon_3+\epsilon_4)K_{1234}^\alpha - \Big[ (1-n_1-n_2)\Delta_{1234}^\alpha + {\rm perms.} \Big] = 4\mu K_{1234}^\alpha , \label{eq:quart}$$ with $\epsilon_i$ the self-consistent single particle energies and $\Delta_{1234}^\alpha =\frac{1}{2} \sum v_{123^\prime 4^\prime} K_{3^\prime 4^\prime 34}^\alpha$ where $K_{1234}$ is the order parameter specified below, and $v_{1234}$ is the antisymmetrised matrix element of the interaction. The occupation numbers $n_i = \langle c_i^\dagger c_i \rangle$ are obtained from the single particle Green’s function $G_1 = (\omega-\epsilon_1 - M_1^\omega)^{-1}$ with the mass operator $M_1^{\omega}=\frac{1}{3!}\sum_{234}|D_{234}^{\alpha}|^2N^0_{234}(\omega+\epsilon_2+\epsilon_3 + \epsilon_4)^{-1}$ with $\Delta_{234}^{\alpha}=N^0_{234}D^{\alpha}_{234}$ and $N^0_{234} = (1-n_2)(1-n_3)(1-n_4) + n_2n_3n_4$ (all energies counted from $\mu$). In the zero density limit Eq. (\[eq:quart\]) goes over into the exact free space $\alpha$-particle Schrödinger equation.
In infinite matter, the order parameter has zero total momentum and we write $$K_{1234} = \langle c_1^\dagger c_2^\dagger c_3^\dagger c_4^\dagger \rangle \rightarrow \delta({\vec k}_1+{\vec k}_2 + {\vec k}_3+ {\vec k}_4)
\Big\langle c^\dagger_{{\vec k}_1}c^\dagger_{{\vec k}_2}c^\dagger_{{\vec k}_3}c^\dagger_{{\vec k}_4}\Big\rangle_{S,T},$$ where $S,T$ stand for the spin-isospin wave function. Thus, the order parameter depends on three momenta ${\vec \kappa} =
{\vec k}_1-{\vec k}_2$; ${\vec \kappa}' = {\vec k}_3-{\vec k}_4$; ${\vec P}={\vec k}_1+{\vec k}_2=-({\vec k}_3+{\vec k}_4)$. In general this makes 9 variables! Nonlinear equations for the order parameter have to be solved what is a very demanding numerical task.
![Spin-isospin saturated $(0S)^4$ mean field configuration of the $\alpha$-particle.[]{data-label="fig:0s"}](alpha_gray.eps)
However, strong simplifications may be possible! For Bose-Einstein condensation of fermion-clusters heavier than a pair, one may proceed to a mean-field description of the cluster with projection on good total momentum ${\vec K}$. The clusters condense in the ${\vec K}=0$ state. Even for the $\alpha$-particle a mean field description is quite a good approximation, under the condition to use effective forces, such as Skyrme or Gogny, and to project, as mentioned, on good linear momentum (see Fig. \[fig:0s\]). $$\Phi^{(\alpha)} \sim\delta({\vec k}_1+{\vec k}_2 + {\vec k}_3+ {\vec k}_4)
\varphi_{0S}(\vec{k}_1)\varphi_{0S}(\vec{k}_2)\varphi_{0S}(\vec{k}_3)\varphi_{0S}(\vec{k}_4).$$
With $$\Phi^{(\alpha)}_{1234}= (YX^{-1})_{1234},$$ and $$XX^\dagger - YY^\dagger =1 \rightarrow XX^\dagger=(1-{\Phi^{(\alpha)}}^2)^{-1},$$ the self-consistent quartet equation can be expressed entirely in terms of $\Phi^{\alpha}$ and then via the product ansatz everything via a single $0S$ wavefunction. The nonlinear equation for $\varphi_{0S}({\vec r})$ should be solvable! This for any number of $\alpha$-particles! The last part of this section is onging work with T. Sogo [@sogo].
Results for Finite Nuclei: $^{12}$C
===================================
As we discussed already, the variational wave function constructed from the Hill-Wheeler equation based on Eqs. (\[eq:2\]), (\[eq:3\]), and (\[eq:4\]) has practically 100 percent overlap with the RGM and GCM wave functions constructed in Refs. [@kamimura] and [@uegaki], once the same Volkov force is used [@cbec]. It is, therefore, not astonishing that our results are very similar to theirs. Nevertheless, let us again discuss the situation in some detail. For $^{12}$C we obtain two eigenvalues in the $0^+$-channel: the ground state and the Hoyle state. Theoretical values for positions, rms values, and transition probabilities are given in Table \[tab:1\] and compared to the data. Inspecting the rms radii, we see that the Hoyle state has a volume 3 to 4 larger than that of the ground state of $^{12}$C. This is the primary aspect of the dilute-gas state we highlighted above. We also can make a deformed calculation in allowing the width parameter $B$ to have different values in the different directions. Projecting on good angular momentum then yields the position of the second $2^+$-state in $^{12}$C which is in good agreement with the experimental value [@ito; @fthsr]. Also its width can be evaluated and one obtains a quite reasonable estimate. Detailed investigation of the wave function of the $2_2^+$-state shows that it can essentially be described in lifting out of the condensate state with the three $\alpha$’s in the $0S$-orbit, one $\alpha$-particle in the next $0D$-orbit. It is tempting to imagine that the $0_3^+$-state which, experimentally, is almost degenerate with the $2_2^+$-state, is obtained by lifting one $\alpha$-particle into the $1S$-orbit. Preliminary theoretical studies [@kato] indicate that this scenario might indeed apply. However, the width of the $0_3^+$ state is very broad ($\sim$ 3 MeV), rendering a theoretical treatment rather delicate. Further investigations are necessary to validate or reject this picture which is shown graphically in Fig. \[fig:exp\_12C\]. At any rate, it would be quite satisfying, if the triplet of states, ($0_2^+,2_2^+, 0_3^+$) could all be explained from the $\alpha$-particle perspective, since those three states are [*precisely*]{} the ones which cannot be explained within a (no core) shell model approach [@nocore].
\[fig:exp\_12C\]
Constructing a pure-state $\alpha$-particle density matrix $\rho(\vec{R},\vec{R}')$ from our wave function, integrating out of the total density matrix all intrinsic $\alpha$-particle coordinates, and diagonalizing this reduced density matrix, we find that the corresponding $0S$ $\alpha$-particle orbit is occupied to 70 percent by the three $\alpha$-particles [@yamada1; @suzuki]. This is a huge percentage, giving vivid support to the view that the Hoyle state is an almost ideal $\alpha$-particle condensate. We will dwell on this point in more detail in sections 7 and 8.
---------------------------------------- --------- ----------------- ----------------------------------- ------------------------
condensate w.f. \[0pt\]\[0pt\][RGM [@kamimura]]{} \[0pt\]\[0pt\][Exp.]{}
(Hill-Wheeler)
\[0pt\]\[0pt\][$E$(MeV)]{} $0_1^+$ $-89.52$ $-89.4$ $-92.2$
$0_2^+$ $-81.79$ $-81.7$ $-84.6$
\[0pt\]\[0pt\][$R_{\rm r.m.s.}$(fm)]{} $0_1^+$ $\ \ \ 2.40$ $\ \ \ 2.40$ $\ \ \ 2.44$
$0_2^+$ $\ \ \ 3.83$ $\ \ \ 3.47$
$M(0_2^+\rightarrow 0_1^+)$(fm$^2$) $\ \ \ 6.45$ $\ \ \ 6.7$ $\ \ \ 5.4$
---------------------------------------- --------- ----------------- ----------------------------------- ------------------------
: Comparison of the binding energies, rms radii $(R_{\rm r.m.s.})$, and monopole matrix elements $(M(0_2^+\rightarrow 0_1^+))$ for $^{12}$C given by solving Hill-Wheeler equation based on Eq. (\[eq:2\]) and by Ref. [@kamimura]. The effective two-nucleon force Volkov No. 2 was adopted in the two cases for which the $3\alpha$ threshold energy is calculated to be $-82.04$ MeV.[]{data-label="tab:1"}
Let us now discuss what to our mind is the most convincing evidence that our description of the Hoyle state is the correct one. Like the authors of Ref. [@kamimura], we reproduce very accurately the inelastic form factor $0_1^+ \rightarrow 0_2^+$ of $^{12}$C, as shown in Fig. \[fig:2\_inel\]. As such, the agreement with experiment is already quite impressive. Additionally, however, the following study was made, results from which are presented in Fig. \[fig:3\_height\]. We artificially varied the extension of the Hoyle state and examined the influence on the form factor. It was found that the overall shape of the form factor shows little variation, for example in the position of the minimum. On the other hand, we found a strong dependence of the absolute magnitude of the form factor; Fig. \[fig:3\_height\] illustrates this behavior with a plot showing the variation of the height of the first maximum of the inelastic form factor as a function of the percentage change of the rms radius of the Hoyle state [@funaki1]. It can be seen that a 20 percent [*increase of the rms radius*]{} produces a remarkable decrease of the maximum by a factor of two! This strong sensitivity of the magnitude of the form factor to the size of the Hoyle state enhances our firm belief that the agreement with the actual measurement is tantamount to a proof that the calculated wide extension of the Hoyle state corresponds to reality.
Summarizing our inquiry into the possible role of $\alpha$ clustering in $^{12}$C, we have accumulated enough facts to be convinced that the Hoyle state is, indeed, what one may call an $\alpha$-particle condensate state. At the same time, we acknowledge that referring to only three particles as a “condensate” constitutes a certain abuse of the word. However, in this regard it should be remembered that also in the case of nuclear Cooper pairing, only a few pairs are sufficient to obtain clear signatures of superfluidity in nuclei!\
Let us now go one step further and investigate the four $\alpha$-particle case.
Alpha-particle condensation in $^{16}$O
=======================================
The establishment of this condensate aspect of the Hoyle state naturally leads us to the speculation about $4\alpha$-particle condensation in $^{16}$O, which is the focus in this section. The situtaion in $^{16}$O is, as compared to $^{12}$C quite a bit more complicated, even in the $0^+$ channel alone. This stems from the fact that while “knocking loose” one $\alpha$-particle in $^{12}$C, necessarily the other two are also almost free ($^8$Be) and all three $\alpha$’s form the gas state. However, exciting one $\alpha$-particle out of the ground state in $^{16}$O, may leave the remaining $^{12}$C core in the ground state or in various excited states of the shell model type. Therefore, in $^{16}$O we need to “knock loose” at least two $\alpha$’s to obtain the $\alpha$ gas state.
The $0^+$ spectrum of $^{16}$O has, in the past, very well been reproduced up to about $13$ MeV excitation energy, including the ground state, with a semi-microscopic cluster model, i.e. the $\alpha + ^{12}$C OCM (Orthogonality Condition Model) [@Suz76]. In particular, this model calculation, as well as that of an $\alpha+^{12}$C Generator-Coordinate-Method one [@baye2], demonstrates that the $0_2^+$ state at $6.05$ MeV and the $0_3^+$ state at $12.05$ MeV have $\alpha + ^{12}$C structures [@Hor68] where the $\alpha$-particle orbits around the $^{12}$C$(0_1^+)$-core in an $S$-wave and around the $^{12}$C$(2_1^+)$-core in a $D$-wave, respectively. Consistent results were later obtained by the $4\alpha$ OCM calculation within the harmonic oscillator basis [@Kat92]. However, the model space adopted in Refs. [@Suz76; @baye2; @Kat92] is not sufficient to account simultaneously for the $\alpha+ ^{12}$C and the $4\alpha$ gas-like configurations. On the other hand, the $4\alpha$-particle condensate state was first investigated in Ref. [@thsr] and its existence was predicted around the $4\alpha$ threshold with the $\alpha$-particle condensate wave function. While this so-called THSR wave function can well describe the dilute $\alpha$ cluster states as well as shell model like ground states, other structures such as $\alpha + ^{12}$C clustering are smeared out and only incorporated in an average way. Since there exists no calculation, so far, which reproduces both the $4\alpha$ gas and $\alpha+^{12}$C cluster structures simultaneously, it is crucial to perform an extended calculation for the simultaneous reproduction of both kinds of structures, which will give a decisive benchmark for the existence of the $4\alpha$-particle condensate state from a theoretical point of view.
Therefore the objective in Ref. [@funaki_4aocm] was to explore the $4\alpha$ condensate state by solving a full OCM four-body equation of motion without any assumption with respect to the structure of the $4\alpha$ system. Here we take the $4\alpha$ OCM with Gaussian basis functions, the model space of which is large enough to cover the $4\alpha$ gas, the $\alpha +^{12}$C cluster, as well as the shell-model configurations. The OCM is extensively described in Ref. [@saitoh]. Many successful applications of OCM are reported in Ref. [@carbon]. The $4\alpha$ OCM Hamiltonian is given as follows:
$${\cal H}=\sum_{i}^{4}T_i - T_{\rm cm}+ \sum_{i<j}^4
\Big[ V_{2\alpha}^{({\rm N})}(i,j)+V^{({\rm C})}_{2\alpha}(i,j)
+ V_{2\alpha}^{({\rm P})}(i,j) \Big] +\sum_{i<j<k}^4
V_{3\alpha}(i,j,k)+ V_{4\alpha}(1,2,3,4), \label{eq:hamil}$$
where $T_i$, $V_{2\alpha}^{({\rm N})}(i,j)$, $V_{2\alpha}^{({\rm C})}(i,j)$, $V_{3\alpha}(i,j,k)$ and $V_{4\alpha}(1,2,3,4)$ stand for the operators of kinetic energy for the $i$-th $\alpha$ particle, two-body, Coulomb, three-body and four-body forces between $\alpha$ particles, respectively. The center-of- mass kinetic energy $T_{\rm cm}$ is subtracted from the Hamiltonian. $V_{2\alpha}^{({\rm P})}(i,j)$ is the Pauli exclusion operator [@kukulin], by which the Pauli forbidden states between two $\alpha$-particles in $0S$, $0D$ and $1S$ states are eliminated, so that the ground state with the shell-model-like configuration can be described correctly. The effective $\alpha$-$\alpha$ interaction $V_{2\alpha}^{\rm (N)}$ is constructed by the folding procedure from two kinds of effective two-nucleon forces. One is the Modified Hasegawa-Nagata (MHN) force [@mhn] and the other is the Schmidt-Wildermuth (SW) force [@sw], see Refs. [@yamada1] and [@kato] for applications, respectively. We should note that the folded $\alpha$-$\alpha$ potentials reproduce the $\alpha$-$\alpha$ scattering phase shifts and energies of the $^8$Be ground state and of the Hoyle state. The three-body force $V_{3\alpha}$ is as in Refs. [@yamada1] and [@kato] where it was phenomenologically introduced, so as to fit the ground state energy of $^{12}$C. In addition, the phenomenological four-body force $V_{4\alpha}$ which is taken to be a Gaussian is adjusted to the ground state energy of $^{16}$O, where the range is simply chosen to be the same as that of the three-body force. The origin of the three- body and four-body forces is considered to derive from the state dependence of the effective nucleon-nucleon interaction and the additional Pauli repulsion between more than two $\alpha$-particles. However, they are short-range, and hence only act in compact configurations. The expectation values of those forces do not exceed 7 percent of the one of the corresponding two-body term, even for the ground state with the most compact structure, i.e. being the most sensitive to those forces.
Employing the Gaussian expansion method [@GEM] for the choice of variational basis functions, the total wave function $\Psi$ of the $4\alpha$ system is expanded in terms of Gaussian basis functions as follows: $$\begin{aligned}
&&\hspace{-0.7cm} \Psi(0_n^+)=\sum_{c, \nu} A^n_{c}(\nu)\Phi_{c}(\nu), \\
&&\hspace{-0.7cm} \Phi_{c}(\nu) ={\cal \widehat S} \Big[[\varphi_{l_1}
(\vec{r}_1,\nu_1)\varphi_{l_2}(\vec{r}_2,\nu_2)]_{l_{12}} \varphi_{l_3}
(\vec{r}_3,\nu_3) \Big]_{J}, \label{eq:30}\end{aligned}$$ where $\vec{r}_1$, $\vec{r}_2$ and $\vec{r}_3$ are the Jacobi coordinates describing internal motions of the $4\alpha$ system. ${\cal \widehat S}$ stands for the symmetrization operator acting on all $\alpha$ particles obeying Bose statistics. $\nu$ denotes the set of size parameters $\nu_1,
\nu_2$ and $\nu_3$ of the normalized Gaussian function, $\varphi_{l}(\vec{r},
\nu_i)=
N_{l,\nu_i}r^l\exp{(-\nu_i r^2)} Y_{l m}(\hat{\vec r})$, and $c$ the set of relative orbital angular momentum channels $[[l_1,l_2]_{l_{12}},l_3]_J$ depending on either of the coordinate type of $K$ or $H$ [@GEM], where $l_1$, $l_2$ and $l_3$ are the orbital angular momenta with respect to the corresponding Jacobi coordinates. The coefficients $A^n_{c}(\nu)$ are determined according to the Rayleigh-Ritz variational principle.
![Comparison of energy spectra between experiment and the present calculation. Two kinds of effective two-body nucleon-nucleon forces MHN and SW are adopted (see text). Dotted and dash-dotted lines denote the $\alpha + ^{12}$C and $4\alpha$ thresholds, respectively. Experimental data are taken from Ref. [@ajze], and from Ref. [@wakasa] for the $0_4^+$ state. The assignments with experiment are tentative, see, however, detailed discussion in the text.[]{data-label="fig:1"}](levels4.eps){width="7.1cm"}
Figure \[fig:1\] shows the energy spectrum with $J^\pi=0^+$, which is obtained by diagonalizing the Hamiltonian, Eq. (\[eq:hamil\]), in a model space as large as given by 5120 Gaussian basis functions, Eq. (\[eq:30\]) (the other multipolarities, needing larger basis sets, are more difficult and shall be studied in future work). It is confirmed that all levels are well converged. With the above mentioned effective $\alpha$-$\alpha$ forces, we can reproduce the full spectrum of $0^+$ states, and tentatively make a one-to-one correspondence of those states with the six lowest $0^+$ states of the experimental spectrum. In view of the complexity of the situation, the agreement is considered to be very satisfactory.
$R_{\rm exp.}$ $M({\rm E}0)_{\rm exp.}$
--------- -- ------- ------- -- ------- ---------------- -------------------------- --------------- ----------------
SW MHN SW MHN
$0_1^+$ $2.7$ $2.7$ $2.71\pm0.02$
$0_2^+$ $3.0$ $3.0$ $4.1$ $3.9$ $3.55\pm 0.21$
$0_3^+$ $2.9$ $3.1$ $2.6$ $2.4$ $4.03\pm 0.09$
$0_4^+$ $4.0$ $4.0$ $3.0$ $2.4$ no data
$0_5^+$ $3.1$ $3.1$ $3.0$ $2.6$ $3.3\pm0.7$
$0_6^+$ $5.0$ $5.6$ $0.5$ $1.0$ no data
: The rms radii $R$ and monopole transition matrix elements to the ground state $M({\rm E}0)$ in units of fm and fm$^2$, respectively. $R_{\rm exp.}$ and $M({\rm E}0)_{\rm exp.}$ are the corresponding experimental data. The finite-size effect of $\alpha$ particle is taken into account in $R$ and $M({\rm E}0)$ (see Ref. [@yamada1] for details).[]{data-label="tab:2"}
We show in Table \[tab:2\] the calculated rms radii and monopole matrix elements to the ground state, together with the corresponding experimental values. The $M({\rm E}0)$ values for the $0_2^+$ and $0_5^+$ states are consistent with the corresponding experimental values. The consistency for the $0_3^+$ state is within a factor of two. As mentioned above, the structures of the $0_2^+$ and $0_3^+$ states are well established as having the $\alpha + ^{12}$C$(0_1^+)$ and $\alpha + ^{12}$C$(2_1^+)$ cluster structures, respectively. These structures of the $0_2^+$ and $0_3^+$ states are confirmed in the present calculation. We also mention that the ground state is described as having a shell-model configuration within the present framework, the calculated rms value agreeing with the observed one ($2.71$ fm).
On the contrary, the structures of the observed $0_4^+$, $0_5^+$ and $0_6^+$ states in Fig. \[fig:1\] have, in the past, not clearly been understood, since they have never been discussed with the previous cluster model calculations [@Suz76; @baye2; @Kat92]. Although Ref. [@thsr], using the THSR wave function, predicts the $4\alpha$ condensate state around the $4\alpha$ threshold, it is not clear to which of those states it corresponds to. We will analyse the situation with the THSR wave function of [@thsr] in a future publication [@thsr_prepare].
As shown in Fig. \[fig:1\], the present calculation succeeded, for the first time, to reproduce the $0_4^+$, $0_5^+$ and $0_6^+$ states, together with the $0_1^+$, $0_2^+$ and $0_3^+$ states. This puts us in a favorable position to discuss the $4\alpha$ condensate state, expected to exist around the $4\alpha$ threshold.
In Table \[tab:2\], the largest rms value of about 5 fm is found for the $0_6^+$ state. Compared with the relatively smaller rms radii of the $0_4^+$ and $0_5^+$ states, this large size suggests that the $0_6^+$ state may be composed of a weakly interacting gas of $\alpha$ particles [@foot] of the condensate type.
![(Color online) $r{\cal Y}(r)$ defined by Eq. (\[eq:rwa\]) for the $0_6^+$ state with the MHN force.[]{data-label="fig:3_red"}](single_orbit_S1_factor.eps)
![(Color online) $r{\cal Y}(r)$ defined by Eq. (\[eq:rwa\]) for the $0_6^+$ state with the MHN force.[]{data-label="fig:3_red"}](rwao09_fbd015_mhn_a151_p8b.eps)
While a large size is generally necessary for forming an $\alpha$ condensate, the best way for its identification is to investigate the single-$\alpha$ orbit and its occupation probability, which can be obtained by diagonalizing the one-body ($\alpha$) density matrix as defined in [@takahashi; @suzuki; @yamada1; @density_matrix]. As a result of the calculation of the $L=0$ case, a large occupation probability of $61 \%$ of the lowest $0S$-orbit is found for the $0_6^+$ state, whereas the other five $0^+$ states all have appreciably smaller values, at most $25 \%$ ($0^+_2$). The corresponding single-$\alpha$ $S$ orbit is shown in Fig. \[fig:2\_occup\]. It has a strong spatially extended behaviour without any node $(0S)$. This indicates that $\alpha$ particles are condensed into the very dilute $0S$ single-$\alpha$ orbit, see also Ref. [@ropke2]. Thus, the $0^+_6$ state clearly has $4\alpha$ condensate character. We should note that the orbit is very similar to the single-$\alpha$ orbit of the Hoyle state [@suzuki; @yamada1]. We also show in Fig. \[fig:2\_occup\] the single-$\alpha$ orbit for the ground state. It has maximum amplitude at around $3$ fm and oscillations in the interior with two nodal $(2S)$ behaviour, due to the Pauli principle and reflecting the shell-model configuration.
In order to further analyze the obtained wave functions, we calculate an overlap amplitude, which is defined as follows: $${\cal Y}(r)= \Big\langle \Big[ \frac{\delta(r^\prime-r)}{r^{\prime 2}}
Y_{L}(\vec{\hat r}^\prime)\Phi_{L}(^{12}{\rm C}) \Big]_{0} \Big| \Psi(0_6^+)
\Big\rangle. \label{eq:rwa}$$ Here, $\Phi_{L}(^{12}{\rm C})$ is the wave function of $^{12}$C, given by the $3\alpha$ OCM calculation [@yamada1], and $r$ is the relative distance between the c.o.m. of $^{12}$C and the $\alpha$ particle. From this quantity we can see how large is the component in a certain $\alpha + ^{12}$C channel which is contained in our wave function (\[eq:30\]) for $0_6^+$. The amplitudes for the $0_6^+$ state are shown in Fig. \[fig:3\_red\]. It only has a large amplitude in the $\alpha + ^{12}$C$(0_2^+)$ channel, whereas the amplitudes in other channels are much suppressed. The amplitude in the Hoyle-state channel has no oscillations and a long tail stretches out to $\sim 20$ fm. This behaviour is very similar to that of the single-$\alpha$ orbit of the $0_6^+$ state discussed above.
The $\alpha$ decay width constitutes a very important information to identify the $0_6^+$ state from the experimental point of view. It can be estimated, based on the $R$-matrix theory, with the overlap amplitude Eq. (\[eq:rwa\]) [@r-matrix]. We find that the total $\alpha$ decay width of the $0_6^+$ state is as small as 50 keV (experimental value: 166 keV). This means that the state can be observed as a quasi-stable state. Thus, the width, as well as the excitation energy, are consistent with the observed data. All the characteristics found from our OCM calculation, therefore, indicate that the calculated 6th $0^+$ state with 4 alpha condensate nature can probably be identified with the experimental $0_6^+$ state at $15.1$ MeV.
Finally we discuss the structures of the $0_4^+$ and $0_5^+$ states. Our present calculations show that the $0_4^+$ and $0_5^+$ states mainly have $\alpha + ^{12}$C$(0_1^+)$ structure with higher nodal behaviour and $\alpha + ^{12}$C$(1^-)$ structure, respectively. Further details will be given in forthcoming work. The calculated width of the $0_4^+$ is $\sim 150$ keV, which is quite a bit larger than that found for the $0_5^+$ state $\sim 50$ keV. Both are qualitatively consistent with the corresponding experimental data, $600$ keV and $185$ keV, respectively. The reason why the width of the $0_4^+$ state is larger than that of the $0_5^+$ state, though the $0_4^+$ state has lower excitation energy, is due to the fact that the former has a much larger component of the $\alpha+ ^{12}$C$(0_1^+)$ decay channel, reflecting the characteristic structure of the $0_4^+$ state. The $4\alpha$ condensate state, thus, should not be assigned to the $0_4^+$ or $0_5^+$ state [@4athsr] but very likely to the $0_6^+$ state.
In conclusion, the investigation of the $0^+$-spectrum with the $4\alpha$ OCM calculation succeeded in describing the structure of the full observed $0^+$ spectrum up to the $0^+_6$ state in $^{16}$O. The $0^+$ spectrum of $^{16}$O up to about 15 MeV is now essentially understood, including the $4\alpha$ condensate state. This is remarkable improvement concerning our knowledge of the structure of $^{16}$O. We find that the $0_6^+$ state above the $4\alpha$ threshold has a very large rms radius of about $5$ fm and has a rather large occupation probability of $61 \%$ of four $\alpha$ particles sitting in a spatially extended single-$\alpha$ $0S$ orbit. The wave function has a large $\alpha + ^{12}$C amplitude only for $^{12}$C$^\ast$, i.e. the Hoyle state. These results are strong evidence of the $0_6^+$ state, which is a new theoretical prediction, for being the $4\alpha$ condensate state, i.e. the analog to the Hoyle state in $^{12}$C. Further experimental information is very much requested to confirm the novel interpretation of this state. Also independent theoretical calculations are strongly needed for confirmation of our results.
In principle, one could go on, inreasing the number of $\alpha$-particles: $^{20}$Ne, $^{24}$Mg, etc. However, one easily imagines that the complexity of the calculations quickly becomes prohibitive. In order to get a rough idea what happens for more $\alpha$-particles, drastic approximations have to be performed. One such approximation is to consider the $\alpha$-particles as ideal inert bosons and to treat them in mean field approximation. This then leads to the Gross-Pitaevskii Equation (GPE) which is widely employed in the physics of cold atoms [@stringari]. One interesting question that can be asked in this connection is: How many $\alpha$’s can maximally exist in a self-bound $\alpha$-gas state? Seeking an answer, we performed a schematic investigation using an effective $\alpha$-$\alpha$ interaction within an $\alpha$-gas mean-field calculation of the Gross-Pitaevskii type [@gross]. The parameters of the force were adjusted to reproduce our microscopic results for $^{12}$C. The corresponding $\alpha$ mean-field potential is shown in Fig. \[fig:4\]. One sees the $0S$-state lying slightly above threshold but below the Coulomb barrier. As more $\alpha$-particles are added, the Coulomb repulsion drives the loosely bound system of $\alpha$-particles farther and farther apart, so that the Coulomb barrier fades away. According to our estimate [@yamada2], a maximum of eight to ten $\alpha$-particles can be held together in a condensate. However, there may be ways to lend additional stability to such systems. We know that in the case of $^8$Be, adding one or two neutrons produces extra binding without seriously disturbing the pronounced $\alpha$-cluster structure. Therefore, one has reason to speculate that adding a few of neutrons to a many-$\alpha$ state may stabilize the condensate. But again, state-of-the-art microscopic investigations are necessary before anything definite can be said about how extra neutrons will influence an $\alpha$-particle condensate. A study in this direction is given in Ref. [@Itagaki] for $^{14}$C.
Alpha-particle condensates in finite nuclei and the alpha-particle occupation numbers
=====================================================================================
As already mentioned, constructing an $\alpha$-particle density matrix $\rho({\vec R}, {\vec R}')$ by integrating out of the total density matrix all intrinsic $\alpha$-particle coordinates and diagonalising the result, one finds that the corresponding $0S$ $\alpha$-particle orbit in $^{12}$C is occupied to more than 70 percent by the three $\alpha$-particles [@suzuki; @yamada1]. This is a huge percentage, affirming the almost ideal $\alpha$-particle condensate nature of the Hoyle state. By contrast, even at zero temperature only ten percent of the particles in superfluid $^4$He belong to the condensate (which is nevertheless a macroscopic supply of condensed particles). To add further perspective to the picture, in the [*ground state*]{} of $^{12}$C the $\alpha$-particle occupations are equally shared between $S,D$ and $G$ orbits, thus invalidating a condensate picture for the ground state. The occupation numbers for the ground and Hoyle states are shown in histogramm format in Fig. \[fig:hist\]. The difference between the Hoyle state and the ground state is seen to be spectacular. In the Hoyle state the $0S$-occupancy is at least an order of magnitude(!) higher than for any other orbit. This is one of the main typical features of Bose-Einstein condensation, even in strongly correlated Bose systems where there may be a strong depletion of the condensate, like in superfluid $^4$He. On the other hand, the ground state occupancies can be explained quite well with the standard shell model [@yamada1]. It should also be noted that the ground state of $^{12}$C is reasonably well reproduced by our theory (see Table \[tab:1\]). A further strong indication of the condensate-like behavior of the $\alpha$-particles in the Hoyle state is their momentum distribution, which is much narrower, almost delta-function-like, than in the ground state, see Fig. \[fig:mom\]. On the same line of investigation one finds, as already mentioned in section 6, that the 6-th $0^+$ state in $^{16}$O has 61% $0S$ occupancy of $\alpha$-particles.
![Radial parts of the single-$\alpha$ $S$ orbits, (a) of the Hoyle state ($^{12}$C), (b) of the $0_6^+$ state in $^{16}$O and of the ground states (c) in $^{12}$C and (d) in $^{16}$O.[]{data-label="fig:1_via"}](single_orbit_S1_12C_16O_hoyle_gd.eps)
In Fig. \[fig:1\_via\](a) and (b), we show radial parts of the single-$\alpha$ $S$ orbits of the Hoyle state [@yamada1] and the $0_6^+$ state in $^{16}$O [@funaki_4aocm], respectively. We see an almost identical shape! Of course, the extension is slightly different because of the smallness of the system. The nodeless character of the wave function is very pronounced and only some oscillations with small amplitude are present in $^{12}$C, reflecting a weak influence of the Pauli principle between the $\alpha$’s! On the contrary, we show in Fig. \[fig:1\_via\](c) and (d), radial parts of the single-$\alpha$ $S$ orbits of the ground states in $^{12}$C [@yamada1] and $^{16}$O [@funaki_4aocm], respectively. Due to its much reduced radius the “$\alpha$-like” clusters strongly overlap, producing strong amplitude oscillations which take care of antisymmetrisation between clusters. Again this example very impressively demonstrates the condensate nature of the Hoyle state and the $0_6^+$ state in $^{16}$O. It is worth noting at this point that the definition of the $\alpha$-particle density matrix, that is of a self bound system, is somewhat ambiguous and that different definitions may lead to different answers for the occupancies. This question has recently been debated in a number of papers [@jensen; @funaki4; @density_matrix]. Our conclusion concerning this point is that one should use for the definition of the internal density matrix an orthogonal systems of independent coordinates, like they are for instance given by the Jacobi coordinates. This definition insures the physically very reasonable boundary condition that, given a system exhibits an ideal Bose condensate in the laboratory system, it so remains in the internal system and both description become equivalent in the thermodynamic limit [@density_matrix]. Jacobi coordinates have been used to evaluate the occupancies of $\alpha$-particles in nuclei mentioned above [@yamada1; @suzuki; @funaki_4aocm; @takahashi].
Reduction of the $\alpha$-condensate with increasing density
============================================================
The properties of $\alpha$ matter can be used to frame the discussion of the structure of $n\,\alpha$ nuclei. As described in the preceding section, computational studies of these nuclei based on THSR cluster states have demonstrated that an $\alpha$ condensate is established at low nucleon density. More specifically, states lying near the threshold for decomposition into $\alpha$ particles, notably the ground state of $^8$Be, $^{12}$C in the $0_2^+$ Hoyle state, and corresponding states in $^{16}$O and other $n \alpha$ nuclei are [*dilute*]{}, being of low mean density and unusually extended for their mass numbers. We have shown quantitatively within a variational approach that $\alpha$-like clusters are well formed, with the pair correlation function of $\alpha$-like clusters predicting relatively large mean distances. For example, in determining the sizes of the $^{12}$C nucleus in its $0_1^+$ (ground) state and in its $0_2^+$ excited state, we obtained rms radii of 2.44 fm and 3.83 fm, respectively. The corresponding mean nucleon densities estimated from $36/4 \pi r^3_{\rm rms}$ are close to the nuclear-matter saturation density $\rho_0= 0.16$ nucleon/fm$^3$ in the former state and 0.03 nucleon/fm$^3$ in the latter. The expected low densities of putative alpha-condensate states are confirmed by experimental measurements of form factors [@funaki1].
All of our considerations indicate that quartetting is possible in the low-density regime of nucleonic matter, and that $\alpha$ condensates can survive until densities of about 0.03 nucleons/fm$^3$ are reached. Here, we are in the region where the concept of $\alpha$ matter can reasonably be applied [@JC80; @SMS06]. It is then clearly of interest to use this model to gain further insights into the formation of the condensate, and especially the reduction or suppression of the condensate due to repulsive interactions. We will show explicitly that in the model of $\alpha$ matter, as in our studies of finite nuclei, condensate formation is diminished with increasing density. Already within an $\alpha$-matter model based on a simple $\alpha - \alpha$ interaction, we can demonstrate that the condensate fraction – the fraction of particles in the condensate – is significantly reduced from unity at a density of 0.03 nucleon/fm$^3$ and essentially disappears approaching nuclear matter-saturation density.
The quantum condensate formed by a homogeneous interacting boson system at zero temperature has been investigated in the classic 1956 paper of Penrose and Onsager [@PO] who characterize the phenomenon in terms of off-diagonal long-range order of the density matrix. Here we recall some of their results that are most relevant to our problem. Asymptotically, i.e., for $|{\vec r} - {\vec r}^\prime | \sim \infty$, the nondiagonal density matrix in coordinate representation can be decomposed as $$\rho ({\vec r}, {\vec r}') \sim \psi_0^* ({\vec r}) \psi_0 ({\vec r}')
+ \gamma ({\vec r}- {\vec r}')\,.$$ In the limit, the second contribution on the right vanishes, and the first approaches the condensate fraction, formally defined by $$\rho_0 = \frac{\langle \Psi |a_0^\dagger a_0^{} | \Psi \rangle }{ \langle
\Psi | \Psi \rangle } \,.$$ Penrose and Onsager showed that in the case of a hard-core repulsion, the condensate fraction is determined by a filling factor describing the ratio of the volume occupied by the hard spheres. They applied the theory to liquid $^4$He, and found that for a hard-sphere model of the atom-atom interaction yielding a filling factor of about 28%, the condensate fraction at zero temperature is reduced from unity (its value for the noninteracting system) to around 8%. (Remarkably, but to some extent fortuitously, this estimate is in rather good agreement with current experimental and theoretical values for the condensate fraction in liquid $^4$He.)
To make a similar estimate of the condensate fraction for $\alpha$ matter, we follow Ref. [@ST] and assume an “excluded volume” for $\alpha$ particles of 20 fm$^3$. At a nucleonic density of $\rho_0/3$, this corresponds to a filling factor of about 28%, the same as for liquid $^4$He. Thus, a substantial reduction of the condensate fraction from unity (for a noninteracting $\alpha$-particle gas at zero temperature) is also expected in low-density $\alpha$ matter.
Turning to a more systematic treatment, we proceed in much the same way as Clark and coworkers [@JC80], referring especially to the most recent study with M. T. Johnson. Adopting the $\alpha-\alpha$ interaction potential $$V_{\alpha}(r) = 475\,\, e^{-(0.7 r/{\rm fm})^2} {\rm MeV} - 130\,\,
e^{-(0.475 r/{\rm fm})^2}{\rm MeV}
\label{AliBodmer}$$ introduced by Ali and Bodmer [@AB66], we calculate the reduction of the condensate fraction as function of density within what is now a rather standard variational approach. Alpha matter is described as an extended, uniform Bose system of interacting $\alpha$ particles, [*disregarding*]{} any change of the internal structure of the $\alpha$ clusters with increasing density. In particular, the dissolution of bound states associated with Pauli blocking (Mott effect) is not taken into account in the present description.
The simplest form of trial wave function incorporating the strong spatial correlations implied by the interaction potential (\[AliBodmer\]) is the familiar Jastrow choice, $$\Psi(\vec r_1, \dots, \vec r_A) = \prod_{i<j}
f(|\vec r_i - \vec r_j|)\,.$$ The normalization condition $$4 \pi \rho_\alpha \int_0^\infty [f^2(r) - 1]\,\, r^2 dr = -1\,,
\label{norm}$$ in which $\rho_{\alpha}$ is the number density of $\alpha$-particles, is imposed as a constraint on the variational wave function, in order to promote the convergence of the cluster expansion used to calculate the energy expectation value [@clark79]. In the low-density limit, the energy functional \[binding energy per $\alpha$ cluster as a functional of the correlation factor $f(r)$\] is given by $$E[f]= 2 \pi \rho_\alpha \int_0^\infty \left\{ \frac{\hbar^2}{m_\alpha}
\, \left( \frac{\partial f(r) }{\partial r} \right)^2 +f^2(r)
V_{\alpha}(r) \right\} r^2 dr \,,
\label{eev}$$ where $m_\alpha$ is the $\alpha$-particle mass, while the condensate fraction is given by $$\rho_0 = \exp \left\{-4 \pi \rho_\alpha \int_0^\infty [f(r) - 1]^2\,\,
r^2 dr \right\}\,.$$ The variational two-body correlation factor $f$ was taken as one of the forms employed by Clark and coworkers [@JC80], namely $$f(r) = (1-e^{-ar})(1+be^{-ar}+ce^{-2ar})\,.$$ At given density $\rho$, the expression for the energy expectation value is minimized with respect to the parameters $a$, $b$, and $c$, subject to the constraint (\[norm\]). It is important to note that these approximations, based on truncated cluster expansions, are reliable only at densities low enough that the length scale associated with decay of $f^2(1)-1$ is sufficiently small compared to the average particle separation, which is inversely proportional to the cubic root of the density [@JC80; @SMS06; @clark79; @Ristig].
To give an example, for the nucleon density $4 \rho_\alpha
= 0.06$ fm$^{-3}$, a minimum of the energy expectation value (\[eev\]) was found at $a=0.616$ fm$^{-1}$, $b=1.221$, and $c=-5.306$, with a corresponding energy per $\alpha$ cluster of $-9.763$ MeV and a condensate fraction of 0.750. The dependence of the condensate fraction on the nucleon density $\rho = 4 \rho_\alpha$ as determined in this exploratory calculation is displayed in Fig. \[fig:cond\_fraction\].
![Occupation of the $S$ orbital as a function of density using the $3\alpha$ OCM for $^{12}$C [@yamada1].[]{data-label="fig:cond_fraction_3alpha"}](condfrac1.eps)
![Occupation of the $S$ orbital as a function of density using the $3\alpha$ OCM for $^{12}$C [@yamada1].[]{data-label="fig:cond_fraction_3alpha"}](ocm_3alpha.eps)
The reduction of the condensate fraction of $\alpha$ matter to roughly 0.8 as given by our calculation at nucleonic density 0.03 fm$^{-3}$ agrees well with results of Suzuki [@suzuki] and Yamada [@yamada1] for $^{12}$C in the Hoyle $0^+_2$ state. Using many-particle approaches to the ground-state wave function and to the THSR ($0_2^+$) state of $^{12}$C, the occupation of the inferred natural $\alpha$ orbitals is found to be quite different in the two cases. Roughly 1/3 shares (approaching equipartition) are found for the $S$, $D$, and, $G$ orbits in the ground ($0_1^+$) state, with $\alpha$-cluster occupations of 1.07, 1.07, and 0.82, respectively. On the other hand, in the Hoyle ($0^+_2$) state, one sees enhanced occupation (2.38) of the $S$ orbit and reduced occupation (0.29, 0.16, respectively) of the $D$ and $G$ orbits. This corresponds to an enhancement of about 70% compared with equipartition.
To get a more extended analysis, OCM calculations have been performed [@yamada1] for studying the density dependence of the $S$-orbit occupancy in the Hoyle state on the different densities $\rho/\rho_0 \sim (R{(0^+_1)}_{\rm exp}/R)^3$, in which the rms radius ($R$) of $^{12}$C is taken as a parameter and $R{(0^+_1)}_{\rm exp} $=2.56 fm. A Pauli-principle respected OCM basis $\Psi^{\rm OCM}_{0^+}(\nu)$ with a size parameter $\nu$ is used, in which the value of $\nu$ is chosen to reproduce a given rms radius $R$ of $^{12}$C, and the $\alpha$ density matrix $\rho(\vec{r},\vec{r}')$ with respect to $\Psi^{\rm OCM}_{0^+}(\nu)$ is diagonalized to obtain the $S$-orbit occupancy in the $0^+$ wave function. The results are shown in Fig. \[fig:cond\_fraction\_3alpha\]. The $S$-orbit occupancy is $70\sim 80$ % around $\rho/\rho_0\sim (R{(0^+_1)}_{\rm exp}/R{(0^+_2)}_{\rm THSR})^3 = 0.21$, while it decreases with increasing $\rho/\rho_0$ and amounts to about $30\sim40$ % in the saturation density region. Figure \[fig:13a-d\] shows the radial behaviours of the $S$-orbit with given densities. A smooth transition of the $S$-orbit is observed, with decreasing $\rho/\rho_0$, from a two-node $S$-wave nature $(\rho/\rho_0\sim 1.18)$ in Fig. \[fig:13a-d\](a) to the zero-node $S$-wave one $(\rho/\rho_0\simeq0.15)$ in Fig. \[fig:13a-d\](d) [@yamada1]. The feature of the decrease of the enhanced occupation of the $S$ orbit is in striking correspondence with the density dependence of the condensate fraction calculated for nuclear matter (see Fig. \[fig:cond\_fraction\]).
![Radial behaviors of the $S$ orbit in the $^{12}$C$(0^+)$ state with (a) $R = 2.42$ fm $(\rho/\rho_0\sim1.18)$, (b) $R = 2.70$ fm $(\rho/\rho_0\sim0.85)$, (c) $R = 3.11$ fm $(\rho/\rho_0\sim0.56)$, and (d) $R = 4.84$ fm $(\rho/\rho_0\sim0.15)$, where $R$ denotes the nuclear radius of the $^{12}$C$(0^+)$ state.[]{data-label="fig:13a-d"}](fig13a-d.eps)
A more accurate and reliable variational description of $\alpha$ matter can be realized within the hypernetted-chain (HNC) approach to evaluation of correlated integrals; this approach [@JC80; @clark79] largely overcomes the limitations of the cluster-expansion treatment, including the need for an explicit normalization constraint. Such an improved approach is certainly required near the saturation density of nuclear matter, where it predicts only a small condensate fraction [@JC80]. Of course, at high densities the simple Ali-Bodmer interaction [@AB66] ceases to be valid, and it becomes crucial to include the effects of Pauli blocking. Once again, this conclusion reinforces the point that we can expect signatures of an $\alpha$ condensate only for dilute nuclei near the threshold of $n \alpha$ decay, but no signatures from configurations with saturated density.
Conclusions, Discussion, Outlook
================================
We have investigated the role that pairing and multiparticle correlations may play in nuclear matter existing in dense astrophysical objects and in finite nuclei. A complete and quantitative description of nuclear matter must allow for the presence of clusters of nucleons, bound or metastable, possibly forming a quantum condensate. In particular, quartetting correlations, responsible for the emergence of $\alpha$-like clusters, are identified as uniquely important in determining the behavior of nuclear matter in the limiting regime of low density and low temperature. We have calculated the transition temperature for the onset of quantum condensates made up of $\alpha$-like and deuteron-like bosonic clusters, and considered in considerable detail the intriguing example of Bose-Einstein condensation of $\alpha$ particles. It turns out that contrary to pairing, quartet condensation primarily exists in the BEC phase at low density. In which way quartet condensation is lost by increasing the density is still an open question. However, it is clear that there can not exist a condensate of quartets with a long coherence length for arbitrarily small attraction as this is the case for pairing in the BCS phase. It is inevitable that under increasing density or pressure, the bound $\alpha$, $d$, or other nuclidic clusters present at low density experience significant modification due to the background medium (and eventually merge with it). We have shown how self-energy corrections and Pauli blocking alter the properties of cluster states, and we have formulated a cluster mean-field approximation to provide an initial description of this process. One result of special interest is the suppression of the $\alpha$-like condensate, which is dominant at lower densities, as the density reaches and exceeds the Mott value, allowing the pairing transition to occur. Even at lower densities $\alpha$-particle condensation may be influenced by neutron excess, i.e. in the case of asymmetric nuclear matter. The study of $\alpha$-particle condensation as a function of asymmetry remains a task for the future.
A truly remarkable manifestation of $\alpha$-particle condensation seems to be present in finite nuclei. Indeed, the so-called Hoyle state ($0_2^+$) in $^{12}$C at 7.654 MeV is very likely a dilute gas of three $\alpha$-particles, held together only by the Coulomb barrier. This view is encouraged by the fact that we can explain all the experimental data in terms of a conceptually simple wave function of the quartet-condensate type. Within the same model, we also systematically predict such states in heavier $n\,\alpha$ nuclei, and the search is on for their experimental identification. In a recent study with OCM (Orthogonality Condition Model) we predicted the 6-th $0^+$ state of $^{16}$O to be a strong candidate for a loosely bound four $\alpha$-particle state [@funaki_4aocm]. The results of that study are presented and discussed here. The condensate feature of the Hoyle state in $^{12}$C and Hoyle-like state in $^{16}$O is born out by the calculation of the bosonic occupation numbers in diagonalising the bosonic density matrix. It is shown that the occupation of the $0S$ state of the $\alpha$-particles is over 70% for the Hoyle state in $^{12}$C and over 60% for the Hoyle-like state in $^{16}$O.
It is quite natural that such loosely bound $\alpha$-particle states should exist up to some maximum number of $\alpha$ particles. We estimate that the phenomenon will terminate at about eight to ten $\alpha$’s as the confining Coulomb barrier fades away. However, there is the possibility that larger condensates could be stabilized by addition of a few neutrons. Indeed, consider $^{9}$Be, which, contrary to $^{8}$Be, is bound by $\sim$ 1.5 MeV, still showing a pronounced two $\alpha$-structure similar to the one of Fig. \[profiles\] (b). One could imagine ten $\alpha$’s or more, stabilised by two or four extra neutrons in a low density phase. However, even without being stabilised, if a compressed hot nuclear blob as e.g. produced in a central Heavy Ion collision expands and cools, it may turn on its way out, at a certain low density, into an expanding $\alpha$ condensed state where all $\alpha$’s are in relative $S$-waves. This would be an analogous situation to an expanding Bose condensate of atoms, once the trapping potential has been switched off. Forthcoming analysis of dedicated experiments with high resolution multiparticle detectors like CHIMERA at LNS, Catania, will tell whether such scenarios can be realised [@Bord]. Other possibilities of loose $\alpha$-gas states may exist on top of particularly stable cores, like $^{16}$O or $^{40}$Ca. Indeed in adding $\alpha$’s to e.g. $^{40}$Ca, one will reach the $\alpha$-particle drip line. Compound states of heavy $N = Z$ nuclei of this kind may be produced in heavy ion reactions and an enhanced $\alpha$-decay rate may reveal the existence of an $\alpha$-particle condensate. Ideas of this type are presently promoted by von Oertzen [@koka; @oertzen], and also M. Brenner [@brenner], and A. Ogloblin [@ogloblin]. However, coincidence measurements of multiple $\alpha$’s of decaying lighter nuclei like $^{16}$O may also be very useful [@freer] [@zarub] to detect at least one additional $\alpha$-condensate state beyond the only one that has been identified so far, namely the $0_2^+$-state in $^{12}$C.
In finite nuclei $\alpha$-particle condensation has to be understood in the same sense as we say nuclei are superfluid inspite of the presence of only a limited number of Cooper pairs. However, in compact stars, formed by Supernova Explosions, macroscopic condensates of $\alpha$-particles may be present. In the study of reference [@ST], $\alpha$-particle phases are predicted for temperatures which can easily be below the critical temperatures obtained by our calculation displayed in Figs. \[fig:trans\_mu\] and \[fig:trans\_dens\]. For the study of such macroscopic condensates, the use of the Gross-Pitaevskii equation for interacting ideal bosons can be useful [@yamada2] but may be hampered by our poor knowledge of the density dependence of the effective $\alpha$-$\alpha$ interaction. A genuine microscopic approach for $\alpha$-particle condensation in infinite matter is demanded, since our wave function, Eqs. (\[eq:2\]) and (\[eq:3\]), only can handle a limited number of $\alpha$’s due to the necessary explicit antisymmetrisation. A preliminary route to this aim is outlined in this presentation.
Another issue which may be raised in the context of $\alpha$-particle condensation is the question, also discussed in condensed matter physics [@noz-jam], whether $\alpha$’s condense as singles or as doubles, i.e. as $^{8}$Be? In microscopic studies of $^{12}$C one, indeed, can see that in the $0_2^+$-state two of the three $\alpha$’s are slightly more closer to one another than to the third one [@feld; @enyo; @neff]. Without Coulomb repulsion $^8$Be may be bound by 2-3 MeV, however, the Coulomb repulsion makes it very slightly unbound. The question is definitely very interesting and deserves future investigation. However, quantitatively, if at all, $\alpha$-$\alpha$ correlations constitute certainly only a slight modification over the present formulation of $\alpha$-condensation. One may even ask whether these $^{8}$Be-like correlations are not an artifact of the calculation because in the RGM calculation of [@kamimura] $^{8}$Be-like correlations are definitely allowed, however, as already mentioned, our pure $\alpha$-particle condensate wave function turned out to have almost 100 percent overlap with the one of the RGM calculation.
As already outlined earlier, nuclear systems exhibit especially strong cluster and few body effects. This stems primarily from the fact that there are four different fermions with in addition more or less equal attraction among one another. However, this situation is not necessarily unique. In the past there have been speculations that in semi-conductors bi-excitons may condense rather than excitons [@nupecc]. A very promising field in this respect may be the possibility to trap in magneto-optical devices fermionic atoms in four different magnetic sub-states [@Salo]. If in addition the interaction between all four fermions could be triggered, eventually with the help of Feshbach resonances, so that they are all attractive, then investigation of quartetting could become a domain of research as rich as is presently the investigation of pairing. Clearly, the study of quartet condensation only is at its beginning and perspectives are rich and manyfold.
In conclusion, we see that the idea of $\alpha$-particle condensation in nuclei has already triggered many new ideas and calculations, in spite of the fact that, so far, a compelling case for such a state has only been made in $^{12}$C. Even so, the possible existence of a completely new nuclear phase in which $\alpha$-particles play the role of quasi-elementary constituents is surely fascinating. Hopefully, many more $\alpha$-particle states of nuclei will be detected in the near future, bringing deeper insights into the role of clustering and quantum condensates in systems of strongly interacting fermions.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank P. Nozières for his interest in quartet condensation, and we are greatful to J. W. Clark, P. Lecheminant and A. Sedrakian for discussions and contributions.
[99]{} S. L. Shapiro and S. A. Teukolsky, [*Black holes, white Dwarfs and Neutron Stars: The Physics of Compact Objects*]{} (Wiley, N.Y., 1983); D. Pines, R. Tamagaki and S. Tsuruta (eds.), [*Neutron Stars*]{} (Addison-Wesley, N.Y., 1992).
H. Shen, H. Toki, K. Oyamatsu, and K. Sumiyoshi, [*Prog. Theor. Phys.*]{} [**100**]{}, 1013 (1998).
Ch. Salomon, M. W. Zwierlein, private communications.
A. S. Stepanenko and J. M. F. Gunn, arXiv: cond-mat/9901317; B. Doucot, J. Vidal, [*Phys. Rev. Lett.*]{} [**88**]{}, 227005 (2002); S. Capponi, G. Roux, P. Lecheminant, P. Azaria, E. Boulat, S.R. White, [*Phys. Rev.*]{} [**A 77**]{}, 013624 (2008).
H. Kamei and K. Miyake, [*J. Phys. Soc. Jpn.*]{}, vol. [**74**]{}, no. [**7**]{}, 1911 (2005).
W. D. Kraeft, D. Kremp, W. Ebeling, and G. Röpke, [*Quantum Statistics of Charged Particle Systems*]{}, Berlin, Akademie-Verlag 1986. M. Beyer, S.A. Sofianos, C. Kurths, G. Röpke, and P. Schuck, [*Phys. Lett.*]{} [**B 478**]{}, 86 (2000). G. Röpke, L. Münchow, and H. Schulz, [*Nucl. Phys.*]{} [**A 379**]{}, 536 (1982). G. Röpke, A. Schnell, P. Schuck, and P. Nozières, [*Phys. Rev. Lett.*]{} [**[80]{}**]{}, 3177 (1998). R. B. Wiringa, S. C. Pieper, J. Carlson, and V. R. Pandharipande, [*Phys. Rev.*]{} [**C 62**]{}, 014001 (2000). I. Bloch, J. Dalibard and W. Zwerger, [*Rev. Mod. Phys.*]{} [**80**]{}, 885 (2008). F. Hoyle, D. N. F. Dunbar, W. A. Wenzel, W. Whaling, [*Phys. Rev.*]{} [**92**]{}, 1095 (1953). C. W. Cook, W. A. Fowler, C. C. Lauritsen, T. B. Lauritsen, [*Phys. Rev.*]{} [**107**]{}, 508 (1957). P. Navrátil, J. P. Vary, and B. R. Barrett, [*Phys. Rev. Lett.*]{} [**84**]{}, 5728 (2000); P. Navrátil, J. P. Vary, and B. R. Barrett, [*Phys. Rev.*]{} [**C 62**]{}, 054311 (2000); B. R. Barrett, B. Mihaila, S. C. Pieper, and R. B. Wiringa, [*Nucl. Phys. News*]{}, [**13**]{}, 17 (2003). Y. Fukushima and M. Kamimura, [*[Proc. Int. Conf. on Nuclear Structure]{}*]{}, Tokyo, 1977, ed. T. Marumori (Suppl. of J. Phys. Soc. Japan, [**44**]{}, 225 (1978)); M. Kamimura, [*Nucl. Phys.*]{} [**A 351**]{}, 456 (1981). E. Uegaki, S. Okabe, Y. Abe, and H. Tanaka, [*Prog. Theor. Phys.*]{} [**[57]{}**]{}, 1262 (1977); E. Uegaki, Y. Abe, S. Okabe, and H. Tanaka, [*Prog. Theor. Phys.*]{} [**59**]{}, 1031 (1978); [**[62]{}**]{}, 1621 (1979). H. Horiuchi, [*Prog. Theor. Phys.*]{} [**51**]{}, 1266 (1974); [**[53]{}**]{}, 447 (1975). S. Saito, [*Prog. Theor. Phys.*]{} [**40**]{} (1968); [**41**]{}, 705 (1969); [*Prog. Theor. Phys. Suppl.*]{} [**62**]{}, 11 (1977). H. Morinaga, [*Phys. Rev.*]{} [**[101]{}**]{}, 254 (1956); [*Phys. Lett.*]{} [**[21]{}**]{}, 78 (1966).
A. Tohsaki, H. Horiuchi, P. Schuck, and G. Röpke, [*Phys. Rev. Lett.* ]{} [**[87]{}**]{}, 192501 (2001). Y. Funaki, A. Tohsaki, H. Horiuchi, P. Schuck, and G. Röpke, [*Phys. Rev.*]{} [**[ C 67]{}**]{}, 051306(R) (2003). A very early attempt to consider for $^{12}$C a three $\alpha$-wavefunction of somewhat similar kind as in Ref. [@thsr], but without the condensate aspect and no RGM solution, has been undertaken in: H. Hutzelmeyer, H. H. Hackenbroich, [*Z. Phys.*]{} [**232**]{}, 356 (1970); see also discussion (p. 80) in “Clustering Phenomena in Nuclei”, K. Wildermuth, P. Kramer, editors, Vol. [**1**]{}: K. Wildermuth, Y. C. Tang “ A Unified Theory of the Nucleus”, Vieweg, 1977.
T. Yamada and P. Schuck, [*Eur. Phys. J.*]{} [**A 26**]{}, 185 (2005).
A. Tohsaki, [*Phys. Rev.*]{} [**C 49**]{}, 1814 (1994). Y. Funaki, H. Horiuchi, A. Tohsaki, P. Schuck and G. Röpke, [*Prog. Theor. Phys.*]{} [**108**]{}, 297 (2002). A. Tohsaki, H. Horiuchi, P. Schuck, and G. Röpke, [*Proc. of the 8th Int. Conf. on Clustering Aspects of Nuclear Structure and Dynamics*]{}, Nara, Japan, 2003, ed. K. Ikeda, I. Tanihata and H. Horiuchi ([*Nucl. Phys.*]{} [**A 738**]{}, 259 (2004)). P. Ring, and P. Schuck, [*The Nuclear Many-Body Problem*]{} (Springer-Verlag, Berlin, 1980).
T. Sogo et al. to be published. M. Itoh et al., in Proceedings of the 8th International Conference on Clustering Aspects of Nuclear Structure and Dynamics, Nara, Japan, 2003, edited by K. Ikeda, I. Tanihata, H. Horiuchi, [*Nucl. Phys.*]{} [**A 738**]{}, 268 (2004). Y. Funaki, A. Tohsaki, H. Horiuchi, P. Schuck and G. Röpke, [*Eur. Phys. J.*]{} [**A 24**]{}, 321 (2005). C. Kurakowa and K. Kat${\rm{\bar{o}}}$, [*Phys. Rev.*]{} [**C 71**]{}, 021301 (2005); [*Nucl. Phys.*]{} [**A 792**]{}, 87 (2007). H. Matsumura and Y. Suzuki, [*Nucl. Phys.*]{} [**A 739**]{}, 238 (2004). Y. Funaki, A. Tohsaki, H. Horiuchi, P. Schuck, and G. Roepke, [*Eur. Phys. J.*]{} [**A 28**]{}, 259 (2006). I. Sick and J. S. McCarthy, [*Nucl. Phys.*]{} **A 150**, 631 (1970); A. Nakada, Y. Torizuka and Y. Horikawa, [*Phys. Rev. Lett.*]{} **27**, 745 (1971); and 1102 (Erratum); P. Strehl and Th. H. Schucan, [*Phys. Lett.*]{} **27 B**, 641 (1968).
Y. Suzuki, [*Prog. Theor. Phys.*]{} [**55**]{}, 1751 (1976); [**56**]{}, 111 (1976). M. Libert-Heinemann, D. Baye, P. -H. Heenen, [*Nucl. Phys.*]{} [**A 339**]{}, 429 (1980). H. Horiuchi and K. Ikeda, [*Prog. Theor. Phys.*]{} [**40**]{}, 277 (1968). K. Fukatsu and K. Kat${\rm{\bar{o}}}$, [*Prog. Theor. Phys.*]{} [**87**]{}, 151 (1992).
Y. Funaki, T. Yamada, H. Horiuchi, G. Röpke, P. Schuck and A. Tohsaki, [*Phys. Rev. Lett.*]{} [**101**]{}, 082502 (2008). K. Ikeda, H. Horiuchi and S. Saito, [*Prog. Theor. Phys. Suppl.*]{} [**68**]{}, 1 (1980).
V. I. Kukulin, V. M. Krasnopol’sky, V. T. Voronchev, P. B. Sazonov, [*Nucl. Phys.*]{} [**A 417**]{}, 128 (1984). A. Hasegawa and S. Nagata, [*Prog. Theor. Phys.*]{} [**45**]{}, 1786 (1971); F. Tanabe, A. Tohsaki and R. Tamagaki, ibid. [**53**]{}, 677 (1975). E. W. Schmid, K. Wildermuth, [*Nucl. Phys.*]{} [**26**]{}, 463 (1961). M. Kamimura, [*Phys. Rev.*]{} [**A 38**]{}, 621 (1988); E. Hiyama, Y. Kino, M. Kamimura, [*Prog. Part. Nucl. Phys.*]{} [**51**]{}, 223 (2003). F. Ajzenberg-Selove, [*Nucl. Phys.*]{} [**A 460**]{}, 1 (1986). T. Wakasa [*et al.*]{}, [*Phys. Lett.*]{} [**B 653**]{}, 173 (2007). Y. Funaki [*et al*]{}., in preparation. The value of the density of this state may be too low, since the extension of states around and above threshold depend very sensitively on their precise position. Y. Suzuki and M. Takahashi, [*Phys. Rev.*]{} [**C 65**]{}, 064318, (2002). T. Yamada, Y. Funaki, H. Horiuchi, G. Röpke, P. Schuck and A. Tohsaki, arXiv:0804.1672. Y. Funaki, H. Horiuchi, G. Röpke, P. Schuck, A. Tohsaki and T. Yamada, [*Phys. Rev.*]{} [**C 77**]{}, 064312 (2008).
A. M. Lane and R. G. Thomas, [*Rev. Mod. Phys.*]{} [**30**]{}, 257 (1958). The reason why we previously [@nupecc] assigned the $0_4^+$ and $0_5^+$ states as being candidates for $\alpha$-condensation will be explained elsewhere. F. Dalfovo, S. Giorgini, L. P. Pitaevskii and S. Stringari, [*Rev. Mod. Phys.*]{} [**71**]{}, 463 (1999). L. P. Pitaevskii, [*Zh. Eksp. Teor. Fiz.*]{} [**40**]{}, 646 (1961) \[[*Sov. Phys. JETP*]{} [**13**]{}, 451 (1961)\]; E. P. Gross, [*Nuovo Cimento*]{} [**20**]{}, 454 (1961); [*J. Math. Phys.*]{} [**4**]{}, 195 (1963).
T. Yamada and P. Schuck, [*Phys. Rev.*]{} [**C 69**]{}, 024309 (2004). N. Itagaki, T. Otsuka, K. Ikeda and S. Okabe, [*Phys. Rev. Lett.*]{} [**92**]{}, 142502 (2004). N.T. Zinner and A.S. Jensen, arXiv:nucl/th0712.1191.
Y. Funaki, H. Horiuchi, G. Röpke, P. Schuck, A. Tohsaki, T. Yamada, W. von Oertzen, arXiv:0805.1526.
M. T. Johnson and J. W. Clark, [*Kinam*]{} [**2**]{}, 3 (1980) (PDF available at Faculty web page of J. W. Clark at http://wuphys.wustl.edu); see also J. W. Clark and T. P. Wang, [*Ann. Phys. (N.Y.)*]{} [**40**]{}, 127 (1966) and G. P. Mueller and J. W. Clark, [*Nucl. Phys.*]{} [**A 155**]{}, 561 (1970). A. Sedrakian, H. Müther, and P. Schuck, [*Nucl. Phys.*]{} [**A 766**]{}, 97 (2006). O. Penrose and L. Onsager, [*Phys. Rev.*]{} [**140**]{}, 576 (1956). S. Ali and A. R. Bodmer, [*Nucl. Phys.*]{} [**A 80**]{}, 99 (1966). J. W. Clark, [*Prog. Nucl. Part. Phys.*]{} [**2**]{}, 89 (1979). R. Pentförder, T. Lindenau, and M. L. Ristig, [*J. Low Temp. Phys.*]{} [**108**]{}, 245 (1997). B. Borderie, M. F. Rivet, private communication.
Tz. Kokalova, N. Itagaki, W. von Oertzen, and C. Wheldon, [*Phys. Rev. Lett.*]{} [**96**]{}, 192502 (2006). W. von Oertzen, [*Eur. Phys. J. A*]{} [**29**]{}, 133 (2006).
M. W. Brenner et al., in Proceedings of the International Conference “Clustering Phenomena in Nuclear Physics”, St. Petersburg, published in Physics of Atomic Nuclei (Yadernaya Fizika), (2000).
A. A. Ogloblin et al., Proceedings of the International Nuclear Physics Conference, Peterhof, Russia, June 28-July 2, (2005).
M. Freer et al., [*Phys. Rev. C*]{} [**70**]{}, 064311, (2004).
P. I. Zarubin, D. A. Artemenkov, G. I. Orlova, T. Cechak et al. (eds.), [*Nuclear Science and Safety in Europe*]{}, 189 (2006), Springer.
P. Nozières, D. Saint James, [*J. Physique*]{} [**43**]{}, 1133 (1982).
H. Feldmeier, T. Neff, [*Physik Journal*]{} [**4**]{}, Nr. [**1**]{}, 29, (2005), www.physik-journal.de.
Y. Kanada-En’yo, [*Prog. Theor. Phys.*]{} [**117**]{}, 655 (2007).
M. Chernykh, H. Feldmeier, T. Neff, P. von Neumann-Cosel, and A. Richter, [*Phys. Rev. Lett.*]{} [**98**]{}, 032501 (2007).
Y. Funaki, H. Horiuchi, G. Röpke, P. Schuck, A. Tohsaki and T. Yamada, [*Nucl. Phys. News*]{}, [**17**]{}(04), 11 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In recent years, various benchmark suites have been developed to evaluate the efficacy of Android security analysis tools. The choice of such benchmark suites used in tool evaluations is often based on the availability and popularity of suites and not on their characteristics and relevance. One of the reasons for such choices is the lack of information about the characteristics and relevance of benchmarks suites.
In this context, we empirically evaluated four Android specific benchmark suites: DroidBench, Ghera, ICCBench, and UBCBench. For each benchmark suite, we identified the APIs used by the suite that were discussed on Stack Overflow in the context of Android app development and measured the usage of these APIs in a sample of 227K real world apps (coverage). We also compared each pair of benchmark suites to identify the differences between them in terms of API usage. Finally, we identified security-related APIs used in real-world apps but not in any of the above benchmark suites to assess the opportunities to extend benchmark suites (gaps).
The findings in this paper can help 1) Android security analysis tool developers choose benchmark suites that are best suited to evaluate their tools (informed by coverage and pairwise comparison) and 2) Android app vulnerability benchmark creators develop and extend benchmark suites (informed by gaps).
author:
- |
Joydeep Mitra Venkatesh-Prasad Ranganath Aditya Narkar\
Kansas State University, USA\
{joydeep,rvprasad,avnarkar}@ksu.edu
bibliography:
- 'references.bib'
date: 'Created: January 29, 2019. Revised: September 19, 2019.'
title: 'BenchPress: Analyzing Android App Vulnerability Benchmark Suites'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Speed of sound waves in gases and liquids is governed by medium compressibility. There exists another type of non-dispersive waves which speed depends on stress instead of medium elasticity. A well-known example is the Alfven wave propagating, with a speed determined by a magnetic tension, in plasma permeated by a magnetic field. Later, an elastic analog of the Alfven waves has been predicted in a flow of dilute polymer solution, where elastic stress engendered by polymer stretching determines the elastic wave speed. Here, we present quantitative evidence of elastic Alfven waves observed in elastic turbulence of a viscoelastic creeping flow between two obstacles hindering a channel flow. The key finding in the experimental proof is a nonlinear dependence of the elastic wave speed $c_{\mathrm{el}}$ on Weissenberg number $\mathrm{Wi}$, which deviates from the prediction based on a model of linear polymer elasticity.'
author:
- 'Atul Varshney$^{1,2}$ and Victor Steinberg$^{1,3}$'
title: Elastic Alfven waves in elastic turbulence
---
A small addition of long-chain, flexible, polymer molecules strongly affects both laminar and turbulent flows of Newtonian fluid. In the former case, elastic instabilities and elastic turbulence (ET) [@larson; @shaqfeh; @groisman; @groisman1; @groisman2] are observed at Reynolds number $\mathrm{Re}\ll1$ and Weissenberg number $\mathrm{Wi}\gg1$, whereas in the latter, turbulent drag reduction (TDR) at $\mathrm{Re}\gg1$ and $\mathrm{Wi}\gg1$ has been found about 70 years ago but its mechanism is still under active investigation [@toms]. Here both $\mathrm{Re}=\rho UD/\eta$ and $\mathrm{Wi}=\lambda U/D$ are defined via the mean fluid speed $U$ and the vessel size $D$, $\rho$ and $\eta$ are the density and the dynamic viscosity of the fluid, respectively, and $\lambda$ is the longest polymer relaxation time. ET is a chaotic, inertialess flow driven solely by nonlinear elastic stress generated by polymers stretched by the flow, which is strongly modified by a feedback reaction of elastic stresses [@lebedev]. The only theory of ET based on a model of polymers with linear elasticity predicts elastic waves that are strongly attenuated in ET, but elastic waves may play a key role in modifying velocity power spectra in TDR [@lebedev; @lebedev1]. Using the Navier-Stokes equation and the equation for the elastic stresses in uniaxial form of the stress tensor approximation, one can write the polymer hydrodynamic equations in the form of the magneto-hydrodynamic (MHD) equations [@lebedev1]. Then, by analogy with the Alfven waves in MHD [@alfven; @landau1], one gets the elastic wave linear dispersion relation as $\omega=(\mathbf{k}\cdot\hat{n})[tr(\sigma_{ij})/\rho]^{1/2}$ with the elastic wave speed [@lebedev; @lebedev1] $c_\mathrm{{el}}=[tr(\sigma_{ij})/\rho]^{1/2}$, where $\omega$ and $\mathbf{k}$ are frequency and wavevector respectively, $\sigma_{ij}$ is the elastic stress tensor, and $\hat{n}$ is the major stretching direction, similar to the director in nematics. Such an evident difference between the elastic stress tensor characterized by the director and the magnetic field that is the vector, however, does not alter the similarity between the elastic and Alfven waves, since only uniaxial stretching independent of a certain direction is a necessary condition for the wave propagation determined by the stress value [@lebedev].
A simple physical explanation of both the Alfven and elastic waves can be drawn from an analogy of the response of either magnetic or elastic tension on transverse perturbations and an elastic string when plucked. As in the case of elastic string, the director is sufficient to define the alignment of the stress. Thus, to excite either Alfven or elastic waves the perturbations should be transverse to the propagation direction, unlike longitudinal sound waves in plasma, gas, and fluid media [@landau]. The detection of the elastic waves is of great importance for a further understanding of ET mechanism and TDR, where turbulent velocity power spectra get modified according to Ref. [@lebedev]. Moreover, $c_\mathrm{{el}}$ provides unique information about the elastic stresses, whereas the wave amplitude is proportional to the transversal perturbations, both of which are experimentally unavailable otherwise [@lebedev1].
Numerical simulations of a two-dimensional Kolmogorov flow of a viscoelastic fluid with periodic boundary conditions reveal filamented patterns in both velocity and stress fields of ET [@berti2]. These patterns propagate along the mean flow direction in a wavy manner with a speed $c_\mathrm{el}\simeq U/2$, nearly independent of $\mathrm{Wi}$. In subsequent studies, extensive three-dimensional Lagrangian simulations of a viscoelastic flow in a wall-bounded channel with a closely spaced array of obstacles show transition to a time-dependent flow, which resembles the elastic waves [@grilli]. Further, the elastic stress field around the obstacles demonstrates similar traveling filamental structures [@berti2; @grilli] in ET, interpreted as elastic waves [@lebedev; @lebedev1]. However, in both studies neither the linear dispersion relation nor the dependence of wave speed $c_\mathrm{el}$ on elastic stress$-$primary signatures of the elastic waves$-$were examined. Moreover, $c_\mathrm{el}$ was found to be close to the flow velocity, contradicting the theory [@lebedev; @lebedev1]. Strikingly, an indication of the elastic waves, in numerical studies, originates from observed frequency peaks in the velocity power spectra above the elastic instability [@berti2; @grilli]. Analogous frequency peaks in the power spectra of velocity and absolute pressure fluctuations above the instability were also reported in experiments of a wall-bounded channel flow in a creeping viscoelastic fluid, obstructed by either a periodic array of obstacles [@arora] or two widely-spaced cylinders [@atul; @atul1]. These observations were in agreement with numerical simulations [@ellero] and were associated with noisy cross-stream oscillations of a pair of vortices engendered due to breaking of time-reversal symmetry.
Our early attempts to excite the elastic waves both in a curvilinear flow and in an elongation flow of polymer solutions at $\mathrm{Re}\ll1$ were unsuccessful [@eldad]. In the ET regime of the curvilinear channel flow, either an excitation amplitude was insufficient and/or an excitation frequency was too high. The reason we chose the elongation flow, realized in a cross-slot micro-fluidic device, is a strong polymer stretching in a well-defined direction along the flow. However, the elongation flow generated in the cross-slot geometry has the highest elastic stresses in a central vertical plane parallel to the flow in the outlet channels$-$analogous to a stretched vertical elastic membrane. The transverse periodic perturbations in the experiment were applied in a cross-stream direction from the top wall [@eldad], however a more effective method would be to perturb it in a span-wise direction that was difficult to realize in a micro-channel. A higher frequency range of perturbations, compared to that found in the current experiment, was used that lead to the wave excitation with wave numbers in the range of high dissipation.
Here we report the first evidence of elastic waves observed in elastic turbulence of a dilute polymer solution flow in a wake between two widely-spaced obstacles, hindering a channel flow. The central finding in the experimental proof of the elastic wave observation is a power-law dependence of $c_\mathrm{el}$ on $\mathrm{Wi}$, which deviates from the prediction based on a model of linear polymer elasticity [@lebedev]. The distinctive feature of the current flow geometry is a two-dimensional nature of the ET flow, in the mid-plane of the device, in contrast to other flow geometries studied earlier.
[**Results**]{}\
[**Flow structure and elastic turbulence.**]{} The schematic of the experimental setup is shown in Fig. \[fig:setup\], where two-widely spaced obstacles hinder the channel flow of a dilute polymer solution (see Methods section for the experimental setup, solution preparation and its characterization). The main feature of the flow geometry used is the occurrence of a pair of quasi-two-dimensional counter-rotating elongated vortices, in the region between the obstacles, as a result of the elastic instability [@atul] at $\mathrm{Re}\ll1$ and $\mathrm{Wi}>1$; $\mathrm{Re}=2R\bar{u}\rho/\eta$ and $\mathrm{Wi}=\lambda\bar{u}/2R$, where obstacles’ diameter $2R$ and average flow speed $\bar{u}$ are defined in Methods section. The frequency power spectra of cross-stream velocity $v$ fluctuations show oscillatory peaks at low frequencies [@atul; @atul1] below $\lambda^{-1}$. Above the elastic instability, the main peak frequency $f_\mathrm{p}$ grows linearly with $\mathrm{Wi}$, characteristic to the Hopf bifurcation [@atul]. The two vortices form two mixing layers with a non-uniform shear velocity profile and with further increase of $\mathrm{Wi}$ their dynamics become chaotic, exhibiting ET properties, with vigorous perturbations that intermittently destroy vortices [@atul1] and seemingly excite the elastic waves. The ET flow in the region between the obstacles is shown through long-exposure particle streaks imaging in Supplementary Movies 1-3 [@sm] for three different $\mathrm{Wi}$.
![Schematic of experimental setup. It consists of a linear channel of dimension $L\times w\times h=45\times 2.5\times1~\mathrm{mm}^3$ with two cylindrical obstacles (shown as two black dots), diameter $2R=0.3~\mathrm{mm}$ and separated by a distance between the obstacles centres $e=1~\mathrm{mm}$, embedded at the center line of the channel. The polymer solution is driven by Nitrogen gas and injected through the inlet into the channel. The fluid exiting the channel outlet is weighed instantaneously as a function of time. An absolute pressure sensor, marked as $P$, after the downstream cylinder is employed to detect pressure fluctuations.[]{data-label="fig:setup"}](Fig1){width="8.5cm"}
{width="15cm"}
![Cross-stream velocity power spectra versus normalized frequency in elastic turbulence. ([**a**]{}) Cross-stream velocity power spectra $S_{f}(v)$ in log-lin coordinates to emphasize an exponential decay of the oscillation peak values at low frequencies $\lambda f\leq40$. An exponential decay is shown by the dashed line, e.g. for the case of $\mathrm{Wi}=197.5$. ([**b**]{}) $S_{ f}(v)$ for different $\mathrm{Wi}$ collapse on to each other upon normalization of $ f$ with $f_\mathrm{d}$. Inset: variation of $f_\mathrm{d}$ with $\mathrm{Wi}$. The error bars on $f_\mathrm{d}$ are estimated based on standard deviation (s.d.) of exponential fit of $S_{ f}(v)$ versus $ f$, and for $\mathrm{Wi}$ they are calculated based on the s.d. from the mean value of fluid discharge rate $Q$ (see Methods section). ([**c**]{}) $S_{ f}(v)$ in log-log coordinates, for different $\mathrm{Wi}$, to demonstrate the power-law decay at high frequencies $\sim10<\lambda f\leq100$. The spectra are obtained at ($x/R,y/R$)=($5.2,0.56$), which is close to the downstream obstacle and to the center of the upper large vortex. The dashed line in ([**c**]{}) is a fit to the data at high frequencies with the power-law exponent $\alpha_{ f}\simeq-3.4\pm0.1$, typical to the ET regime. $S_{ f}(v)$ of steady flow is shown by grey lines in ([**a**]{}) and ([**c**]{}). []{data-label="fig:spectra"}](Fig3){width="8.7cm"}
[**Characterization of low frequency oscillations.**]{} To investigate the nature of these oscillations we present time series of the streamwise $u(t)$ and cross-stream $v(t)$ velocity components and their temporal auto-correlation functions $A(u)=\langle u(t)u(t+\tau) \rangle_t/\langle |u(t)|^2 \rangle_t$ and $A(v)=\langle v(t)v(t+\tau) \rangle_t/\langle |v(t)|^2 \rangle_t$ in Fig. \[fig:elastwaves\]a-d. Distinct oscillations in $v(t)$ contrary to weak noisy oscillations in $u(t)$ indicate flow anisotropy. Further, the cross-stream velocity power spectra $S_{ f}(v)$ as a function of normalized frequency $\lambda f$ for five $\mathrm{Wi}$ values in the ET regime are shown in log-lin and log-log coordinates in Figs. \[fig:spectra\]a and b, respectively. The power spectra $S_{ f}(v)$ exhibit the oscillation peaks at low frequencies up to $\lambda f\leq40$ with an exponential decay of the peak values (Fig. \[fig:spectra\]a). These low frequency oscillations look much more pronounced on a linear scale (Supplementary Fig. 1(a) [@sm]). Further, these oscillations are also observed in the power spectra of pressure fluctuations $S(P)$ versus $\lambda f$, though not so regular (Supplementary Fig. 1(b) [@sm]). The exponential decay of $S_{ f}(v)$ at $\lambda f\leq40$ implies that only a single frequency (or time) scale is identified for each $\mathrm{Wi}$ (Fig. \[fig:spectra\]a). This frequency $f_\mathrm{d}$, for each $\mathrm{Wi}$, is obtained by an exponential fit to the data, i.e. $S_{ f}(v)\sim \exp{(- f/f_\mathrm{d})}$. The variation of $f_\mathrm{d}$ with $\mathrm{Wi}$ is shown in the inset in Fig. \[fig:spectra\]b; it varies from 0.7 to 2.5 Hz in the range of $\mathrm{Wi}$ from 75 to 200, which is comparable to oscillation peak frequency $ f_\mathrm{p}$ (Fig. \[fig:freq\]) and larger than $\lambda^{-1}$. Strikingly, on normalization of $f$ with $f_\mathrm{d}$ for each $\mathrm{Wi}$, $S_{ f}(v)$ for all $\mathrm{Wi}$ collapse on to each other (Fig. \[fig:spectra\]b). At higher frequencies up to $\lambda f\leq100$, $S_{ f}(v)$ decay as the power-law with the exponent $\alpha_{ f}=-3.4\pm0.1$ typical for ET [@groisman2] (Fig. \[fig:spectra\]c). Contrary to a general case, where the power-law decay of $S_{ f}(v)$ corresponding to ET [@groisman; @groisman1; @groisman2] commences at $\lambda f\approx 1$, the low frequency oscillations cause the power-law spectra start to decay at higher frequencies $10<\lambda f<40$, perhaps due to an additional mechanism of energy pumping into ET associated with the low frequency oscillations. In addition, $S(P)$ exhibit the power spectra decay in the high frequency range $10<\lambda f<100$ with the exponent close to -3 (see the bottom inset in Fig. 2 in Ref. [@atul1]), characteristic to the ET regime [@jun]. It should be emphasized that the power spectra of the streamwise velocity $S_{ f}(u)$ do not show the low frequency oscillations and decays with a power-law exponent $\alpha\leq 2$.
Figure \[fig:freq\] shows the dependence of $ f_\mathrm{p}$ in a wide range of $\mathrm{Wi}$. The first elastic instability, characterized as the Hopf bifurcation, occurs at low $\mathrm{Wi}$, where $ f_\mathrm{p}$ grows linearly with $\mathrm{Wi}$$-$in accord with our early results [@atul]. At higher $\mathrm{Wi}$ in the ET regime, $ f_\mathrm{p}(\mathrm{Wi})$ dependence becomes nonlinear at $\mathrm{Wi}\geq 60$. In the inset in Fig. \[fig:freq\], we present the same data for $ f_\mathrm{p}$ as a function of $\mathrm{Wi}_\mathrm{int}$. Here, the Weissenberg number of the inter-obstacle velocity field is defined as $\mathrm{Wi}_\mathrm{int}=\lambda\dot{\gamma}$ and $\dot{\gamma}(=\langle\partial u/\partial y \rangle_t)$ is the time-averaged shear-rate in the cross-stream direction in the inter-obstacle flow region. The parameter $\mathrm{Wi}_\mathrm{int}$ is relevant to the description of elastic waves in ET flow between the obstacles’ region. The inset in Fig. \[fig:speed\]b shows a linear dependence of $\mathrm{Wi}_\mathrm{int}$ on $\mathrm{Wi}$.
![Dependence of oscillation peak frequency on $\mathrm{Wi}$. Dashed line is a linear fit to the data in the regime above the elastic instability. Inset: $ f_\mathrm{p}$ as a function of $\mathrm{Wi}_\mathrm{int}$. The error bars on $ f_\mathrm{p}$ are estimated from the spectral width of the oscillatory peaks of $S_{ f}(v)$, and for $\mathrm{Wi}_\mathrm{int}$ they are calculated based on the s.d. from the mean value of ($\partial u/\partial y$).[]{data-label="fig:freq"}](Fig4){width="8.4cm"}
[**Dependence of elastic wave speed on $\mathbf{Wi}_\mathbf{int}$.**]{} Figure \[fig:speed\]a shows a family of temporal cross-correlation functions $C_v(\Delta x,\tau)= \langle v(x,t)v(x+\Delta x,t+\tau) \rangle_t/\langle v(x,t)v(x+\Delta x,t) \rangle_t$ of $v$ between two spatially separated points, with their distance being $\Delta x$, located on a horizontal line at $y/R=0.18$ for $\mathrm{Wi}=148.4$. A gaussian fit to $C_v(\Delta x,\tau)$ in the vicinity of $\tau=0$ yields the peak value $\tau_\mathrm{p}$ at a given $\Delta x$. A linear dependence of $\Delta x$ on $\tau_\mathrm{p}$ (e.g. Fig. \[fig:speed\]a inset for $\mathrm{Wi}=148.4$) provides the perturbation propagation velocity as $c_\mathrm{el}=\Delta x/\tau_\mathrm{p}$. The variation of $c_\mathrm{el}$ as a function of $\mathrm{Wi}_\mathrm{int}$ is presented in Fig. \[fig:speed\]b together with nonlinear fit of the form $c_\mathrm{el}=A(\mathrm{Wi}_\mathrm{int}-\mathrm{Wi}_\mathrm{int}^\mathrm{c})^{\beta}$, where $A=8.9\pm1.2$ mm s$^{-1}$, $\beta=0.73\pm0.12$, and onset value $\mathrm{Wi}_\mathrm{int}^\mathrm{c}=1.75\pm0.2$. The same data of $c_\mathrm{el}$ is plotted against $\mathrm{Wi}$ (see Supplementary Fig. 3 [@sm]) and fitted as $c_\mathrm{el}\sim (\mathrm{Wi}-\mathrm{Wi}_\mathrm{c})^{\beta}$ that yields the onset value $\mathrm{Wi}_\mathrm{c}=59.7\pm1.8$.
![Elastic wave speed versus $\mathrm{Wi}_\mathrm{int}$. ([**a**]{}) Cross-correlation functions of the cross-stream velocity $C_v(\Delta x,\tau)$ versus lag time $\tau$ for different values of $\Delta x$, obtained at $y/R=0.18$ and for $\mathrm{Wi}=148.4$. Inset: $\Delta x$ versus $\tau_\mathrm{p}$ for $\mathrm{Wi}=148.4$, and a slope of linear fit to it (shown by dashed line) provides $c_\mathrm{el}$. The error bars on $\Delta x$ are determined by the spatial resolution of measurements, and for $\tau_\mathrm{p}$ they are estimated based on the s.d. of gaussian fit of $C_v(\Delta x,\tau)$. ([**b**]{}) Dependence of $c_\mathrm{el}$ on $\mathrm{Wi}_\mathrm{int}$, where the dashed line is a fit of the form $c_\mathrm{el}=A(\mathrm{Wi}_\mathrm{int}-\mathrm{Wi}_\mathrm{int}^\mathrm{c})^{\beta}$, where $A=8.9\pm1.2$ mm s$^{-1}$, $\beta=0.73\pm0.12$, and onset value $\mathrm{Wi}_\mathrm{int}^\mathrm{c}=1.75\pm0.2$. Inset: $\mathrm{Wi}_\mathrm{int}$ versus $\mathrm{Wi}$. The error on $c_\mathrm{el}$ is estimated based on the s.d. of the linear fit of $\Delta x$ versus $\tau_\mathrm{p}$.[]{data-label="fig:speed"}](Fig5){width="8.6cm"}
[**[Discussion]{}**]{}
In the light of the predictions [@lebedev], it is surprising to observe the elastic waves in the ET regime due to their anticipated strong attenuation. An estimate of the wave number $k=\omega/ c_\mathrm{el}=2\pi f_\mathrm{p}/c_\mathrm{el}$ from $c_\mathrm{el}$ (Fig. \[fig:speed\]b) and $ f_\mathrm{p}$ (Fig. \[fig:freq\]) provides $k$ in the range between 0.63 and 1.3 $\mathrm{mm}^{-1}$ (Supplementary Fig. 2 [@sm]). The corresponding wavelengths ($\sim2\pi/k$) are significantly larger than the inter-obstacle spacing $e-2R=0.7$ mm. The spatial velocity power spectra $S_{k}$ is limited by a size of the observation window of about 0.7 mm that gives $k_x\approx 9$ mm$^{-1}$, much larger than the wave numbers calculated above. Thus, the low $k_x$ part of $S_{k}(v)$, where the elastic wave peaks can be anticipated, is not resolved by the spatial velocity spectra (Supplementary Fig. 4(b) [@sm]). The power-law decay with $\alpha_k\approx-3.3$ is found at low $k_x$ followed by a bottleneck part and a consequent gradual power-law decay with an exponent $\sim -0.5$ at higher $k_x$ (Supplementary Fig. 4(b) [@sm]), unlike $S_{ f}(v)$, where the peaks appear at low $ f$ and the steep power-law decay with the exponent $\alpha_{ f}=-3.4$ at higher $ f$ (see Fig. \[fig:spectra\]b). The spatial streamwise velocity power spectra $S_{k}(u)$, obtained at the same $\mathrm{Wi}$ and near the center line $y/R=0.01$, are similar to $S_\mathrm{k}(v)$ at low $k_x$ and decays gradually with exponent $\sim -0.3$ at higher $k_x$ (Supplementary Fig. 4(a) [@sm]).
The observed nonlinear dependence of $c_\mathrm{el}$ on $\mathrm{Wi}_\mathrm{int}$ differs from the theoretical prediction based on the Oldroyd-B model [@lebedev; @lebedev1]. The expression for the elastic wave speed in the model [@bird] gives $c_\mathrm{el}=[tr(\sigma_{ij})/\rho]^{1/2}\approx (N_{1}/\rho)^{1/2}$, where $N_{1}=2\mathrm{Wi}_\mathrm{int}^2\eta/\lambda$ is the first normal stress difference. Then one obtains $c_\mathrm{el}=(2\eta/\rho\lambda)^{1/2}\mathrm{Wi}_\mathrm{int}$. First, $c_\mathrm{el}$ is proportional to $\mathrm{Wi}_\mathrm{int}$ and second, the coefficient in the expression for the parameters used in the experiment is estimated to be $(2\eta/\rho\lambda)^{1/2}=4.5$ mm s$^{-1}$. Taking into account that the model [@lebedev; @lebedev1] and the estimate of elastic stress are based on linear polymer elasticity [@bird], whereas in experiments polymers in ET flow are stretched far beyond the linear limit [@liu], thus it is not surprising to find the quantitative discrepancies between them. Indeed, the value of the coefficient found from the fit ($8.9$ mm s$^{-1}$) and estimated theoretical value ($4.5$ mm s$^{-1}$) differ almost by a factor of two (see Fig. \[fig:speed\]b). Moreover, for the maximal value of $c_\mathrm{el}=17$ mm s$^{-1}$ (at $\mathrm{Wi}_\mathrm{int}\approx4$) obtained in the experiment, an estimate of elastic stress gives $\langle\sigma\rangle= c_\mathrm{el}^2\rho=0.37$ Pa that is lower but comparable with $\langle\sigma\rangle\approx1$ Pa obtained from the experiment on stretching of a single polymer T4DNA molecule at similar concentrations [@liu]. Thus, both the $c_\mathrm{el}$ dependence on $\mathrm{Wi}_\mathrm{int}$ and the coefficient value indicate that the Oldroyd-B model based on linear polymer elasticity cannot quantitatively describe the elastic wave speed and so the elastic stresses. Another aspect of this result is the Mach number $\mathrm{Ma}\equiv\bar{u}/c_\mathrm{el}$; the maximum value achieved in the experiment is $\mathrm{Ma}_\mathrm{max}=\bar{u}_\mathrm{max}/c_\mathrm{el}\approx 0.3$, contrast to what is claimed in [@rodd; @kenney] due to a wrong definition based on the elasticity $\mathrm{El}=\mathrm{Wi}/\mathrm{Re}$ instead of elastic stress $\sigma$ used for the estimation of $c_\mathrm{el}$ and $\mathrm{Ma}$.
We discuss two possible reasons related to the detection of the elastic waves. As indicated in the introduction, the key feature of the current geometry is a two-dimensional nature of the chaotic flow, at least in the mid-plane of the device (see Fig. 4SM in Supplemental Material of Ref. [@atul1]), that makes it analogous to a stretched elastic membrane. This flow structure is different from three-dimensional elastic turbulence in other studied flow geometries and thus may explain the failure in the earlier attempts to observe the elastic waves. Another qualitative discrepancy with the theory [@lebedev; @lebedev1] is the predicted strong attenuation of the elastic waves in ET. Below we estimate the range of the wave numbers with low attenuation for the elastic waves and compare with the observed values.
There are two mechanisms of the elastic wave attenuation, namely polymer (or elastic stress) relaxation and viscous dissipation [@lebedev; @lebedev1]. The former has scale-independent attenuation $\lambda^{-1}$, which at the weak attenuation satisfies the relation $\omega\lambda>1$, and the latter provides low attenuation [@teo] at $\eta k^2/\rho\omega<1$. The first condition leads to $ks>1$, where $s=\mathrm{Wi}_\mathrm{int}(2\eta\lambda/\rho)^{1/2}$ that provides a minimum wave number in the ET regime as $k_\mathrm{min}>s^{-1}=6.3\times 10^{-3}~\mathrm{mm}^{-1}$ for $\mathrm{Wi}_\mathrm{int}=4$. The maximum value of $k_\mathrm{max}$ follows from the second condition that gives $k\Lambda<1$ at $\Lambda=(\mathrm{Wi}_\mathrm{int})^{-1}(\eta\lambda/2\rho)^{1/2}$. Thus, one obtains in the ET regime $k_\mathrm{max}<\Lambda^{-1}=0.2~ \mathrm{mm}^{-1}$ for $\mathrm{Wi}_\mathrm{int}=4$ and therefore, the range of the wave numbers with the low attenuation is rather broad $6.3\times 10^{-3}<k<0.2~ \mathrm{mm}^{-1}$ and lies far outside of the $k$-range of $S_\mathrm{k}(u)$ and $S_\mathrm{k}(v)$ presented in Supplementary Fig. 4 [@sm], where the range of the wave numbers of the elastic waves is not resolved. However, the range of the observed wave number $0.63\leq k\leq 1.3~\mathrm{mm}^{-1}$ of the elastic waves, shown in Supplementary Fig. 2 [@sm], is sufficiently close to the estimated upper bound of $k$.
[**Methods**]{}
[**Experimental setup.**]{} The experiments are conducted in a linear channel of $L\times w \times h=45\times2.5\times 1$ $\mathrm{mm}^3$, shown schematically in Fig. \[fig:setup\]. The channel is prepared from transparent acrylic glass (PMMA). The fluid flow is hindered by two cylindrical obstacles of $2R=0.30$ mm made of stainless steel separated by a distance of $e=1$ mm and embedded at the center of the channel. Thus the geometrical parameters of the device are $2R/w=0.12$, $h/w=0.4$ and $e/2R=3.3$ (see Fig. \[fig:setup\]). The longitudinal and transverse coordinates of the channel are $x$ and $y$, respectively, with ($x,y$)=($0,0$) lies at the center of the upstream cylinder. The fluid is driven by $N_2$ gas at a pressure up to $\sim 10$ psi and is injected via an inlet into the channel.
[**Preparation and characterization of polymer solution.**]{} As a working fluid, a dilute polymer solution of high molecular weight polyacrylamide (PAAm, $M_w=18$ MDa; Polysciences) at concentration $c=80$ ppm ($c/c^*\simeq0.4$, where $c^*=200$ ppm is the overlap concentration for the polymer used [@liu2]) is prepared using a water-sucrose solvent with sucrose weight fraction of $60\%$. The solvent viscosity, $\eta_\mathrm{s}$, at $20^{\circ}\mbox{C}$ is measured to be $100~\mathrm{mPa}\cdot \mathrm{s}$ in a commercial rheometer (AR-1000; TA Instruments). An addition of the polymer to the solvent increases the solution viscosity, $\eta$, of about $30\%$. The stress-relaxation method [@liu2] is employed to obtain longest relaxation time ($\lambda$) of the solution and it yields $\lambda=10\pm 0.5$ s.
[**Flow discharge measurement.**]{} The fluid exiting the channel outlet is weighed instantaneously $W(t)$ as a function of time $t$ by a PC-interfaced balance (BA210S, Sartorius) with a sampling rate of $5~\mbox{Hz}$ and a resolution of $0.1~\mbox{mg}$. The time-averaged fluid discharge rate $\bar{Q}$ is estimated as $\overline{\Delta W/\Delta t}$. Thus, Weissenberg and Reynolds numbers are defined as $\mathrm{Wi}=\lambda\bar{u}/2R$ and $\mathrm{Re}=2R\bar{u}\rho/\eta$, respectively; here $\bar{u}=\bar{Q}/\rho wh$ and fluid density $\rho=1286$ Kg m$^{-3}$.
[**Imaging system.**]{} For flow visualisation, the solution is seeded with fluorescent particles of diameter $1~\mu m$ (Fluoresbrite YG, Polysciences). The region between the obstacles is imaged in the mid-plane [*[via]{}*]{} a microscope (Olympus IX70), illuminated uniformly with LED (Luxeon Rebel) at $447.5~\mbox{nm}$ wavelength, and two CCD cameras attached to the microscope: (i) GX1920 Prosilica with a spatial resolution $1000\times 500$ pixel at a rate of $65$ fps and (ii) a high resolution CCD camera XIMEA MC124CG with a spatial resolution $4000\times 2200$ pixel at a rate of $35$ fps, are used to acquire images with high temporal and spatial resolutions, respectively. We perform micro particle image velocimetry [@piv] ($\mu$PIV) to obtain the spatially-resolved velocity field $\mathbf{U}=(u,v)$ in the region between the cylinders. Interrogation windows of $16\times16$ pixel$^2$ ($26\times26~\mu m^2$) for high temporal resolution images and $64\times64$ pixel$^2$ ($10\times 10~\mu m^2$) for high spatial resolution images, with $50\%$ overlap are chosen to procure $\mathbf{U}$.
[10]{}
Larson, R. G. Instabilities in viscoelastic flows. , 213–263 (1992).
Shaqfeh, E. S. G. Purely [Elastic]{} [Instabilities]{} in [Viscometric]{} [Flows]{}. , 129–185 (1996).
Groisman, A. and Steinberg, V. Elastic turbulence in a polymer solution flow. , 53–55 (2000).
Groisman, A. and Steinberg, V. Efficient mixing at low [Reynolds]{} numbers using polymer additives. , 905–908 (2001).
Groisman, A. and Steinberg, V. Elastic turbulence in curvilinear flows of polymer solutions. , 29 (2004).
Toms, B. A. Some [Observation]{} on the [Flow]{} of [Linear]{} [Polymer]{} [Solutions]{} [Through]{} [Straight]{} [Tubes]{} at [Large]{} [R]{}eynolds [Numbers]{}. volume 2, 135–141, (1948).
Balkovsky, E., Fouxon, A., and Lebedev, V. Turbulence of polymer solutions. , 056301 (2001).
Fouxon, A. and Lebedev, V. Spectra of turbulence in dilute polymer solutions. , 2060–2072 (2003).
Alfv$\acute{e}$n, H. Existence of [Electromagnetic]{}-[Hydrodynamic]{} [Waves]{}. , 405–406 (1942).
Landau, L. D., Lifshitz, E. M., and Pitaevskii, L. P. . Elsevier Ltd., 2$^{nd}$ edition, (1984).
Landau, L. D. and Lifshitz, E. M. . Elsevier Ltd., 2$^{nd}$ edition, (1987).
Berti, S. and Boffetta, G. Elastic waves and transition to elastic turbulence in a two-dimensional viscoelastic [Kolmogorov]{} flow. , 036314 (2010).
Grilli, M., V$\acute{a}$zquez-Quesada, A., and Ellero, M. Transition to turbulence and mixing in a viscoelastic fluid flowing inside a channel with a periodic array of cylindrical obstacles. , 174501 (2013).
Arora, K., Sureshkumar, R., and Khomami, B. Experimental investigation of purely elastic instabilities in periodic flows. , 209–226 (2002).
Varshney, A. and Steinberg, V. Elastic wake instabilities in a creeping flow between two obstacles. , 051301(R) (2017).
Varshney, A. and Steinberg, V. Mixing layer instability and vorticity amplification in a creeping viscoelastic flow. , 103303 (2018).
V$\acute{a}$zquez-Quesada, A. and Ellero, M. simulations of a viscoelastic flow around a periodic array of cylinders confined in a channel. , 1–8 (2012).
Afik, E. Measuring elastic properties of flow in dilute polymer solutions. c. [T]{}hesis, Weizmann Institute of Science, (2009).
.
Jun, Y. and Steinberg, V. Power and [Pressure]{} [Fluctuations]{} in [Elastic]{} [Turbulence]{} over a [Wide]{} [Range]{} of [Polymer]{} [Concentrations]{}. , 124503 (2009).
Bird, R. B., Armstrong, R. C., and Hassager, O. . Wiley-Interscience, 2$^{nd}$ edition, (1987).
Liu, Y. and Steinberg, V. Molecular sensor of elastic stress in a random flow. , 44002 (2010).
Rodd, L. E., Cooper-White, J. J., Boger, D. V., and McKinley, G. H. Role of the elasticity number in the entry flow of dilute polymer solutions in micro-fabricated contraction geometries. , 170–191 (2007).
Shi, X., Kenney, S., Chapagain, G., and Christopher, G. F. Mechanisms of onset for moderate [Mach]{} number instabilities of viscoelastic flows around confined cylinders. , 805–815 (2015).
Burghelea, T., Steinberg, V., and Diamond, P. H. Internal viscoelastic waves in a circular [Couette]{} flow of a dilute polymer solution. , 704 (2002).
Liu, Y., Jun, Y., and Steinberg, V. Concentration dependence of the longest relaxation times of dilute and semi-dilute polymer solutions. , 1069–1085 (2009).
Thielicke, W. and Stamhuis, E. -[T]{}owards user-friendly, affordable and accurate digital particle image velocimetry in [MATLAB]{}. , p.e30 (2014).
[**Acknowledgments**]{}\
We thank Guy Han and Yuri Burnishev for technical support. A.V. acknowledges support from the European Union’s Horizon 2020 research and innovation programme under the Marie Sk[ł]{}odowska-Curie grant agreement No. 754411. This work was partially supported by the Israel Science Foundation (ISF; grant \#882/15) and the Binational USA-Israel Foundation (BSF; grant \#2016145).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Genetic algorithms (GAs) emulate the process of biological evolution, in a computational setting, in order to generate good solutions to difficult search and optimisation problems. GA-based optimisers tend to be extremely robust and versatile compared to most traditional techniques used to solve optimisation problems. This paper provides a very brief introduction to GAs and outlines their utility in astronomy and astrophysics.'
address:
- '$^1$ Astrophysics, Cosmology and Gravity Centre (ACGC), Department of Astronomy, University of Cape Town, Private Bag X3, Rondebosch 7701, South Africa'
- '$^2$ South African Astronomical Observatory, PO Box 9, Observatory, 7935, South Africa'
author:
- 'Vinesh Rajpaul$^{1,2}$'
bibliography:
- 'SAIP.bib'
title: Genetic algorithms in astronomy and astrophysics
---
Introduction
============
Many interesting mathematical problems can be reformulated as global optimisation problems; the solution of systems of algebraic or even differential equations, for example, can be cast quite naturally in terms of optimisation. The same holds true for the all-important inverse problems that are ubiquitous in the physical sciences, i.e. problems where one seeks to transform experimental data into model parameters in order infer properties of the physical systems being studied (a simple example: choosing parameters to minimise a $\chi^2$-statistic when fitting a Voigt-profile to a spectral line).
The goal of a global optimisation problem is, given a so-called *cost function*[^1] $f:\Omega \subseteq {\mathbb{R}^n} \to \mathbb{R}$, to try to find a point (more generally, a set of points) $\vec x^*\in\Omega$ such that: $$\forall \vec x \in \Omega:f(\vec x) \geqslant f({ \vec x^*});$$ $f(\vec x^*)$ is called the global minimum[^2]. A local minimum, $f(\hat {\vec x})$, is defined by the condition: $$\forall \vec x \in \Omega,\exists \delta > 0:\left\| {\vec x - \hat{\vec x}} \right\| < \delta \Rightarrow f(\vec x) \geqslant f(\hat{\vec x}).$$ Whereas finding an arbitrary local minimum of a function is a relatively straightforward task, especially if one has a good “first guess” – extremely efficient techniques exist to solve such local optimisation problems – finding *global* minima is a far more challenging problem. Real-world cost functions tend to be nonlinear, discontinuous and/or hugely multimodal, and there is no foolproof approach to locating their global minima.
Most established approaches – whether deterministic, stochastic or (meta)heuristic – to solving global optimisation problems yield excellent results on a limited class of problems, but have drawbacks that tend to cripple them when faced with certain (reasonably) difficult problems. For example, they might get stuck too easily in local minima, they might be thwarted by discontinuous functions or they might be too slow to be of practical value when faced with enormous search spaces [@Charbonneau:2002b].
Evolutionary algorithms, inspired by biological evolution, are metaheuristic optimisation algorithms that tend to yield “good enough” results on a very wide range of (even extremely difficult) optimisation problems. So-called *genetic algorithms* form one of the most successful subsets, and certainly the most popular subset, of evolutionary algorithms.
Genetic algorithms: the basic idea
==================================
Genetic algorithms, or GAs for short, draw inspiration from population genetics (and, like all evolutionary algorithms, from evolutionary biology in general) and they incorporate, in a computational setting, notions such as natural selection/survival of the fittest, genetic recombination, inheritance and mutation. The first GA-based optimiser was proposed in the mid-1970s [@Holland:1975], and since then many modifications and improvements to the basic algorithm have been developed, including mechanisms without any direct biological analogues [@Haupt:2004].
In spite of the rich variety of their potential incarnations, most GAs share a basic working scheme: they start with a population of many candidate solutions (called individuals or phenotypes), associate with each solution an encoded version of the phenotype (called a chromosome, genotype or an individual’s genetic material) and also a measure of the solution’s fitness (quality). This fitness function is often simply the additive inverse of the cost function to be optimised. Then, by repeated application of “genetic operators” mainly at the genotypic level, they cause the population as a whole to increase in phenotypic fitness, i.e. they cause the solutions to evolve towards optimality.
A typical (though simplistic and by no means general or optimal) working scheme for a genetic algorithm is as follows:
1. construct a random initial population of genotypes;
2. decode the genotypes and evaluate their phenotypic fitness; if the fittest phenotype matches the user-defined target fitness (or other termination criterion), `break`, otherwise continue;
3. produce offspring by randomly selecting and recombining genetic material from the current population, favouring individuals with high phenotypic fitness;
4. introduce, with some low probability, random changes (copying errors) into the genetic material of the offspring;
5. replace low-fitness members of the old population with the offspring created in the previous step, and `goto` step 2.
The selective recombination of the population’s genetic material exploits good solutions to build even better ones, and the random mutations serve to inject entirely new and potentially favourable material into the gene pool that could not be obtained simply by recombining the genetic material of existing individuals.
Fig. \[fig:schematic\] illustrates the working scheme of a simple GA where the solutions are encoded as binary strings[^3]. It may be shown that given enough time, and subject to a few reasonable assumptions, a GA will always converge to the global optimum of a cost function[^4] [@Michalewicz:1996; @Eiben:1990].
GAs: pros and cons
==================
Relative to more conventional optimisation algorithms, GA-based optimisers offer a number of striking advantages, some of which are outlined below.
*Robustness*. GA-based optimisers can handle – with aplomb – problems with multimodal or low-contrast objective functions, multiple objectives and/or problems where the parameter spaces have a very high dimensionality [@Charbonneau:2002b].
*Simplicity*. In order to solve a given optimisation problem, most “off-the-shelf” GAs requires only a single, unambiguous measure of the quality (fitness) of candidate solutions. They do not require, for example, gradients or Hessian matrices, the computation of which might be prohibitively difficult or impossible in some problems. Moreover the ideas underpinning GAs are intuitively accessible and it is a relatively easy task to develop a working GA from scratch.
*Speed*. Apart from the intrinsically high speed with which GAs tend to explore large parameter spaces [@Michalewicz:1996], they are embarrassingly parallel: very little effort is required to transform a serial GA-implementation to a parallel implementation. Thus they are well-suited to exploiting high-performance hardware (multi-core workstations, graphics processing units, clusters etc.).
*Versatility*. A single GA-based optimiser can be expected to yield “good enough” results on a very wide class of problems – from a problem as simple as fitting a three-parameter Gaussian to some data, to one as complex as choosing a molecular configuration to minimise a Buckingham potential with hundreds of parameters – and it is easy to incorporate problem-specific knowledge into a GA-based solver. The widespread adoption of GAs in fields such as engineering, chemistry, biology and economics bears testimony to their great versatility [@Haupt:2004].
To illustrate the great robustness and versatility of a typical GA, consider the following cost function proposed by Charbonneau [@Charbonneau:1995]: $$\label{eq:Charbonneau}
f(x,y;n) = - {\left[ {16x(1 - x)y(1 - y)\sin (n\pi x)\sin (n\pi y)} \right]^2},$$ where $x,y \in [0,1]$ and $n \in \{ 2k + 1;\forall k \in {\mathbb{N}}\}$. For $n=13$, say, it may be shown that $f(x,y)$ has $169$ (in general $n^2$) local minima on its domain, only one of which is the global minimum; moreover, the minima are separated by steep walls and there is little contrast between many of the minima (see fig. \[fig:landscape\]).
![Performance of a GA-based optimiser applied to the $n=13$ case of $f(x,y;n)$; the thick lines denote median performance in $10000$ trials, and the thin lines, upper and lower $3\sigma$ limits.[]{data-label="fig:performance"}](plot1)
![Performance of a GA-based optimiser applied to the $n=13$ case of $f(x,y;n)$; the thick lines denote median performance in $10000$ trials, and the thin lines, upper and lower $3\sigma$ limits.[]{data-label="fig:performance"}](plot2)
Fig. \[fig:performance\] illustrates how a GA-based optimiser fared on the (rather challenging) $n=13$ problem: in 10000 trials, the algorithm converged to the global minimum every single time, with the minimum location determined to a median accuracy of about one part in a billion after only $\sim10^4$ function evaluations (or a fraction of a second on a modern workstation). For comparison, a blind random search would require $\sim10^{18}$ evaluations to guarantee similar accuracy!
Although this performance is impressive in its own right, it is worth emphasising that it took mere minutes to adapt an existing GA-based optimiser[^5] to solve this problem, and that the algorithm control parameters were not optimised in any way for this new problem.
An obvious question arises: *why* do GAs work as well as they do? This topic is far beyond the scope of this survey paper but suffice it to say that a universally-accepted explanation has not yet been developed. Holland’s famous *Schema Theorem* has long been touted as providing an explanation for GAs’ success [@Holland:1975], although more recently it has become apparent that this theorem provides insight only into the workings of simplistic GAs; and even then, it is not clear whether the assumptions underlying the theorem are tenable [@Syswerda:1989; @Wright:2003].
Despite all their attractive features, GAs also have their share of disadvantages (more or less in accordance with Wolpert and Macready’s famous “no free lunch” theorem [@NFL]). GAs might be called “Jacks of all problems, but masters of none”: optimising a GA’s performance on a given problem is often difficult or impossible, and in order to achieve near-optimal performance it is usually necessary to hybridise a GA with problem-specific heuristics. For example, they tend to be better at locating than at fine-tuning solutions: once a GA is in the vicinity of a global optimum, it is usually a good idea to let a local optimiser take over [@Charbonneau:2002b]. GAs can be inefficient on simple problems where the computational expense of applying the genetic operators outweighs that of evaluating the function to be optimised; conversely, on problems where each cost function evaluation is extremely expensive – for example, where each evaluation requires a long simulation to be run – a GA-based (or indeed any) forward modelling approach could be impractical.
Finally the (currently) limited theoretical understanding of GAs is regarded by some, quite understandably, as a drawback and this might explain their relatively slow uptake in the physical sciences [@Charbonneau:1995].
Applications: astronomy and astrophysics
========================================
This section presents a sample of the numerous and diverse applications that genetic algorithms have found in astronomy and astrophysics. For brevity’s sake, only one or two short but representative examples have been drawn from different subfields.
*Astrophysical dynamics*. Wahde and Donner developed a method for reliably determining the orbital parameters of interacting galaxies and applied their method to both artificial and real data [@Wahde:2001]. Their method is based on a GA that searches very efficiently through the large space of possible orbits; indeed, the authors argue that GAs are ideally suited for investigations of tidally interacting galaxies, where large multimodal spaces must be searched in order to constrain a large number of model parameters. Cantó *et al*. devised an interesting variant of the canonical GA which they applied successfully to various problems, including the challenging task of finding the orbital parameters of the planets orbiting 55 Cancri, based on radial velocity measurements of the aforesaid stellar system [@Canto:2009].
*Physical and observational cosmology*. Although Monte Carlo methods seem to predominate in cosmology, GAs have already found a number of applications in the field. To mention just a few: Nesseris and Shafieloo used GAs to reconstruct the expansion history of the universe in a model-independent manner and thence, in conjunction with the so-called *Om statistic*, they derived a null test on the cosmological constant model $\Lambda$CDM [@Nesseris:2010]; via GAs, Allanach *et al*. were able to answer some important questions related to the discrimination of SUSY-breaking models, and in particular to quantify the measurements necessary to tell different SUSY-breaking scenarios apart [@Allanach:2004]; and Bogdanos and Nesseris used GAs to analyse Type a SNe data and to extract model-independent constraints on the evolution of the dark energy equation of state [@Bogdanos:2009]. The latter authors note that as a non-parametric method, GAs provide a convenient model-independent platform for cosmological data analysis that can minimise bias due to premature choice of e.g. a dark energy model.
*Gravitational lens modelling*. Gravitational microlensing is an ideal technique for probing the galactic population of faint or dark objects such as substellar objects, stellar remnants and exoplanets. Though very successful, theoretical microlensing models tend to be complex and their associated inverse modelling problems are notoriously difficult. The author of this paper has recently been developing GAs to speed up this difficult modelling, with a view to being able to model ongoing events approximately in real time (i.e. on a timescale of minutes rather than weeks or months!) and thereby to facilitate better-informed observations and thus more useful observational data. Results of this work are expected to be published in early 2012. As another example, Liesenborgs *et al*. presented a GA-based, non-parametric technique for inferring the projected lensing-mass distributions in strongly lensed systems [@Liesenborgs:2006].
*Stellar spectrum fitting*. Performing fits to stellar spectra is a nontrivial but important undertaking; from fitted models one can infer a veritable multitude of stellar properties. Baier *et al*. were able to combine radiative transfer codes with a GA to produce an automated procedure for fitting the dust spectra of AGB stars. Their GA-based routine dramatically improved extant fits made with more traditional methods and provided a quantitative platform from which to compare different models [@Baier:2010]. In a similar vein, Mokiem *et al*. used a parallelised GA as the basis for an autonomous fitter of spectra of massive stars with stellar winds [@Mokiem:2005].
*Stellar structure modelling*. Metcalfe and Charbonneau implemented a highly-parallelised and distributed GA to determine the globally optimal parameters for stellar models. The efficient, parallel exploration of parameter space made possible by their GA-based optimisation led to some important results in the field of white dwarf astroseismology, including the unexpected resolution of a then-puzzling discrepancy between stellar evolution model and astroseismic inferences of He-layer masses in DBV white dwarfs [@Metcalfe:2000].
*Telescope scheduling*. Autonomous telescope scheduling is a difficult task that requires dynamic adjustment of numerous observational constraints whilst trying to ensure the efficient achievement of many different scientific objectives. Kubanek developed an easy-to-implement yet robust approach to a robotic telescope scheduling problem, based on a GA that seeks out Pareto-optimal solutions (telescope schedules) [@Kubanek:2010].
Conclusions
===========
This paper introduced genetic algorithms, mentioned some of their strengths (and weaknesses) and finally illustrated their utility in astronomy and astrophysics. For those who would like to learn more about GAs or other evolutionary algorithms, there are many fine books and papers on the subject: to mention just a few, Michalewicz’s book [@Michalewicz:1996] gives an excellent introduction with a theoretical leaning, and Haupt’s book provides an equally good though more “hands-on” treatment [@Haupt:2004]. Goldberg’s seminal book [@Goldberg:1989], one of the most widely-cited works in all of computer science, serves as an outstanding tutorial-style reference, and finally Charbonneau’s note [@Charbonneau:2002b] provides a straightforward discussion of how standard statistical methods can be used to construct confidence intervals for GA-estimated model parameters.
The author of this paper would welcome correspondence from anyone who would like to discuss evolutionary algorithms, perhaps with a view to applying them in their own work. The author is grateful to the University of Cape Town and the National Research Foundation for the provision of financial support, and also to the anonymous referees for their helpful comments.
References {#references .unnumbered}
==========
[^1]: Depending on the context, cost functions are also referred to as *energy functions* or *objective functions*.
[^2]: Maximisation of $f(\vec x)$ is, of course, equivalent to minimisation of $g(\vec x):=-f(\vec x)$.
[^3]: Most early GAs encoded solutions as binary strings, both for the sake of simplicity and supposed theoretical optimality; a large body of empirical evidence, however, indicates that it is preferable to work directly with floating-point representations of solutions when solving numerical optimisation problems [@Charbonneau:2002b; @Haupt:2004; @Michalewicz:1996; @Wright:1991].
[^4]: Of course this knowledge is of little practical value; of more importance is the *rate* of convergence to the global optimum, though unfortunately with GAs this rate is highly problem-dependent and difficult to estimate *a priori*.
[^5]: This GA, coded by the author, used floating-point encoding, dynamically-adjusted mutation rates and tournament-style selection of reproducing partners.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Generalizing Block and Weinberger’s characterization of amenability we introduce the notion of uniformly finite homology for a group action on a compact space and use it to give a homological characterization of topological amenability for actions. By considering the case of the natural action of $G$ on its Stone-Čech compactification we obtain a homological characterization of exactness of the group.'
address:
- 'School of Mathematics, University of Southampton, Highfield, Southampton, SO17 1SH, England'
- 'School of Mathematics, University of Southampton, Highfield, Southampton, SO17 1SH, England'
- 'Department of Mathematics, Texas A& M University, College Station, TX 77843'
- 'School of Mathematics, University of Southampton, Highfield, Southampton, SO17 1SH, England'
author:
- Jacek Brodzki
- 'Graham A. Niblo'
- 'Piotr W. Nowak'
- Nick Wright
title: A homological characterization of topological amenability
---
There are two well known homological characterizations of amenability for a countable discrete group $G$. One, given by Johnson [@Johnson], states that a group is amenable if and only if a certain cohomology class in the first bounded cohomology $H^1_b(G,\ell_0^1(G)^{**})$ vanishes, where $\ell_0^1(G)$ is the augmentation ideal. By contrast Block and Weinberger [@block-weinberger] described amenability in terms of the non-vanishing of a homology class in the $0$-dimensional uniformly finite homology of $G$, $H_0^{uf}(G,\mathbb{R})$. The relationship between these characterizations is explored in [@bnw].
Amenable actions on a compact space were extensively studied by Anantharaman-Delaroche and Renault in [@AD-R] as a generalization of amenability which is sufficiently strong for applications and yet is exhibited by almost all known groups. A group is amenable if and only if the action on a point is amenable and it is exact if and only if it acts amenably on its Stone-Čech compactification, $\beta G$, [@HR; @guentner-kaminker; @ozawa]. It is natural to consider the question of whether or not the Johnson and Block-Weinberger characterizations of amenability can be generalized to this much broader context. In particular Higson asked for such a characterization of exactness.
In [@bnnw] we showed how to generalize Johnson’s result in terms of bounded cohomology with coefficients in a specific module $N_0(G, X)^{**}$ associated to the action. In this paper we turn our attention to the Block-Weinberger theorem, studying a related module $W_0(G,X)$ (the *standard module of the action*), and define the *uniformly finite homology of the action*, ${H}^{uf}_*(G \curvearrowright X)$ as the group homology with coefficients in $W_0(G, X)^*$. The modules $N_0(G,X)^{**}$ and $W_0(G,X)^*$ should be thought of as analogues of the modules $(\ell^{\infty}(G)/{\mathbb{R}})^*$ and $\ell^{\infty}(G)$ respectively, which play a key role in the definition of the uniformly finite homology for groups. The two characterizations are intimately related, and we consider this relationship in section \[pairing\].
In the case of Block and Weinberger’s uniformly finite homology the vanishing of the $0$-dimensional homology group is equivalent to vanishing of a fundamental class $[\sum\limits_{g\in G}g]\in H_0^{uf}(G,{\mathbb{R}})$, however the homology group ${H}^{uf}_0(G \curvearrowright X)$ is rarely trivial even when the action is topologically non-amenable. Indeed if $X$ is a compactification of $G$ then the homology group is always non-zero, see Theorem \[properaction\] below. A similar phenomenon can be observed for controlled coarse homology [@nowak-spakula], which is another generalization of uniformly finite homology: only the vanishing of the fundamental class has geometric applications. Here we show that topological amenability is detected by a fundamental class ${\hbox{$[G\curvearrowright X]$}}\in {H}^{uf}_0(G \curvearrowright X)$ for the action, and we obtain a homological characterization of topological amenability generalizing the Block-Weinberger theorem, Theorem \[amenable\], which may be summarized as follows:
\[homological\] Let $G$ be a finitely generated group acting by homeomorphisms on a compact Hausdorff topological space $X$. The action of $G$ on $X$ is topologically amenable if and only if the fundamental class ${\hbox{$[G\curvearrowright X]$}}$ is non-zero in ${H}^{uf}_0(G \curvearrowright X)$.
When the space $X$ is a point, the uniformly finite homology of the action $H_n^{uf}(G\curvearrowright X)$ reduces to $H_n^{uf}(G,\mathbb{R})$, the uniformly finite homology of $G$ with real coefficients [@block-weinberger], recovering the characterization proved by Block and Weinberger.
The uniformly finite homology of an action {#sect1}
==========================================
Let $G$ be a group generated by a *finite* set $S=S^{-1}$, acting by homeomorphisms on a compact Hausdorff space $X$.
The space $C(X, \ell^1(G))$ of continuous $\ell^1(G)$ valued functions on $X$ is equipped with the $\sup-\ell^1$ norm $$\Vert \xi\Vert=\sup\limits_{x\in X}\sum_{g\in G}\vert \xi(x)(g)|.$$
The summation map on $\ell^1(G)$ induces a continuous map $\sigma:C(X, \ell^1(G))\rightarrow C(X)$, where $C(X)$ is equipped with the $\ell^\infty$ norm. The space $N_0(G, X)$ is defined to be the pre-image $\sigma^{-1}(0)$ which we identify as $C(X, \ell^1_0(G))$, while, identifying $\mathbb R$ with the constant functions on $X$ we define $W_0(G,X)$ to be the subspace $N_0(G,X)+\mathbb R = \sigma^{-1}(\mathbb R)$. Restricting $\sigma$ to the subspace $W_0(G,X)$ we can regard it as a map $W_0(G,X)\to {\mathbb{R}}$, and with this convention we may regard $\sigma$ as an element of the dual space $W_0(G,X)^*$.
Given an element $\xi\in C(X, \ell^1(G))$ we obtain a family of functions $\xi_g\in C(X)$ indexed by the elements of $G$ by setting $\xi_g(x)=\xi(x)(g)$.
In this notation, the Banach space $C(X, \ell^1(G))$ is equipped with a natural action of $G$, $$(g\cdot \xi)_h=g*\xi_{g^{-1}h},$$ for each $g,h\in G$, where $*$ denotes the translation action of $G$ on $C(X)$: $g*f(x)=f(g^{-1}x)$ for $f\in C(X)$. We note that with these actions on $C(X, \ell^1(G))$ and $C(X)$, the map $\sigma$ is equivariant which implies that $N_0, W_0$ are $G$-invariant subspaces.
We call ${W_{0}}(G,X)$, with the above action of $G$, the *standard module of the action* of $G$ on $X$.
We have the following short exact sequence of $G$-modules: $$0\longrightarrow N_0(G,X) \xrightarrow{i} W_0(G,X) \xrightarrow{\sigma} \mathbb{R} \longrightarrow 0.$$
It is also worth pointing out that when $X$ is a point we have $W_0(G,X)=\ell^1(G)$ and $N_0(G,X)=\ell^1_0(G)$. The above modules and decompositions were introduced, with a slightly different but equivalent description, in [@bnnw] for a compact $X$ and in [@douglas-nowak] in the case when $X=\beta G$, the Stone-Čech compactification of $G$.
Recall that if $V$ is a $G$-module then $V^*$ is a also a $G$-module with the action of $G$ given by $(g\psi)(\xi)=\psi(g^{-1}\xi)$ for $\psi\in V^*$ and $\xi\in V$. With this definition we introduce the notion of uniformly finite homology for a group action.
Let $G$ be a finitely generated group acting by homeomorphisms on a compact space $X$. We define the *uniformly finite homology of the action* by setting $$H_n^{uf}(G\curvearrowright X)=H_n(G,W_0(G,X)^*),$$ for every $n\ge 0$, where $H_n$ denotes group homology.
A certain homology class in the uniformly finite homology of the action will be of particular importance to us.
Let $G$ act by homeomorphisms on a compact space $X$. The *fundamental class of the action*, denoted ${\hbox{$[G\curvearrowright X]$}}$, is the homology class in ${H}^{uf}_0(G \curvearrowright X)$ represented by $\sigma$.
As noted above, when $X$ is a point we have $W_0(G,X)=\ell^1(G)$, so $W_0(G,X)^*=\ell^{\infty}(G)$, ${\hbox{$[G\curvearrowright X]$}}=[\sum\limits_{g\in G}g]$, and $$H_0^{uf}(G,{\mathbb{R}})\simeq H_0(G,\ell^{\infty}(G))\simeq H^{uf}_0(G,W_0(G,\operatorname{pt})^*)={H}^{uf}_0(G \curvearrowright \operatorname{pt}).$$
Consider the dual of the short exact sequence of coefficients above: $$0\rightarrow \mathbb R^*\xrightarrow{\sigma^*} W_0(G,X)^*\rightarrow N_0(G,X)^*\rightarrow 0.$$ The map $\sigma$ is always split as a vector space map, and hence its dual $\sigma^*$ is also split. We now consider the question of when we can split the map $\sigma^*$ equivariantly. Identifying ${\mathbb{R}}^*$ with ${\mathbb{R}}$, the map $\sigma^*$ takes $1$ to $\sigma$, hence the condition that $\mu:W_0(G,X)^*\to \mathbb R$ splits $\sigma^*$ is the condition $\mu(\sigma)=1$. Hence a $G$-equivariant splitting of $\sigma^*$ can be regarded as a $G$-invariant functional $\mu\in W_0(G, X)^{**}$ such that $\mu(\sigma)=1$. But this is precisely an invariant mean for the action as described in [@bnnw definition 13], so we obtain:
Let $G$ be a group acting by homeomorphisms on a compact Hausdorff space $X$. Then the action is topologically amenable if and only if there is a $G$-equivariant splitting of the map $\sigma^*$ in the short exact sequence
$$0\rightarrow \mathbb R^*\xrightarrow{\sigma^*} W_0(G,X)^*\rightarrow N_0(G,X)\rightarrow 0.$$
Applying this lemma to the long exact sequence in group homology arising from the short exact sequence above we obtain:
\[longexactcoeffs\] If the group $G$ acts topologically amenably on the compact Hausdorff space $X$, then for each $n$ there is a short exact sequence
$$0\rightarrow H_n(G, \mathbb R)\rightarrow H_n(G, W_0(G,X)^*)\rightarrow H_n(G, N_0(G, X)^*)\rightarrow 0,$$ mapping the fundamental class $[1]\in H_0(G, {\mathbb{R}})$ to the fundamental class ${\hbox{$[G\curvearrowright X]$}}$ of the action. This gives us an isomorphism
$$H_n^{uf}(G\curvearrowright X)\cong H_n(G, \mathbb R) \oplus H_n(G, N_0(G,X)^*).$$
In Theorem \[amenable\] we characterize topological amenability in terms of the $0$-dimensional homology. In particular when the action is not topologically amenable we will show (Corollary \[decomp\]) that $H_0^{uf}(G\curvearrowright X)$ is isomorphic to $H_0(G, N_0(G,X)^*)$.
Non-vanishing elements in $H_0^{uf}(G\curvearrowright X)$
=========================================================
Unlike the Block-Weinberger case, vanishing of the fundamental class does not in general imply the vanishing of ${H}^{uf}_0(G \curvearrowright X)$.
\[properaction\] Let $X$ be a compact $G$ space containing an open $G$-invariant subspace $U$ on which $G$ acts properly. Then ${H}^{uf}_0(G \curvearrowright X)$ is non-zero. In particular ${H}^{uf}_0(G \curvearrowright \overline{G})$ is non-zero for any compactification $\overline{G}$ of $G$.
If $G$ is finite, and the action of $G$ on $X$ is trivial, then ${H}^{uf}_0(G \curvearrowright X)=W_0(G,X)^*$ which is non-zero.
Otherwise we may assume that the action of $G$ on $U$ is non-trivial, replacing $U$ with $X$ if $G$ is finite. Thus we may pick a point $x_0$ in $U$, and $x_1=g_1x_0$ in $Gx_0$ with $x_0\neq x_1$. Let $f\in C(X)$ be a positive function of norm 1, with $f(x_0)=1$ and with the support $K$ of $f$ contained in $U\setminus\{x_1\}$. By construction $x_0\notin g_1^{-1}K$.
Define $\xi\in W_0(G,X)$ by $\xi_e=f, \xi_{g_1}=-f,$ and $\xi_{g}=0$ for $g\neq e,g_1$. We note that $\xi$ is in $W_0(G,X)$ as required, indeed it is in $N_0(G,X)$, since $\sum_{g\in G}\xi_g$ is identically zero. We now form the sequence $$\xi^n=\sum_{k\in G}\phi_n(k) k\cdot \xi,\text{ where }\phi_n(k)=\max\left\{\frac{n-d(e,k)}{n},0\right\}.$$ If $\xi^n_g(x)$ is non-zero then $x$ is in $gK$ or $gg_1^{-1}K$. By properness of the action there are only finitely many $h\in G$ such that $hK$ meets $K$. Let $N$ be the number of such $h$. If $x\in hK$, then $x\in gK \cup gg_1^{-1}K$ for at most $2N$ values of $g$, hence for each $x\in X$, the set of $g$ with $\xi^n_g(x)\neq 0$ has cardinality at most $2N$. Since $|\xi^n_g(x)|\leq 2$ for each $g, n, x$ it follows that $\|\xi^n\|\leq 4N$ for all $n$.
For $s\in S$ consider $$\xi^n-s\cdot\xi^n=\sum_{g\in G}\phi_n(g) (g\cdot \xi-sg\cdot \xi)=\sum_{g\in G}(\phi_n(g)-\phi_n(s^{-1}g)) g\cdot \xi.$$ Since $|(g\cdot\xi)_h(x)|\leq 1$ for all $x$ and $|\phi_n(g)-\phi_n(s^{-1}g)|\leq \frac 1n$ it follows that $|(\xi^n-s\cdot\xi^n)_h(x)|\leq \frac 1n$ for all $h,x$. On the other hand, for a given $x$, $(\xi^n-s\cdot\xi^n)_h(x)$ is non-zero for at most $4N$ values of $h$, hence $\|\xi^n-s\cdot\xi^n\|\leq \frac {4N}{n}$. We thus have a sequence $\xi^n$ in $W_0(G,X)$ with $\|\delta \xi^n\|\to 0$. It follows that if $\zeta$ is a weak-\* limit point of $\xi^n$ in $W_0(G,X)^{**}$ then $\delta^{**}\zeta=0$, so $\zeta$ is a cocycle defining a class $[\zeta]$ in $H^0(G,W_0(G,X)^{**})$.
Let ${\mathrm{ev}}_{e,x_0}\in W_0(G,X)^*$ be the evaluation functional $\eta\mapsto \eta_e(x_0)$, and consider the homology class $[{\mathrm{ev}}_{e,x_0}]\in H^{uf}_0(G \curvearrowright X)$. We have $${\mathrm{ev}}_{e,x_0}(\xi^n)=\xi^n_e(x_0)=\phi_n(e) (e\cdot\xi)_{e}(x_0)+\phi_n(g_1^{-1}) (g_1^{-1}\cdot\xi)_e(x_0)$$ since the other terms in the sum vanish. The first term is $\phi_n(e)f(x_0)=1$, while $(g_1^{-1}\cdot\xi)_e(x_0)=(g_1^{-1}*\xi_{g_1})(x_0)=0$ since $x_0$ is not in $g_1^{-1}K$. Thus ${\mathrm{ev}}_{e,x_0}(\xi^n)=1$ for all $n$. It follows that the pairing of $[{\mathrm{ev}}_{e,x_0}]$ with $[\zeta]$ is 1, hence $[{\mathrm{ev}}_{e,x_0}]$ is a non-trivial element of $H^{uf}_0(G \curvearrowright X)$.
We remark that there is a surjection from $H^{uf}_0(G \curvearrowright X)$ onto $H_0(G,N_0(G,X)^*)$, induced by the surjection $W_0(G,X)^*\to N_0(G,X)^*$, and the non-trivial elements constructed in the proposition remain non-trivial after applying this map.
Characterizing amenability
==========================
We recall the definition of a (topologically) amenable action.
\[action\] Let $G$ be a group acting by homeomorphisms on a compact Hausdorff space. The action of $G$ on $X$ is said to be *topologically amenable* if there exists a sequence of elements $\xi^n\in W_{00}(G,X)$ such that
1. $\xi^n_g\ge0$ in $C(X)$ for every $n\in {\mathbb{N}}$ and $g\in G$,
2. $\sigma(\xi^n)=1$ for every $n$,
3. $\sup_{s\in S}\Vert \xi^n-s\cdot \xi^n\Vert\to 0$.
Universality of the Stone-Čech compactification leads to the observation that a group acts amenably on some compact space if and only if it acts amenably on $\beta G$, which is equivalent to exactness. Amenable actions on compact spaces (lying between the point and $\beta G$) form a spectrum of generalized amenability properties interpolating between amenability and exactness. We will return to this point later.
Now consider the coboundary map
$$\begin{CD}
W_0(G,X) @>\delta>>\left(\displaystyle{\bigoplus_{s\in S}}W_0(G,X)\right)_{\infty},
\end{CD}$$ where $$(\delta\xi)_s=\xi-s\cdot \xi,$$ for $\xi\in W_0(G,X)$, where the (finite) direct sum is equipped with a supremum norm. The operator $\delta$ is clearly bounded. Since $S$ is finite the dual of $\delta$ is $$\begin{CD}
W_0(G,X)^*@>\delta^*>>\left(\displaystyle{\bigoplus_{s\in S}}W_0(G,X)^*\right)_{1},
\end{CD}$$ where the direct sum is equipped with an $\ell^1$-norm and the adjoint map is given by $$\delta^*\psi=\sum_{s\in S} \psi_s-s^{-1}\cdot \psi_s.$$ The functional $\sigma$ can be used to detect amenability of the action.
\[theorem : characterization by sigma in the image\] Let $G$ be a finitely generated group acting on a compact space $X$ by homeomorphisms. The following conditions are equivalent:
1. the action of $G$ on $X$ is topologically amenable,
2. $\sigma\notin\overline{\operatorname{Image}(\delta^*)}^{\Vert \cdot \Vert}$,
3. $\sigma\notin \operatorname{Image}(\delta^*)$,
$(1)\implies (2)$. Assume first that the action is amenable. Take $\mu$ to be the weak-\* limit of a convergent subnet of $\xi_{\beta}$ as in the definition of amenable actions. Then $$\mu(\sigma)=\lim_{\beta} \sigma( \xi_{\beta}) =1,$$ and in particular $\sigma$ is not in the kernel of $\mu$. On the other hand $$|\mu(\delta^*\psi)|=\lim_{\beta} |\delta^*\psi (\xi_{\beta})|=\lim_{\beta} |\psi(\delta \xi_{\beta})|\le
\lim_{\beta}\left(\Vert \psi\Vert\ \sup_{s\in S}\Vert \xi_{\beta}-s\cdot \xi_{\beta}\Vert\right)=0,$$ for every $\psi\in \bigoplus_{s\in S} W_0(G,X)^*$. Thus $$\operatorname{Image}(\delta^*)\subseteq \ker \mu.$$ Since $\ker \mu$ is norm-closed, we conclude $$\overline{\operatorname{Image}(\delta^*)}^{\Vert \cdot \Vert}\subseteq \ker \mu.$$ Thus $\sigma \notin \overline{\operatorname{Image}(\delta*)}^{\Vert \cdot \Vert}$ and (2) follows.
$(2)\implies (3)$ is obvious.
To prove $(3)\implies (1)$ we suppose there exists a constant $D>0$ such that
()&&&D ()&&
for all $\xi$, and seek a contradiction. Consider a functional $\psi:\delta(W_0(G,X))\to {\mathbb{R}}$, defined by $$\psi(\delta\xi)=\sigma(\xi).$$ This is well defined, since $\delta: W_0(G,X)\to \bigoplus_{s\in S} W_0(G,X)$ is injective. By inequality ($\dagger$), $\psi$ is continuous on $\delta(W_0(G,X))$ and, by the Hahn-Banach theorem, we can extend it to a continuous functional $\Psi$ on $\bigoplus_{s\in S} W_0(G,X)$. By definition, for $\xi\in W_0(G,X)$ we have $$[\delta^*(\Psi)](\xi)=\Psi(\delta\xi)=\psi(\delta\xi)=\sigma(\xi),$$ hence $\sigma$ is in the image of $\delta^*$, contradicting (3).
It follows that there is no $D>0$ such that inequality ($\dagger$) holds for all $\xi\in W_0(G,X)$, hence there exists a sequence $\xi^n\in W_0(G,X)$ such that $\sigma(\xi^n)=1$ for all $n$, and $\Vert\delta \xi^n\Vert\to 0$. Since $W_{00}(G,X)$ is dense in $W_0(G,X)$, we may assume without loss of generality that $\xi^n\in W_{00}(G,X)$, and applying the standard normalization argument we deduce that the action is amenable.
We are now in the position to prove the main theorem, which is stated here in a more general form. The *reduced homology* $\overline{H}^{uf}_n(G \curvearrowright X) =\overline{H}_n(G,W_0(G,X)^*)$ in the statement is defined, as in the context of $L^2$-(co)homology, by taking the closure of the images in the chain complex.
\[amenable\] Let $G$ be a finitely generated group acting by homeomorphisms on a compact space $X$. The following conditions are equivalent
1. the action of $G$ on $X$ is topologically amenable,
2. ${\hbox{$[G\curvearrowright X]$}}\neq 0$ in $\overline{H}^{uf}_0(G \curvearrowright X)$,
3. \[vanishingclass\] ${\hbox{$[G\curvearrowright X]$}}\neq 0$ in ${H}^{uf}_0(G \curvearrowright X)$,
4. the map $(i^*)_*:H_0^{uf}(G\curvearrowright X)\to H_0(G, N_0(G,X)^*)$ is not injective,
5. the map $(i^*)_*:H_1^{uf}(G\curvearrowright X)\to H_1(G, N_0(G,X)^*)$ is surjective.
The equivalence (1)$\Longleftrightarrow$(2)$\Longleftrightarrow$(3) follows from Theorem \[theorem : characterization by sigma in the image\]. Indeed, we have $H_0(G,M)=M_G$, where $M_G$ is the coinvariant module, namely the quotient of $M$ by the module generated by elements of the form $g\cdot m-m$. Since $G$ is finitely generated it is enough to consider only sums of elements of the form $s\cdot m-m$, where $s$ are the generators. Indeed, if $g=s_1s_2\dots s_n$ for $s_i\in S$, we can write $$g\cdot m-m=\left(\sum_{i=1}^{n-1} s_i\cdot m_i-m_i\right)+s_n\cdot m-m,$$ where $m_i=(s_{i+1}\dots s_n)\cdot m$ for $i\leq n$. Hence $W_0(G,X)^*_G$ is exactly the quotient $W_0(G,X)^*$ by the image of $\delta^*$.
As in the proof of Corollary \[longexactcoeffs\] the short exact sequence of coefficients yields a long exact sequence which terminates as
$$\rightarrow H_0(G, \mathbb R^*)\xrightarrow{\sigma^*} H_0(G, W_0(G,X)^*)\xrightarrow{i^*} H_0(G, N_0(G, X)^*)\rightarrow 0,$$ and in which the fundamental class $[1]\in H_0(G, \mathbb R^*)$ maps to the class ${\hbox{$[G\curvearrowright X]$}}$. Thus ${\hbox{$[G\curvearrowright X]$}}\neq 0$ if and only if the map $\sigma^*$ is non-zero, or equivalently the kernel of $i^*$ is non-zero. Thus it follows that (3) is equivalent to (4).
Also by exactness of the sequence ${\hbox{$[G\curvearrowright X]$}}\neq 0$ if and only if $[1]$ is not in the image of the connecting map, or equivalently the connecting map is zero, and we obtain the equivalence of (3) and (5).
Combining this with Corollary \[longexactcoeffs\] we obtain:
\[decomp\] Let $G$ be a group acting by homeomorphisms on a compact Hausdorff topological space $X$.
$$H_0^{uf}(G \curvearrowright X)\cong
\begin{cases}
H_0(G, \mathbb R)\oplus H_0(G, N_0(G,X)^*) &\text{ when the action is amenable,}\\
H_0(G, N_0(G,X)^*) & \text{ when the action is not amenable.}\\
\end{cases}$$
Functoriality
=============
We return to the remark that we made earlier that the actions of $G$ on compact spaces form a spectrum, with the single point at one end of the spectrum and the Stone-Čech compatification of $G$ at the other end. We can make sense of this statement homologically as follows.
Suppose that $G$ is a finitely generated group acting by homeomorphisms on two compact spaces $X, Y$. Given a continuous, equivariant map $X \rightarrow Y$ of compact $G$-spaces we obtain induced continuous maps $f^*: C(Y,\ell^1(G))\to C(X,\ell^1(G))$ and $f^*: C(Y)\to C(X)$ defined by $f^*(\xi)=\xi\circ f$. Let $\sigma_X:C(X,\ell^1(G))\to C(X),\sigma_Y:C(Y,\ell^1(G))\to C(Y)$ denote the summation maps. Summation is compatible with the pull-backs in the sense that $\sigma_X\circ f^*=f^*\circ\sigma_Y$, hence $f^*$ restricts to maps $W_0(G,Y)\to W_0(G,X)$ and $N_0(G,Y)\to N_0(G,X)$. Note that equivariance of $f$ implies equivariance of $f^*$.
Let $\xi\in C(Y,\ell^1(G))$ we have $${{\|{f^*\xi}\|}}=\sup_{x\in X}\sum_{g\in G} |\xi_g(f(x))|\leq \sup_{y\in Y}\sum_{g\in G} |\xi_g(y)|={{\|{\xi}\|}}$$ so when $f$ is surjective, we have equality, and $f^*$ is an isometry onto its image. Dualising the restriction of $f^*$ to $W_0(G,Y)\to W_0(G,X)$ we obtain a continuous linear map which we denote by $f_*:W_0(G,X)^*\rightarrow W_0(G, Y)^*$. Equivariance of this map follows from equivariance of $f^*$.
As the map $f_*$ is equivariant, it induces a map on group homology (also denoted $f_*$): $$f_*:H^{uf}_n(G \curvearrowright X)\to H^{uf}_n(G \curvearrowright Y)$$
In the special case that $f$ is surjective, as $f^*$ is an isometry onto its image it follows that $f_*$ is surjective, so we obtain a short exact sequence of $G$-modules
$$0\rightarrow W_0(G,f)^*\rightarrow W_0(G,X)^*\xrightarrow {f_*}W_0(G, Y)^*\rightarrow 0$$ where $W_0(G,f)$ denotes the quotient space $W_0(G,X)/f^*W_0(G, Y)$.
This induces a long exact sequence in group homology from which we extract the following fragment. $$\cdots \rightarrow H^{uf}_0(G \curvearrowright X)\xrightarrow{f_*} H^{uf}_0(G \curvearrowright Y)\rightarrow 0.$$ Thus surjectivity of $f$ implies surjectivity of the map $f_*$ on homology in dimension 0.
In general, whether $f$ is surjective or not, the fundamental class ${\hbox{$[G\curvearrowright Y]$}}$ is in the image of $f_*$. Specifically we have $f_*{\hbox{$[G\curvearrowright X]$}}={\hbox{$[G\curvearrowright Y]$}}$ which follows from the identity $\sigma_X\circ f^*=f^*\circ\sigma_Y$.
It follows that if ${\hbox{$[G\curvearrowright Y]$}}$ is non-trivial then so is ${\hbox{$[G\curvearrowright X]$}}$, recovering the statement that if the action on $Y$ is topologically amenable then so is the action on $X$.
Now suppose that $X$ is an arbitrary compact space on which $G$ acts by homeomorphisms so by universality there are equivariant continuous maps
$$\beta G\rightarrow X\rightarrow {\{*\}}.$$ It follows that if $G$ is amenable then the action on $X$ is topologically amenable. On the other hand if the action on $X$ is topologically amenable then the action on $\beta G$ is also topologically amenable, hence $G$ is exact. Hence we recover two well known facts.
Consider again the general situation of a continuous $G$-map $f:X\rightarrow Y$. We have seen that topological amenability automatically transfers from $Y$ to $X$, but in general it does not transfer in the opposite direction. In order to transfer it from $X$ to $Y$ we need to place additional constraints on the map $f$.
Let $G$ be a group and $X,Y$ be compact Hausdorff topological spaces on which $G$ acts by homeomorphisms. A continuous $G$-equivariant map $f:X\rightarrow Y$ induces a $G-C(Y)$-module structure on $C(X)$ by pullback. The map $f$ is said to be *amenable* if there is a bounded $C(Y)$-linear $G$-equivariant map $\mu: C(X) \rightarrow C(Y)$ with $\mu({\bf 1}_X)={\bf 1}_Y$.
Amenability of the map $f$ implies that $f$ is surjective, hence $f^*$ is topologically injective and $\mu$ is a splitting of $f^*$.
When $G$ is the trivial group this reduces to the classical definition of an amenable map, while if $Y$ is a point then the map $X\rightarrow Y$ is amenable if and only if the action of $G$ on $X$ is co-amenable.
Let $G$ be a group and $X,Y$ be compact Hausdorff topological spaces on which $G$ acts by homeomorphisms. Let $f:X\rightarrow Y$ be an amenable continuous $G$-equivariant map. If the action of $G$ on $X$ is topologically amenable then so is the action of $G$ on $Y$.
We use the isomorphism between the space $C(X,\ell^1(G))$ and the completed injective tensor product $ C(X)\, \widehat{\otimes}_\epsilon \,\ell^1(G)$ (see, e.g., [@Treves Theorem 44.1]) to identify $W_0(G,X)$ as a subspace of $C(X)\,\widehat{\otimes}_\epsilon \,\ell^1(G)$ and $W_0(G,Y)$ as a subspace of $C(Y)\,\widehat{\otimes}_\epsilon \, \ell^1(G)$. Since $f$ is amenable we have a $G$-equivariant splitting $\mu:C(X) \rightarrow C(Y)$ of the map $f^*$, giving a map $\mu\otimes_\epsilon 1:C(X)\,\widehat{\otimes}_\epsilon \, \ell^1(G)\rightarrow C(Y)\,\widehat{\otimes}_\epsilon \, \ell^1(G)$. This restricts to a map $W_0(G, X)\rightarrow W_0(G,Y)$ since $\mu$ takes constant functions on $X$ to constant functions on $Y$.
The corresponding dual map $W_0(G,Y)^*\rightarrow W_0(G,X)^*$ induces a map on homology that, abusing notation, we will denote $\mu^*:H_0(G\curvearrowright Y)\rightarrow H_0(G\curvearrowright X)$. By construction this splits the map $f_*:H_0(G\curvearrowright X)\rightarrow H_0(G\curvearrowright Y)$, and since $\mu({\bf 1}_X )= {\bf 1}_Y$, $\mu^*({\hbox{$[G\curvearrowright Y]$}})={\hbox{$[G\curvearrowright X]$}}$. It follows that if the fundamental class ${\hbox{$[G\curvearrowright X]$}}$ is not trivial then neither is ${\hbox{$[G\curvearrowright Y]$}}$, and so topological amenability of the action on $X$ implies topological amenability for the action on $Y$ as required.
The interaction between uniformly finite homology and bounded cohomology {#pairing}
========================================================================
We conclude with some remarks concerning the interaction of the uniformly finite homology of an action and the bounded cohomology with coefficients introduced in [@bnnw]. These illuminate the special role played by the Johnson class in $H_b^1(G, N_0(G,X)^{**}$ and the fundamental class in $H_0^{uf}(G\curvearrowright X)$ and extend the results in [@bnw] which considered the special case of the action of $G$ on a point.
In [@bnnw] we showed that topological amenability of the action is encoded by triviality of an element $[J]$ in $H^1_b(G, N_0(G,X)^{**})$, which we call the Johnson class for the action. This class is the image of the class $[1]\in H^0_b(G, {\mathbb{R}})$ under the connecting map arising from the short exact sequence of coefficients
$$0\rightarrow N_0(G,X)^{**}\rightarrow W_0(G,X)^{**}\rightarrow {\mathbb{R}}\rightarrow 0$$
which is dual to the short exact sequence appearing in the proof of Theorem \[amenable\].
By applying the forgetful functor from bounded to ordinary cohomology, we obtain a pairing of $H^1_b(G, N_0(G,X)^{**})$ with $H_1(G, N_0(G,X)^{*})$, and clearly if the Johnson class $[J]$ is trivial then its pairing with any $[c]\in H_1(G,N_0^*)$ is zero.
Now suppose that every $[c]\in H_1(G,N_0(G,X)^*)$ pairs trivially with the Johnson class. Since the Johnson class $[J]$ is obtained by applying the connecting map to the generator $[1]$ of $H^0_b(G, {\mathbb{R}})={\mathbb{R}}$, pairing $[J]$ with $[c]\in H_1(G,N_0(G,X)^*)$ is the same as pairing $[1]$ with the image of $[c]$ under the connecting map in homology. As this pairing (between $H^0(G,{\mathbb{R}})=H^0_b(G,{\mathbb{R}})$ and $H_0(G,{\mathbb{R}})$) is faithful, it follows that the image of $[c]$ under the connecting map is trivial for all $[c]$, so the connecting map is zero, which we have already noted is equivalent to amenability of the action. Thus in the case when the group is non-amenable, the non-triviality of the Johnson element must be detected by the pairing.
On the other hand, we can run a similar argument in the opposite direction: if pairing ${\hbox{$[G\curvearrowright X]$}}$ with every element $[\phi]\in H^0_b(G,W_0(G,X)^
{**})$ we get zero, then since ${\hbox{$[G\curvearrowright X]$}}=(\sigma^*)_*[1]$, we have that the pairing of $(\sigma^{**})_*[\phi]\in H^0_b(G,{\mathbb{R}})$ with $[1]\in H_0(G,{\mathbb{R}})$ is trivial, whence $(\sigma^{**})_*[\phi]=0$ (again by faithfulness of the pairing). Thus, by exactness, the connecting map on cohomology is injective and the Johnson class is non-trivial. So when the action is amenable, (and hence the Johnson class is trivial), non-triviality of ${\hbox{$[G\curvearrowright X]$}}$ must be detected by the pairing.
[aaa]{}
<span style="font-variant:small-caps;">C. Anantharaman-Delaroche</span>, *J. Renault*. Amenable groupoids. Monographies de L’Enseignement Mathématique, 36. L’Enseignement Mathématique, Geneva, 2000.
<span style="font-variant:small-caps;">J. Block, S. Weinberger</span>, *Aperiodic tilings, positive scalar curvature and amenability of spaces*. J. Amer. Math. Soc. 5 (1992), no. 4, 907–918.
<span style="font-variant:small-caps;">J. Brodzki, G. A. Niblo and N. J. Wright</span> *Pairings, duality, amenability and bounded cohomology*. arXiv:1003.2584v1.
<span style="font-variant:small-caps;">J. Brodzki, G.A. Niblo, P. W. Nowak, N. Wright</span>, *Amenable actions, invariant means and bounded cohomology*. arXiv:1004.0295v1.
<span style="font-variant:small-caps;">R. G. Douglas, P. W. Nowak</span>, *Invariant expectations and vanishing of bounded cohomology for exact groups*. arXiv:1001.0718v3.
<span style="font-variant:small-caps;">E. Guentner, J. Kaminker</span>, *Exactness and the Novikov conjecture*. Topology 41 (2002), no. 2, 411–418.
<span style="font-variant:small-caps;">N. Higson, J. Roe</span>, *Amenable group actions and the Novikov conjecture*. J. Reine Angew. Math. 519 (2000), 143–153.
<span style="font-variant:small-caps;">B. E. Johnson</span>, *Cohomology of Banach Algebras*. Memoirs of the AMS Number 127, 1972, AMS, Providence, Rhode Island.
<span style="font-variant:small-caps;">P. W. Nowak, J. Špakula</span>, *Controlled coarse homology and isoperimetric inequalities*, J. Topol. 3 (2010), no. 2, 443–462.
<span style="font-variant:small-caps;">N. Ozawa</span>, *Amenable actions and exactness for discrete groups*. C. R. Acad. Sci. Paris Sr. I Math. 330 (2000), no. 8, 691–695.
<span style="font-variant:small-caps;">F. Treves</span>, Topological vector spaces, distributions and kernels, Academic Press, New York, 1967.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
=0.6 cm
[**Abstract**]{}
Based on the Sturm-Liouville eigenvalue problem, Banerjee *et al.* proposed a perturbative approach to analytically investigate the properties of the ($2+1$)-dimensional superconductor with Born-Infeld electrodynamics \[Phys. Rev. D [**87**]{}, 104001 (2013)\]. By introducing an iterative procedure, we will further improve the analytical results and the consistency with the numerical findings, and can easily extend the analytical study to the higher-dimensional superconductor with Born-Infeld electrodynamics. We observe that the higher Born-Infeld corrections make it harder for the condensation to form but do not affect the critical phenomena of the system. Our analytical results can be used to back up the numerical computations for the holographic superconductors with various condensates in Born-Infeld electrodynamics.
author:
- 'Chuyu Lai$^{1,2}$, Qiyuan Pan$^{1,2,3}$[^1], Jiliang Jing$^{1,2}$[^2] and Yongjiu Wang$^{1,2}$[^3]'
title: '[**On analytical study of holographic superconductors with Born-Infeld electrodynamics**]{}'
---
=0.8 cm
Introduction
============
As one of the most significant developments in fundamental physics in the last one decade, the anti-de Sitter/conformal field theories (AdS/CFT) correspondence [@Maldacena; @Witten; @Gubser1998] allows to describe the strongly coupled conformal field theories through a weakly coupled dual gravitational description. A recent interesting application of such a holography is constructing of a model of a high $T_{c}$ superconductor, for reviews, see Refs. [@HartnollRev; @HerzogRev; @HorowitzRev; @CaiRev] and references therein. It was found that the instability of the bulk black hole corresponds to a second order phase transition from normal state to superconducting state which brings the spontaneous U(1) symmetry breaking [@GubserPRD78], and the properties of a ($2+1$)-dimensional superconductor can indeed be reproduced in the ($3+1$)-dimensional holographic dual model based on the framework of usual Maxwell electrodynamics [@HartnollPRL101]. In order to understand the influences of the $1/N$ or $1/\lambda$ ($\lambda$ is the ’t Hooft coupling) corrections on the holographic dual models, it is of great interest to consider the holographic superconductor models with the nonlinear electrodynamics since the nonlinear electrodynamics essentially implies the higher derivative corrections of the gauge field [@HendiJHEP]. Jing and Chen introduced the first holographic superconductor model in Born-Infeld electrodynamics and observed that the nonlinear Born-Infeld corrections will make it harder for the scalar condensation to form [@JS2010]. Along this line, there have been accumulated interest to study various holographic dual models with the nonlinear electrodynamics [@JLQS2012; @JingJHEP; @PJWPRD; @SDSL2012JHEP156; @LeeEPJC; @LPW2012; @Roychowdhury; @BGQX; @JPCPLB; @ZPCJNPB; @YaoJing; @DLAP; @SGMPLA].
In most cases, the holographic dual models were studied numerically. In order to back up numerical results and gain more insights into the properties of the holographic superconductors, Siopsis *et al.* developed the variational method for the Sturm-Liouville (S-L) eigenvalue problem to analytically calculate the critical exponent near the critical temperature and found that the analytical results obtained by this way are in good agreement with the numerical findings [@Siopsis; @SiopsisBF]. Generalized to study the holographic insulator/superconductor phase transition [@Cai-Li-Zhang], this method can clearly present the condensation and critical phenomena of the system at the critical point in AdS soliton background.
More recently, Gangopadhyay and Roychowdhury extended the S-L method to investigate the properties of the ($2+1$)-dimensional superconductor with Born-Infeld electrodynamics by introducing a perturbative technique, and observed that the analytical results agree well with the existing numerical results for the condensation operator $\langle\mathcal{O}_{-}\rangle$ [@SDSL2012]. For the operator $\langle\mathcal{O}_{+}\rangle$, Banerjee *et al.* improved the perturbative approach and explored the effect of the Born-Infeld electrodynamics on the ($2+1$)-dimensional superconductor [@BGRLPRD2013]. However, comparing with the case of $\langle\mathcal{O}_{-}\rangle$ [@SDSL2012], we find that for the operator $\langle\mathcal{O}_{+}\rangle$ the agreement of the analytical result with the numerical calculation is not so good, for example in the case of the Born-Infeld parameter $b=0.3$ [@BGRLPRD2013], the difference between the analytical and numerical values is $22.1\%$! Furthermore, this perturbative approach is not very valid to study the higher-dimensional superconductor with Born-Infeld electrodynamics. Thus, the motivation for completing this work is two fold. On one level, it is worthwhile to reduce the disparity between the analytical and numerical results for the operator $\langle\mathcal{O}_{+}\rangle$, and further improve the analytical results and the consistency with the numerical findings. On another more speculative level, it would be important to develop a more general analytical technique which can be used to study systematically the $d$-dimensional superconductors with Born-Infeld electrodynamics and see some general features for the effects of the higher derivative corrections to the gauge field on the holographic dual models. In order to avoid the complex computation, in this work we will concentrate on the probe limit where the backreaction of matter fields on the spacetime metric is neglected.
The plan of the work is the following. In Sec. II we will introduce the holographic superconductor models with Born-Infeld electrodynamics in the $(d+1)$-dimensional AdS black hole background. In Sec. III we will improve the perturbative approach proposed in [@BGRLPRD2013] and give an analytical investigation of the holographic superconductors with Born-Infeld electrodynamics by using the S-L method. We will conclude in the last section with our main results.
Holographic superconductors with Born-Infeld electrodynamics
============================================================
We begin with the background of the $(d+1)$-dimensional planar Schwarzschild-AdS black hole $$ds^2=-r^{2}f(r)dt^2+\frac{dr^2}{r^2f(r)}+r^2\sum_{i=1}^{d-1}dx_i^2,$$ where $f(r)=1-r_+^{d}/r^{d}$ with the radius of the event horizon $r_{+}$. For convenience, we have set the AdS radius $L=1$. The Hawking temperature of the black hole is determined by $$\begin{aligned}
T=\frac{dr_{+}}{4\pi},\end{aligned}$$ which will be interpreted as the temperature of the CFT.
Working in the probe limit, we consider the Born-Infeld electrodynamics and the charged complex scalar field coupled via the action $$\begin{aligned}
\label{System}
S=\int
d^{d+1}x\sqrt{-g}\left[\frac{1}{b}\left(1-\sqrt{1+\frac{1}{2}bF^{2}}\right)
-|\nabla\psi-iA\psi|^{2}-m^2|\psi|^2\right],\end{aligned}$$ with the quadratic term $F^{2}=F_{\mu\nu}F^{\mu\nu}$. When the Born-Infeld parameter $b\rightarrow0$, the model (\[System\]) reduces to the standard holographic superconductors investigated in [@HartnollPRL101; @HorowitzPRD78].
With the ansatz of the matter fields as $\psi=|\psi|$, $A_{t}=\phi$ where $\psi$ and $\phi$ are both real functions of $r$ only, we can arrive at the following equations of motion for the scalar field $\psi$ and the gauge field $\phi$ $$\begin{aligned}
&&\psi^{\prime\prime}+\left(
\frac{1+d}{r}+\frac{f^\prime}{f}\right)\psi^\prime
+\left(\frac{\phi^2}{r^4f^2}-\frac{m^2}{r^2f}\right)\psi=0,
\label{BHPsir}\end{aligned}$$ $$\begin{aligned}
\phi^{\prime\prime}+\frac{d-1}{r}\left(1-b\phi^{\prime
2}\right)\phi^\prime-\frac{2\psi^{2}}{r^2f}\left(1-b\phi^{\prime
2}\right)^{3/2}\phi=0,\label{BHPhir}\end{aligned}$$ where the prime denotes the derivative with respect to $r$.
Applying the S-L method to analytically study the properties of the holographic superconductors with Born-Infeld electrodynamics, we will introduce a new variable $z=r_{+}/r$ and rewrite the equations of motion (\[BHPsir\]) and (\[BHPhir\]) into $$\begin{aligned}
&&\psi^{\prime\prime}+\left(
\frac{1-d}{z}+\frac{f^\prime}{f}\right)\psi^\prime
+\left(\frac{\phi^2}{r_{+}^2f^2}-\frac{m^2}{z^2f}\right)\psi=0,
\label{BHPsiz}\end{aligned}$$ $$\begin{aligned}
\phi^{\prime\prime}+\frac{1}{z}\left[(3-d)+\frac{b(d-1)z^{4}}{r_{+}^{2}}\phi^{\prime
2}\right]\phi^\prime-\frac{2\psi^{2}}{z^2f}\left(1-\frac{bz^{4}}{r_{+}^{2}}\phi^{\prime
2}\right)^{3/2}\phi=0,\label{BHPhiz}\end{aligned}$$ with $f=1-z^{d}$. Here and hereafter the prime denotes the derivative with respect to $z$.
In order to get the solutions in the superconducting phase, we have to impose the appropriate boundary conditions for $\psi$ and $\phi$. At the event horizon $z=1$ of the black hole, the regularity gives the boundary conditions $$\begin{aligned}
\psi(1)=-\frac{d}{m^{2}}\psi^\prime(1)\,,\hspace{0.5cm} \phi(1)=0\,.
\label{horizon}\end{aligned}$$ Near the AdS boundary $z\rightarrow0$, the asymptotic behaviors of the solutions are $$\begin{aligned}
\psi=\frac{\psi_{-}}{r_{+}^{\Delta_{-}}}z^{\Delta_{-}}+\frac{\psi_{+}}{r_{+}^{\Delta_{+}}}z^{\Delta_{+}}\,,\hspace{0.5cm}
\phi=\mu-\frac{\rho}{r_{+}^{d-2}}z^{d-2}\,, \label{infinity}\end{aligned}$$ where $\Delta_\pm=(d\pm\sqrt{d^{2}+4m^{2}})/2$ is the conformal dimension of the scalar operator dual to the bulk scalar field, $\mu$ and $\rho$ are interpreted as the chemical potential and charge density in the dual field theory respectively. It should be pointed out that, provided $\Delta_{-}$ is larger than the unitarity bound, both $\psi_{-}$ and $\psi_{+}$ can be normalizable and they can be used to define operators in the dual field theory according to the AdS/CFT correspondence, $\psi_{-}=\langle\mathcal{O}_{-}\rangle$ and $\psi_{+}=\langle\mathcal{O}_{+}\rangle$, respectively. Just as in Refs. [@HartnollPRL101; @HorowitzPRD78], we will impose boundary condition that either $\psi_{-}$ or $\psi_{+}$ vanishes. In this work, we impose boundary condition $\psi_{-}=0$ since we concentrate on the condensate for the operator $\langle\mathcal{O}_{+}\rangle$. For clarity, we set $\langle\mathcal{O}\rangle=\langle\mathcal{O}_{+}\rangle$ and $
\Delta=\Delta_{+}$ in the following discussion.
Analytical study of holographic superconductors with Born-Infeld electrodynamics
================================================================================
Here we will improve the perturbative approach proposed in [@BGRLPRD2013] and use the S-L method [@Siopsis] to analytically discuss the properties of the $d$-dimensional superconductor phase transition with Born-Infeld electrodynamics. We will investigate the relation between critical temperature and charge density as well as the critical exponent of condensation operators, and examine the effect of the Born-Infeld parameter.
Critical temperature
--------------------
At the critical temperature $T_{c}$, the scalar field $\psi=0$. Thus, near the critical point the equation of motion (\[BHPhiz\]) for the gauge field $\phi$ becomes $$\begin{aligned}
\phi^{\prime\prime}+\frac{1}{z}\left[(3-d)+\frac{b(d-1)z^{4}}{r_{+c}^{2}}\phi^{\prime
2}\right]\phi^\prime=0,\label{NESWPhiCritical}\end{aligned}$$ where $r_{+c}$ is the radius of the horizon at the critical point. Defining $\xi(z)=\phi'(z)$, we can obtain $$\begin{aligned}
\xi^{\prime}+\frac{3-d}{z}\xi=\frac{b(1-d)z^{3}}{r_{+c}^{2}}\xi^{3},\label{XiCritical}\end{aligned}$$ which is the special case of Bernoulli’s Equation $y'(x)+f(x)y=g(x)y^{n}$ [@Chow] for $n=3$. Considering that the boundary condition (\[infinity\]) for $\phi$, we can get the solution to Eq. (\[XiCritical\]) $$\begin{aligned}
\xi(z)=\phi'(z)=-\frac{\lambda
r_{+c}(d-2)z^{d-3}}{\sqrt{1+(d-2)^{2}b\lambda^{2}z^{2(d-1)}}},
\label{XiSolution}\end{aligned}$$ which leads to the expression $$\begin{aligned}
\phi(z)=\lambda r_{+c}\zeta(z), \label{PhiSolution}\end{aligned}$$ with $$\begin{aligned}
\zeta(z)=\int^{1}_{z}\frac{(d-2)\tilde{z}^{d-3}}{\sqrt{1+(d-2)^{2}b\lambda^{2}\tilde{z}^{2(d-1)}}}d\tilde{z},
\label{ZetaSolution}\end{aligned}$$ where we have set $\lambda=\rho/r^{d-1}_{+c}$ and used the fact that $\phi(1)=0$.
Obviously, the integral in (\[ZetaSolution\]) is not doable exactly. Just as in Refs. [@SDSL2012; @BGRLPRD2013], we will perform a perturbative expansion of $(d-2)^{2}b\lambda^{2}$. In order to simplify the following calculation, we will express the Born-Infeld parameter $b$ as $$\begin{aligned}
\label{BIbn}
b_{n}=n\Delta b,~~~n=0,1,2,\cdot\cdot\cdot,\end{aligned}$$ where $\Delta b=b_{n+1}-b_{n}$ is the step size of our iterative procedure. Considering the fact that $$\begin{aligned}
\label{bLambda}
(d-2)^{2}b\lambda^{2}=(d-2)^{2}b_{n}\lambda^{2}=(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})+0[(\Delta
b)^{2}],\end{aligned}$$ where we have set $b_{-1}=0$ and $\lambda^{2}|_{b_{-1}}=0$, we will discuss the following two cases (note that the variable $z$ has a range $0\leq z\leq1$):
*Case 1.* If $(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})<1$, we have $$\begin{aligned}
&\zeta(z)&=\zeta_{1}(z)\approx\int^{1}_{z}(d-2)\tilde{z}^{d-3}\left[1-\frac{(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})\tilde{z}^{2(d-1)}}{2}\right]d\tilde{z}
\nonumber\\
&&=(1-z^{d-2})+\frac{(d-2)^{3}b_{n}(\lambda^{2}|_{b_{n-1}})}{2(4-3d)}(1-z^{3d-4}).
\label{ZetaCase1}\end{aligned}$$
*Case 2.* If $(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})>1$, we set $(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})\Lambda^{2(d-1)}=1$ for $z=\Lambda$. Obviously, we find that $(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})z^{2(d-1)}<1$ for $z<\Lambda<1$, which results in $$\begin{aligned}
&\zeta(z)&=\zeta_{2A}(z)\approx\int^{\Lambda}_{z}(d-2)\tilde{z}^{d-3}\left[1-\frac{(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})\tilde{z}^{2(d-1)}}{2}\right]d\tilde{z}
\nonumber\\
&&\qquad\qquad+\int^{1}_{\Lambda}\frac{1}{\sqrt{b_{n}}(\lambda|_{b_{n-1}})\tilde{z}^{2}}\left[1-
\frac{1}{2(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})\tilde{z}^{2(d-1)}}\right]d\tilde{z}
\nonumber\\
&&=-z^{d-2}+\frac{(d-2)z^{3d-4}}{2(3d-4)\Lambda^{2(d-1)}}+\frac{3(d-1)[6+d(4d-9)]}{2(2d-1)(3d-4)}\Lambda^{d-2}
+(d-2)\Lambda^{d-1}\left[\frac{\Lambda^{2(d-1)}}{2(2d-1)}-1\right],
\label{ZetaCase2A}\end{aligned}$$ and $(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})z^{2(d-1)}>1$ for $\Lambda< z\leq1$, which leads to $$\begin{aligned}
&\zeta(z)&=\zeta_{2B}(z)\approx\int^{1}_{z}\frac{1}{\sqrt{b_{n}}(\lambda|_{b_{n-1}})\tilde{z}^{2}}\left[1-
\frac{1}{2(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})\tilde{z}^{2(d-1)}}\right]d\tilde{z}
\nonumber\\
&&=(d-2)\Lambda^{d-1}\left[\frac{\Lambda^{2(d-1)}}{2(2d-1)}\left(1-z^{1-2d}\right)+\frac{1}{z}-1\right].
\label{ZetaCase2B}\end{aligned}$$ It should be noted that in both cases we observe that $\zeta(1)=0$ from (\[ZetaCase1\]) and (\[ZetaCase2B\]), which is consistent with the boundary condition $\phi(1)=0$ given in (\[horizon\]).
Introducing a trial function $F(z)$ near the boundary $z=0$ as $$\begin{aligned}
\psi(z)\sim\frac{\langle\mathcal{O}\rangle}{r_{+}^{\Delta}}
z^{\Delta}F(z), \label{BintroduceF}\end{aligned}$$ with the boundary conditions $F(0)=1$ and $F'(0)=0$, from Eq. (\[BHPsiz\]) we can obtain the equation of motion for $F(z)$ $$\begin{aligned}
\label{BFEoM}
(TF^{\prime})^{\prime}+T\left(P+\lambda^2Q\zeta^{2}\right)F=0,\end{aligned}$$ with $$\begin{aligned}
T=z^{1+2\Delta-d}(1-z^{d}),~~P=\frac{\Delta(\Delta-d)}{z^{2}}+\frac{\Delta
f'}{z f}-\frac{m^{2}}{z^{2}f},~~ Q=\frac{1}{f^{2}}.\end{aligned}$$ According to the S-L eigenvalue problem [@Gelfand-Fomin], we deduce the eigenvalue $\lambda$ minimizes the expression $$\begin{aligned}
\label{lambdaeigenvalueCase1}
\lambda^{2}=\frac{\int^{1}_{0}T\left(F'^{2}-PF^{2}\right)dz}{\int^{1}_{0}TQ\zeta_{1}^{2}F^2dz}\
, & \quad {\rm for} \ (d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})<1,\end{aligned}$$ and $$\begin{aligned}
\label{lambdaeigenvalueCase2}
\lambda^{2}=\frac{\int^{1}_{0}T\left(F'^{2}-PF^{2}\right)dz}
{\int^{\Lambda}_{0}TQ\zeta_{2A}^{2}F^2dz+\int^{1}_{\Lambda}TQ\zeta_{2B}^{2}F^2dz}\
, & \quad {\rm for} \ (d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})>1.\end{aligned}$$ Using Eqs. (\[lambdaeigenvalueCase1\]) and (\[lambdaeigenvalueCase2\]) to compute the minimum eigenvalue of $\lambda^{2}$, we can obtain the critical temperature $T_{c}$ for different Born-Infeld parameter $b$, spacetime dimension $d$ and mass of the scalar field $m$ from the following relation $$\begin{aligned}
\label{CTTc}
T_{c}=\frac{d}{4\pi}\left(\frac{\rho}{\lambda_{min}}\right)^{\frac{1}{d-1}}.\end{aligned}$$ In the following calculation, we will assume the trial function to be $F(z)=1-az^{2}$ with a constant $a$.
As an example, we will study the case for $d=3$ and $m^{2}L^2=-2$ with the chosen values of the Born-Infeld parameter $b$. Setting $\Delta b=0.1$, for $b_{0}=0$ we use Eq. (\[lambdaeigenvalueCase1\]) and get $$\begin{aligned}
\lambda^{2}=\frac{4(15-20a+12a^{2})}{10(9-\sqrt{3}\pi-3\ln3)+10(13-12\ln3)a+(10\sqrt{3}\pi-21-30\ln3)a^{2}},\end{aligned}$$ whose minimum is $\lambda^{2}|_{b_{0}}=17.31$ at $a=0.6016$. According to Eq. (\[CTTc\]), we can easily obtain the critical temperature $T_{c}=0.1170\rho^{1/2}$, which is in good agreement with the numerical result $T_{c}=0.1184\rho^{1/2}$ [@HartnollPRL101]. For $b_{1}=0.1$, we can easily have $b_{1}(\lambda^{2}|_{b_{0}})>1$ and $\Lambda=[b_{1}(\lambda^{2}|_{b_{0}})]^{-1/4}=0.8718$. Using Eq. (\[lambdaeigenvalueCase2\]) we arrive at $$\begin{aligned}
\lambda^{2}=\frac{1-\frac{4a}{3}+\frac{4a^{2}}{5}}{0.02060-0.01199a+0.002659a^{2}},\end{aligned}$$ whose minimum is $\lambda^{2}|_{b_{1}}=33.84$ at $a=0.6532$. So the critical temperature $T_{c}=0.09898\rho^{1/2}$, which also agrees well with the numerical finding $T_{c}=0.1007\rho^{1/2}$ [@JS2010]. For $b_{1}=0.2$, we still have $b_{2}(\lambda^{2}|_{b_{1}})>1$ and $\Lambda=[b_{2}(\lambda^{2}|_{b_{1}})]^{-1/4}=0.6200$. With the help of Eq. (\[lambdaeigenvalueCase2\]) we obtain $$\begin{aligned}
\lambda^{2}=\frac{1-\frac{4a}{3}+\frac{4a^{2}}{5}}{0.01176-0.006582a+0.001450a^{2}},\end{aligned}$$ whose minimum is $\lambda^{2}|_{b_{2}}=58.19$ at $a=0.6640$. Therefore the critical temperature $T_{c}=0.08644\rho^{1/2}$, which is again consistent with the numerical result $T_{c}=0.08566\rho^{1/2}$ [@JS2010]. For other values of $b$, the similar iterative procedure can be applied to give the analytical result for the critical temperature.
$b$ 0 0.1 0.2 0.3
-------------------------- ------------------------------------ -------------------------------------- ------------------------------------- --------------------------------- -- --
$Analytical$ $0.1170\rho^{1/2}$ $0.09898\rho^{1/2}$ $0.08644\rho^{1/2}$ $0.07586\rho^{1/2}$
$Numerical$ $0.1184\rho^{1/2}$ $0.1007\rho^{1/2}$ $0.08566\rho^{1/2}$ $0.07292\rho^{1/2}$
: \[BICriticalTcD3\] The critical temperature $T_{c}$ obtained by the analytical S-L method and from numerical calculation [@JS2010] for the chosen values of the Born-Infeld parameter $b$ in the case of 4-dimensional AdS black hole background. Here we fix the mass of the scalar field by $m^{2}L^2=-2$ and the step size by $\Delta b=0.1$.
In Table \[BICriticalTcD3\], we provide the critical temperature $T_{c}$ of the chosen parameter $b$ with the scalar operator $\langle\mathcal{O}\rangle=\langle\mathcal{O}_{+}\rangle$ for the ($2+1$)-dimensional superconductor if we fix the mass of the scalar field by $m^{2}L^2=-2$ and the step size by $\Delta b=0.1$. From Table \[BICriticalTcD3\], we observe that the differences between the analytical and numerical values are within $4.1\%$. Compared with the analytical results given in Table 1 of Ref. [@BGRLPRD2013], the iterative procedure can further improve our analytical results and improve the consistency with the numerical findings.
Extending the investigation to the ($3+1$)-dimensional superconductor, in Table \[BICriticalTcD4\] we also give the critical temperature $T_{c}$ for the scalar operator $\langle\mathcal{O}\rangle=\langle\mathcal{O}_{+}\rangle$ when we fix the mass of the scalar field $m^{2}L^2=-3$ for different Born-Infeld parameter $b$ by choosing the step size $\Delta b=0.05$ and $0.025$, respectively. Obviously, for the case of $\Delta
b=0.025$ the agreement of the analytical results derived from S-L method with the numerical calculation is impressive. Thus, we argue that, even in the higher dimension, the analytical results derived from the S-L method are in very good agreement with the numerical calculation. Furthermore, reducing the step size $\Delta b$ reasonably, we can improve the analytical result and get the critical temperature more consistent with the numerical result.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------
$b$ 0 0.1 0.2 0.3
------------------------ -------------------------------- --------------------------------- --------------------------------- ------------------------------- -- --
$Analytical(\Delta $0.1962\rho^{1/3}$ $0.1460\rho^{1/3}$ $0.1091\rho^{1/3}$ $0.07866\rho^{1/3}$
b=0.05)$
$Analytical(\Delta $0.1962\rho^{1/3}$ $0.1329\rho^{1/3}$ $0.08754\rho^{1/3}$ $0.05195\rho^{1/3}$
b=0.025)$
$Numerical$ $0.1980\rho^{1/3}$ $0.1275\rho^{1/3}$ $0.08298\rho^{1/3}$ $0.05292\rho^{1/3}$
-------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[BICriticalTcD4\] The critical temperature $T_{c}$ with the chosen values of the Born-Infeld parameter $b$ and the step size $\Delta b$ in the case of 5-dimensional AdS black hole background. Here we fix the mass of the scalar field by $m^{2}L^2=-3$.
From Tables \[BICriticalTcD3\] and \[BICriticalTcD4\], we point out that the critical temperature $T_{c}$ decreases as the Born-Infeld parameter $b$ increases for the fixed scalar field mass and spacetime dimension, which supports the numerical computation found in Refs. [@JS2010; @JLQS2012; @ZPCJNPB]. It is shown that the higher Born-Infeld electrodynamics corrections will make the scalar hair more difficult to be developed. On the other hand, the consistency between the analytical and numerical results indicates that the S-L method is a powerful analytical way to investigate the holographic superconductor with various condensates even when we take the Born-Infeld electrodynamics into account.
Critical phenomena
------------------
Since the condensation for the scalar operator $\langle\mathcal{O}\rangle$ is so small when $T \rightarrow T_c$, we can expand $\phi(z)$ in $\langle\mathcal{O}\rangle$ near the boundary $z=0$ as $$\begin{aligned}
\label{PhiExpandNearTc}
\frac{\phi(z)}{r_+}=\lambda\zeta(z)+\frac{\langle\mathcal{O}\rangle^2}{r_+^{2\Delta}}\chi(z)+\cdot\cdot\cdot,\end{aligned}$$ with the boundary conditions $\chi(1)=0$ and $\chi'(1)=0$ [@Siopsis; @ZPJ2015; @Li-Cai-Zhang]. Thus, substituting the functions (\[BintroduceF\]) and (\[PhiExpandNearTc\]) into (\[BHPhiz\]), we keep terms up to $0(b)$ [@BGRLPRD2013] to get the equation of motion for $\chi(z)$ $$\begin{aligned}
\label{BHChizEoM}
(U\chi^\prime)^\prime=\frac{2\lambda z^{1+2\Delta-d}F^{2}\zeta}{f},\end{aligned}$$ where we have introduced a new function $$\begin{aligned}
U(z)=\frac{e^{3b\lambda^{2}z^{4}\zeta'^{2}/2}}{z^{d-3}}.\end{aligned}$$ Making integration of both sides of Eq. (\[BHChizEoM\]), we have $$\begin{aligned}
\label{Chi0}
\left[\frac{\chi'(z)}{z^{d-3}}\right]\bigg|_{z\rightarrow 0}=
\left\{
\begin{array}{rl}
-\lambda\alpha_{1}\ ,~~~~~~ & \quad {\rm for}\
(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})<1, \\ \\
-\lambda(\alpha_{2A}+\alpha_{2B}) \ , & \quad {\rm for}\
(d-2)^{2}b_{n}(\lambda^{2}|_{b_{n-1}})>1,
\end{array}\right.\end{aligned}$$ with $$\begin{aligned}
\alpha_{1}=\int_{0}^{1}\frac{2z^{1+2\Delta-d}F^{2}\zeta_{1}}{f}dz,~~
\alpha_{2A}=\int_{0}^{\Lambda}\frac{2z^{1+2\Delta-d}F^{2}\zeta_{2A}}{f}dz,~~
\alpha_{2B}=\int_{\Lambda}^{1}\frac{2z^{1+2\Delta-d}F^{2}\zeta_{2B}}{f}dz.\end{aligned}$$
For clarity, we will fix the spacetime dimension $d$ in the following discussion. Considering the case of $d=3$ and the asymptotic behavior (\[infinity\]), for example, near $z\rightarrow0$ we can arrive at $$\begin{aligned}
\label{BHPhiExpandD3}
\frac{\rho}{r^{2}_+}(1-z)=\lambda\zeta(z)+\frac{\langle
\mathcal{O}\rangle^2}{r_+^{2\Delta}}\left[\chi(0)+\chi^\prime(0)z+\cdot\cdot\cdot\right].\end{aligned}$$ From the coefficients of the $z^1$ terms in both sides of the above formula, we can obtain $$\begin{aligned}
\label{BHRhoExpD3}
\frac{\rho}{r^{2}_{+}}=\lambda-\frac{\langle
\mathcal{O}\rangle^2}{r_+^{2\Delta}}\chi'(0),\end{aligned}$$ where $\chi^\prime(0)$ can be easily calculated by using Eq. (\[Chi0\]). Therefore we will know that $$\begin{aligned}
\label{D3OExp}
\langle\mathcal{O}\rangle=\beta
T_{c}^{\Delta}\left(1-\frac{T}{T_c}\right)^{\frac{1}{2}},\end{aligned}$$ where the coefficient $\beta$ is given by $$\begin{aligned}
\label{D3Beta}
\beta=\left\{
\begin{array}{rl}
\left(\frac{4\pi}{3}\right)^{\Delta}\sqrt{\frac{2}{\alpha_{1}}}\
,~~~~ & \quad {\rm for}\ b_{n}(\lambda^{2}|_{b_{n-1}})<1,
\\ \\
\left(\frac{4\pi}{3}\right)^{\Delta}\sqrt{\frac{2}{\alpha_{2A}+\alpha_{2B}}}\
, & \quad {\rm for}\ b_{n}(\lambda^{2}|_{b_{n-1}})>1.
\end{array}\right.\end{aligned}$$ Obviously, the expression (\[D3OExp\]) is valid for different values of the Born-Infeld parameter and scalar field mass in the case of the ($2+1$)-dimensional superconductor. For concreteness, we will focus on the case for the mass of the scalar field $m^{2}L^2=-2$ and the step size $\Delta b=0.1$. Since in Ref. [@JS2010] the scalar operator is given by $\langle\mathcal{O}_{+}\rangle=\sqrt{2}\psi_{+}$ which is different from $\langle\mathcal{O}_{+}\rangle=\psi_{+}$ in this work, we present the condensation value $\gamma=\sqrt{2}\beta$ obtained by the analytical S-L method and from numerical calculation with the chosen values of the Born-Infeld parameter $b$ for the ($2+1$)-dimensional superconductor in Table \[D3BetaValue\]. We see that the condensation value $\gamma$ increases as the Born-Infeld parameter $b$ increases for the fixed scalar field mass and spacetime dimension, which indicates the consistent picture shown in $T_{c}$ that the higher Born-Infeld electrodynamics corrections make the condensation to be formed harder. On the other hand, comparing with the analytical results shown in Table II of Ref. [@BGRLPRD2013], we find that the iterative procedure indeed reduces the disparity between the analytical and numerical results.
$b$ 0 0.1 0.2 0.3
--------------------------- ---------------------------- ----------------------------- ---------------------------- ----------------------- -- --
$Analytical$ $92.80$ $117.92$ $137.22$ $161.14$
$Numerical$ $139.24$ $207.36$ $302.76$ $432.64$
: \[D3BetaValue\] The condensation value $\gamma=\sqrt{2}\beta$ obtained by the analytical S-L method and from numerical calculation [@JS2010] with the chosen values of the Born-Infeld parameter $b$ in the case of 4-dimensional AdS black hole background. Here we fix the mass of the scalar field by $m^{2}L^2=-2$ and the step size by $\Delta b=0.1$.
As another example, let us move on to the case of $d=4$. From the asymptotic behavior (\[infinity\]), we can expand $\phi$ when $z\rightarrow0$ as $$\begin{aligned}
\label{BHPhiExpandD4}
\frac{\rho}{r^{3}_+}(1-z^{2})=\lambda\zeta(z)+\frac{\langle
\mathcal{O}\rangle^2}{r_+^{2\Delta}}\left[\chi(0)+\chi^\prime(0)z
+\frac{1}{2}\chi^{\prime\prime}(0)z^2+\cdot\cdot\cdot\right].\end{aligned}$$ Considering the coefficients of $z^1$ terms in above equation, we observe that $\chi^\prime(0)\rightarrow 0$ if $z\rightarrow 0$, which is consistent with Eq. (\[Chi0\]). Comparing the coefficients of the $z^2$ terms, we have $$\begin{aligned}
\label{BHRhoExpD4}
\frac{\rho}{r^{3}_{+}}=\lambda-\frac{\langle
\mathcal{O}\rangle^2}{2r_+^{2\Delta}}\chi''(0),\end{aligned}$$ where $\chi''(0)$ can be computed by using Eq. (\[Chi0\]). So we can deduce the same relation (\[D3OExp\]) for the ($3+1$)-dimensional superconductor with the different condensation coefficient $$\begin{aligned}
\label{D4Beta}
\beta=\left\{
\begin{array}{rl}
\pi^{\Delta}\sqrt{\frac{6}{\alpha_{1}}}\ ,~~~~ & \quad {\rm for}\
4b_{n}(\lambda^{2}|_{b_{n-1}})<1,
\\ \\
\pi^{\Delta}\sqrt{\frac{6}{\alpha_{2A}+\alpha_{2B}}}\ , & \quad {\rm
for}\ 4b_{n}(\lambda^{2}|_{b_{n-1}})>1.
\end{array}\right.\end{aligned}$$ In Table \[D4BetaValue\], we give the condensation value $\beta$ obtained by the analytical S-L method with the chosen values of the Born-Infeld parameter $b$ and step size $\Delta b$ for the ($3+1$)-dimensional superconductor. In both cases we find again that, for the fixed scalar field mass and spacetime dimension, the condensation value $\beta$ increases as the Born-Infeld parameter $b$ increases, just as the observation obtained in the ($2+1$)-dimensional superconductor with Born-Infeld electrodynamics.
-------------------------------------------------------------------------------------------------------------------------------------------
$b$ 0 0.1 0.2 0.3
------------------- ---------------------------- ----------------------------- ----------------------------- ------------------------ -- --
$\Delta $238.91$ $418.95$ $697.64$ $1195.56$
b=0.05$
$\Delta $238.91$ $496.06$ $1005.38$ $2303.28$
b=0.025$
-------------------------------------------------------------------------------------------------------------------------------------------
: \[D4BetaValue\] The condensation value $\beta$ obtained by the analytical S-L method with the chosen values of the Born-Infeld parameter $b$ and step size $\Delta b$ in the case of 5-dimensional AdS black hole background. Here we fix the mass of the scalar field by $m^{2}L^2=-3$.
It should be noted that one can easily extend our discussion to the higher-dimensional superconductor and get our expression (\[D3OExp\]), although the coefficient $\beta$ is different. Thus, near the critical point, the scalar operator $\langle\mathcal{O}\rangle$ will satisfy $$\begin{aligned}
\label{DOExp}
\langle\mathcal{O}\rangle\sim\left(1-T/T_c\right)^{1/2},\end{aligned}$$ which holds for various values of the Born-Infeld parameter $b$, spacetime dimension $d$ and mass of the scalar field $m$. It shows that the phase transition is of the second order and the critical exponent of the system always takes the mean-field value $1/2$. The Born-Infeld electrodynamics will not influence the result.
Conclusions
===========
We have generalized the variational method for the S-L eigenvalue problem to analytically investigate the condensation and critical phenomena of the $d$-dimensional superconductors with Born-Infeld electrodynamics, which may help to understand the influences of the $1/N$ or $1/\lambda$ corrections on the holographic superconductor models. We found that the S-L method is still powerful to disclose the properties of the holographic superconductor with various condensates even when we take the Born-Infeld electrodynamics into account. Using the iterative procedure in the perturbative approach proposed by Banerjee *et al.* [@BGRLPRD2013], we further improved the analytical results and the consistency with the numerical findings for the ($2+1$)-dimensional superconductor. Furthermore, extending the investigation to the higher-dimensional superconductor with Born-Infeld electrodynamics, we observed again that the analytical results derived from this method with a reasonable step size are in very good agreement with those obtained from numerical calculation. Our analytical result shows that the Born-Infeld parameter makes the critical temperature of the superconductor decrease, which can be used to back up the numerical findings as shown in the existing literatures that the higher Born-Infeld electrodynamics corrections can hinder the condensation to be formed. Moreover, with the help of this analytical method, we interestingly noted that the Born-Infeld electrodynamics, spacetime dimension and scalar mass cannot modify the critical phenomena, and found that the holographic superconductor phase transition belongs to the second order and the critical exponent of the system always takes the mean-field value. It should be noted that one can easily extend our technique to the holographic superconductor models with the logarithmic form [@JPCPLB] and exponential form [@ZPCJNPB] of nonlinear electrodynamics. More recently, a model of p-wave holographic superconductors from charged Born-Infeld black holes [@CSJHEP2015] via a Maxwell complex vector field model [@CaiPWave-1; @CaiPWave-2; @CaiPWave-3] was studied numerically. It would be of interest to generalize our study to this p-wave model and analytically discuss the effect of the Born-Infeld electrodynamics on the system. We will leave it for further study.
This work was supported by the National Natural Science Foundation of China under Grant Nos. 11275066, 11175065 and 11475061; Hunan Provincial Natural Science Foundation of China under Grant Nos. 12JJ4007 and 11JJ7001; and FAPESP No. 2013/26173-9.
[99]{}
J. Maldacena, Adv. Theor. Math. Phys. [**2**]{}, 231 (1998) \[Int. J. Theor. Phys. [**38**]{}, 1113 (1999)\].
E. Witten, Adv. Theor. Math. Phys. [**2**]{}, 253 (1998).
S.S. Gubser, I.R. Klebanov, and A.M. Polyakov, Phys. Lett. B [**428**]{}, 105 (1998).
S.A. Hartnoll, Class. Quant. Grav. [**26**]{}, 224002 (2009).
C.P. Herzog, J. Phys. A [**42**]{}, 343001 (2009).
G.T. Horowitz, Lect. Notes Phys. [**828**]{} 313, (2011); arXiv:1002.1722 \[hep-th\].
R.G. Cai, L. Li, L.F. Li, and R.Q. Yang, Sci. China Phys. Mech. Astron. [**58**]{}, 060401 (2015); arXiv:1502.00437 \[hep-th\].
S.S. Gubser, Phys. Rev. D [**78**]{}, 065034 (2008).
S.A. Hartnoll, C.P. Herzog, and G.T. Horowitz, Phys. Rev. Lett. [**101**]{}, 031601 (2008).
S.H. Hendi, J. High Energy Phys. [**03**]{}, 065 (2012).
J.L. Jing and S.B. Chen, Phys. Lett. B [**686**]{}, 68 (2010).
J.L. Jing, L.C. Wang, Q.Y. Pan, and S.B. Chen, Phys. Rev. D [**83**]{}, 066010 (2011).
J.L. Jing, Q.Y. Pan, and S.B. Chen, J. High Energy Phys. [**11**]{}, 045 (2011).
Q.Y. Pan, J.L. Jing, and B. Wang, Phys. Rev. D [**84**]{}, 126020 (2011).
S. Gangopadhyay and D. Roychowdhury, J. High Energy Phys. [**05**]{}, 156 (2012).
Chong Oh Lee, Eur. Phys. J. C [**72**]{}, 2092 (2012).
Y.Q. Liu, Y. Peng, and B. Wang, arXiv:1202.3586 \[hep-th\].
D. Roychowdhury, Phys. Rev. D [**86**]{}, 106009 (2012); Phys. Lett. B [**718**]{}, 1089 (2013).
N. Bai, Y.H. Gao, G.B. Qi, and X.B. Xu, arXiv:1212.2721 \[hep-th\].
J.L. Jing, Q.Y. Pan, and S.B. Chen, Phys. Lett. B [**716**]{}, 385 (2012).
Z.X. Zhao, Q.Y. Pan, S.B. Chen, and J.L. Jing, Nucl. Phys. B [**871**]{}, 98 (2013).
W.P. Yao and J.L. Jing, J. High Energy Phys. [**05**]{}, 101 (2013); Nucl. Phys. B [**889**]{}, 109 (2014).
S. Dey and A. Lala, Ann. Phys. [**354**]{}, 165 (2014).
S. Gangopadhyay, Mod. Phys. Lett. A [**29**]{}, 1450088 (2014).
G. Siopsis and J. Therrien, J. High Energy Phys. [**05**]{}, 013 (2010).
G. Siopsis, J. Therrien, and S. Musiri, Class. Quant. Grav. [**29**]{}, 085007 (2012); arXiv:1011.2938 \[hep-th\].
R.G. Cai, H.F. Li, and H.Q. Zhang, Phys. Rev. D [**83**]{}, 126007 (2011); arXiv:1103.5568 \[hep-th\].
S. Gangopadhyay and D. Roychowdhury, J. High Energy Phys. [**05**]{}, 002 (2012).
R. Banerjee, S. Gangopadhyay, D. Roychowdhury, and A. Lala, Phys. Rev. D [**87**]{}, 104001 (2013); arXiv:1208.5902 \[hep-th\].
G.T. Horowitz and M.M. Roberts, Phys. Rev. D [**78**]{}, 126008 (2008).
T.L. Chow, *Mathematical Methods for Physicists-A concise introduction*, Cambridge University Press, Cambridge (2000).
I.M. Gelfand and S.V. Fomin, Calculaus of Variations, Revised English Edition, Translated and Edited by R.A. Silverman, Prentice-Hall, Inc. Englewood Cliffs, New Jersey (1963).
L. Zhang, Q.Y. Pan, and J.L. Jing, Phys. Lett. B [**743**]{}, 104 (2015).
H.F. Li, R.G. Cai, and H.Q. Zhang, J. High Energy Phys. [**04**]{}, 028 (2011).
P. Chaturvedi and G. Sengupta, J. High Energy Phys. [**04**]{}, 001 (2015).
R.G. Cai, S. He, L. Li, and L.F. Li, J. High Energy Phys. [**12**]{}, 036 (2013); arXiv:1309.2098 \[hep-th\].
R.G. Cai, L. Li, and L.F. Li, J. High Energy Phys. [**01**]{}, 032 (2014); arXiv:1309.4877 \[hep-th\].
L.F. Li, R.G. Cai, L. Li, and C. Shen, Nucl. Phys. B [**894**]{}, 15 (2015); arXiv:1310.6239 \[hep-th\].
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, we prove that a noncompact complete hypersurface with finite weighted volume, weighted mean curvature vector bounded in norm, and isometrically immersed in a complete weighted manifold is proper. In addition, we obtain an estimate for $f$-stability index of a constant weighted mean curvature hypersurface with finite weighted volume and isometrically immersed in a shrinking gradient Ricci soliton that admits at least one parallel field globally defined. For such hypersurface, we still give a necessary condition for equality to be achieved in the estimate obtained.\
author:
- Hilário Alencar
- Adina Rocha
date: 'December 20, 2016'
title: 'The $f$-Stability Index of the Constant Weighted Mean Curvature Hypersurfaces in Gradient Ricci Solitons'
---
Introduction {#intro}
============
Consider a $(n+1)$-dimensional Riemannian manifold $(\overline M^{n+1},g)$ endowed with a weighted measure of the form $e^{-f}d\mu$, where $f$ is a smooth function on $M$ and $d\mu$ is the volume element induced by the Riemannian metric $g$. A [*weighted manifold*]{} is a triple $${\overline M}_f^{n+1}=( \overline M^{n+1},g,e^{-f}d\mu).$$ A natural extension of the Ricci tensor to this new context is the Bakry-Émery Ricci tensor, see [@BakryEmery1985], given by $$\overline{{Ric}}_f=\overline{{Ric}}+\overline{\nabla}^{2} f,$$ where $\overline{\nabla}^{2} f$ is the Hessian of $f$ on $\overline M^{n+1}$. It is known that a complete weighted manifold satisfying $\overline{{Ric}}_f\geq\displaystyle{k}g$ for some constant $k>0$ is not necessarily compact. One of the examples is the Gaussian shrinking soliton $\left({\mathbb{R}}^{n+1},g_{can},e^{-\frac{|x|^2}{4}}d\mu\right)$ with the canonical metric $g_{can}$ and $\overline{{Ric}}_f=\displaystyle{\frac{1}{2}g_{can}}.$
The gradient Ricci solitons are natural generalizations of the Einstein metrics and was introduced by Hamilton in [@Hamilton1988]. Indeed, a complete Riemannian metric $g$ on a smooth manifold $\overline M^{n+1}$ is a [*shrinking gradient Ricci soliton*]{} if there exists a potential function $f$, and a real constant $k>0$ such that the Ricci tensor $\overline{{Ric}}$ of the metric $g$ satisfies the equation $$\overline{{Ric}}+\overline{\nabla}^2f=kg,$$ where $\overline{\nabla}^2f$ denotes the Hessian of $f$. In this context, the gradient Ricci solitons are complete weighted manifolds with $\overline{{Ric}}_f=kg$ for some constant real $k$.
Observe that when the potential function is a constant, the gradient Ricci solitons are simply Einstein metrics. It is still important to mention that gradient Ricci solitons plays an important role in Hamilton’s Ricci flow and correspond to self-similar solutions, and often arise as Type I singularity models. For more details see [@Hamilton1982].
Let $x:M^n\rightarrow\overline M^{n+1}_f$ be an isometric immersion of a Riemannian orientable manifold $M^n$ into weighted manifold $\overline M^{n+1}_f.$ The function $f:\overline M \rightarrow{\mathbb{R}}$, restricted to $M$, induces a weighted measure $e^{-f}d\sigma$ on $M$. Thus, we have an induced weighted manifold $M_f^n=(M,\langle\,,\,\rangle,e^{-f}d\sigma)$.
The [*second fundamental form*]{} $A$ of $x$ is defined by $$A(X,Y)=(\overline\nabla_XY)^{\perp}, \ \ \ \ \ \ X,Y\in T_pM, \ \ p\in M,$$ where $\perp$ symbolizes the projection above the normal bundle of $M$. The [*weighted mean curvature vector*]{} of $M$ is defined by $${\bf H}_f={\bf H}+(\overline\nabla f)^{\perp},$$ and its the [*weighted mean curvature*]{} $H_f$ is given by $${\bf H}_f=-H_f\eta,$$ where ${\bf H}={tr}A$ and $\eta$ is unit outside normal vector field. The hypersurface $M$ is called [*$f$-minimal*]{} when its weighted mean curvature vector ${\bf H}_f$ vanishes identically, and when there exists real constant $C$ such that ${\bf H}_f=-C\eta$, we say the hypersurface $M$ has [*constant weighted mean curvature*]{}.
The [*weighted volume*]{} of a measurable set $\Omega\subset M$ is given by $${Vol}_f(\Omega)=\int_{\Omega}e^{-f}d\sigma.$$
Let $B^{\overline M}_r$ be the geodesic ball of $\overline M$ with center in a fixed point $o\in \overline M$ and radius $r>0$. It is said that the weighted volume of $M$ has [*polynomial growth*]{} if there exists positive numbers $\alpha$ and $C$ such that $$\label{equation10}
{Vol}_f(B^{\overline M}_r\cap M)\leq Cr^{\alpha}$$ for any $r\geq1.$ When $\alpha=n$ in (\[equation10\]), $M$ is said to have Euclidean volume growth.
We can consider either $f$-minimal or constant weighted mean curvature hypersurfaces in gradient Ricci solitons. In particular, a self-shrinker to the mean curvature flow is a $f$-minimal hypersurface of the shrinking Gaussian soliton. In [@ChengZhou2013], Cheng and Zhou showed that for $f$-minimal hypersurfaces in the shrinking Gaussian soliton $\left({\mathbb{R}}^{n+1},g_{can},e^{-\frac{|x|^2}{4}}dx\right)$, the properness of its immersion, its polynomial volume growth, and its finite weighted volume are equivalent to each other. Those equivalences are still being valid for $f$-minimal hypersurfaces immersed in a complete shrinking gradient Ricci soliton $\overline M_f$ satisfying $\overline{{Ric}}_f=\dfrac{1}{2}g$, where $g$ is Riemannian metric and $f$ is a convex function (see [@ChengMejiaZhou2015]). In this direction, the following result was obtained:
\[ip4\] Let $M^n$ be a noncompact complete hypersurface isometrically immersed in a complete weighted manifold $\overline M^m_f$, with weighted mean curvature vector bounded in norm. If $M^n$ has finite weighted volume, then $M^n$ is proper.
\[ip1\] Let $f\in C^{\infty}(\overline M)$ be a convex function, $\overline M^m_f$ a gradient Ricci soliton with $\overline{{Ric}}_f=\dfrac{1}{2}g$, and $x:M^n\rightarrow\overline M^m_f$ a noncompact complete immersion with weighted mean curvature vector satisfying $$\sup_{x\in M}\langle {\bf H}_f,\overline{\nabla}f\rangle<\infty.$$ If $M^n$ is proper, then it has finite weighted volume and Euclidean volume growth.
Now it is known that the [*weighted Laplacian operator*]{} $\Delta_f$, defined by $$\Delta_fu:=\Delta u-\langle\nabla f,\nabla u\rangle,$$ is associated to $e^{-f}d\sigma$ as well as $\Delta$ is associated to $d\sigma.$ Moreover, $\Delta_f$ is a self-adjoint operator on the $L^{2}_f$ space of square integrable functions on $M$ with respect to the measure $e^{-f}d\sigma,$ and therefore, the $L^{2}_f$ spectrum of $\Delta_f$ on $M$, denoted by $\sigma(-\Delta_f)$, is a subset of $[0,+\infty).$
Next, let $F:(-\varepsilon,\varepsilon)\times M\rightarrow \overline M_f$, $F_f(p)=F(t,p)$ for all $t\in(-\varepsilon,\varepsilon)$ and $p\in M$ be a variation of the immersion $x$ associated with the normal vector field $u\eta$, where $u\in C^{\infty}_c(M).$ The corresponding variation of the [*functional weighted area*]{} $\mathcal A_f(t)={Vol}_f(F_t(M))$ satisfies $$\label{ie1}
\mathcal A_f'(0)=\int_M H_f ue^{-f}d\sigma,$$ where $H_f$ is such that ${\bf H}_f=-H_f\eta$. The expression (\[ie1\]) is known as [*first variation formula*]{}.
The $f$-minimal hypersurfaces are critical points of the functional weighted area. Yet, the hypersurfaces with constant weighted mean curvature can be viewed as critical points of the functional weighted area restricted to variations which preserve the [*enclose weighted volume,*]{} i.e., for functions $u\in C^{\infty}_c(M)$ which satisfy the additional condition $$\int_Mue^{-f}d\sigma=0.$$ For such critical points, the [*second variation*]{} of the functional weighted area is given by $$\mathcal A_f''(0)=-\int_M\left(u\Delta_fu+\left(|A|^2+\overline{{Ric}_f}(\eta,\eta)\right)u^2\right)d\sigma,$$ where $\overline{{Ric}}_f$ is the Bakry-' Emery Ricci curvature and $A$ is the second fundamental form. For more details, see [@Rocha2016].
When $f$ is a constant function, the first and second variation formula were given by Barbosa and do Carmo [@BarbosadoCarmo1984] and Barbosa, do Carmo and Eschenburg [@BarbosadoCarmoEschenburg1988].
The operator $$L_f=\Delta_f+|A|^2+\overline{{Ric}_f}(\eta,\eta)$$ is called the [*$f$-stability operator*]{} of the immersion $x$. In the $f$-minimal case, the $f$-stability operator is viewed as acting on $\mathcal F=C_c^{\infty}(M)$; in the case of the hypersurfaces with constant weighted mean curvature, the $f$-stability operator is viewed as acting on $$\mathcal F=C_c^{\infty}(M)\cap \left\{u\in C^{\infty}_c(M); \ \int_Mue^{-f}d\sigma=0\right\}.$$ Associated with $L_f$ is the quadratic form $$I_f(u,u)=-\int_MuL_fue^{-f}d\sigma.$$ For each compact domain $\Omega\subset M$, define the index, ${Ind}_f\Omega$, of $L_f$ in $\Omega$ as the maximal dimension of a subspace of $\mathcal F$ where $I_f$ is a negative definite. The [*index*]{}, ${Ind}_fM$, of $L_f$ in $M$ (or simply, the index of $M$) is then defined by $${Ind}_fM=\sup_{\Omega\subset M}{Ind}_f\Omega,$$ where the supreme is taken over all compact domains $\Omega\subset M.$ For more details, see [@Fischer-Colbrie1985] and [@ChengZhou2015].
Let $M\subset{\mathbb{R}}^{n+1}$ be a proper, non-planar, two-sided hypersurface satisfying ${Vol}_f(M)<\infty$, $H=\dfrac{1}{2}\langle x,\eta\rangle+C$ and ${Ind}_f(M)\leq n,$ where $H$ is the mean curvature, $x$ is the position vector of ${\mathbb{R}}^{n+1}$, $\eta$ is the unit normal field of the hypersurface, ${Ind}_f(M)$ is the $f$-stability index and $C$ is a real constant. McGonagle and Ross ([@McGR2015], Theorem 5.6) showed that exists a natural number $i$ such that $n+1-{Ind}_f(M)\leq i\leq n$ e $\Sigma=\Sigma_0\times{\mathbb{R}}^i.$ In addition, they obtained ${Ind}_f(M)\geq2$.
It is important to mention that the properness hypothesis can be removed of the Theorem 5.6 of [@McGR2015]. In fact, by Proposition \[ip4\], the finite weighted volume implies in the properness of its immersion.
Next, we obtain an estimate for $f$-stability index of a constant weighted mean curvature hypersurface with finite weighted volume and isometrically immersed in a gradient Ricci soliton that admit at least one parallel field globally defined. If fact,
\[it2\] Let $\overline M_f$ be a shrinking gradient Ricci soliton with $\overline{{Ric}}_f=kg$. Let $M^n$ be a constant weighted mean curvature hypersurface with finite weighted volume and isometrically immersed in $\overline M_f$. Denote by $\mathcal P_{\overline M_f}$ the set of parallel fields globally defined on $\overline M_f^{n+1}$ and $\eta$ the unit normal field to $M$.
1. If the unit function $1\not\in \{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\}$, $$\label{iequation8}
{Ind}_f(M)\geq\dim \mathcal P_{\overline M_f}-\dim\{X\in P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}.$$
2. If the unit function $1\in \{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\}$, $M$ is totally geodesic.
As a consequence of Theorem \[it2\], we have
\[ic1\] Let $M^n$ be a constant weighted mean curvature hypersurface with finite weighted volume and isometrically immersed in a shrinking gradient Ricci soliton $\overline M_f$. Denote by $\mathcal P_{\overline M_f}$ the set of parallel fields globally defined on $\overline M_f^{n+1}$ and let $\eta$ the unit normal field to $M$. If the unit function $1\not\in \{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\}$ and there exist a parallel field $X_0$ such that $\langle X_0,\eta\rangle\not\equiv0$, then $${Ind}_f(M)\geq1.$$ Moreover, $$\dim\{X\in \mathcal P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}=\dim\mathcal P_{\overline M_f}-1$$ whenever ${Ind}_f(M)=1.$
A necessary condition for equality to be achieved in the estimate (\[iequation8\]) of Theorem \[it2\], is given by
\[itheorem2\] Let $\overline M_f$ be a shrinking gradient Ricci soliton with $\overline{{Ric}}_f=kg$. Let $M^n$ be a constant weighted mean curvature hypersurface isometrically immersed in $\overline M_f$. Denote by $\mathcal P_{\overline M_f}$ the set of parallel fields globally defined on $\overline M_f^{n+1}$ and $\eta$ the unit normal field to $M$. Suppose that ${Vol}_f(M)<\infty$, ${Ind}_f(M)<\infty$, $\dim \mathcal P_{\overline M_f}>0$, and $$\label{ieq5}
{Ind}_f(M)=\dim \mathcal P_{\overline M_f}-\dim\{X\in P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}.$$
1. If ${Ind}_f(M)=\dim \mathcal P_{\overline M_f}$, $M$ is totally geodesic and the bottom $\mu_1(M)$ of the spectrum of $f$-stability operator satisfies $\mu_1(M)=-k.$
2. If ${Ind}_f(M)\neq\dim \mathcal P_{\overline M_f},$ either $M$ is diffeomorphic to the product of a Euclidian space with some other manifold or there is a circle action on $M$ whose orbits are not real homologous to zero.
Properness and Finite Weighted Volume of a Constant Weighted Mean Curvature Hypersurface
========================================================================================
We will begin by proving Proposition \[ip4\] which states that a noncompact complete hypersurface with finite weighted volume, weighted mean curvature vector bounded in norm, and isometrically immersed in a complete weighted manifold is proper. Furthermore, long after we will prove Proposition \[ip1\], it give some conditions to that the a proper noncompact complete hypersurface have finite weighted volume.
We supposed that $M$ is not proper. Thus, there exists a positive real number $R$ such that $\overline B_R^{\overline M}(o)\cap M$ is no compact in $M$, where $\overline B_R^{\overline M}(o)$ denotes the closure of the $B_R^{\overline M}(o)$. Then, for any $a>0$ sufficiently small with $a<2R$, there exists a sequence $\{p_k\}$ of the points in $B_R^{\overline M}(o)\cap M$ with ${{\rm dist}\,}_{M}(p_k,p_j)\geq a>0$ for any different $k$ and $j$.
Since $B_{a/2}^{M}(p_k)\cap B_{a/2}^{M}(p_j)=\emptyset$ for any $k\neq j$, we obtain $B_{a/2}^{M}(p_j)\subset B_{2R}^{\overline M}(o)$, where $B_{a/2}^{M}(p_k)$ and $B_{a/2}^{M}(p_j)$ denote the intrinsic balls of $M$ of radius $a/2$, center in $p_k$ and $p_j$, respectively.
Let $\{e_1,e_2,\ldots,e_n\}$ be a orthonormal basis of $T_xM$. If $x\in B_{a/2}^{M}(p_j)$, then the function extrinsic distance to $p_j$, denoted by $r_j(x)={{\rm dist}\,}_{\overline M}(x,p_j)$, satisfies $$\begin{aligned}
\overline{\nabla}^2r_j(e_i,e_i)&=&\langle\overline{\nabla}_{e_i}\overline{\nabla} r_j,e_i\rangle=\langle \overline{\nabla}_{e_i}\nabla r_j,e_i\rangle+\langle \overline{\nabla}_{e_i}(\overline{\nabla}r_j)^{\perp},e_i\rangle\\
&=&\langle {\nabla}_{e_i}\nabla r_j,e_i\rangle+\langle(\overline{\nabla}_{e_i}\nabla r_j)^{\perp},e_i\rangle-\langle A(e_i,e_i), \overline{\nabla}r_j\rangle\\
&=&{\nabla}^2r_j(e_i,e_i)-\langle {\bf H}, \overline{\nabla}r_j\rangle.
\end{aligned}$$ Observe that $\overline M$ has bounded locally geometry, this is, there exists positive real numbers $k$ and $i_0$ so that the sectional curvature of $\overline M$ is bounded above by $k$ and the injectivity radius of $\overline M$ is bounded below by $i_0$ in a neighborhood of a point $o\in \overline M$. By choosing $R>0$ such that $2R<\min\{ i_0,1/\sqrt k\}$, it follows from Hessian comparison of the distance (see Lemma 7.1, [@ColdingMinicozzi2011]), that $$\overline\nabla^2r_j(e_i,e_i)\geq-\sqrt k+\dfrac{1}{r_j}|e_i-\langle e_i,\overline{\nabla}r_j\rangle\overline{\nabla}r_j|^2$$ in $\overline B_{2R}^{\overline M}(o).$ Hence, in $\overline B_{2R}^{\overline M}(o)\cap M$, $$\begin{aligned}
\Delta r_j&=&\sum_{i=1}^{n}\nabla^2r_j(e_i,e_i)=\sum_{i=1}^{n}\overline\nabla^2r_j(e_i,e_i)+\langle {\bf H}, \overline{\nabla}r_j\rangle\\
&\geq&\sum_{i=1}^{n}\left(-\sqrt k+\dfrac{1}{r_j}|e_i-\langle e_i,\overline{\nabla}r_j\rangle\overline{\nabla}r_j|^2\right)+\langle {\bf H}, \overline{\nabla}r_j\rangle+\langle (\overline{\nabla}f)^{\perp}, \overline{\nabla}r_j\rangle-\langle (\overline{\nabla}f)^{\perp}, \overline{\nabla}r_j\rangle\\
&=&-n\sqrt k+\dfrac{n}{r_j}-\dfrac{|\nabla r_j|^2}{r_j}+\langle {\bf H}_f, \overline{\nabla}r_j\rangle-\langle (\overline{\nabla}f)^{\perp}, \overline{\nabla}r_j\rangle\\
&\geq&-n\sqrt k+\dfrac{n}{r_j}-\dfrac{|\nabla r_j|^2}{r_j}-|{\bf H}_f|-|(\overline\nabla f)^{\perp}|.
\end{aligned}$$
By hypothesis, the norm of ${\bf H}_f$ is bounded above and $\sup_{p\in\overline B_{2R}^{\overline M}(o)}|\overline\nabla f(p)|<\infty$ for each $R>0$. Thus, $$\begin{aligned}
\Delta r_j
&\geq&-n\sqrt k+\dfrac{n}{r_j}-\dfrac{|\nabla r_j|^2}{r_j}-\sup_{p\in B_{2R}^{\overline M}(o)\cap M}|{\bf H}_f(p)|-\sup_{p\in B_{2R}^{\overline M}(o)\cap M}|(\overline\nabla f(p))^{\perp}|\\
&\geq&\dfrac{n}{r_j}-\dfrac{|\nabla r_j|^2}{r_j}-C,
\end{aligned}$$ where $$C=n\sqrt k+\sup_{p\in B_{2R}^{\overline M}(o)\cap M}|{\bf H}_f(p)|+\sup_{p\in \overline B_{2R}^{\overline M}(o)}|\overline\nabla f(p)|.$$ Therefore, in $B_{2R}^{\overline M}(o)\cap M$, $$\Delta r_j^2=2r_j\Delta r_j+2|\nabla r_j|^2\geq2r_j\left(\dfrac{n}{r_j}-\dfrac{|\nabla r_j|^2}{r_j}-C\right)+2|\nabla r_j|^2
= 2n-2C r_j.$$ By choosing $a<\min\{2n/C,2R\},$ we have for $0<\zeta\leq a/2,$ $$\begin{aligned}
\label{e1}
\int_{B_{\zeta}^{M}(p_j)}(2n-2C r_j)\,d\sigma&\leq&\int_{B_{\zeta}^{M}(p_j)}\Delta r^2_j\,d\sigma=\int_{\partial B_{\zeta}^{M}(p_j)}\langle\nabla r^2_j,\nu\rangle\,d\sigma\nonumber\\
&\leq&\int_{\partial B_{\zeta}^{M}(p_j)} 2r_j|\nabla r_j||\nu| dA\leq\int_{\partial B_{\zeta}^{M}(p_j)} 2r_j\,dA\\
&\leq& 2\zeta A_j
(\zeta)\nonumber,
\end{aligned}$$ where $\nu$ denotes the unit normal vector field pointing out of $\partial B_{\zeta}^{M}(p_j)$ and $A_j(\zeta)$ denotes the area of $\partial B_{\zeta}^{M}(p_j)$. By using co-area formula in (\[e1\]), we have $$\begin{aligned}
\int_{B_{\zeta}^{M}(p_j)}(n-C r_j)\,d\sigma&=&\int_0^\zeta\int_{\partial B_{t}^{M}(p_j)}(n-C r_j)|\nabla r_j|^{-1}dA_t\,dt\\
&\geq&\int_0^\zeta\int_{\partial B_{t}^{M}(p_j)}(n-C r_j)dA_t\,dt\\
&\geq&\int_0^\zeta(n-C t)\int_{\partial B_{t}^{M}(p_j)}dA_t\,dt\\
&\geq&(n-C\zeta)V_j(\zeta),
\end{aligned}$$ where $V_j(\zeta)$ denotes the volume of $B_{\zeta}^{M}(p_j)$. Therefore, $$\begin{aligned}
(n-C\zeta)V_j(\zeta)\leq\zeta A_j(\zeta).
\end{aligned}$$ Since $$V_j'(\zeta)=\dfrac{d}{d\sigma}\int_{B_{\zeta}^{M}(p_j)}\,d\sigma=\dfrac{d}{d\sigma}\int_0^\zeta\int_{\partial B_{t}^{M}(p_j)}|\nabla r_j|^{-1}dA_t\,dt=\int_{\partial B_{\zeta}^{M}(p_j)}|\nabla r_j|^{-1}dA\geq A_j(\zeta),$$ then $$(n-C\zeta)V_j(\zeta)\leq \zeta V_j'(\zeta).$$ Thus, $$\label{e2}
\dfrac{d}{d\sigma}\log V_j(\zeta)=\dfrac{V_j'(\zeta)}{V_j(\zeta)}\geq\dfrac{n}{\zeta}-C.$$ By integrating (\[e2\]) from $\varepsilon>0$ to $\zeta$, we obtain $$\log V_j(\zeta)-\log V_j(\varepsilon)\geq n\log\zeta-n\log\varepsilon-C(\zeta-\varepsilon),$$ that is $$\dfrac{V_j(\zeta)}{V_j(\varepsilon)}\geq\dfrac{\zeta^n}{\varepsilon^n}e^{-C(\zeta-\varepsilon)}.$$ Now observing that $$\lim_{\varepsilon\rightarrow0^+}\dfrac{V_j(\varepsilon)}{\varepsilon^n}=\omega_n,$$ we obtain $$V_j(\zeta)\geq\omega_n\zeta^ne^{-C\zeta}$$ for $0<\zeta\leq a/2.$ Thus, we conclude that $$\begin{aligned}
{Vol}_f(M)&=&\int_{M}e^{-f}\,d\sigma\geq\sum_{j=1}^{\infty}\int_{B_{a/2}^{M}(p_j)}e^{-f}\,d\sigma\\
&\geq&\left(\inf_{\overline B_{2R}^{\overline M}(o)}e^{-f}\right)\sum_{j=1}^{\infty}V_j(a/2)=+\infty.
\end{aligned}$$ This contradicts the assumption of the finite weighted volume of $M$. Therefore, $M^n$ is a proper hypersurface of $\overline M^m_f.$
The function $f\in C^{\infty}(\overline M)$ is said [*convex*]{} if the Hessian of $f$ is non-negative, this is $\overline{\nabla}^2f\geq0.$
\[remark1\] Let $\overline M_f$ be a complete gradient Ricci soliton satisfying $\overline{{Ric}}_f=\dfrac{1}{2} g.$ In this case, Cao and Zhou [@CaoZhou2010] showed that by translating $f$ $$\label{ae} \overline R+|\overline\nabla f|^2-f=0 \ \ \ \ \text{e}\ \ \ \ \ \overline R\geq0.$$ Thus, it follows from the equations in (\[ae\]) that $$\label{equ1}
|\overline\nabla f|^2\leq f.$$ In addition, there exists constants $c_1,c_2\in{\mathbb{R}}$ such that $$\label{equa3} \dfrac{1}{4}(r(x)-c_1)^2\leq f(x)\leq \dfrac{1}{4}(r(x)+c_2)^2,$$ where $r(x)={{\rm dist}\,}_{\overline M}(x,o)$ is the distance of $x\in \overline M$ to a fixed point $o\in \overline M$. The constant $c_2$ depends only on the dimension of the manifold and $c_1$ depends on the geometry of $g$ on unit ball center in $o$ (see [@CaoZhou2010], Lemma 2.1, Lemma 2.2 and Theorem 1.1). In [@MunteanuWang2012], Munteanu and Wang showed that the inequalities in (\[equa3\]) are true only assuming that $\overline{{Ric}}_f\geq\dfrac{1}{2}g$ and $|\overline{\nabla}f|^2\leq f.$
By hypothesis, $\overline{{Ric}}_f=\dfrac{1}{2}g.$ Thus, it follows from Remark \[remark1\] that $$\overline R+|\overline\nabla f|^2-f=0, \ \ \ \ \ \overline R+\overline{\Delta} f=\dfrac{m}{2}, \ \ \ \ \ \text{and} \ \ \ \ \ \overline R\geq0,$$ and we have that $$\overline{\Delta}f-|\overline{\nabla} f|^2+f=\dfrac{m}{2} \ \ \ \ \text{and} \ \ \ \ |\overline\nabla f|^2\leq f.$$
By being $f$ a convex function, we obtain $$\begin{aligned}
\Delta_ff+f&=&\Delta f-|\nabla f|^2+f=\overline\nabla^2f(e_i,e_i)+\langle{\bf H},\overline\nabla f^{\perp}\rangle-|\nabla f|^2+f\\
&=&\overline{\Delta}f-\sum_{i=n+1}^m\overline{\nabla}^2 f(\eta_i,\eta_i)+\langle{\bf H}_f,\overline\nabla f^{\perp}\rangle-|\overline{\nabla}f^{\perp}|^2-|\nabla f|^2+f\\
&=&
\overline{\Delta}f-\sum_{i=n+1}^m\overline{\nabla}^2 f(\eta_i,\eta_i)+\langle{\bf H}_f,\overline\nabla f^{\perp}\rangle-|\overline{\nabla}f|^2+f\\
&\leq&\dfrac{m}{2}+C,\\
\end{aligned}$$ where $C=\sup_{x\in M}\langle{\bf H}_f,\overline\nabla f^{\perp}\rangle<\infty.$
Observe that $$\label{equa33}
\dfrac{1}{4}(r(x)-c)^2\leq f(x)\leq \dfrac{1}{4}(r(x)+c)^2,$$ where $c$ is a constant (see Remark \[remark1\], inequalities in (\[equa3\])). Hence, we can conclude that $f$ is proper on $\overline M$. Since, by hypothesis, $x:M\rightarrow\overline M_f$ is a proper immersion, then $f|_M:M\rightarrow{\mathbb{R}}$ is a proper function.
Therefore, it follows from Theorem 1.1 de [@ChengZhou2013] that $M$ has finite weighted volume and Euclidean volume growth of the sub-level set of the potential function $f$.
The $f$-stability Index of the Constant Weighted Mean Curvature Hypersurfaces
=============================================================================
In this section, we will prove Theorem \[it2\], Corollary \[ic1\], and Theorem \[itheorem2\]. Which are results about the $f$-stability index of a constant weighted mean curvature hypersurface with finite weighted volume and isometrically immersed in a gradient Ricci soliton that admits at least one parallel field globally defined. For this, we are going to give some definitions and known results.
We say that a vector field $X\in T\overline M$ is [*parallel*]{} if $$\overline\nabla_YX=0$$ for all vector fields $Y\in T\overline M.$
\[proposition1\] Let $\overline M_f^{n+1}$ be a gradient Ricci soliton satisfying $\overline{{Ric}}_f=kg$ and $X$ a parallel vector field on $\overline M_f^{n+1}$. If $M^n$ is a constant weighted mean curvature hypersurface immersed isometrically immersed in $\overline M_f^{n+1},$ then $$L_f\langle X,\eta\rangle=k\langle X,\eta\rangle$$ and $$\Delta_f\langle X,\eta\rangle^{2}=-2|A|^{2}\langle X,\eta\rangle^{2}+2| AX^{\top}|^2,$$ where $\eta$ is the unit normal vector field to $M$.
Let $\{e_1,e_2,\ldots,e_n\}$ be a geodesic orthonormal frame on $M$. By hypothesis, $H_f=C$, this is $H=\langle\overline{\nabla }f,\eta\rangle+C,$ where $C$ is a real constant. Thus, $$\begin{aligned}
\label{equation12}
\nabla H&=&\sum_{i=1}^ne_i(H)e_i=\sum_{i=1}^ne_i( \langle \overline{\nabla}f,\eta\rangle) e_i=\sum_{i=1}^n\langle \overline{\nabla}_{e_i}\overline{\nabla}f,\eta\rangle e_i+\sum_{i=1}^n\langle\overline{\nabla }f,\overline{\nabla }_{e_i}\eta\rangle e_i\nonumber\\
&=&\sum_{i=1}^n\langle \overline{\nabla}_{e_i}\overline{\nabla}f,\eta\rangle e_i-\sum_{i=1}^n\langle A(e_i,e_j),\eta\rangle\langle\overline{\nabla }f,e_j\rangle e_i.
\end{aligned}$$ To $u=\langle X,\eta\rangle$ and $a_{ij}=\langle Ae_i,e_j\rangle$, we have $$\label{equation11}
\nabla u=\sum_{j=1}^ne_j(u)e_j=\sum_{j=1}^n\langle\overline{\nabla }_{e_j}\eta,X\rangle e_j=-\sum_{i,j=1}^na_{ji}\langle e_i,X\rangle e_j.$$ It follows from (\[equation12\]) and (\[equation11\]), $$\langle \nabla H,X\rangle=\sum_{i=1}^n\langle \overline{\nabla}_{e_i}\overline{\nabla}f,\eta\rangle \langle e_i,X\rangle-\sum_{i,j=1}^na_{ij}\langle\overline{\nabla }f,e_j\rangle \langle e_i,X\rangle=\langle \overline{\nabla}_{ {X^{\top}}}\overline{\nabla}f,\eta\rangle +\langle\overline{\nabla}f,\nabla u\rangle.$$ Moreover, $$e_i(u)=\langle\overline{\nabla }_{e_i}\eta,X\rangle=-\sum_{j=1}^na_{ij}\langle e_j,X\rangle$$ and, by deriving the previous expression and observing that $\nabla_{e_k}e_j=0$, we obtain $$e_k(e_i(u))=-\sum_{j=1}^n\left(a_{ij,k}\langle e_j,X\rangle+a_{ij}\langle X,\overline{\nabla}_{e_k}e_j\rangle\right)=-\sum_{j=1}^na_{ij,k}\langle e_j,X\rangle-\sum_{j=1}^na_{ij}a_{kj}\langle X,\eta\rangle.$$ It follows from Codazzi equation that $$\overline R(e_j,e_k)e_i^{\perp}=\left(a_{ki,j}-a_{ji,k}\right)\eta,$$ that is $$\langle \overline R(e_j,e_k)e_i,\eta\rangle=a_{ki,j}-a_{ji,k}.$$ Hence, $$\begin{aligned}
e_k(e_i(u))&=&-\sum_{j=1}^na_{ki,j}\langle e_j,X\rangle+\sum_{j=1}^n\langle \overline R(e_j,e_k)e_i,\eta\rangle\langle e_j,X\rangle-\sum_{j=1}^na_{ij}a_{kj}\langle X,\eta\rangle
\end{aligned}$$ and $$\begin{aligned}
\label{equation13}
\Delta u&=&\sum_{i=1}^ne_i(e_i(u))=-\sum_{i,j=1}^na_{ii,j}\langle e_j,X\rangle+\sum_{i,j=1}^n\langle \overline R(e_j,e_i)e_i,\eta\rangle\langle e_j,X\rangle-\sum_{i,j=1}^na_{ij}a_{ij}\langle X,\eta\rangle\nonumber\\
&=&\langle\nabla H,X\rangle+\sum_{i=1}^n\langle\overline R(X^{\top},e_i)e_i,\eta\rangle-|A|^{2}\langle X,\eta\rangle\nonumber\\
&=&\langle \overline{\nabla}_{ {X^{\top}}}\overline{\nabla}f,\eta\rangle +\langle\overline{\nabla}f,\nabla u\rangle+\overline{Ric}(X^{\top},\eta)-|A|^{2}u.
\end{aligned}$$ On the other hand, $$\label{equation14}0=\dfrac{1}{2}\langle X^{\top},\eta\rangle=\overline{Ric}_f(X^{\top},\eta)=\overline{Ric}(X^{\top},\eta)+\overline{\nabla}^{2}f(X^{\top},\eta)=\overline{Ric}(X^{\top},\eta)+\langle \overline{\nabla}_{ {X^{\top}}}\overline{\nabla}f,\eta\rangle.$$ Therefore, it follows from (\[equation13\]) and (\[equation14\]), $$\begin{aligned}
\label{equ11}
\Delta_fu&=&\Delta u-\langle\nabla f,\nabla u\rangle\nonumber\\
&=&\langle \overline{\nabla}_{ {X^{\top}}}\overline{\nabla}f,\eta\rangle +\langle\overline{\nabla}f,\nabla u\rangle-\langle \overline{\nabla}_{ {X^{\top}}}\overline{\nabla}f,\eta\rangle-|A|^{2}u-\langle\nabla f,\nabla u\rangle\nonumber\\
&=&-|A|^{2}u,
\end{aligned}$$ by implying that $$L_fu=\Delta_fu+|A|^2u+ku=ku.$$ Moreover, by using the equality (\[equ11\]), we have that $$\Delta_fu^{2}=2u\Delta_fu+2|\nabla u|^{2}=-2|A|^{2}u^{2}+2|AX^{\top}|^{2}.$$
Let $\mathcal P_{\overline M}$ be the set of all tangent vector fields to $\overline M$ which are parallel and globally defined.
O shrinking Gaussian soliton $\left({\mathbb{R}}^{n+1},g_{can},e^{-|x|^2/2}\right)$ has exactly $n+1$ parallel vector fields linearly independent and globally defined on ${\mathbb{R}}^{n+1}$. Other example is the shrinking cylinder soliton $\left({\mathbb{S}}^{n+1-k}\times{\mathbb{R}}^k,g,e^{-f}\right)$, $k\geq1$, with metric $$g=2(n-k-1)g_{{\mathbb{S}}^{n-k}}+g_{{\mathbb{R}}^{k}}$$ and potential function $$f(\theta,x)=\frac{|x|^2}{2}, \ \theta\in {\mathbb{S}}^{n-k}, \ x\in{\mathbb{R}}^k.$$ In this example, we have that $$\dim\mathcal P_{{\mathbb{S}}^{n+1-k}\times{\mathbb{R}}^k}=k.$$
The vector subspace of $C^{\infty}(M)$ [*generated*]{} by $E\subset C^{\infty}(M)$, denoted by ${{\rm Span}\,}E,$ is the set of all the linear combinations of the elements of $E$.
\[lema2\] Let $\overline M_f^{n+1}$ be a shrinking gradient Ricci soliton satisfying $\overline{{Ric}}_f=kg$ and $M^n$ be a constant weighted mean curvature hypersurface isometrically immersed in $\overline M_f^{n+1}.$ If $M$ is compact, then $I_f$ is negative defined in the $${{\rm Span}\,}\{1,\langle X,\eta\rangle: \ X \in\mathcal P_{\overline M_f}\}.$$ Moreover, $$\int_M|A|^2\langle X,\eta\rangle e^{-f}d\sigma=0.$$
It follows from Proposition \[proposition1\] that the function $u=\langle X,\eta\rangle$, with $X\in\mathcal P_{\overline M_f}$, satisfies $L_fu=ku.$ Thus, since $M$ is compact, we have $$\begin{aligned}
\int_Mkue^{-f}d\sigma&=&\int_ML_fue^{-f}d\sigma=\int_M\left(\Delta_fu+|A|^2u+ku\right)e^{-f}d\sigma\\
&=&\int_M|A|^2ue^{-f}d\sigma+\int_Mkue^{-f}d\sigma.
\end{aligned}$$ Therefore, $$\label{eeee7}
\int_M|A|^2ue^{-f}d\sigma=0.$$ Observe that $$\label{eeee8}
I_f(1,1)=-\int_M1L_f1e^{-f}d\sigma=-\int_M\left(|A|^2+k\right)e^{-f}d\sigma$$ and $$\label{eeee9}I_f(u,u)=-\int_MuL_fue^{-f}\,d\sigma=-k\int_Mu^2e^{-f}\,d\sigma.$$ Therefore, $$\begin{aligned}
I_f(c_0+u,c_0+u)&=& I_f(c_0,c_0)+I_f(u,u)+2I_f(c_0,u)\\
&=& -\int_M\left[c_0^2|A|^2+kc_0^2+ku^2+2c_0\left(\Delta_fu+|A|^2u+ku\right)\right]e^{-f}d\sigma\\
&=& -\int_M\left[c_0^2|A|^2+kc_0^2+ku^2+2c_0ku\right]e^{-f}d\sigma\\
&=& -c_0^2\int_M|A|^2e^{-f}d\sigma-k\int_M(c_0+u)^{2}e^{-f}d\sigma<0,
\end{aligned}$$ where $u=\langle X,\eta\rangle.$ This shows that $I_f$ is negative defined in ${{\rm Span}\,}\{1,\langle X,\eta\rangle: \ X\in\mathcal P_{\overline M_f}\}.$
Supposing that $M^n$ has finite weighted volume and putting $\alpha=-\int_M\langle X,\eta\rangle e^{-f}d\sigma,$ we can conclude that $$\int_M(\alpha+\langle X,\eta\rangle )e^{-f}d\sigma=0.$$ Therefore, $$\mathcal F\cap{{\rm Span}\,}\{1,\langle X,\eta\rangle: X \in\mathcal P_{\overline M_f}\}\not=\emptyset.$$
Now, let’s look at the noncompact manifolds. For this, we will consider the functions that have compact support in $M$.
\[proposition3\] Let $\overline M_f^{n+1}$ be a weighted manifold and $M^n$ be a noncompact hypersurface isometrically immersed in $\overline M_f^{n+1}.$ Then $$I_f(\phi u,\phi u)=-\int_M\phi^2uL_fue^{-f}d\sigma+\int_M|\nabla\phi|^2u^2e^{-f}d\sigma,$$ where $\phi\in C_c^{\infty}(M)$, $u\in C^{\infty}(M)$, and $\eta$ denotes the unit normal field on $M$. Moreover, if $\overline{{Ric}}_f=kg$ and $M$ has constant weighted mean curvature, then $$\label{equation7}\int_M\phi^2|A|^2\langle X,\eta\rangle e^{-f}d\sigma=-2\int_M\phi \langle\nabla\phi,AX^{\top}\rangle e^{-f}d\sigma,$$ where $X$ is a parallel vector field on $\overline M_f^{n+1}$.
Note that $$\begin{aligned}
I_f(\phi u,\phi u)&=&-\int_M(\phi u)L_f(\phi u)e^{-f}d\sigma\\
&=&-\int_M\left[(\phi u)\Delta_f(\phi u)+\left(|A|^2+\overline{{Ric}}_f(\eta,\eta)\right)\phi^2 u^2\right]e^{-f}d\sigma\\
&=&-\int_M\left[\phi^2u\Delta_fu+\phi u^2\Delta_f\phi+2\phi u\langle\nabla\phi,\nabla u\rangle+\left(|A|^2+\overline{{Ric}}_f(\eta,\eta)\right)\phi^2 u^2\right]e^{-f}d\sigma,
\end{aligned}$$ As $\phi$ has compact support, then $$0=\int_M{div}(\phi u^2e^{-f}\nabla\phi)d\sigma=\int_M\left(\phi u^2\Delta_f\phi+2u\phi\langle\nabla u,\nabla\phi\rangle+u^2|\nabla\phi|^2\right)e^{-f}d\sigma.$$ Therefore, by using the last two expressions, we obtain $$\begin{aligned}
I_f(\phi u,\phi u)&=&-\int_M\left[\phi^2u\Delta_fu-|\nabla\phi|^2u^2+\left(|A|^2+\overline{{Ric}}_f(\eta,\eta)\right)\phi^2 u^2\right]e^{-f}d\sigma\\
&=&-\int_M\phi^2uL_fue^{-f}d\sigma+\int_M|\nabla\phi|^2u^2e^{-f}d\sigma.
\end{aligned}$$ Moreover, supposing $\overline{{Ric}}_f=kg$, $H_f$ is constant, and $X$ is a parallel field on $\overline M_f$, we obtain $$\begin{aligned}
\int_M\phi^2k\langle X,\eta\rangle e^{-f}d\sigma&=&\int_M\phi^2 L_f\langle X,\eta\rangle e^{-f}d\sigma\\
&=&-\int_M\langle \nabla\phi^2,\nabla\langle X,\eta\rangle\rangle e^{-f}d\sigma+\int_M\left(|A|^2+k\right)\phi^2\langle X,\eta\rangle e^{-f}d\sigma.
\end{aligned}$$ Hence, $$\int_M|A|^2\phi^2\langle X,\eta\rangle e^{-f}d\sigma=\int_M\langle \nabla\phi^2,\nabla\langle X,\eta\rangle\rangle e^{-f}d\sigma=-2\int_M \phi \langle\nabla\phi,AX^{\top}\rangle e^{-f}d\sigma.$$
\[lema5\] Let $\overline M_f^{n+1}$ be a shrinking gradient Ricci soliton satisfying $\overline{{Ric}}_f=kg$ and $M^n$ be a constant weighted mean curvature hypersurface isometrically immersed in $\overline M_f^{n+1}.$ If $M$ is noncompact and finite weighted volume, there exists $\phi\in C^{\infty}_c(M)$ such that $I_f$ is negative defined in $\phi V$ and $\dim(\phi V)=\dim V$, where $$V={\rm {{\rm Span}\,}}\left\{1,\langle X,\eta\rangle: \ X \ \text{is a parallel field on }\overline M^{n+1}_f\right\}.$$ Moreover, by assuming that $\int_M|A|^2e^{-f}d\sigma<\infty$, we have $$\int_M|A|^2\langle X,\eta\rangle e^{-f}d\sigma=0.$$
Let $u=c_0+\langle X,\eta\rangle$, where $c_0$ is a real constant and $X$ is a parallel field on $\overline M^{n+1}_f$. As $H_f$ is constant, then by Proposition \[proposition1\], we have $$\begin{aligned}
\label{equation15}
L_f u&=&\Delta_f u+\left(|A|^2+k\right)u=L_f\langle X,\eta\rangle+\left(|A|^2+k\right)c_0\nonumber\\
&=&k\langle X,\eta\rangle+\left(|A|^2+k\right)c_0.
\end{aligned}$$ It follows from Proposition \[proposition3\] and equality (\[equation15\]) that $$\begin{aligned}
I_f(\phi u,\phi u)&=&-\int_{M}\phi^2uL_fue^{-f} d\sigma+\int_{M}|\nabla\phi|^2u^2 e^{-f}d\sigma\\
&=&-\int_{M}\phi^2u\left(k\langle X,\eta\rangle+\left(|A|^2+k\right)c_0\right) e^{-f}d\sigma+\int_{M}|\nabla\phi|^2u^2 e^{-f}d\sigma\\
&=&-k\int_{M}\phi^2u^2e^{-f}d\sigma-\int_{M}\phi^2|A|^2c_0^2e^{-f}d\sigma-\int_{M}\phi^2|A|^2c_0\langle X,\eta\rangle e^{-f}d\sigma\\
&&+\int_{M}|\nabla\phi|^2u^2 e^{-f}d\sigma.
\end{aligned}$$ Now, by using once more the Proposition \[proposition3\] and Cauchy-Schwarz’s inequality, $$\begin{aligned}
\left|\int_M\phi^2|A|^2c_0\langle X,\eta\rangle e^{-f}d\sigma\right|&=& 2\left|\int_M\phi \langle\nabla\phi, AX^{\top}\rangle c_0 \ e^{-f}d\sigma\right|\\
&\leq& 2\int_M|\phi|\,| \nabla\phi|\,|A|\,|X^{\top}|\, |c_0| \ e^{-f}d\sigma\\
&\leq&\int_M\phi^2|A|^2 c_0^2\, e^{-f}d\sigma+\int_M|\nabla\phi|^2|X^{\top}|^2e^{-f}d\sigma.
\end{aligned}$$ Therefore, $$\label{eq6}
I_f(\phi u,\phi u)\leq-k\int_M\phi^2 u^2e^{-f}d\sigma+\int_M|\nabla \phi|^2(u^2+|X^{\top}|^2)e^{-f}d\sigma.$$ Let $r(x)$ be the extrinsic distance from $x\in M$ to a fixed point $o\in \overline M_f$. For $R>0$ sufficiently large, define the function $\phi_R:M\rightarrow{\mathbb{R}}$ such that $$\label{equation5}
\phi_R(x)=\left\{
\begin{array}{lcl}
1, & r(x)\leq R;\\
\dfrac{2R-r(x)}{R}, & R\leq r(x)\leq2R;\\
0,& r(x)\geq2R.
\end{array}
\right.$$ Observe that $|\nabla\phi_R|\leq1/R$ and $\phi_R\in C^{\infty}_c(M)$ because $M$ is proper (see Proposition \[ip4\]). Now, by substituting $\phi=\phi_R$ in the inequality (\[eq6\]), we obtain $$I_f(\phi_R u,\phi_R u)\leq-k\int_M\phi^2_R u^2e^{-f}d\sigma+\dfrac{|X|^2+c_0^2+2|c_0|\,|X|}{R^2}\int_{M\cap(B_{2R}\setminus B_R)}e^{-f}d\sigma,$$ (recall that $|X|$ is constant because $X$ is parallel). Since ${Vol}_f(M)<\infty$ and $|X|$ is constant, then for each $u\in V$ there exists $R_u$ sufficiently large such that $$I_f(\phi_{R_u}u,\phi_{R_u}u)<0.$$
Let’s find a function $\phi\in C^{\infty}_c(M)$ that is not dependent of the function $u$. In fact, we consider the subset $$S=\left\{u\in V: \ \int_Mu^2e^{-f}d\sigma=1\right\}.$$ Note that $V\subset L_f^2(M)$ is a subspace of finite dimension smaller than or equal to $\dim \mathcal P_{\overline M_f}+1$. Hence, $S$ is a compact set in $L_f^2(M)$. Thus, there exists a positive real number $R_0$ such that any function $u\in S$ vanishes identically on $M\cap B_{R_0}^{\overline M_f}(o)$. Otherwise, we could get a sequence ${R}_j\rightarrow\infty$ of positive numbers so that for each $j$ exists $u_j\in S$ with $u_j\equiv0$ on $M\cap B_{R_0}^{\overline M_f}(o)$. Hence, we would have $$u=\lim_{j\rightarrow\infty}u_j \ \in \ S$$ and $u\equiv0$ on $M.$ However, this is not possible because if $u\in S$, then $$\int_M u^2e^{-f}d\sigma=1.$$ Therefore, to $R$ sufficiently large and $R\geq R_0$, and for any function $u\in S$, we have $$I_f(\phi_R u,\phi_R u)\leq-k\int_M\phi^2_R u^2e^{-f}d\sigma+\dfrac{|X|^2+c_0^2+2|c_0||X|}{R^2}\int_{M\cap(B_{2R}\setminus B_R)}e^{-f}d\sigma<0$$ because $$M(R)=\int_M\phi_R^2u^2e^{-f}d\sigma >0$$ is an increasing function on $R$, and as ${Vol}_f(M)<\infty$, $$\lim_{R\rightarrow\infty}\dfrac{|X|^2+c_0^2+2|c_0||X|}{R^2}\int_{M\cap(B_{2R}\setminus B_R)}e^{-f}d\sigma=0.$$ Now, we can find $\phi=\phi_R$ independent of $u$ such that $I_f(\phi u,\phi u)<0$ for all $u\in S$. If $u\in V$, then $\dfrac{1}{|u|_{L^{2}_f}}u\in S$, and thus, for $u\not\equiv0$, $$I_f(\phi u,\phi u)=|u|^{2}_{L^{2}_f}I_f\left(\phi \dfrac{u}{|u|_{L^{2}_f}},\phi \dfrac{u}{|u|_{L^{2}_f}}\right)<0.$$
Now, we will show that $\dim V=\dim(\phi V).$ In fact, let $\{u_1,u_2,\ldots,u_s\}$ be a orthonormal basis to the vector subspace $V\subset L^{2}_f(M)$. For the function $\phi$ built here, we have that $u_i\not\equiv0$ on $M\cap B_R^{\overline M_f}(o).$ Therefore, $\{\phi u_1,\phi u_2,\ldots,\phi u_s\}$ is linearly independent, so we can conclude that $\dim(\phi V)=\dim V.$
Finally, it follows from Proposition \[proposition3\], equation (\[equation7\]) with $\phi=\sqrt\phi_j$, and Cauchy–Schwarz inequality that $$\begin{aligned}
\label{equation9}\left|2\int_M\phi_j|A|^2\langle X,\eta\rangle e^{-f}d\sigma\right|&=&\left|-2\int_M \langle\nabla\phi_j,AX^{\top}\rangle e^{-f}d\sigma\right|\\
&\leq&\left(\int_M|\nabla \phi_j|^2e^{-f}d\sigma\right)^{\frac{1}{2}}\left(\int_M |AX^{\top}|^2e^{-f}d\sigma\right)^{\frac{1}{2}}.
\end{aligned}$$ Now, putting $R=j$ in the function defined in (\[equation5\]) and reviewing that $|X|$ is constant, we get $$\label{equation6}\left|2\int_M\phi_j|A|^2\langle X,\eta\rangle e^{-f}d\sigma\right|
\leq\frac{|X|}{j}\left(\int_{M\cap(B_{2j}\setminus B_j)}e^{-f}d\sigma\right)^{\frac{1}{2}}\left(\int_M |A|^2e^{-f}d\sigma\right)^{\frac{1}{2}}.$$ By hypothesis $M$ has finite weighted volume and $\int_M|A|^2e^{-f}d\sigma<\infty,$ hence the right-hand side of (\[equation6\]) tends to zero as $j\rightarrow\infty$, and we can conclude that $$\label{equation8}
\lim_{j\rightarrow\infty}\int_M\phi_j|A|^2\langle X,\eta\rangle e^{-f}d\sigma=0.$$ Since $$\lim_{j\rightarrow\infty}(\phi_j|A|^2\langle X,\eta\rangle)(x)=(|A|^2\langle X,\eta\rangle)(x)$$ for each $x\in M$, $$|\phi_j|A|^2\langle X,\eta\rangle|\leq|A|^2|\langle X,\eta\rangle|$$ and $$0\leq\int_M|A|^2|\langle X,\eta\rangle| e^{-f}d\sigma\leq|X|^2\int_M|A|^2e^{-f}d\sigma<\infty,$$ then using the dominated convergence theorem and expression (\[equation8\]), we get $$\int_M|A|^2\langle X,\eta\rangle e^{-f}d\sigma=0.$$
Let $V\equiv {{\rm Span}\,}\{1,\langle X,\eta\rangle: \ X\in\mathcal P_{\overline M_f}\}$. Since $1\not\in \{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\}$ by hypothesis, $$\dim V=1+\dim \{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\}.$$ By Lemmas \[lema2\] and \[lema5\], there exists a function $\phi\in C_c^{\infty}(M)$ such that $\dim\phi V=\dim V$ and $I_f$ is negative defined in $\phi V.$ In compact hypersurfaces, choose $\phi\equiv1$. Recall that the ${Ind}_f(M)$ is the maximal dimension of a subspace of $\mathcal F\cap C_c^{\infty}(M)$ which $I_f$ is negative defined, with $\mathcal F=C_c^{\infty}(M)$ if $M$ is a $f$-minimal and $$\mathcal F=\left\{u\in C^{\infty}_c(M); \ \int_Mue^{-f}d\sigma=0\right\}$$ if $M$ has constant weighted mean curvature.
Now, observe that $$\dim\{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\}\leq\dim (\mathcal F\cap \phi V).$$ In fact, $|X|$ is constant because $X$ is a parallel field, and since the weighted volume of $M$ is finite, we obtain that $\int_M\phi\langle X,\eta\rangle e^{-f}d\sigma\leq|X|\int_M\phi e^{-f}d\sigma<\infty$. Hence, there exists a real number $c_0$ satisfying $$\int_M\phi(c_0+\langle X,\eta\rangle)e^{-f}d\sigma=0.$$ Therefore, $\phi(c_0+\langle X,\eta\rangle)\in \mathcal F\cap \phi V$ whenever $\langle X,\eta\rangle\in\{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\}$. Since $I_f$ is negative defined in $\mathcal F\cap \phi V$, we conclude that $$\dim\{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\}\leq\dim (\mathcal F\cap \phi V)\leq{Ind}_f(M).$$
Consider the linear transformation $T:\mathcal P_{\overline M_f}\rightarrow C^{\infty}(M)$ defined by $T(X)=\langle X,\eta\rangle$. Now applying the kernel and image theorem, that turns $$\begin{aligned}
\label{b}
\dim \mathcal P_{\overline M_f}&=&\dim\{X\in\mathcal P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}+\dim\{\langle X,\eta\rangle; \ {X\in\mathcal P_{\overline M_f}}\}\nonumber\\
&\leq&\dim\{X\in\mathcal P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}+{Ind}_f(M).
\end{aligned}$$ Otherwise, if $1\in\{\langle X,\eta\rangle: \ {X\in\mathcal P_{\overline M_f}}\},$ $L_f1=k$ by Proposition \[proposition1\], and by definition $L_f1=|A|^2+k$. Therefore, $|A|^2\equiv0.$
Note that $$\label{a} \dim\{X\in\mathcal P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}\leq \dim\mathcal P_{\overline M_f}-1$$ as long as we assume that exists a field $X_0\in \mathcal P_{\overline M_f}$ such that $\langle X_0,\eta\rangle\not\equiv0$. It follows from Theorem \[it2\] and from the inequality (\[a\]) that $$\dim\mathcal P_{\overline M_f}-{Ind}_f M\leq \ \dim\{X\in\mathcal P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\} \ \leq \dim\mathcal P_{\overline M_f}-1.$$ Therefore, $${Ind}_f M\geq1.$$ Now supposing ${Ind}_f M=1$, we have $$\dim\mathcal P_{\overline M_f}-1\leq \dim\{X\in \mathcal P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}\leq \dim\mathcal P_{\overline M_f}-1.$$ Therefore, $$\dim\{X\in \mathcal P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}=\dim\mathcal P_{\overline M_f}-1.$$
Now, let’s obtain a necessary condition for a constant weighted mean curvature hypersurface $M^n$ with finite weighted volume satisfies the following equality: $${Ind}_f(M)=\dim \mathcal P_{\overline M_f}-\dim\{X\in P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}.$$ For this, we will prove some lemmas, they are adaption of known results.
The weighted Laplacian operator $\Delta_f$ is associated to $e^{-f}d\sigma$ as well as $\Delta$ is associated to $d\sigma.$ This is viewed in the following Green’s theorem version for wighted Laplacian:
\[Green\] Let $\Omega\subset M$ be a compact set and let $u,v\in C^{\infty}(\Omega)$, then $$\label{G1}
\int_{\Omega}(u\Delta_fv)e^{-f}\,d\sigma+\int_{\Omega}\langle\nabla u,\nabla v\rangle e^{-f}\,d\sigma=\int_{\partial \Omega}u\nu(v) e^{-f}\,d\partial \Omega, \ \ \ \text{and}$$ $$\label{G2}
\int_{\Omega}(u\Delta_fv-v\Delta_fu)e^{-f}\,d\sigma=\int_{\partial \Omega}(u\nu(v)-v\nu(u)) e^{-f}\,d\partial \Omega,$$ where $\nu$ denotes the exterior unit normal to $\Omega$ along $\partial\Omega$.
In fact, $$\begin{aligned}
{div}(e^{-f}u\nabla v)&=& e^{-f}u\,{div}(\nabla v)+\langle\nabla(e^{-f}u),\nabla v\rangle\\
&=& e^{-f}u\Delta v-e^{-f}u\langle\nabla f,\nabla v\rangle+e^{-f}\langle\nabla u,\nabla v\rangle\\
&=& e^{-f}u\Delta_f v+e^{-f}\langle\nabla u,\nabla v\rangle.
\end{aligned}$$ Integrating both sides from above inequality and applying the divergent theorem to the field $X=e^{-f}u\nabla v$, it follows the equality (\[G1\]). The equality (\[G2\]) is obtained by integrating the difference $u\Delta_f v-v\Delta_f u$ and by applying the equality (\[G1\]).
\[lemma2\] Suppose that $M$ is a complete hypersurface without boundary. If $u,$ $v$ are $C^2$ functions with $$\label{equation1}
\int_M\left(|u\nabla v|+|\nabla u|\,|\nabla v|+|u\Delta_f v|\right)e^{-f}d\sigma<\infty,$$ then $$\int_M u(\Delta_f v)e^{-f}d\sigma=-\int_M\langle\nabla v,\nabla u\rangle e^{-f}d\sigma.$$
Let $W^{1,2}(e^{-f}d\sigma)$ be the weighted Sobolev space, which is the set of the functions $u$ on $M$ satisfying $$\int_M(u^2+|\nabla u|^2)e^{-f}d\sigma<\infty$$ with the norm $$\|u\|_{W^{1,2}_f}:=\left(\int_M(u^2+|\nabla u|^2)e^{-f}d\sigma\right)^{\frac{1}{2}}.$$ Fixed the notation $W^{1,2}_f=W^{1,2}_f(e^{-f}d\sigma)$ and $L^{2}_f=L^{2}_f(e^{-f}d\sigma)$, we get
\[l2\] Let $M^n$ be a constant weighted mean curvature hypersurface isometrically immersed in a gradient Ricci soliton $\overline M_f$ that satisfies $\overline{{Ric}}_f=kg$. Suppose that $h$ is a $C^2$ function with $L_fh=-\mu h$ for $\mu\in{\mathbb{R}}.$
1. If $h\in W^{1,2}_f$, then $|A|h\in L^2_f$ and $$\label{equation0}
\int_M|A|^2h^2e^{-f}d\sigma\leq\int_M\left((1-k-\mu)h^2+2|\nabla h|^2\right)e^{-f}d\sigma.$$
2. If $h>0$ and $\phi\in W^{1,2}_f,$ then $$\label{equation3}
\int_M\phi^2\left(2|A|^2+|\nabla\log h|^2\right)e^{-f}d\sigma\leq\int_M\left[4|\nabla\phi|^2-2(\mu+k)\phi^2\right]e^{-f}d\sigma.$$
Let $\phi$ be a smooth function with compact support. Note that $$\Delta_f h^2=2|\nabla h|^2+2h\Delta_f h$$ and $$\label{equation19}
\Delta_f h=\left(L_f-|A|^2-k\right)h=-\left(\mu+|A|^2+k\right)h.$$ By Lemma \[Green\] and equality (\[equation19\]), $$\begin{aligned}
\label{equation2}
\int_M\langle\nabla\phi^2,\nabla h^2\rangle e^{-f}d\sigma&=&-\int_M\phi^2\Delta_f h^2 e^{-f}d\sigma\nonumber\\
&=&-2\int_M\phi^2\left[|\nabla h|^2-\left(\mu+|A|^2+k\right)h^2\right]e^{-f}d\sigma.
\end{aligned}$$ Assume now that $\phi\leq1$ and $|\nabla\phi|\leq1$. Rearranging the terms in (\[equation2\]) and using the inequality $$0\leq|\phi\nabla h-h\nabla\phi|^2=\phi^2|\nabla h|^2+h^2|\nabla\phi|^2-2\phi h\langle\nabla\phi,\nabla h\rangle\leq |\nabla h|^2+h^2-2\phi h\langle\nabla\phi,\nabla h\rangle,$$ we have $$\begin{aligned}
\label{equation16}
\int_M\phi^2\left(2k+2\mu+2|A|^2\right)h^2e^{-f}d\sigma&=&4 \int_M\phi h\langle\nabla\phi,\nabla h\rangle e^{-f}d\sigma+2\int_M\phi^2|\nabla h|^2 e^{-f}d\sigma\nonumber\\
&\leq&2\int_M h^2e^{-f}d\sigma+4\int_M|\nabla h|^2e^{-f}d\sigma.
\end{aligned}$$ Finally, consider the intrinsic ball $B_j=B_j(p)$ in $M$ of radius $j$ and center at a fixed point $p\in M$. By applying (\[equation16\]) with $\phi=\phi_j$, where $\phi_j$ is one on $B_j$ and cuts off linearly to zero from $\partial B_j$ to $\partial B_{j+1},$ letting $j\rightarrow\infty,$ and using the monotone convergence theorem we obtain $$\int_M|A|^2h^2e^{-f}d\sigma\leq\int_M\left((1-k-\mu)h^2+2|\nabla h|^2\right)e^{-f}d\sigma.$$ Now, we will prove the inequality (\[equation3\]). In fact, $\log h$ is well-defined and $$\begin{aligned}
\label{equation17}
\Delta_f\log h&=&\frac{1}{h}\Delta_fh-|\nabla\log h|^2=\frac{1}{h}L_fh-|A|^2-k-|\nabla\log h|^2\nonumber\\
&=&-\mu-|A|^2-k-|\nabla\log h|^2.
\end{aligned}$$ Let $\psi$ be a function with compact support. Thus, it follows from Lemma \[Green\] and expression (\[equation17\]) that $$\begin{aligned}
\label{equation20}
\int_M\langle\nabla\psi^2,\nabla\log h\rangle e^{-f}d\sigma&=&-\int_M\psi^2(\Delta_f\log h)e^{-f}d\sigma\nonumber\\
&=&\int_M\psi^2\left(\mu+|A|^2+k+|\nabla\log h|^2\right)e^{-f}d\sigma.
\end{aligned}$$ Combining (\[equation20\]) with the following inequality $$\langle\nabla\psi^2,\nabla\log h\rangle\leq2|\nabla\psi|^2+\frac{1}{2}\psi^2|\nabla\log h|^2,$$ gives us this $$\label{equation4}
\int_M\psi^2\left(2|A|^2+|\nabla\log h|^2\right)e^{-f}d\sigma\leq\int_M\left(4|\nabla\psi|^2-2\mu\psi^2-2k\psi^2\right)e^{-f}d\sigma.$$ Let $\psi_j$ be one on $B_j$ and cut off linearly to zero from $\partial B_j$ to $\partial B_{j+1}$. Since $\phi\in W^{1,2}$, applying (\[equation4\]) with $\psi=\psi_j\phi,$ and letting $j\rightarrow\infty,$ using the monotone convergence theorem, we get $$\int_M\phi^2\left(2|A|^2+|\nabla\log h|^2\right)e^{-f}d\sigma\leq\int_M\left[4|\nabla\phi|^2-2(\mu+k)\phi^2\right]e^{-f}d\sigma.$$
Lemma \[l2\] was obtain by Colding and Minicozzi II (see [@ColdingMinicozzi2012], Lemma 9.15) to a complete noncompact hypersurface $\Sigma\subset{\mathbb{R}}^{n+1}$ without boundary that satisfies $$H=\frac{\langle x,\eta\rangle }{2},$$ where $x$ is position vector.
Through use the Lemmas \[lemma2\] and \[l2\], and ideas as the proof of Lemma 9.25 from [@ColdingMinicozzi2012] we have
\[proposition4\] Let $\mu_1(M)$ be the bottom $L^2_f$ spectrum of $L_f$. If $\mu_1(M)\neq-\infty$, then there exists a positive $C^2$ function $u$ on $M$ with $L_f u=-\mu_1(M)u.$ Moreover, if $w\in W^{1,2}_f$ and $L_f w=-\mu_1(M) w,$ then $w=Cu$ for some $C\in{\mathbb{R}}.$
\[lemma6\] Let $M$ be a complete oriented constant weighed mean curvature hypersurface in gradient Ricci soliton $\overline M_f$. If $\mu_1(M)\neq-\infty$ and ${Vol}_f(M)<\infty$, then $$\int_M|A|^2e^{-f}d\sigma<\infty.$$
Since $\mu_1(M)\neq-\infty$, there is a $C^2$ positive function $h$ on $M$ satisfying $L_fh=-\mu_1(M)h$ by Lemma \[proposition4\]. Let $\phi_j$ be the cut off functions such that $|\nabla\phi_j|\leq1$ and $\phi_j\leq1$. So $\phi_j\in W^{1,2}(e^{-f}d\sigma)$ because ${Vol}_f(M)<\infty$. By Lemma \[l2\], equality (\[equation3\]), we have $$\begin{aligned}
2\int_M\phi_j^2|A|^2e^{-f}d\sigma&\leq&\int_M\phi_j^2\left(2|A|^2+|\nabla\log h|^2\right)e^{-f}d\sigma\\
&\leq&\int_M\left(4|\nabla\phi_j|^2-2(\mu_1(M)+k)\phi_j^2\right)e^{-f}d\sigma\\
&\leq&\left(4+2|\mu_1(M)+k|\right)\int_Me^{-f}d\sigma<\infty.
\end{aligned}$$ By letting $j\rightarrow\infty,$ we obtain the conclusion this lemma by monotone convergence theorem.
The compact manifolds which admit a parallel vector field with respect to some metric was characterized by Welsh in [@Welsh1986a]. Namely, they are the compact fibre bundles over tori with finite structural group. The dimension of the torus can be assumed to be the number of linearly independent parallel vector fields.
The noncompact case has also been solved by Welsh in [@Welsh1986]. If fact,
\[proposition0\] If a Riemannian manifold $M$ admits a complete parallel vector field, then either $M$ is diffeomorphic to the product of a Euclidian space with some other manifold or there is a circle action on $M$ whose orbits are not real homologous to zero.
Now suppose that $\dim \mathcal P_{\overline M_f}=l>0$ and by hypothesis $${Ind}_f(M)=\dim \mathcal P_{\overline M_f}-\dim\{X\in P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}.$$ We can have two cases: either ${Ind}_f(M)=\dim \mathcal P_{\overline M_f}$ or ${Ind}_f(M)\neq\dim \mathcal P_{\overline M_f}.$
Initially, suppose that ${Ind}_f(M)=\dim \mathcal P_{\overline M_f}$, that is $\dim\{X\in\mathcal P_{\overline M_f}: \ \langle X,\eta\rangle\equiv0\}=0$ and there exists $l$ independent linearly parallel unit vector fields $X_1,X_2, \ldots,X_l$ such that $\langle X_j,\eta\rangle\not\equiv0$ for all $j=1,2,\ldots, l$. Put $$u_j=\langle X_j,\eta\rangle \ \ \ \text{for all} \ j=1,2,\ldots,l.$$ By Proposition \[proposition1\], $$L_fu_j=ku_j,$$ where $\overline{{Ric}}_f=kg$. Note that $$\label{eq4}
\int_Mu_j^{2}e^{-f}d\sigma\leq\int_Me^{-f}d\sigma<\infty$$ because $$u_j=\langle X_j,\eta\rangle\leq|X_j||\eta|=1.$$ Hence, $u_1,u_2,\ldots,u_l$ are $L^2(e^{-f}d\sigma)$ eigenfunctions with negative eigenvalue $-k$. Since, by hypothesis, $${Ind}_f(M)=l>0,$$ then the bottom $\mu_1(M)$ of the spectrum of $L_f$ satisfies $\mu_1(M)=-k.$
Now, consider a local orthonormal frame $\{e_1,e_2,\ldots,e_n\}$ on $M$. Observe that $$\begin{aligned}
\label{equation18}
|\nabla u_j|^2=\sum_{i=1}^n|\nabla_{e_i}u_j|^2\geq \sum_{i=1}^n\left(\sum_{k=1}^n a_{ik}^2\right)\left(\sum_{k=1}^n\langle e_k,X_j\rangle^2\right)\leq |A|^2.
\end{aligned}$$ Hence, it follows from Lemma \[lemma6\] and inequality (\[equation18\]) that $$\int_M|\nabla u_j|^{2}e^{-f}d\sigma\leq\int_M|A|^2e^{-f}d\sigma<\infty$$ and by using (\[eq4\]), we get $u_j\in W^{1,2}(e^{-f}d\sigma)$. By Lemma \[proposition4\], $u_j>0$ on $M$ without lost of generality. Therefore, it follows from Lemma \[lema5\] and the integrability of $|A|^2$, that $$\int_M|A|^2u_je^{-f}d\sigma=0.$$ Thus, $|A|\equiv0$ on $M$.
In the case which ${Ind}_f(M)\neq\dim \mathcal P_{\overline M_f}$, we have there exists a $u_0\equiv0,$ where $u_0=\langle X_0,\eta\rangle$ and $X_0\in T\overline M_f$ is parallel field with respect to Riemannian connection of $(\overline{M}^{n+1},\overline g)$, in particular, $X_0\in TM$ is parallel with respect to Riemannian connection of $(M,g)$. By Proposition \[proposition0\], either $M$ is diffeomorphic to the product of a Euclidian space with some other manifold, or there is a circle action on $M$ whose orbits are not real homologous to zero.
[**Hilário Alencar**]{}\
Instituto de Matemática\
Universidade Federal de Alagoas\
Maceió, AL, 57072-900, Brazil\
[**Adina Rocha**]{}\
Instituto de Matemática\
Universidade Federal de Alagoas\
Maceió, AL, 57072-900, Brazil\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The AN.ON-Next project aims to integrate privacy-enhancing technologies into the internet’s infrastructure and establish them in the consumer mass market.\
The technologies in focus include a basis protection at internet service provider level, an improved overlay network-based protection and a concept for privacy protection in the emerging 5G mobile network. A crucial success factor will be the viable adjustment and development of standards, business models and pricing strategies for those new technologies.
author:
- David Harborth
- Dominik Herrmann
- 'Stefan K[ö]{}psell'
- Sebastian Pape
- Christian Roth
- Hannes Federrath
- Dogan Kesdogan
- Kai Rannenberg
bibliography:
- 'quellen.bib'
title: 'Integrating Privacy-Enhancing Technologies into the Internet Infrastructure'
---
Introduction {#intro}
============
Related Work {#relwork}
============
ISP-based Anonymization {#isp}
=======================
Network Overlay-based Anonymization {#networkoverlay}
===================================
Anonymization Techniques for the 5G Network {#5g}
===========================================
Business Models for Privacy-Enhancing Technologies in the Internet Infrastructure {#businessmodels}
=================================================================================
Conclusion
==========
Acknowledgements
================
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We construct a family of irreducible representations of the quantum plane and of the quantum Weyl algebra over an arbitrary field, assuming the deformation parameter is not a root of unity. We determine when two representations in this family are isomorphic, and when they are weight representations, in the sense of [@vB97].'
author:
- 'Samuel A. Lopes[^1] and João N. P. Lourenço[^2]'
title: 'A multiparameter family of non-weight irreducible representations of the quantum plane and of the quantum Weyl algebra'
---
Introduction
============
Assume throughout that ${\mathbb{F}}$ is a field of arbitrary characteristic, not necessarily algebraically closed, with group of units ${\mathbb{F}}^*$. Fix $q\in{\mathbb{F}}^*$ with $q\neq 1$. The *quantum plane* is the unital associative algebra $${{\mathbb{F}}_q[x, y]}={\mathbb{F}}\{ x, y\}/(yx-qxy)$$ with generators $x$ and $y$ subject to the relation $yx=qxy$.
Consider the operators $\tau_q$ and $\partial_q$ defined on the polynomial algebra ${\mathbb{F}}[t]$ by $$\tau_q (p)(t)=p(qt), \quad \mbox{and} \quad \partial_q (p)(t)=\frac{p(qt)-p(t)}{qt-t}, \quad \mbox{for $p\in{\mathbb{F}}[t]$}.$$ Then the assignment $x\mapsto \tau_q$, $y\mapsto \partial_q$ yields a (reducible) representation ${{\mathbb{F}}_q[x, y]}\rightarrow \mathrm{End}_{\mathbb{F}}({\mathbb{F}}[t])$ of ${{\mathbb{F}}_q[x, y]}$, which is faithful if and only if $q$ is not a root of unity. The operators $\tau_q$ and $\partial_q$ are central in the theory of linear $q$-difference equations and $\partial_q$ is also known as the *Jackson derivative*, as it appears in [@fJ10]. See e.g. [@yM88], [@cK95 Chap. IV] and references therein for further details.
The irreducible representations of the quantum plane ${{\mathbb{F}}_q[x, y]}$ have been classified in [@vB97] using results from [@BvO97]. Following [@vB97] we say that a representation of ${{\mathbb{F}}_q[x, y]}$ is a *weight representation* if it is semisimple as a representation of the polynomial subalgebra ${\mathbb{F}}[H]$ generated by the element $H=xy$. When $q$ is a root of unity all irreducible representations of ${{\mathbb{F}}_q[x, y]}$ are finite-dimensional weight representations, and these are well understood. For example, if ${\mathbb{F}}$ is algebraically closed and $q$ is a primitive $n$-th root of unity then the irreducible representations of ${{\mathbb{F}}_q[x, y]}$ are either $1$ or $n$ dimensional. When $q$ is not a root of unity there are irreducible representations of ${{\mathbb{F}}_q[x, y]}$ that are not weight representations, and in particular are not finite dimensional. These turn out to be the *${\mathbb{F}}[H]$-torsionfree* irreducible representations of ${{\mathbb{F}}_q[x, y]}$, as they remain irreducible (i.e. nonzero) upon localizing at the nonzero elements of ${\mathbb{F}}[H]$. In [@vB97 Cor. 3.3] the torsionfree representations of ${{\mathbb{F}}_q[x, y]}$ are classified in terms of elements satisfying certain conditions, but no explicit construction of these representations is given.
We assume $q$ is not a root of unity, and we give an explicit construction of a 3-parameter family ${\mathsf{V\,}^{m, n}_{f}}$ of infinite-dimensional representations of ${{\mathbb{F}}_q[x, y]}$ having the following properties (compare Propositions \[P:isoclass\], \[P:dec\] and \[P:weight\]):
- $m$ and $n$ are positive integers, and $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies condition below, which essentially encodes $n$ independent parameters from ${\mathbb{F}}^*$;
- ${\mathsf{V\,}^{m, n}_{f}}$ is irreducible if and only if ${\mathsf{gcd}}(m, n)=1$;
- if $(m, n)\neq (m', n')$ then ${\mathsf{V\,}^{m, n}_{f}}$ and ${\mathsf{V\,}^{m', n'}_{f'}}$ are not isomorphic;
- ${\mathsf{V\,}^{m, n}_{f}}$ is a weight representation if and only if $m=n$;
- if ${\mathbb{F}}$ is algebraically closed and $V$ is an irreducible weight representation of ${{\mathbb{F}}_q[x, y]}$ that is infinite dimensional, then $V\simeq{\mathsf{V\,}^{1, 1}_{f}}$ for some $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$.
Thus, in some sense weight and non-weight representations of ${{\mathbb{F}}_q[x, y]}$ are rejoined in the family ${\mathsf{V\,}^{m, n}_{f}}$.
The localization of ${{\mathbb{F}}_q[x, y]}$ at the multiplicative set generated by $x$ contains a copy of the *$q$-Weyl algebra*, which is the algebra $$\label{E:qwa}
\mathbb{A}_1(q)={\mathbb{F}}\{ X, Y\}/(YX-qXY-1)$$ with generators $X$ and $Y$ subject to the relation $YX-qXY=1$ (see for details about this embedding). This is used in Subsection \[SS:qwa\] to regard the representations ${\mathsf{V\,}^{m, n}_{f}}$ as infinite-dimensional irreducible representations of $\mathbb{A}_1(q)$. In contrast with the action of ${{\mathbb{F}}_q[x, y]}$ on ${\mathsf{V\,}^{m, n}_{f}}$ when $m=n$, it turns out that ${\mathsf{V\,}^{m, n}_{f}}$ is never a weight representation of $\mathbb{A}_1(q)$ in the sense of [@vB97]. In Subsection \[SS:restriction\] we pursue a dual approach by constructing representations ${\mathsf{W}^{n}_{g_{}}}$ of $\mathbb{A}_1(q)$ and then restricting the action from the $q$-Weyl algebra to two distinct subalgebras of $\mathbb{A}_1(q)$ isomorphic to ${{\mathbb{F}}_q[x, y]}$.
A family ${\mathsf{V\,}^{m, n}_{f}}$ of infinite-dimensional irreducible representations of ${{\mathbb{F}}_q[x, y]}$ for $q$ not a root of unity {#S:qnru}
================================================================================================================================================
Assume $q\in{\mathbb{F}}^*$ is not a root of unity. We introduce a family ${\mathsf{V\,}^{m, n}_{f}}$ of infinite-dimensional representations of ${{\mathbb{F}}_q[x, y]}$ which are not in general weight representations in the sense of [@vB97], but which includes all irreducible infinite-dimensional weight representations of ${{\mathbb{F}}_q[x, y]}$ if we further assume ${\mathbb{F}}$ to be algebraically closed.
Structure of the representations ${\mathsf{V\,}^{m, n}_{f}}$
------------------------------------------------------------
Fix positive integers $m, n\in{\mathbb{Z}}_{>0}$ and a function $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfying $$\label{prop}
f(i+n)=qf(i), \quad \quad \text{for all $i\in{\mathbb{Z}}$.}$$ Such functions are in one-to-one correspondence with elements of $\left({\mathbb{F}}^*\right)^n$. Let ${\mathsf{V\,}^{m, n}_{f}}$ denote the representation of ${{\mathbb{F}}_q[x, y]}$ on the space ${\mathbb{F}}[t^{\pm 1}]$ of Laurent polynomials in $t$ given by $$\label{action}
x.t^{i}=t^{i+n}, \quad \quad y.t^{i}=f(i)t^{i-m},\quad \quad \text{for all $i\in{\mathbb{Z}}$.}$$ Condition ensures that the expressions do define an action of ${{\mathbb{F}}_q[x, y]}$ on ${\mathbb{F}}[t^{\pm 1}]$ as, for all $i\in{\mathbb{Z}}$, $$(yx-qxy).t^{i}=(f(i+n)-qf(i))t^{i+n-m}=0.$$
\[Ex:floor\] Fix $\mu\in{\mathbb{F}}^{*}$ and $m, n\in{\mathbb{Z}}_{>0}$. For $i\in{\mathbb{Z}}$ let $f(i)=\mu q^{\left\lfloor \frac{i}{n}\right\rfloor}$, where $\left\lfloor \frac{i}{n}\right\rfloor$ denotes the largest integer not exceeding $\frac{i}{n}$. Then $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies condition and thus there is a representation ${\mathsf{V\,}^{m, n}_{f}}$ of ${{\mathbb{F}}_q[x, y]}$ on ${\mathbb{F}}[t^{\pm 1}]$ with action $$x.t^{i}=t^{i+n}, \quad \quad y.t^{i}=\mu q^{\left\lfloor \frac{i}{n}\right\rfloor}t^{i-m},\quad \quad \text{for all $i\in{\mathbb{Z}}$.}$$
We begin the study of the representations ${\mathsf{V\,}^{m, n}_{f}}$ by first considering the case that the parameters $m$ and $n$ are coprime. The following consequence of will be helpful.
\[L:NT\] Assume ${\mathsf{gcd}}(m, n)=1$ and $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies . For $k\in{\mathbb{Z}}$ define $${\mathsf{s}_{f}}(k)=\prod_{i=0}^{n-1} f(k-im).$$ Then ${\mathsf{s}_{f}}(k)={\mathsf{s}_{f}}(0)q^{k}$.
For $j\in{\mathbb{Z}}$ let $0\leq \overline{\jmath}<n$ be the unique integer such that $\overline{\jmath}\equiv j \modd n$. Then the formula $f(j)=f(\overline{\jmath}) q^{\frac{j-\overline{\jmath}}{n}}$ can be verified by induction on $\left| \frac{j-\overline{\jmath}}{n} \right|$. Thus, $${\mathsf{s}_{f}}(k)=\prod_{i=0}^{n-1} f(k-im)=\prod_{i=0}^{n-1} f\left(\overline{k-\imath m}\right) \prod_{i=0}^{n-1}q^{\frac{k-im-\overline{k-\imath m}}{n}}.$$ Since $m$ and $n$ are coprime, the set $\left\{ \overline{k-\imath m} \mid 0\leq i<n \right\}$ consists of all the integers from $0$ to $n-1$, and is thus independent of $k$. Moreover, $$\sum_{i=0}^{n-1}{\frac{k-im-\overline{k-\imath m}}{n}} = k+ \sum_{i=0}^{n-1}{\frac{-im-\overline{k-\imath m}}{n}} = k+ \sum_{i=0}^{n-1}{\frac{-im-\overline{(-\imath m)}}{n}}.$$ Hence, $${\mathsf{s}_{f}}(k) =q^{k}\prod_{i=0}^{n-1} f\left(\overline{-\imath m}\right) \prod_{i=0}^{n-1}q^{\frac{-im-\overline{(-\imath m)}}{n}}=q^{k}{\mathsf{s}_{f}}(0).$$
\[P:irred\] Assume ${\mathsf{gcd}}(m, n)=1$ and $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies . Then the representation ${\mathsf{V\,}^{m, n}_{f}}$ defined by is an irreducible representation of ${{\mathbb{F}}_q[x, y]}$.
We begin with a computation: for $k\in{\mathbb{Z}}$ we have, by Lemma \[L:NT\], $$\label{E:eigen}
x^m y^n . t^k=x^m\left(\prod_{i=0}^{n-1}f(k-im)\right) t^{k-nm}={\mathsf{s}_{f}}(k) t^{k}={\mathsf{s}_{f}}(0)q^{k} t^{k}.$$ Hence, $x^m y^n . p(t)={\mathsf{s}_{f}}(0)p(qt)$ for all $p\in{\mathbb{F}}[t^{\pm 1}]$.
Let $\mathsf{W}\subseteq {\mathsf{V\,}^{m, n}_{f}}$ be a nonzero subrepresentation. If $p(t)\in\mathsf{W}$ then also $p(qt)\in\mathsf{W}$, by . As $q$ is not a root of unity, the latter implies that $t^{\ell}\in\mathsf{W}$ for some $\ell\in{\mathbb{Z}}$. The coprimeness of $m$ and $n$ shows the existence of integers $a$ and $b$ so that $an-bm=1$. By replacing $a$ and $b$ with $a+jm$ and $b+jn$ for a sufficiently large integer $j$, we can assume $a, b\in{\mathbb{Z}}_{>0}$. Then $x^{a} y^{b}.t^{k}=\lambda_{k}t^{k+1}$ for some $\lambda_{k}\in{\mathbb{F}}^{*}$, showing that $t^{k}\in\mathsf{W}$ for all $k\geq \ell$. A similar argument shows that $t^{k}\in\mathsf{W}$ for all $k\leq \ell$. Hence $\mathsf{W}={\mathsf{V\,}^{m, n}_{f}}$, establishing the irreducibility of ${\mathsf{V\,}^{m, n}_{f}}$.
Next we describe ${\mathsf{V\,}^{m, n}_{f}}$ in terms of a maximal left ideal of ${{\mathbb{F}}_q[x, y]}$. Recall that for a representation $\mathsf{V}$ of ${{\mathbb{F}}_q[x, y]}$ and an element $v\in\mathsf{V}$, the annihilator of $v$ in ${{\mathbb{F}}_q[x, y]}$ is $\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(v)=\{ r\in{{\mathbb{F}}_q[x, y]}\mid r.v=0 \}$, a left ideal of ${{\mathbb{F}}_q[x, y]}$.
\[P:isoclass\] Assume ${\mathsf{gcd}}(m, n)=1$ and $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies .
1. For $1\in{\mathsf{V\,}^{m, n}_{f}}$, $\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)={{\mathbb{F}}_q[x, y]}\left(x^{m}y^{n}-{\mathsf{s}_{f}}(0)\right)$ and $${\mathsf{V\,}^{m, n}_{f}}\simeq {{\mathbb{F}}_q[x, y]}/{{\mathbb{F}}_q[x, y]}\left(x^{m}y^{n}-{\mathsf{s}_{f}}(0)\right).$$
2. For positive integers $m', n'$, and $f':\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfying (with $n$ replaced by $n'$), we have ${\mathsf{V\,}^{m, n}_{f}}\simeq {\mathsf{V\,}^{m', n'}_{f'}}$ if and only if $m=m'$, $n=n'$ and ${\mathsf{s}_{f'}}(0)=q^{k}{\mathsf{s}_{f}}(0)$ for some $k\in{\mathbb{Z}}$.
(a) Let $\theta=x^m y^n$. First we show that $$\label{E:claim_ann}
\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)={{\mathbb{F}}_q[x, y]}\left({\mathbb{F}}[\theta]\cap\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1) \right).$$ The inclusion $\supseteq$ is clear, so suppose $u\in\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)$. Write $u=\sum_{i\geq 0}\mu_i x^{a_i}y^{b_i}=\sum_{k\in{\mathbb{Z}}}u_k$, where $\displaystyle u_k=\sum_{na_i-mb_i=k}\mu_i x^{a_i}y^{b_i}$. Since $u_k.1$ is in ${\mathbb{F}}t^k$, it follows that $u_k\in\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)$ for all $k\in{\mathbb{Z}}$, and it suffices to prove $u_k\in{{\mathbb{F}}_q[x, y]}\left({\mathbb{F}}[\theta]\cap\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1) \right)$.
If $na_i-mb_i=na_j-mb_j$ then, as ${\mathsf{gcd}}(m, n)=1$, we deduce that $(a_i, b_i)=(a_j, b_j)+\xi (m, n)$ for some $\xi\in{\mathbb{Z}}$. Thus, by the normality of $x$ and $y$, there are $a, b\geq 0$ with $na-mb=k$ such that $u_k=x^a y^b w_0$, where $w_0=\sum_{j\geq 0}\nu_j x^{\xi_j m}y^{\xi_j n}\in{\mathbb{F}}[\theta]$. Notice that for any $\ell\in{\mathbb{Z}}$, $x^a y^b.t^\ell$ is a nonzero scalar multiple of $t^{\ell+k}$, so $x^a y^b w_0=u_k\in\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)$ implies that $w_0\in\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)$. Hence, $u_k\in{{\mathbb{F}}_q[x, y]}\left({\mathbb{F}}[\theta]\cap\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1) \right)$ and is established.
Now implies that $\theta-{\mathsf{s}_{f}}(0)\in{\mathbb{F}}[\theta]\cap\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)$. Since ${\mathbb{F}}[\theta]\left( \theta-{\mathsf{s}_{f}}(0) \right)$ is a maximal ideal of ${\mathbb{F}}[\theta]$ it follows that ${\mathbb{F}}[\theta]\cap\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)={\mathbb{F}}[\theta]\left( \theta-{\mathsf{s}_{f}}(0) \right)$ and $\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(1)={{\mathbb{F}}_q[x, y]}\left(\theta-{\mathsf{s}_{f}}(0)\right)$. This proves (a) as $1\in{\mathsf{V\,}^{m, n}_{f}}$ generates ${\mathsf{V\,}^{m, n}_{f}}$.
(b) We observe that the arguments above also show that for $t^k\in{\mathsf{V\,}^{m, n}_{f}}$, $\mathsf{ann}_{{{\mathbb{F}}_q[x, y]}}(t^k)={{\mathbb{F}}_q[x, y]}\left(\theta-q^k {\mathsf{s}_{f}}(0)\right)$ and $${\mathsf{V\,}^{m, n}_{f}}\simeq {{\mathbb{F}}_q[x, y]}/{{\mathbb{F}}_q[x, y]}\left(x^{m}y^{n}-q^k{\mathsf{s}_{f}}(0)\right),$$ for any $k\in{\mathbb{Z}}$. This establishes the *if* part of (b). For the direct implication, suppose ${\mathsf{V\,}^{m, n}_{f}}\simeq {\mathsf{V\,}^{m', n'}_{f'}}$. We have, for $a, b\geq 0$ and $t^k\in{\mathsf{V\,}^{m, n}_{f}}$, $$x^a y^b.t^k=\left( \prod_{i=0}^{b-1}f(k-im)\right)t^{k+na-mb}$$ and $\prod_{i=0}^{b-1}f(k-im)\neq 0$. This implies that $x^a y^b$ is diagonalizable on ${\mathsf{V\,}^{m, n}_{f}}$ if and only if $na=mb$. As ${\mathsf{gcd}}(m, n)=1$ this amounts to having $(a, b)=\xi(m, n)$ for some $\xi\geq 0$.
Since ${\mathsf{V\,}^{m, n}_{f}}\simeq {\mathsf{V\,}^{m', n'}_{f'}}$, then $x^{m'}y^{n'}$ is diagonalizable on ${\mathsf{V\,}^{m, n}_{f}}$ and similarly $x^{m}y^{n}$ is diagonalizable on ${\mathsf{V\,}^{m', n'}_{f'}}$. By the relation above we conclude that $(m, n)=(m', n')$. Moreover, the eigenvalues of $x^{m}y^{n}$ on ${\mathsf{V\,}^{m, n}_{f}}$ are of the form $q^{k}{\mathsf{s}_{f}}(0)$, whereas ${\mathsf{s}_{f'}}(0)$ is an eigenvalue of $x^{m'}y^{n'}=x^{m}y^{n}$ on ${\mathsf{V\,}^{m', n'}_{f'}}$. Hence ${\mathsf{s}_{f'}}(0)=q^{k}{\mathsf{s}_{f}}(0)$ for some $k\in{\mathbb{Z}}$, which concludes the proof.
\[R:f\] By Proposition \[P:isoclass\] above, for ${\mathsf{gcd}}(m, n)=1$ and $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfying , the isomorphism class of ${\mathsf{V\,}^{m, n}_{f}}$ depends only on $m$, $n$ and ${\mathsf{s}_{f}}(0)\in{\mathbb{F}}^*$.
Fix $\lambda\in{\mathbb{F}}^*$. Since ${\mathsf{gcd}}(m, n)=1$ there is a unique $f_\lambda:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ such that holds and $f_\lambda(km)=\lambda$ if $k=0$ and $f_\lambda(km)=1$ if $-(n-1)\leq k\leq -1$. Then ${\mathsf{s}_{f_\lambda}}(0)=\lambda$, ${\mathsf{V\,}^{m, n}_{f_\lambda}}\simeq {{\mathbb{F}}_q[x, y]}/{{\mathbb{F}}_q[x, y]}\left(x^{m}y^{n}-\lambda\right)$ and, for $\lambda'\in{\mathbb{F}}^*$, ${\mathsf{V\,}^{m, n}_{f_\lambda}}\simeq {\mathsf{V\,}^{m, n}_{f_{\lambda'}}}$ if and only if $\lambda/\lambda'\in \langle q\rangle$, where $\langle q\rangle$ is the subgroup of ${\mathbb{F}}^*$ generated by $q$.
If ${\mathbb{F}}$ contains an $n$-th root of $\lambda$, say $\mu$, there is a more natural construction for the irreducible representation ${{\mathbb{F}}_q[x, y]}/{{\mathbb{F}}_q[x, y]}\left(x^{m}y^{n}-\lambda\right)$. Define $f^{\mu}(i)=\mu q^{\left\lfloor \frac{i}{n}\right\rfloor}$, as in Example \[Ex:floor\]. Then ${\mathsf{s}_{f^\mu}}(0)=q^k \mu^n=q^k \lambda$, for some $k\in{\mathbb{Z}}$. It follows from Proposition \[P:isoclass\] that ${\mathsf{V\,}^{m, n}_{f^\mu}}\simeq {{\mathbb{F}}_q[x, y]}/{{\mathbb{F}}_q[x, y]}\left(x^{m}y^{n}-\lambda\right)$ and ${\mathsf{V\,}^{m, n}_{f^\mu}}$ depends only on $m$, $n$ and $\lambda$, and not on the particular $n$-th root of $\lambda$ that was chosen.
Finally we consider the general case of arbitrary $m, n\in{\mathbb{Z}}_{>0}$.
\[P:dec\] Let $m, n\in{\mathbb{Z}}_{>0}$ be arbitrary, with $d={\mathsf{gcd}}(m, n)$, and assume $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies . Then there is a direct sum decomposition $$\label{dec}
{\mathsf{V\,}^{m, n}_{f}}\simeq \bigoplus_{k=0}^{d-1}{\mathsf{V\,}^{m/d, n/d}_{f_k}}$$ into irreducible representations, where $f_k(i)=f(k+id)$, for $0\leq k<d$ and $i\in{\mathbb{Z}}$.
Moreover, suppose $m', n'\in{\mathbb{Z}}_{>0}$, and $f':\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies (with $n$ replaced by $n'$). If ${\mathsf{V\,}^{m, n}_{f}}\simeq {\mathsf{V\,}^{m', n'}_{f'}}$ then $m=m'$ and $n=n'$.
For $0\leq k<d$, the subspace $t^k {\mathbb{F}}[t^{\pm d}]$ of ${\mathsf{V\,}^{m, n}_{f}}$ is readily seen to be invariant under the actions of $x$ and $y$, and we have ${\mathsf{V\,}^{m, n}_{f}} = \bigoplus_{k=0}^{d-1}t^k {\mathbb{F}}[t^{\pm d}]$. Thus, next we argue that the subrepresentation $t^k {\mathbb{F}}[t^{\pm d}]$ is isomorphic to ${\mathsf{V\,}^{m/d, n/d}_{f_k}}$, where $f_k(i)=f(k+id)$ for all $i\in{\mathbb{Z}}$. First notice that $f_k(i+n/d)=f(k+id+n)=qf(k+id)=qf_k(i)$, so ${\mathsf{V\,}^{m/d, n/d}_{f_k}}$ is defined. Consider the map $\phi: {\mathsf{V\,}^{m/d, n/d}_{f_k}}\rightarrow t^k {\mathbb{F}}[t^{\pm d}]$ given by $\phi(p)(t)=t^kp(t^d)$, for all $p\in{\mathbb{F}}[t^{\pm 1}]$. In particular, $\phi(t^i)=t^{k+id}$ for $i\in{\mathbb{Z}}$. Still viewing $t^k {\mathbb{F}}[t^{\pm d}]$ as a subrepresentation of ${\mathsf{V\,}^{m, n}_{f}}$, we have: $$\begin{aligned}
\phi(x.t^i) &=\phi(t^{i+n/d})=t^{k+id+n}=x.t^{k+id}=x.\phi(t^i),\\
\phi(y.t^i) &=\phi(f_k(i)t^{i-m/d})=f(k+id)t^{k+id-m}=y.t^{k+id}=y.\phi(t^i).\end{aligned}$$ Since $\phi$ is clearly bijective, the calculations above show that $\phi$ is an isomorphism of representations, and ${\mathsf{V\,}^{m, n}_{f}}\simeq \bigoplus_{k=0}^{d-1}{\mathsf{V\,}^{m/d, n/d}_{f_k}}$. The fact that each summand ${\mathsf{V\,}^{m/d, n/d}_{f_k}}$ is irreducible follows from ${\mathsf{gcd}}(m/d, n/d)=1$ and Proposition \[P:irred\], which will be established independently.
Finally, assume ${\mathsf{V\,}^{m, n}_{f}}\simeq {\mathsf{V\,}^{m', n'}_{f'}}$ for positive integers $m'$ and $n'$, and $f':\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfying $f'(i+n')=qf'(i)$, for all $i\in{\mathbb{Z}}$. Then, up to isomorphism, ${\mathsf{V\,}^{m, n}_{f}}$ and ${\mathsf{V\,}^{m', n'}_{f'}}$ have the same composition factors, and in particular the same composition length. This proves that $d={\mathsf{gcd}}(m, n)={\mathsf{gcd}}(m', n')$ and that ${\mathsf{V\,}^{m/d, n/d}_{f_0}}\simeq {\mathsf{V\,}^{m'/d, n'/d}_{f'_k}}$ for some $k$. By Proposition \[P:isoclass\], which will also be established independently, we have $m/d=m'/d$ and $n/d=n'/d$, so $m=m'$ and $n=n'$.
Weight representations of the form ${\mathsf{V\,}^{m, n}_{f}}$ {#SS:weight}
--------------------------------------------------------------
Let us now determine when ${\mathsf{V\,}^{m, n}_{f}}$ is a weight representation in the sense of [@vB97]. Recall that this occurs when ${\mathsf{V\,}^{m, n}_{f}}$ is semisimple as a representation over the polynomial subalgebra ${\mathbb{F}}[H]$, where $H=xy$. Assume first that $m=n=1$ and fix $\lambda\in{\mathbb{F}}^*$. The map $f_\lambda$ defined in Remark \[R:f\] is given by $f_\lambda (i)=\lambda q^{i}$ for all $i\in{\mathbb{Z}}$, and the corresponding representation ${\mathsf{V\,}^{1, 1}_{f_\lambda}}\simeq {{\mathbb{F}}_q[x, y]}/{{\mathbb{F}}_q[x, y]}\left(H-\lambda\right)$ is irreducible. Since $H.t^i=xy.t^i=\lambda q^i t^i$ for all $i$, the decomposition ${\mathsf{V\,}^{1, 1}_{f_\lambda}}=\bigoplus_{i\in{\mathbb{Z}}}{\mathbb{F}}\, t^i$ shows that ${\mathsf{V\,}^{1, 1}_{f_\lambda}}$ is semisimple over ${\mathbb{F}}[H]$. Moreover, for $\nu\in{\mathbb{F}}^*$, ${\mathsf{V\,}^{1, 1}_{f_\lambda}}\simeq {\mathsf{V\,}^{1, 1}_{f_\nu}}$ if and only if $\lambda/\nu\in\langle q\rangle$, the multiplicative subgroup of ${\mathbb{F}}^*$ generated by $q$, by Proposition \[P:isoclass\]. In case ${\mathbb{F}}$ is algebraically closed, these are all the infinite-dimensional irreducible weight representations of ${{\mathbb{F}}_q[x, y]}$, by [@vB97 Cor. 3.2]. Combined with Proposition \[P:isoclass\](b) the above yields the classification of irreducible weight representations in the family ${\mathsf{V\,}^{m, n}_{f}}$.
\[P:weight\] Assume ${\mathsf{gcd}}(m, n)=1$ and $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies . Then ${\mathsf{V\,}^{m, n}_{f}}$ is a weight representation if and only if $m=n=1$.
For completeness, we include a brief and direct proof of Proposition \[P:weight\] not assuming that ${\mathbb{F}}$ is algebraically closed, a condition that was used implicitly at the end of the previous paragraph.
Assume first that $m=n=1$. Then since $f$ satisfies we have $f=f_\lambda$ for $\lambda=f(0)$ and the discussion above shows that ${\mathsf{V\,}^{m, n}_{f}}$ is a weight representation of ${{\mathbb{F}}_q[x, y]}$. Conversely, suppose ${\mathsf{V\,}^{m, n}_{f}}$ is a weight representation of ${{\mathbb{F}}_q[x, y]}$. Then clearly $\dim_{\mathbb{F}}{\mathbb{F}}[H].v<+\infty$ for any $v\in{\mathsf{V\,}^{m, n}_{f}}$. Notice that, for all $i\in{\mathbb{Z}}$, $H.t^i=xy.t^i=f(i)t^{i+n-m}$. Thus, for $\ell\in{\mathbb{Z}}$, $H^\ell.t^i=\zeta t^{i+\ell(n-m)}$ for some $\zeta\in{\mathbb{F}}^*$. But then the condition $\dim_{\mathbb{F}}{\mathbb{F}}[H].1<+\infty$ immediately implies $m=n$, and hence $m=n=1$, as ${\mathsf{gcd}}(m, n)=1$.
Given arbitrary positive integers $m$ and $n$, and $f$ satisfying , the representation ${\mathsf{V\,}^{m, n}_{f}}$ is a weight representation if and only if $m=n$. The direct implication follows from the proof of Proposition \[P:weight\]. For the converse implication, recall that ${\mathsf{V\,}^{m, m}_{f}}$ is the direct sum of $m$ representations of the form ${\mathsf{V\,}^{1, 1}_{f_k}}$, for $0\leq k<m$, by Proposition \[P:dec\], so the claim follows as each of these is a weight representation.
Connections with the representation theory of the $q$-Weyl algebra $\mathbb{A}_1(q)$ {#S:conn}
====================================================================================
We continue to assume $q\in{\mathbb{F}}^*$ is not a root of unity. Let $\mathbb{A}_1(q)$ be the $q$-Weyl algebra given by generators $X$ and $Y$ and defining relation $YX-qXY=1$, as in . It is straightforward to show that $\{x^k \mid k\geq 0\}$ is a right and left Ore set consisting of regular elements of ${{\mathbb{F}}_q[x, y]}$, and we denote the corresponding localization by ${\mathbb{F}}_q[x^{\pm 1}, y]$. The calculation $$\big(x^{-1}(y-1)\big)x-qx\big(x^{-1}(y-1)\big)=x^{-1}yx -q(y-1)-1= qy -q(y-1)-1=q-1$$ shows that there is an algebra map $$\label{E:qwainqp}
\mathbb{A}_1(q)\rightarrow {\mathbb{F}}_q[x^{\pm 1}, y], \quad \mbox{with \quad $X\mapsto x,\quad Y\mapsto \frac{1}{q-1}x^{-1}(y-1)$.}$$ To see that the map in is injective we can argue as follows. The multiplicative subset $\{X^k \mid k\geq 0\}$ of $\mathbb{A}_1(q)$ is a right and left Ore set of regular elements and we denote the corresponding localization by $\widehat{\mathbb{A}}_1(q)$. Then the map in extends to a map $\widehat{\mathbb{A}}_1(q)\rightarrow {\mathbb{F}}_q[x^{\pm 1}, y]$, which has an inverse ${\mathbb{F}}_q[x^{\pm 1}, y] \rightarrow \widehat{\mathbb{A}}_1(q)$ with $x^{\pm 1}\mapsto X^{\pm 1}$ and $y\mapsto (q-1)XY+1$. It follows that induces an isomorphism $\widehat{\mathbb{A}}_1(q)\simeq {\mathbb{F}}_q[x^{\pm 1}, y]$, and in particular is injective. In view of the above we will identify $X$ with $x$, $Y$ with $\frac{1}{q-1}x^{-1}(y-1)$ and $\mathbb{A}_1(q)$ with the corresponding subalgebra of ${\mathbb{F}}_q[x^{\pm 1}, y]$. Since $y=(q-1)XY+1=YX-XY$, we have the embeddings $$\label{E:embed}
{{\mathbb{F}}_q[x, y]}\subseteq \mathbb{A}_1(q)\subseteq {\mathbb{F}}_q[x^{\pm 1}, y]=\widehat{\mathbb{A}}_1(q).$$
Extension of the representations ${\mathsf{V\,}^{m, n}_{f}}$ to $\mathbb{A}_1(q)$ {#SS:qwa}
---------------------------------------------------------------------------------
Our aim in this subsection is to extend the action of ${{\mathbb{F}}_q[x, y]}$ on ${\mathsf{V\,}^{m, n}_{f}}$ to an action of the $q$-Weyl algebra $\mathbb{A}_1(q)$. Assume thus that $m, n$ are positive integers and $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies . If $\rho^{m, n}_{f}:{{\mathbb{F}}_q[x, y]}\rightarrow \mathrm{End}_{\mathbb{F}}({\mathsf{V\,}^{m, n}_{f}})$ is the representation of ${{\mathbb{F}}_q[x, y]}$ on ${\mathsf{V\,}^{m, n}_{f}}$, we first observe that $\rho^{m, n}_{f}(x)$ is an invertible linear map on ${\mathsf{V\,}^{m, n}_{f}}$, a fact which is clear from . Therefore $\rho^{m, n}_{f}$ extends to the localization ${\mathbb{F}}_q[x^{\pm 1}, y]$, and ${\mathsf{V\,}^{m, n}_{f}}$ can be seen as a representation of ${\mathbb{F}}_q[x^{\pm 1}, y]$ with $x^{-1}.t^{i}=t^{i-n}$ for all $i\in{\mathbb{Z}}$. Now we get an action of $\mathbb{A}_1(q)$ on ${\mathsf{V\,}^{m, n}_{f}}={\mathbb{F}}[t^{\pm 1}]$ by restricting $\rho^{m, n}_{f}$: $$\begin{aligned}
\label{E:qwaaction}\nonumber
&X.t^{i} =x.t^{i}=t^{i+n},\\
&Y.t^{i} =\frac{1}{q-1}x^{-1}(y-1).t^{i}=\frac{1}{q-1}(f(i)t^{i-m-n}-t^{i-n}),\quad \quad \text{for all $i\in{\mathbb{Z}}$.}\end{aligned}$$
In our next result we view ${\mathsf{V\,}^{m, n}_{f}}$ as a representation of $\mathbb{A}_1(q)$, as above.
Assume ${\mathsf{gcd}}(m, n)=1$ and $f:\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies . Then:
1. ${\mathsf{V\,}^{m, n}_{f}}$ defined by is an irreducible representation of $\mathbb{A}_1(q)$.
2. For positive integers $m', n'$, and $f':\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfying (with $n$ replaced by $n'$), we have ${\mathsf{V\,}^{m, n}_{f}}\simeq {\mathsf{V\,}^{m', n'}_{f'}}$ as representations of $\mathbb{A}_1(q)$ if and only if $m=m'$, $n=n'$ and ${\mathsf{s}_{f'}}(0)=q^{k}{\mathsf{s}_{f}}(0)$ for some $k\in{\mathbb{Z}}$.
3. ${\mathsf{V\,}^{m, n}_{f}}$ is not semisimple as a representation over the polynomial subalgebra of $\mathbb{A}_1(q)$ generated by $XY$; hence, ${\mathsf{V\,}^{m, n}_{f}}$ is not a weight representation of $\mathbb{A}_1(q)$ in the sense of [@vB97].
Part (a) and the direct implication in (b) follow from the embedding , and from Propositions \[P:irred\] and \[P:isoclass\].
Suppose now $f':\mathbb{Z} \rightarrow {\mathbb{F}}^{*}$ satisfies , and there is $k\in{\mathbb{Z}}$ so that ${\mathsf{s}_{f'}}(0)=q^{k}{\mathsf{s}_{f}}(0)$. Then by Proposition \[P:isoclass\] there is an isomorphism $\phi : {\mathsf{V\,}^{m, n}_{f}}\rightarrow {\mathsf{V\,}^{m, n}_{f'}}$ as representations of ${{\mathbb{F}}_q[x, y]}$. For $v\in{\mathsf{V\,}^{m, n}_{f}}$ we have $\phi(v)=\phi(xx^{-1}.v)=x.\phi(x^{-1}.v)$, thus $\phi(x^{-1}.v)=x^{-1}.\phi(v)$. Whence $\phi$ is an isomorphism of representations of ${\mathbb{F}}_q[x^{\pm 1}, y]$. The other implication in (b) now follows from .
Observe that $XY=\frac{1}{q-1}(y-1)$, so the polynomial subalgebra of $\mathbb{A}_1(q)$ generated by $XY$ is just ${\mathbb{F}}[y]$. Given $0\neq v\in{\mathsf{V\,}^{m, n}_{f}}$, the formula $y.t^{i}=f(i)t^{i-m}$ for $i\in{\mathbb{Z}}$ implies $\dim_{\mathbb{F}}{\mathbb{F}}[y].v=+\infty$. Hence, ${\mathsf{V\,}^{m, n}_{f}}$ is not semisimple over ${\mathbb{F}}[y]={\mathbb{F}}[XY]$, and therefore it is not a weight representation of $\mathbb{A}_1(q)$ in the sense of [@vB97].
In [@BO09] the authors introduce Whittaker representations for generalized Weyl algebras. For the cases covered in this note, a representation $\mathsf{V}$ is a *Whittaker representation* for ${{\mathbb{F}}_q[x, y]}$ (respectively, for $\mathbb{A}_1(q)$) if $\mathsf{V}$ is generated by an element $v\in \mathsf{V}$ which is an eigenvector for the action of $x\in{{\mathbb{F}}_q[x, y]}$ (respectively, for the action of $X\in \mathbb{A}_1(q)$). Since $m, n\geq 1$, it is immediate that the operators $x, y\in{{\mathbb{F}}_q[x, y]}$ (respectively, $X, Y\in\mathbb{A}_1(q)$) have no eigenvectors in ${\mathsf{V\,}^{m, n}_{f}}$, so ${\mathsf{V\,}^{m, n}_{f}}$ is not a Whittaker representation for the quantum plane (respectively, for the $q$-Weyl algebra).
The representations ${\mathsf{W}^{n}_{g_{}}}$ of $\mathbb{A}_1(q)$ and their restriction to ${{\mathbb{F}}_q[x, y]}$ {#SS:restriction}
--------------------------------------------------------------------------------------------------------------------
We will now use a similar idea to construct representations of the $q$-Weyl algebra on the Laurent polynomial algebra ${\mathbb{F}}[t^{\pm 1}]$. Fix positive integers $m, n\in{\mathbb{Z}}_{>0}$ and a function $g:\mathbb{Z} \rightarrow {\mathbb{F}}$. Then the formulas $$\label{qwa:action}
X.t^{i}=t^{i+n}, \quad \quad Y.t^{i}=g(i)t^{i-m},\quad \quad \text{for all $i\in{\mathbb{Z}}$}$$ yield a representation of $\mathbb{A}_1(q)$ on ${\mathbb{F}}[t^{\pm 1}]$ if and only if $m=n$ and $g$ satisfies $$\label{qwa:prop}
g(i+n)=qg(i)+1, \quad \quad \text{for all $i\in{\mathbb{Z}}$.}$$ We denote the corresponding representation of $\mathbb{A}_1(q)$ by ${\mathsf{W}^{n}_{g_{}}}$. Notice that for all $i\in{\mathbb{Z}}$ $$\label{qwa:xy:com}
XY.t^i=g(i)t^i, \quad \quad (YX-XY).t^i=\left(g(i+n)-g(i)\right)t^i=\left((q-1)g(i)+1\right)t^i,$$ so ${\mathsf{W}^{n}_{g_{}}}$ is a weight representation of $\mathbb{A}_1(q)$ in the sense of [@vB97].
It follows from the computations at the beginning of Section \[S:conn\] that the element $YX-XY$ is normal in $\mathbb{A}_1(q)$ and it is sometimes referred to as a Casimir element, in spite of not being central. The equality $YX-XY=(q-1)XY+1$ shows that $YX-XY$ and $(q-1)XY+1$ generate the same subalgebra of $\mathbb{A}_1(q)$ and thus a weight representation of $\mathbb{A}_1(q)$ could be defined in an equivalent manner as a representation which is semisimple over the subalgebra generated by the Casimir element $YX-XY$.
Our first observation is the analogue of Proposition \[P:dec\].
\[L:qwa:dec\] Let $n\in{\mathbb{Z}}_{>0}$ and assume $g:\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfies . There is a direct sum decomposition $$\label{qwa:dec}
{\mathsf{W}^{n}_{g_{}}}\simeq \bigoplus_{k=0}^{n-1}{\mathsf{W}^{1}_{g_{k}}},$$ where $g_k(i)=g(k+in)$, for $0\leq k<n$ and $i\in{\mathbb{Z}}$.
For $0\leq k<n$, the subspace $t^k {\mathbb{F}}[t^{\pm n}]$ is invariant under the actions of $X$ and $Y$ and ${\mathsf{W}^{n}_{g_{}}}=\bigoplus_{k=0}^{n-1}t^k {\mathbb{F}}[t^{\pm n}]$. Moreover, the map $\phi: {\mathsf{W}^{1}_{g_{k}}}\rightarrow t^k {\mathbb{F}}[t^{\pm n}]$ given by $\phi(p)(t)=t^kp(t^n)$, for all $p\in{\mathbb{F}}[t^{\pm 1}]$ is easily checked to be an isomorphism.
In view of the above, it is enough to study the structure of the representations ${\mathsf{W}^{1}_{g_{}}}$, where $g:\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfies $g(i+1)=qg(i)+1$ for all $i\in{\mathbb{Z}}$. Equivalently, $g(i)=g(0)q^i + [i]_q$, where $[i]_q=\frac{q^i-1}{q-1}$ for all $i\in{\mathbb{Z}}$.
\[P:qwa:w:irriso\] Let $g, g':\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfy with $n=1$. Then:
1. ${\mathsf{W}^{1}_{g_{}}}\simeq\mathsf{W}^{1}_{g'}$ if and only if $g(0)=g'(i)$ for some $i\in{\mathbb{Z}}$;
2. ${\mathsf{W}^{1}_{g_{}}}$ is irreducible if and only if $g(0)\notin \{ [i]_q \mid i\in{\mathbb{Z}}\}\cup\left\{ -\frac{1}{q-1}\right\}$.
For (a), suppose ${\mathsf{W}^{1}_{g_{}}}\simeq\mathsf{W}^{1}_{g'}$. By the eigenvalues of $XY$ on ${\mathsf{W}^{1}_{g_{}}}$ are $g(i)$, for $i\in{\mathbb{Z}}$ and similarly the eigenvalues of $XY$ on $\mathsf{W}^{1}_{g'}$ are $g'(i)$, for $i\in{\mathbb{Z}}$. Thus, $g$ and $g'$ must have the same image and in particular $g(0)=g'(i)$ for some $i\in{\mathbb{Z}}$. Conversely, if the latter holds then the map $\phi:{\mathsf{W}^{1}_{g_{}}}\rightarrow\mathsf{W}^{1}_{g'}$ given by $\phi(p)(t)=t^ip(t)$ for all $p\in{\mathbb{F}}[t^{\pm 1}]$ is an isomorphism.
For (b), first observe that for $i\in{\mathbb{Z}}$ we have $g(0)=[i]_q\iff g(-i)=0$. Thus, if $g(0)=[i]_q$ for some $i\in{\mathbb{Z}}$, then $t^{-i}{\mathbb{F}}[t]$ is invariant under the actions of $X$ and $Y$, so ${\mathsf{W}^{1}_{g_{}}}$ is not irreducible in this case. Next observe that $g(0)=-\frac{1}{q-1}\iff g$ is not injective $\iff g$ is constant. It follows that if $g(0)=-\frac{1}{q-1}$, then $(t-1){\mathbb{F}}[t^{\pm 1}]$ is a proper subrepresentation and hence ${\mathsf{W}^{1}_{g_{}}}$ is not irreducible. This proves the direct implication in (b). For the converse, by the observations above, we can assume that $g(i)\neq 0$ for all $i\in{\mathbb{Z}}$ and that $g$ is injective. Let $\mathsf{S}$ be a nonzero subrepresentation of ${\mathsf{W}^{1}_{g_{}}}$. By repeatedly applying the operator $X$ to a chosen nonzero element of $\mathsf{S}$, we will obtain a nonzero element of $\mathsf{S}\cap{\mathbb{F}}[t]$. Let $p$ be one such element, chosen so that it has minimum degree, say $p=\sum_{k=0}^d a_k t^k$, with $a_d\neq 0$. Since $g(i)\neq 0$ for all $i\in{\mathbb{Z}}$, the minimality of $p$ implies that $a_0\neq 0$. Then $$\mathsf{S}\cap{\mathbb{F}}[t]\ni (XY-g(d)).p=\sum_{k=0}^{d-1} (g(k)-g(d))a_k t^k.$$ By the minimality of $p$ we must have $(XY-g(d)).p=0$. Hence, $g(0)=g(d)$ and the injectivity of $g$ gives $d=0$. It follows that $t^0\in\mathsf{S}$ and thus $\mathsf{S}={\mathsf{W}^{1}_{g_{}}}$.
Now that we understand the representations ${\mathsf{W}^{n}_{g_{}}}$, we will consider their restriction to ${{\mathbb{F}}_q[x, y]}$ via each of the two embeddings $$\begin{aligned}
\label{E:emb:sigma}
\sigma &: {{\mathbb{F}}_q[x, y]}\rightarrow \mathbb{A}_1(q), \quad\quad x\mapsto X, \quad y\mapsto YX-XY=(q-1)XY+1;\\ \label{E:emb:tau}
\tau &: {{\mathbb{F}}_q[x, y]}\rightarrow \mathbb{A}_1(q), \quad\quad x\mapsto YX-XY=(q-1)XY+1, \quad y\mapsto Y.\end{aligned}$$
We consider first the restriction relative to $\sigma$. In this case, the action of ${{\mathbb{F}}_q[x, y]}$ on ${\mathsf{W}^{n}_{g_{}}}$ is given by $$\label{rest:w:qp}
x.t^{i}=t^{i+n}, \quad \quad y.t^{i}=\left((q-1)g(i)+1\right)t^i,\quad \quad \text{for all $i\in{\mathbb{Z}}$.}$$
\[L:qwa:rest:sigma\] Consider the restriction map $\sigma$ given in to view the representations ${\mathsf{W}^{n}_{g_{}}}$ as representations of ${{\mathbb{F}}_q[x, y]}$.
1. Let $n\in{\mathbb{Z}}_{>0}$ and assume $g:\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfies . Then ${\mathsf{W}^{n}_{g_{}}}\simeq \bigoplus_{k=0}^{n-1}{\mathsf{W}^{1}_{g_{k}}}$ as representations of ${{\mathbb{F}}_q[x, y]}$, where $g_k(i)=g(k+in)$, for $0\leq k<n$ and $i\in{\mathbb{Z}}$.
2. Let $g, g':\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfy with $n=1$. Then ${\mathsf{W}^{1}_{g_{}}}\simeq\mathsf{W}^{1}_{g'}$ as representations of ${{\mathbb{F}}_q[x, y]}$ if and only if $g(0)=g'(i)$ for some $i\in{\mathbb{Z}}$.
3. Assume $g:\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfies with $n=1$. Then ${\mathsf{W}^{1}_{g_{}}}$ has trivial socle as a representation of ${{\mathbb{F}}_q[x, y]}$, i.e., it has no irreducible ${{\mathbb{F}}_q[x, y]}$-subrepresentations.
Part (a) follows directly from Lemma \[L:qwa:dec\] and part (b) follows from the proof of Proposition \[P:qwa:w:irriso\](a), as the argument for the direct implication in Proposition \[P:qwa:w:irriso\](a) used only the restriction of the action to the subalgebra generated by $XY$, which coincides with the subalgebra generated by $YX-XY$.
For part (c), suppose by way of contradiction that $\mathsf{S}$ is an irreducible ${{\mathbb{F}}_q[x, y]}$-subrepresentation of ${\mathsf{W}^{1}_{g_{}}}$. Let $0\neq s\in\mathsf{S}$. Then $x.s\neq 0$ and thus ${{\mathbb{F}}_q[x, y]}x.s=\mathsf{S}$, which is a contradiction as $s\notin {{\mathbb{F}}_q[x, y]}x.s$.
In the conditions of Lemma \[L:qwa:rest:sigma\], it can be checked that ${\mathsf{W}^{1}_{g_{}}}$ has maximal ${{\mathbb{F}}_q[x, y]}$-subrepresentations if and only if $g$ is constant.
Now we consider the restriction of ${\mathsf{W}^{n}_{g_{}}}$ to ${{\mathbb{F}}_q[x, y]}$ relative to the map $\tau$ defined in . In this case, the action of ${{\mathbb{F}}_q[x, y]}$ on ${\mathsf{W}^{n}_{g_{}}}$ is given by $$\label{rest:w:qp:tau}
x.t^{i}=\left((q-1)g(i)+1\right)t^i, \quad \quad y.t^{i}=g(i)t^{i-n},\quad \quad \text{for all $i\in{\mathbb{Z}}$.}$$
\[L:qwa:rest:tau\] Consider the restriction map $\tau$ given in to view the representations ${\mathsf{W}^{n}_{g_{}}}$ as representations of ${{\mathbb{F}}_q[x, y]}$.
1. Let $n\in{\mathbb{Z}}_{>0}$ and assume $g:\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfies . Then ${\mathsf{W}^{n}_{g_{}}}\simeq \bigoplus_{k=0}^{n-1}{\mathsf{W}^{1}_{g_{k}}}$ as representations of ${{\mathbb{F}}_q[x, y]}$, where $g_k(i)=g(k+in)$, for $0\leq k<n$ and $i\in{\mathbb{Z}}$.
2. Let $g, g':\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfy with $n=1$. Then ${\mathsf{W}^{1}_{g_{}}}\simeq\mathsf{W}^{1}_{g'}$ as representations of ${{\mathbb{F}}_q[x, y]}$ if and only if $g(0)=g'(i)$ for some $i\in{\mathbb{Z}}$.
3. Assume $g:\mathbb{Z} \rightarrow {\mathbb{F}}$ satisfies with $n=1$. If $g(0)\notin \{ [i]_q \mid i\in{\mathbb{Z}}\}$ then ${\mathsf{W}^{1}_{g_{}}}$ has trivial socle as a representation of ${{\mathbb{F}}_q[x, y]}$, i.e., it has no irreducible ${{\mathbb{F}}_q[x, y]}$-subrepresentations. If $g(0)=[i]_q$ for some $i\in{\mathbb{Z}}$ then ${\mathbb{F}}t^{-i}$ is the unique irreducible ${{\mathbb{F}}_q[x, y]}$-subrepresentation of ${\mathsf{W}^{1}_{g_{}}}$.
The proof is the same as the proof of Lemma \[L:qwa:rest:sigma\], except for part (c). For this part, suppose that $\mathsf{S}$ is an irreducible ${{\mathbb{F}}_q[x, y]}$-subrepresentation of ${\mathsf{W}^{1}_{g_{}}}$. If there is $0\neq s\in\mathsf{S}$ such that $y.s\neq 0$, then we obtain a contradiction as in the proof of Lemma \[L:qwa:rest:sigma\](c), showing that no such irreducible ${{\mathbb{F}}_q[x, y]}$-subrepresentation of ${\mathsf{W}^{1}_{g_{}}}$ exists. If $g(0)\notin \{ [i]_q \mid i\in{\mathbb{Z}}\}$ then $g(i)\neq 0$ for all $i\in{\mathbb{Z}}$, so $y.s\neq 0$ for all $s\neq 0$ and the first claim follows. Now suppose $g(0)=[i]_q$ for some $i\in{\mathbb{Z}}$. Then $g(k)=0\iff k=-i$. In particular, $x.t^{-i}=t^{-i}$ and $y.t^{-i}=0$, so that ${\mathbb{F}}t^{-i}$ is an irreducible ${{\mathbb{F}}_q[x, y]}$-subrepresentation of ${\mathsf{W}^{1}_{g_{}}}$. If $\mathsf{S}$ is any irreducible ${{\mathbb{F}}_q[x, y]}$-subrepresentation of ${\mathsf{W}^{1}_{g_{}}}$, then the argument above implies that $y.s=0$ for all $s\in\mathsf{S}$, and this in turn implies that $\mathsf{S}\subseteq{\mathbb{F}}t^{-i}$, which establishes the second claim in (c).
**Acknowledgments.** The authors wish to thank G. Benkart and M. Ondrus for helpful comments and suggestions on a preliminary version of this manuscript. They would also like to thank the anonymous referees for their valuable comments and for suggesting the approach in Subsection \[SS:restriction\].
[1]{}
V. Bavula, *Classification of the simple modules of the quantum [W]{}eyl algebra and the quantum plane*, Quantum groups and quantum spaces ([W]{}arsaw, 1995), Banach Center Publ., vol. 40, Polish Acad. Sci., Warsaw, 1997, pp. 193–201.
V. Bavula and F. van Oystaeyen, *The simple modules of certain generalized crossed products*, J. Algebra **194** (1997), no. 2, 521–566.
G. Benkart and M. Ondrus, *Whittaker modules for generalized [W]{}eyl algebras*, Represent. Theory **13** (2009), 141–164.
F.H. Jackson, *[$q$]{}-difference equations*, Amer. J. Math. **32** (1910), no. 4, 305–314.
C. Kassel, *Quantum [G]{}roups*, Graduate Texts in Mathematics, vol. 155, Springer-Verlag, New York, 1995.
Yu.I. Manin, *Quantum groups and noncommutative geometry*, Université de Montréal, Centre de Recherches Mathématiques, Montreal, QC, 1988.
<span style="font-variant:small-caps;">Samuel A. Lopes</span>\
*CMUP, Faculdade de Ciências, Universidade do Porto, Rua do Campo Alegre 687\
4169-007 Porto, Portugal*\
`[email protected]`\
<span style="font-variant:small-caps;">João N. P. Lourenço</span>\
*Faculdade de Ciências, Universidade do Porto, Rua do Campo Alegre 687\
4169-007 Porto, Portugal*\
`[email protected]`
[^1]: The author was partially funded by the European Regional Development Fund through the programme COMPETE and by the Portuguese Government through the FCT – Fundação para a Ciência e a Tecnologia under the project PEst-C/MAT/UI0144/2013.
[^2]: The author was supported by Fundação Calouste Gulbenkian through the undergraduate research programme *Novos Talentos em Matemática*.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Results on two-particle angular correlations are presented in proton-proton collisions at center of mass energies of 7 TeV, over a broad range of pseudorapidity and azimuthal angle. In very high multiplicity events at 7 TeV, a pronounced structure emerges in the two-dimensional correlation function for particle pairs with intermediate $p_{\mathrm{T}}$ of 1–3 GeV/c, in the kinematic region $2.0 < |\Delta\eta| < 4.8$ and small $\Delta\phi$. This structure, which has not been observed in pp collisions before, is similar to what is known as the “ridge” in heavy ion collisions. It is not predicted by commonly used proton-proton Monte Carlo models and is not seen in lower multiplicity pp collisions. Updated studies of this new effect as a function of particle transverse momentum, rapidity and event characteristics are shown.'
address:
author:
- 'Dragos Velicanu for the CMS collaboration[^1]'
title: Ridge correlation structure in high multiplicity pp collisions with CMS
---
=1
Long-range, near-side ($\Delta\phi \approx 0$) ridge-like azimuthal correlations for $2.0$$<$$|\Delta\eta|$$<$$4.8$ have recently been observed for the first time in high multiplicity pp collisions at [$\sqrt{s}$]{} = 7 TeV [@Khachatryan:2010gv]. The novel structure resembles similar features observed in relativistic heavy-ion experiments. This striking feature is most evident in the intermediate transverse momentum range of both $1$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$3$ GeV/c and $1$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$3$ GeV/c. A steep increase of the near-side associated yield with multiplicity has been found in the data.
Following up the first observation of the ridge correlation structure in high multiplicity pp collisions at [$\sqrt{s}$]{} = 7 TeV, new results are presented in this paper to study the detailed event multiplicity, transverse momentum and pseudorapidity gap ($\Delta\eta$) dependence of the ridge effect using the full statistics data collected in 2010. With the nearly 4$\pi$ solid-angle acceptance of the silicon tracker and dedicated high multiplicity high-level trigger (HLT) setup, the CMS experiment has a unique capability in studying this novel effect [@JINST].
The same analysis procedure, as used in the ridge measurement of the central heavy-ion collisions [@ref:HIN-11-001-PAS], is applied in order to make direct comparison to the heavy-ion results, for full details see [@ref:HIN-11-006-PAS]. The pp data used in this extended analysis are collected under almost the same condition as those used in the publication of the first pp ridge observation [@Khachatryan:2010gv] with about a factor of 2 increase in statistics, 660K $N \geq 110$ events.
![ Two-dimensional (2-D) per-trigger-particle associated yield of charged hadrons as a function of $\Delta\eta$ and $\Delta\phi$ with jet peak cutoff for better demonstration of the ridge from high multiplicity ($N \geq 110$) pp collisions at [$\sqrt{s}$]{} = 7 TeV, for (a) $2 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$ 3$ GeV/c and $1 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$ 2$ GeV/c and (b) $5 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$ 6$ GeV/c and $1 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$ 2$ GeV/c []{data-label="fig:corr2D_N110_20110420"}](corr2D_N110_20110420 "fig:"){width="0.49\linewidth"} ![ Two-dimensional (2-D) per-trigger-particle associated yield of charged hadrons as a function of $\Delta\eta$ and $\Delta\phi$ with jet peak cutoff for better demonstration of the ridge from high multiplicity ($N \geq 110$) pp collisions at [$\sqrt{s}$]{} = 7 TeV, for (a) $2 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$ 3$ GeV/c and $1 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$ 2$ GeV/c and (b) $5 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$ 6$ GeV/c and $1 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$ 2$ GeV/c []{data-label="fig:corr2D_N110_20110420"}](corr2D_N110_highpT_20110510 "fig:"){width="0.49\linewidth"}
The per-trigger-particle associated yield distribution of charged hadrons as a function of $\Delta\eta$ and $\Delta\phi$ in high multiplicity ($N \geq 110$) pp collisions at [$\sqrt{s}$]{} = 7 TeV with trigger particles with $2$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$3 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$ and associated particles with $1$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$2 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$ is shown in Fig. \[fig:corr2D\_N110\_20110420\] obtained with the full statistics data in 2010. The ridge-like structure is clearly visible at $\Delta\phi \approx 0$ extending to $|\Delta\eta|$ of at least 4 units as previously observed in Ref. [@Khachatryan:2010gv]. However, at higher [$p_\mathrm{T}^{\mathrm{trig}}$]{} of 5–6[${\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}$]{} as presented in Fig. \[fig:corr2D\_N110\_20110420\], the ridge almost disappears. The absolute values of $\Delta\eta$ and $\Delta\phi$ are used in the analysis, thus the resulting distributions are symmetric about ($\Delta\eta$,$\Delta\phi$)=(0,0) by construction.
In order to fully explore the detailed properties of both short-range jet-like correlations and long-range ridge-like structure, especially its dependence on event multiplicity, transverse momentum and $|\Delta\eta|$, the associated yield distributions are obtained in eight bins ($2 \leq N < 35$, $35 \leq N < 45$, $45 \leq N < 60$, $60 \leq N < 90$, $N \geq 90$, $N \geq 110$, $N \geq 130$, $N \geq 150$) of charged particle multiplicity and six bins (0.1–1, 1–2, 2–3, 3–4, 4–5 and 5–6$ {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$) of particle transverse momentum. The 1-D $\Delta\phi$ azimuthal correlation functions are calculated by integrating over the $0.0<|\Delta\eta|<1.0$ and $2.0<|\Delta\eta|<4.0$ region, defined as the jet region and ridge region, respectively.
![ Integrated near-side ($|\Delta\phi| $$<$$ \Delta\phi_{\rm ZYAM}$) associated yields for $2 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$ 3 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$ and $1 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$ 2 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$, above the minimum level found by the ZYAM procedure, as a function of $|\Delta\eta|$ for the high multiplicity ($N \geq 110$) pp collisions at [$\sqrt{s}$]{} = 7 TeV. The statistical uncertainties are shown as bars, while the brackets denote the systematic uncertainties. []{data-label="fig:yieldvseta_pp_N110_trg2_ass1"}](yieldvseta_pp_N110_trg2_ass1_20110422){width="60.00000%"}
The near-side (small $\Delta\phi$ region) integrated associated yield is calculated for both jet and ridge regions relative to the constant background, details in Ref. [@ref:HIN-11-001-PAS]. Fig. \[fig:yieldvseta\_pp\_N110\_trg2\_ass1\] presents the resulting near-side associated yield as a function of $|\Delta\eta|$ (in slices of 0.6 units) in high multiplicity ($N \geq 110$) pp collisions at [$\sqrt{s}$]{} = 7 TeV with trigger particles with $2$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$3 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$ and associated particles with $1$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$2 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$. The high multiplicity data exhibit a jet-like correlation peak in the yield for small $|\Delta\eta|$ and show significant and roughly constant yield out to the highest $|\Delta\eta|$ regions. This is qualitatively similar to what has been observed in central PbPb collisions at [$\sqrt{s_{_{NN}}}$]{} = 2.76 TeV [@ref:HIN-11-001-PAS] but is completely absent in minimum bias pp collisions as well as pp MC models.
![ Integrated near-side ($|\Delta\phi| $$<$$ \Delta\phi_{\rm ZYAM}$) associated yields for the long-range ridge region ($2<|\Delta\eta|<4$) with $1 $$<$$ {\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$ 2 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$, above the minimum level found by the ZYAM procedure, as a function of [$p_\mathrm{T}^{\mathrm{trig}}$]{} for five multiplicity bins ($2 \leq N < 35$, $35 \leq N < 90$, $N \geq 90$, $N \geq 110$, $N \geq 130$) of pp collisions at [$\sqrt{s}$]{} = 7 TeV. The statistical uncertainties are shown as bars, while the brackets denote the systematic uncertainties. []{data-label="fig:yieldvspt_pp_ass1_eta24"}](yieldvspt_pp_ass1_eta2-4_final){width="105.00000%"}
Figure \[fig:yieldvspt\_pp\_ass1\_eta24\] shows the integrated near-side associated yield of the ridge region correlations with $1$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$2 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$ (the [$p_{\mathrm{T}}$]{} range where the ridge effect appears to be strongest) as a function of [$p_\mathrm{T}^{\mathrm{trig}}$]{} in five bins of event multiplicity. The ridge yield is almost zero for the first two low multiplicity bins in Fig. \[fig:yieldvspt\_pp\_ass1\_eta24\]. In the high multiplicity region ($N \geq 90$), the ridge yield first increases steadily with [$p_\mathrm{T}^{\mathrm{trig}}$]{}, reaches a maximum around [$p_\mathrm{T}^{\mathrm{trig}}$]{} $\sim$ 2–3 GeV/c and drops at higher [$p_\mathrm{T}^{\mathrm{trig}}$]{} .
![ Integrated near-side ($|\Delta\phi| $$<$$ \Delta\phi_{\rm ZYAM}$) associated yields for the short-range jet region ($0<|\Delta\eta|<1$) and the long-range ridge region ($2<|\Delta\eta|<4$), with $2$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$3 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$ and $1 $$<$${\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$ 2 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$, above the minimum level found by the ZYAM procedure, as a function of event multiplicity from pp collisions at [$\sqrt{s}$]{} = 7 TeV. The statistical uncertainties are shown as bars, while the brackets denote the systematic uncertainties. []{data-label="fig:yieldvsmult_pp_trg2_ass1"}](yieldvsmult_pp_trg2_ass1_eta0-1_final "fig:"){width="0.50\linewidth"} ![ Integrated near-side ($|\Delta\phi| $$<$$ \Delta\phi_{\rm ZYAM}$) associated yields for the short-range jet region ($0<|\Delta\eta|<1$) and the long-range ridge region ($2<|\Delta\eta|<4$), with $2$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$3 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$ and $1 $$<$${\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$ 2 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$, above the minimum level found by the ZYAM procedure, as a function of event multiplicity from pp collisions at [$\sqrt{s}$]{} = 7 TeV. The statistical uncertainties are shown as bars, while the brackets denote the systematic uncertainties. []{data-label="fig:yieldvsmult_pp_trg2_ass1"}](yieldvsmult_pp_trg2_ass1_eta2-4_final "fig:"){width="0.50\linewidth"}
The multiplicity dependence of the near-side associated yield in the jet and ridge region is illustrated in Fig. \[fig:yieldvsmult\_pp\_trg2\_ass1\] for one transverse momentum bin of $2$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{trig}}}}$$<$$3 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$ and $1$$<$${\ensuremath{p_\mathrm{T}^{\mathrm{assoc}}}}$$<$$2 {\ensuremath{{\,\text{Ge\hspace{-.08em}V\hspace{-0.16em}/\hspace{-0.08em}}c}}\xspace}$, the [$p_{\mathrm{T}}$]{} bin where the ridge effect appears to be strongest. The ridge effect gradually turns on with event multiplicity around $N \sim$ 50–60 and shows a tendency to saturate when it reaches $N \sim 120$, although this is not yet conclusive with current statistical and systematic uncertainties.
In summary, comprehensive studies of the ridge correlation structure in high multiplicity pp events, as a function of event multiplicity, particle transverse momentum and $\Delta\eta$ using full statistics pp data in 2010, are presented in this paper, which provides further information on the properties of the novel ridge phenomena. This is an important step forward toward understanding the physical origin of the ridge in high multiplicity pp collisions and put additional constraints on various theoretical models.
References {#references .unnumbered}
==========
[^1]: For the full list of CMS authors and acknowledgments, see appendix “Collaborations”.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[*Making use of an universal quantum network – QCPU proposed by me, it is obtained that the whole quantum network which can implement some the known quantum algorithms including Deutsch algorithm, quantum Fourier transformation, Shor’s algorithm and Grover’s algorithm*]{}.'
address: |
CCAST(World Laboratory) P.O.Box 8730, Beijing 100080$^1$\
and Laboratory of Quantum Communication and Quantum Computing\
University of Science and Technology of China$^2$\
Department of Modern Physics, University of Science and Technology of China\
P.O. Box 4, Hefei 230027, People’s Republic of China$^3$
author:
- 'An Min WANG$^{1,2,3}$'
title: ' Quantum CPU and Quantum Algorithm [^1]'
---
\#1 \#1
Quantum computing has attracted a lot of physicists since Shor found his quantum algorithm of factorization of a large number. This is because that Shor’s algorithm can speed up exponentially the factorization computation in a quantum computer than doing this in a classical computer. From then, quantum computer appears very promising and quantum algorithm plays a vital role. Actually, the high efficiency of quantum computer is just profited by the good quantum algorithms.
Both classical and quantum algorithms consist of a series of quantum computing steps. In the quantum algorithm, a computing step can be a transformation or a quantum measurement. The final quantum computing task made of these steps is represented usually by the summation and product of them, which formed an total unitary transformation. In order to implement a given quantum algorithm including quantum simulating procedure, one needs to design, assemble and scale these quantum computing steps so that one finally can obtain a whole quantum network.
Roughly speaking, the quantum network can be divided into three levels. First, the fundamental quantum network consists of a few basic quantum gates. Usually it can be constructed in the different methods. Secondly, the simple quantum network corresponds to a given computing step, which consists of the quantum gates and/or the fundamental quantum networks acting synchronously in time. The last is the complicated quantum network is made of the fundamental and simple quantum networks. A quantum algorithm is usually implemented by a complicated quantum network.
As is well known, Barenco [*et.al*]{}’s proposed a beautiful method to design a quantum network in terms of the elementary gates. However, so far one has not found how to assemble the quantum network for the summation of simultaneous transformations into a multiplication form. For example, although we have the quantum network for quantum Fourier transformation, it only can act on one state. Moreover, for the product of a series of successive transformations such as the time evolution operator, one doesn’t know how to scale them into a whole quantum network. At present, except for the simplest Deustch’s algorithm, I have not seen there is the whole quantum networks for the known main quantum algorithms. It is perhaps difficult to arrive at this aim only by Barenco [*et.al*]{}’s method. Thus, it seems to me that it is worth finding the new method since the integrality of the quantum network is necessary in practice. In this letter, by making use of an universal quantum network – QCPU proposed by me, I obtain the whole quantum network which can implement some the known quantum algorithms including Deutsch algorithm, quantum Fourier transformation, Shor’s algorithm and Grover’s algorithm. In fact, QCPU also can be used in quantum simulating.
An universal quantum network for quantum computing $U$ is defined as: $$Q(U)=\prod_{m,n=0}^{2^k-1}\exp\{(U_{mn}{\mbox{$|{m}\rangle$}}{\mbox{$\langle{n}|$}}\otimes I_A)\cdot C_A^\dagger\}\label{UQN}$$ where $I_R$ and $I_A$ are the identity matrices in the register space and the auxiliary qubit space respectively, and $C_A^\dagger=I_R\otimes c_A^\dagger=I_R\otimes {\mbox{$|{1}\rangle$}}{}_A{}_A{\mbox{$\langle{0}|$}}$. If the graphics rules for the factor with form $\exp\{(U_{mn}{\mbox{$|{m}\rangle$}}{\mbox{$\langle{n}|$}})\cdot C_A^\dagger\}$ are given out, the picture of quantum network $Q(U)$ can be drawn easily. It is easy to verify that $Q(U)$ can implement a general quantum computing. In this sense, it can be called the Quantum CPU. It is the most important that $Q(U)$ has two very useful new properties $$Q(U_1+U_2+\cdots+U_r)=Q(U_1)Q(U_2)\cdots Q(U_r),$$ $$Q(U_1U_2\cdots U_r)= I_R\otimes I_A+ C_A^\dagger\left(\prod_{j=1}^{r} C_AQ(U_j)\right) C_AC_A^\dagger,\label{CQC}$$ where $C_A$ is the so-called “[*Connector*]{}" defined by $
C_A=I_R\otimes c_A=I_R\otimes {\mbox{$|{0}\rangle$}}{}_A{}_A{\mbox{$\langle{1}|$}}$. It is used to the preparing transformed state so that this prepared state can be used in the successive transformation. Furthermore, note that there are the relations $c_A^2=c_A^{\dagger 2}=0; c_A c_A^\dagger+c_A^\dagger c_A=I_A$, thus $c_A$ and $c_A^\dagger$ can be thought of as the fermionic annihilate and create operators respectively in the auxiliary qubit. In order to give out the realization of QCPU for the product of transformation in a form of full multiplication, eq.(\[CQC\]) can be rewritten as $$\bar{Q}(U_1U_2\cdots U_r)=(I_R)_{input}\otimes \left[C_A^\dagger\left(\prod_{j=1}^{r} C_AQ(U_j)\right) C_AC_A^\dagger\right]_{out},\label{CQCP}$$ while the initial state is now prepared as $({\mbox{$|{\Psi(t)}\rangle$}})_{input}\otimes ({\mbox{$|{\Psi(t)}\rangle$}}\otimes {\mbox{$|{0}\rangle$}}_A)_{out}$. Therefore, it seems to me this new construction of the universal quantum network is scalable easily.
This QCPU is universal because it can implement a general quantum computing task, including the general quantum algorithm and the quantum simulating procedure. This QCPU and its realizations are standard and easy-assemble because they only have two kinds of basic elements and two auxiliary elements. This QCPU and its realizations are scalable because they can easily connect each other. This QCPU is favor to the design of quantum algorithm including quantum simulating because it gives out the standard and explicit realization for each computing step. This QCPU is possible helpful for programming it since its simplicity in design. In terms of the QCPU, it is easy to obtain the whole quantum networks for the known main quantum algorithms as seen in the following.
Deutsch’s problem is the first quantum algorithm and is the simplest one. However, it explains some important properties in quantum computing. Only taking one input qubit $x$ and one output qubit $f(x)$ (unknown function), we altogether have four possible values of the function $ f_1(0)=f_1(1)=0; f_2(0)=f_2(1)=1; f_3(0)=0$ and $f_3(1)=1;f_4(0)=1$ and $ f_4(1)=0$, because each $f(0)$ and $f(1)$ have two possible values. This problem is to judge whether $f$ is constant ($f(0)=f(1)$, such as $f_1$ or $f_2$) or balanced ($f(0)\neq f(1)$, such as $f_3$ or $f_4$). Intuitively, the best classical strategy is to calculate $f$ explicitly for input $0$ and $1$, and then to compare them. Therefore, in order to solve this problem, one has to two times calculations for the function $f$. However, quantum algorithm only needs single calculation to get the result. This algorithm was first proposed by Deutsch. Now it has been improved and extended. Deutsch’s algorithm can be described by the following four steps. First, prepare the first and second qubits respectively in ${\mbox{$|{0}\rangle$}}$ and ${\mbox{$|{1}\rangle$}}$. The total quantum state is ${\mbox{$|{01}\rangle$}}$. Secondly, act on each qubit by Hadamard transformation $H=(\sigma_1+\sigma_3)/\sqrt{2}$, where $\sigma_1,\sigma_3$ are usual Pauli matrix. Thus, it leads that $H\otimes H{\mbox{$|{01}\rangle$}}=({\mbox{$|{00}\rangle$}}-{\mbox{$|{01}\rangle$}}+{\mbox{$|{10}\rangle$}}-{\mbox{$|{11}\rangle$}})/2$. Thirdly, define a two-qubit gate $U_f:
{\mbox{$|{i,j}\rangle$}}\rightarrow {\mbox{$|{i,j\oplus f(i)}\rangle$}}$, where $i,j=0,1$ and $\oplus$ means addition and mod 2. It acts on the above superposition state. The last, again act on each qubit by Hadamard transformation. It is easy to verify that the final state is ${\mbox{$|{01}\rangle$}}$ (if $f=f_1$); $-{\mbox{$|{01}\rangle$}}$ (if $f=f_2$); ${\mbox{$|{11}\rangle$}}$ (if $f=f_3$); $-{\mbox{$|{11}\rangle$}}$ (if $f=f_4$). Therefore, to make a measurement to the first qubit will reveal if the function is constant (output 0) or balanced (output 1).
Obviously, Deutsch’s algorithm can be written as the following: $$\begin{aligned}
U({\rm Deutsch})&=& \left(
\begin{array}{cccc}
1&0&0&0\\
0&(-)^{f(0)}\delta_{f(0)f(1)}&0&\epsilon_{f(0)f(1)}\\
0&0&1&0\\
0&\epsilon_{f(0)f(1)}&0&(-)^{f(0)}\delta_{f(0)f(1)}
\end{array}\right)\\[10pt]
&=&\left(\begin{array}{cc}
1&0\\
0&1\end{array}\right)\otimes
\left(\begin{array}{cc}
1&0\\
0&(-)^{f(0)}\delta_{f(0)f(1)}\end{array}\right)+
\left(\begin{array}{cc}
0&1\\
1&0\end{array}\right)\otimes
\left(\begin{array}{cc}
0&0\\
0&\epsilon_{f(0)f(1)}\end{array}\right), \label{QCPU1}\end{aligned}$$ where $\delta_{f(0)f(1)}= 1$ if $f(0)=f(1)$, otherwise it is zero, and $\epsilon_{f(0)f(1)}$ is an usual antisymmetric tensor defined by $\epsilon_{01}=-\epsilon_{10}=1$, the others are zero. Furthermore, Deutsch’s algorithm can be simplified as an unitary transformation acting on the single qubit: $$V({\rm Deutsch})=\left(\begin{array}{cc}
(-)^{f(0)}\delta_{f(0)f(1)}&\epsilon_{f(0)f(1)}\\
\epsilon_{f(0)f(1)}&
(-)^{f(0)}\delta_{f(0)f(1)}\end{array}\right).$$ Thus, from this matrix it follows that eq.(\[QCPU1\]) is equivalent to the realization of QCPU $$Q(V({\rm Deutsch}))=\prod_{m,n=0}^1 \exp\{[V_{mn}({\rm Deutsch}){\mbox{$|{x_m}\rangle$}}{\mbox{$\langle{x_n}|$}}\otimes I_A]\cdot C^\dagger\}.$$ Its action on ${\mbox{$|{1}\rangle$}}\otimes {\mbox{$|{0}\rangle$}}_A$ will give the expected result. To make a measurement in basis ${\mbox{$|{11}\rangle$}}$ will reveal if the function is constant (output 0) or balanced (output 1).
Quantum Fourier transformation plays an important role in quantum algorithm including factorization, search and simulating. Its matrix $F$ reads $$F=\frac{1}{\sqrt{N}}\sum_{m,n=0}^{N-1}\e^{2\pi\I mn/N}{\mbox{$|{x_m}\rangle$}}{\mbox{$\langle{p_n}|$}},\quad F^{-1}=\frac{1}{\sqrt{N}}\sum_{m,n=0}^{N-1}\e^{-2\pi\I mn/N}{\mbox{$|{p_m}\rangle$}}{\mbox{$\langle{x_n}|$}}.$$ Quantum Fourier transformation can be rewritten as $$F=\sum_{n=0}^{2^k-1}\prod^k_{j=1}\frac{1}{\sqrt{2}}({\mbox{$|{0}\rangle$}}+\e^{2^j\I\pi n/(2^k-1)}{\mbox{$|{1}\rangle$}}){\mbox{$\langle{p_n}|$}}=\sum_{n=0}^{2^k-1}\prod_{\otimes,j=1}^k B_j(2^j\pi n/(2^k-1)H_j{\mbox{$|{0}\rangle$}}{\mbox{$\langle{p_n}|$}}.$$ From the definition of QCPU (\[UQN\]), it is easy to get $$Q(F)=\prod_{n=0}^{2^k-1}Q[B(n)HM_{0n}]= \prod_{m=0}^{2^k-1}\prod_{n=0}^{2^k-1}\exp\{[(B(n)H)_{m0}{\mbox{$|{m}\rangle$}}{\mbox{$\langle{n}|$}}\otimes I_A] \cdot C^\dagger\},$$ where $B(n)H=\prod_{\otimes,j=0}^{2^k-1}B_j[2^j\pi n/(2^k-1)]]H_j$ and $M_{0n}={\mbox{$|{0}\rangle$}}{\mbox{$\langle{x_n}|$}}$.
Shor’s algorithm is used to the factorization of a large number $N$. It can speed up exponentially computing in a quantum computer than doing this in a classical computer. Shor’s algorithm can be described by the following five steps. First, start with two $k-$qubit registers in ${\mbox{$|{0}\rangle$}}{\mbox{$|{0}\rangle$}}$, then prepare the first register into a superposition with the equal weight in terms of Fourier transformation or $k-$qubit Hadamard gate denoted by $H$: $$H{\mbox{$|{0}\rangle$}}{\mbox{$|{0}\rangle$}}=\sum_{n=0}^{2^k-1}{\mbox{$|{n}\rangle$}}{\mbox{$|{0}\rangle$}}.$$ Secondly, select randomly a factor $a$ and make a mapping $$\sum_{n=0}^{2^k-1}{\mbox{$|{n}\rangle$}}{\mbox{$|{0}\rangle$}}\stackrel{G}{\longrightarrow} \sum_{n=0}^{2^k-1}{\mbox{$|{n}\rangle$}}{\mbox{$|{a^n {\rm mod} N}\rangle$}}.$$ Obviously $$G=\sum_{n=0}^{2^k-1}{\mbox{$|{n}\rangle$}}{\mbox{$|{a^n {\rm mod}N}\rangle$}}{\mbox{$\langle{n}|$}}{\mbox{$\langle{0}|$}}.$$ Thirdly, measure the second register by $I_1\otimes {\mbox{$|{a^m {\rm mod}N}\rangle$}}{\mbox{$\langle{a^m {\rm mod}N}|$}}$ and obtain the result: $\sum_{j=0}^{[2^k/r]-1} |jr + l\rangle |u\rangle$. Fourthly, do Fourier transformation $F$ to the first register so that $$U_{\rm DFT}{\mbox{$|{jr+l}\rangle$}}= \frac{1}{\sqrt{2^{k}}}\sum_{y=0}^{2^k-1} \exp \{2\pi {\rm i}(jr+l)y/2^k\} {\mbox{$|{y}\rangle$}}.$$ and obtain the final state $\displaystyle \frac{1}{\sqrt{r}}\sum_{m=0}^{r-1} \exp (2\pi {\rm i}lm/r) |m2^k/r\rangle $. The last, measure the first register in the basis $y=m2^k/r$. If one obtains one values $y$, then solve equation $y/2^k=m/r$ to find the period. Once $r$ is known the factors of $N$ are obtained by calculating the greatest common divisor of $N$ and $a^{r/2}\pm 1$.
Thus, the main steps in Shor’s algorithm can be represented by one total transformation matrix: $$U({\rm Shor})= (F\otimes I_2)M(a^m {\rm mod}N)GH;\quad M(n)=I_1\otimes {\mbox{$|{n}\rangle$}}{\mbox{$\langle{n}|$}}.$$ The product of the serval matrices is an easy problem. After we know the form of $U({\rm Shor})$, we are able to obtain the parameters to determine the realization of QCPU which can implement Shor’s algorithm. But, it is more convenient, sometime more efficient, to connect the several realizations of QCPU for quantum computing steps and quantum measurement corresponding this algorithm together. Since we have had the realization of QCPU for quantum Fourier transformation, we only need the realizations of QCPU for $H$ and $G$. Obviously, starting from the initial state ${\mbox{$|{0}\rangle$}}_1\otimes{\mbox{$|{0}\rangle$}}_2{\mbox{$|{0}\rangle$}}_A$, it follows that $Q(H)$ is $$Q(H)= C_A^\dagger\prod_{j=1}^k (CQ(H_j\otimes I_2))\cdot C_A C_A^\dagger.$$ Note that $H_j$ is a Hadamard transformation only acting on the $j-$ qubit in the first register and the second register has kept the original state. We do not need to add the other input register. The realization of QCPU for Mapping $G$ can read $$Q(G)=\prod_{n=0}^{2^k-1}\exp\{({\mbox{$|{n}\rangle$}}{\mbox{$|{a^n {\rm mod}N}\rangle$}}{\mbox{$\langle{n}|$}}{\mbox{$\langle{0}|$}})\otimes I_A)\cdot C^\dagger\}.$$ Therefore the whole quantum network for Shor’s factorization can be obtained: $$\bar{Q}({\rm Shor})=(I_R)_{input}\otimes\left(C_A^\dagger C_AQ(F\otimes I_2)M(a^m{\rm mod}N)\otimes I_AC_A Q(G)C_A Q(H) C_AC_A^\dagger\right)_{out}.$$ From this example about Shor’s algorithm, it seem to me to design a quantum algorithm is to seek an appropriate summation and product form of a series of simultaneous and successive quantum transformations as well as quantum measurements, and to implement a quantum algorithm is to give the realizations of QCPU for every transformations and then connect them including quantum measurements together. It is so standard and convenient.
Grover’s algorithm is used to search the expected term in the unstructured data. It can be described by the following four steps. First, start with a $k-$qubit registers in ${\mbox{$|{0}\rangle$}}$, then prepare it into a superposition with the equal weight in terms of Fourier transformation or $k-$Hadamard gate, that is $H{\mbox{$|{0}\rangle$}}=\sum_{n=0}^{2^k-1}{\mbox{$|{n}\rangle$}}$. Secondly, do a reflection: $$R_2=I-2{\mbox{$|{j}\rangle$}}{\mbox{$\langle{j}|$}}=\sum_{n=0}^{2^k-1}(-1)^{\delta_{jm}}{\mbox{$|{x_m}\rangle$}}{\mbox{$\langle{x_m}|$}},$$ where $j$ corresponds to the expected data. Thirdly, make the following operation: $$R_1=F^{-1}R_0F=F^{-1}[2{\mbox{$|{0}\rangle$}}{\mbox{$\langle{0}|$}}-I]F=-F^{-1}\sum_{m=0}^{2^k-1}(-)^{\delta_{0 m}}{\mbox{$|{x_m}\rangle$}}{\mbox{$\langle{x_m}|$}}F,$$ where $I$ is an identity matrix, $F$ is a quantum Fourier transformation, $F^{-1}$ is its inverse and $R_0=2{\mbox{$|{0}\rangle$}}{\mbox{$\langle{0}|$}}-I$. The last, repeat $R_1R_2$ $\sqrt{N}\pi/4$ times and then do all measurements.
Since the quantum network for quantum Fourier transformation has been obtained and its inverse has the similar realization but its parameters with a negative sign. While $R_0$ and $R_2$ is diagonal, it is very easy to get from our definition of QCPU $$\begin{aligned}
Q(R_0)&=&\exp\{(2{\mbox{$|{0}\rangle$}}{\mbox{$\langle{0}|$}}\otimes I_A)\cdot C^\dagger\}\exp\{-C_A^\dagger\}\\
Q(R_2)&=&Q(I_R)\exp\{(-2{\mbox{$|{x_j}\rangle$}}{\mbox{$\langle{x_j}|$}}\otimes I_A)\cdot C^\dagger\}\exp\{C_A^\dagger\}.\end{aligned}$$ Thus the quantum network for Grover’s algorithm is just obtained $$Q({\rm Grover})=(I_R)_{input}\otimes\left(C_A^\dagger C_AQ(F^{-1})CQ(R_0)CQ(F)CQ(R_2)CQ(H)C_AC_A^\dagger\right)_{out}.$$
In conclusion, it is showed that the whole quantum networks of the known main quantum algorithms can be described by the realizations of QCPU. Actually, it is also able to obtain the quantum network to simulate Schrödinger equation. Therefore, QCPU is possible to play a useful role in searching for new quantum algorithm. This research is on progressing.
I would like to thank Artur Ekert for his great help and for his hosting my visit to center of quantum computing in Oxford University.
R. P. Feynman. [*Int. J. Theor. Phys.*]{} [**21**]{} (1982)467 D. Deutsch. [*Proc. R. Soc. Lond. A*]{} [**400**]{} (1985)97 P. W. Shor. in [*Proceedings of the 35th Annual Symposium on the Foundations of Computer Science*]{}, IEEE Computer Society Press, Los Alamitos, CA. ed. S. Goldwasser. 1994. 20; [*SIAM Journal of Computation*]{} [**26**]{} (1997)1484 A. Barenco [*et. al*]{}, “Elementary gates for quantum computing“, quant-ph/9503016 A. Ekert, Phys. Scripta, [**T76**]{} (1998)218 An Min Wang, ”An Universal Quantum Network – Quantum CPU”, preprint quant-ph/9910089 L. K. Grover. in [*Proceedings, 28th Annual ACM Symposium on the Theory of Computing*]{}. 1996. 122, preprint quant-ph/9605034; [*Phys. Rev. Lett.*]{} [**79**]{} (1997)325 An Min Wang, ”Quantum CPU and quantum simulating", preprint quant-ph/9910090
[^1]: Supported by the National Natural Science Foundation of China under Grant No. 69773052 and the Fellowship of China Academy of Sciences
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe ways to define and calculate $L_1$-norm signal subspaces which are less sensitive to outlying data than $L_2$-calculated subspaces. We start with the computation of the $L_1$ maximum-projection principal component of a data matrix containing $N$ signal samples of dimension $D$. We show that while the general problem is formally NP-hard in asymptotically large $N$, $D$, the case of engineering interest of fixed dimension $D$ and asymptotically large sample size $N$ is not. In particular, for the case where the sample size is less than the fixed dimension ($N<D$), we present in explicit form an optimal algorithm of computational cost $2^N$. For the case $N \geq D$, we present an optimal algorithm of complexity $\mathcal O(N^D)$. We generalize to multiple $L_1$-max-projection components and present an explicit optimal $L_1$ subspace calculation algorithm of complexity ${\mathcal O}(N^{DK-K+1})$ where $K$ is the desired number of $L_1$ principal components (subspace rank). We conclude with illustrations of $L_1$-subspace signal processing in the fields of data dimensionality reduction, direction-of-arrival estimation, and image conditioning/restoration.'
author:
- |
\
EDICS: MLR-ICAN, MLR-LEAR, MDS-ALGO, ASP-ANAL, SSP-SSAN\
Submitted: June 27, 2013 - Revised: April 4, 2014
title: |
\
Optimal Algorithms for $L_1$-subspace Signal Processing[^1]
---
[***Index Terms* —**]{} Dimensionality reduction, direction-of-arrival estimation, eigen-decomposition, erroneous data, faulty measurements, $L_1$ norm, $L_2$ norm, machine learning, outlier resistance, subspace signal processing.
Introduction
============
A general intention of subspace signal processing is to partition the vector space of the observed data and isolate the subspace of the signal component(s) of interest from the disturbance (noise) subspace. Subspace signal processing theory and practice rely, conventionally, on the familiar $L_2$-norm based singular-value decomposition (SVD) of the data matrix. The SVD solution traces its origin to the fundamental problem of $L_2$-norm low-rank matrix approximation [@Eckart1936], which is equivalent to the problem of maximum $L_2$-norm data projection with as many projection (“principal”) components as the desired low-rank value [@Golub1996]. Among the many strengths of $L_2$-norm principal component analysis (PCA), one may point out the simplicity of the solution, scalability (new principal directions add on to the previous ones), and correspondence to maximum-likelihood estimation (MLE) under the assumption of additively Gaussian-noise corrupted data.
Practitioners have long observed, however, that $L_2$-norm PCA is sensitive to the presence of outlier values in the data matrix, that is, erroneous values that are away from the nominal data, appear only few times in the data matrix, and are not to appear again under normal system operation upon design. Recently, there has been an –arguably small but growing– interest in pursuing $L_1$-norm based approaches to deal with the problem of outliers in principal-components design \[\]-\[\].[^2] The growth in interest can also be credited incidentally to the popularity of compressed sensing methods \[\]-\[\] that rely on $L_1$-based calculations in signal reconstruction.
This paper makes a case for $L_1$-subspace signal processing. Interestingly, in contrast to $L_2$, subspace decomposition under the $L_1$ error minimization criterion and the $L_1$ projection maximization criterion are not the same. A line of recent research pursues calculation of $L_1$ principal components under error minimization \[\]-\[\]. The error surface is non-smooth and the problem non-convex resisting attempts to guaranteed optimization even with exponential computational cost. Suboptimal algorithms may be developed by viewing the minimization function as a convex nondifferentiable function with a bounded Lipschitz constant [@Combettes2002], [@Nesterov2013]. A different approach is to calculate subspace components by $L_1$ projection maximization \[\]-\[\].[^3] No algorithm has appeared so far with guaranteed convergence to the criterion-optimal subspace and no upper bounds are known on the expended computational effort.
In this present work, given any data matrix ${\mathbf X}\in \mathbb R^{D \times N}$ of $N$ signal samples of dimension $D$, we show that the general problem of finding the maximum $L_1$-projection principal component of $\mathbf X$ is formally NP-hard for asymptotically large $N$, $D$. We prove, however, that the case of engineering interest of fixed given dimension $D$ is not NP-hard. In particular, for the case where $N<D$, we present in explicit form an algorithm to find the optimal component with computational cost $2^N$. For the case where the sample size exceeds the data dimension ($N \geq D$) –which is arguably of higher interest in signal processing applications– we present an algorithm that computes the $L_1$-optimal principal component with complexity $\mathcal O\big(N^{\text{rank}(\mathbf X)}\big)$, $\text{rank}(\mathbf X) \leq D$. We generalize the effort to the problem of calculating $K$, $1<K \leq \text{rank}( \mathbf X)$, $L_1$ components (necessarily a joint computational problem) and present an explicit optimal algorithm for multi-component subspace design of complexity ${\mathcal O}(N^{\text{rank}(\mathbf X)K-K+1})$. We conclude with illustrations of the developed $L_1$ subspaces in problems from the fields of dimensionality reduction, direction-of-arrival estimation, and image reconstruction that demonstrate the inherent outlier resistance of $L_1$ subspace signal processing.
The rest of the paper is organized as follows. Section II presents the problem statement and establishes notation. Section III is devoted to the optimal computation of the $L_1$ principal component. Section IV generalizes to optimal $L_1$-subspace calculation (joint multiple $L_1$ components). Experimental illustrations are given in Section V and a few concluding remarks are drawn in Section VI.
Problem Statement
=================
Consider $N$ real-valued measurements $\mathbf x_1, \mathbf x_2, \ldots, \mathbf x_N$ of dimension $D$ that form the $D\times N$ data matrix $${\bf X}=[{\bf x}_1\;\;{\bf x}_2\;\ldots\;{\bf x}_N].$$ In the common version of the low-rank approximation problem, one seeks to describe (approximate) data matrix ${\bf X}$ by a rank-$K$ product ${\bf R}{\bf S}^T$ where ${\bf R} \in \mathbb R^{D \times K}$, ${\bf S} \in \mathbb R^{N \times K}$, $K \leq\min(D,N)$. Given the observation data matrix ${\bf X}$, $L_2$-norm matrix approximation minimizes the sum of the element-wise squared error between the original matrix and its rank-$K$ surrogate in the form of Problem $\mathcal P_1^{L_2}$ defined below, $$\begin{split}
{\mathcal P}^{L_2}_1:\;\;\;\;&\left({\bf R}_{L_2},{\bf S}_{L_2}\right)=\operatorname*{arg\,min}_{{\bf R}\in{\mathbb R}^{D\times K},\;{\bf S}\in{\mathbb R}^{N\times K}}\left\|{\bf X}-{\bf R}{\bf S}^T\right\|_2
\end{split}
\label{eq:RS}$$ where $\|{\bf A}\|_2=\sqrt{\sum_{i,j}|A_{i,j}|^2}$ is the $L_2$ matrix norm (that is, Frobenius norm) of a matrix $\mathbf A$ with elements $A_{i,j}$. Problem ${\mathcal P}^{L_2}_1$ is our most familiar $K$-singular-value-decomposition ($K$-SVD) problem solved with computational complexity $\mathcal O \big((D+N)\min^2(D,N)\big)$ [@Golub1996]. $\mathcal P_1^{L_2}$ corresponds also to the statistical problem of maximum-likelihood estimation (MLE) of an unknown rank-$K$ matrix corrupted by additive element-wise independent Gaussian noise [@Vantrees].
We may expand (\[eq:RS\]) to $\displaystyle\min_{{\bf R}\in{\mathbb R}^{D\times K}}\min_{{\bf S}\in{\mathbb R}^{N\times K}}\left\|{\bf X}-{\bf R}{\bf S}^T\right\|_2$ and inner minimization results to ${\bf S}={\bf X}^T{\bf R}$ for any fixed ${\bf R}$, $\mathbf R^T \mathbf R = \mathbf I_K$, by the Projection Theorem [@Golub1996]. Hence, we obtain the equivalent problem $$\begin{split}
{\mathcal P}^{L_2}_2:\;\;\;\;&{\bf R}_{L_2}=\operatorname*{arg\,min}_{
{\bf R}\in{\mathbb R}^{D\times K},\,{\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}-{\bf R}{\bf R}^T{\bf X}\right\|_2
\end{split}
\label{eq:RR}$$ frequently referred to as left-side $K$-SVD. Since $\left\|{\bf A}\right\|_2^2=\text{tr}\left({\bf A}^T{\bf A}\right)$ where $\text{tr}(\cdot)$ denotes the trace of a matrix, the $L_2$ error minimization problem ${\mathcal P}^{L_2}_2$ is also equivalent to the $L_2$ projection (energy) maximization problem $$\begin{split}
{\mathcal P}^{L_2}_3:\;\;\;\;&{\bf R}_{L_2}=\operatorname*{arg\,max}_{{\bf R}\in{\mathbb R}^{D\times K},\,{\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}^T{\bf R}\right\|_2.
\end{split}
\label{eq:R}$$ The optimal ${\bf R}_{L_2}$ (in ${\mathcal P}^{L_2}_1$, ${\mathcal P}^{L_2}_2$, and ${\mathcal P}^{L_2}_3$) is known simply as the $K$ dominant-singular-value left singular vectors of the original data matrix or $K$ dominant-eigenvalue eigenvectors of $\mathbf X \mathbf X^T$ [@Eckart1936], [@Golub1996]. Note that, if $K<D$ and we possess the solution ${\bf R}_{L_2}^{(K)}$ for $K$ singular/eigen vectors in (\[eq:RS\]), (\[eq:RR\]), (\[eq:R\]), then the solution for rank $K+1$ is derived readily by ${\bf R}_{L_2}^{(K+1)}=\left[{\bf R}_{L_2}^{(K)}\;\;{\bf r}_{L_2}^{(K+1)}\right]$ with $${\bf r}_{L_2}^{(K+1)}=\operatorname*{arg\,max}_{{\bf r}\in{\mathbb R}^D,\;\left\|{\bf r}\right\|_2=1}\left\|{\bf X}^T \left({\bf I}_D-{\bf R}_{L_2}^{(K)}{{\bf R}_{L_2}^{(K)}}^T\right){\bf r} \right\|_2.$$ This is known as the PCA scalability property.
$L_2$ PCA, as reviewed above in ${\mathcal P}^{L_2}_1$, ${\mathcal P}^{L_2}_2$, and ${\mathcal P}^{L_2}_3$, has a simple solution, is scalable (new principal directions add on to the previous ones), and corresponds to MLE under the assumption of Gaussian additively corrupted data. Practitioners, however, have long noticed a drawback. By minimizing the sum of squared errors, $L_2$ principal component calculation becomes sensitive to extreme error value occurrences caused by the presence of outlier measurements in the data matrix (measurements that are numerically distant from the nominal data, appear only few times in the data matrix, and are not to appear under normal system operation upon design). Motivated by this observed drawback of $L_2$ subspace signal processing, in this work we study and pursue subspace-decomposition approaches that are based on the $L_1$ norm, $$\left\|{\bf A}\right\|_1=\sum_{i,j}\left|A_{i,j}\right|.$$ We may “translate" the three equivalent $L_2$ optimization problems (\[eq:RS\]), (\[eq:RR\]), (\[eq:R\]) to new problems that utilize the $L_1$ norm as follows, $$\begin{aligned}
\begin{split}
&{\mathcal P}^{L_1}_1:\;\;\;\;\left({\bf R}_{L_1},{\bf S}_{L_1}\right)=\operatorname*{arg\,min}_{{\bf R}\in{\mathbb R}^{D\times K},\;{\bf R}^T{\bf R}={\bf I}_K,\;{\bf S}\in{\mathbb R}^{N\times K}}\left\|{\bf X}-{\bf R}{\bf S}^T\right\|_1,
\end{split} \\
\label{eq:RSL1}
\begin{split}
&{\mathcal P}^{L_1}_2:\;\;\;\;{\bf R}_{L_1}=\operatorname*{arg\,min}_{{\bf R}\in{\mathbb R}^{D\times K},\; {\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}-{\bf R}{\bf R}^T{\bf X}\right\|_1,
\end{split} \\
\begin{split}
&{\mathcal P}^{L_1}_3:\;\;\;\;{\bf R}_{L_1}=\operatorname*{arg\,max}_{{\bf R}\in{\mathbb R}^{D\times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}^T{\bf R}\right\|_1.
\end{split}
\label{eq:RL1}\end{aligned}$$ A few comments appear useful at this point: (i) ${\mathcal P}^{L_1}_1$ corresponds to MLE when the additive noise disturbance follows a Laplacian distribution [@Vantrees]. (ii) The optimal metric value in ${\mathcal P}^{L_1}_3$ with a single dimension ($K=1$) is the complexity parameter for saddle-point methods when used to provide an approximate solution to the $\ell_1$/nuclear-norm Dantzig selector problem [@Nesterov2013]. (iii) Under the $L_1$ norm, the three optimization problems ${\mathcal P}^{L_1}_1$, ${\mathcal P}^{L_1}_2$, and ${\mathcal P}^{L_1}_3$ are no longer equivalent. (iv) Under $L_1$, the PCA scalability property does not hold (due to loss of the Projection Theorem). (v) Even for reduction to a single dimension (rank $K=1$ approximation), the three problems are difficult to solve. (vi) As of today, it is unknown which of the subspaces defined in $\mathcal P_{1}^{L_1}$, $\mathcal P_{2}^{L_1}$, and $\mathcal P_{3}^{L_1}$ exhibits stronger resistance against faulty measurements; indeed, none of these problems had been solved optimally so far for general $D,K$.
In this present work, we focus exclusively on ${\mathcal P}^{L_1}_3$. In Section III, we seek to find efficiently the principal maximum $L_1$ projection component of ${\bf X}$. In Section IV, we investigate the problem of calculating (jointly necessarily) multiple ($K>1$) $L_1$ projection components that maximize the $L_1$ “energy" of the data on the projection subspace.
The $L_1$-norm Principal Component {#sec:OneComponent}
==================================
In this section, we concentrate on the calculation of the $L_1$-maximum-projection component of a data matrix $\mathbf X \in \mathbb R^{D \times N}$ (Problem $\mathcal P_3^{L_1}$ in (\[eq:RL1\]), $K=1$). First, we show that the problem is in general NP-hard and review briefly suboptimal techniques from the literature. Then, we prove that, if the data dimension $D$ is fixed, the principal $L_1$-norm component is in fact computable in polynomial time and present an algorithm that calculates the $L_1$ principal component of $\mathbf X$ with complexity ${\mathcal O}\left(N^{\text{rank}({\bf X})}\right)$, $\text{rank}({\bf X})\leq D$.
Hardness of Problem and an Exhaustive-search Algorithm Over the Binary Field
----------------------------------------------------------------------------
We present a fundamental property of Problem ${\mathcal P}^{L_1}_3$, $K=1$, that will lead us to an efficient solution. The property is presented in the form of Proposition \[prop:quad\] below and interprets ${\mathcal P}^{L_1}_3$ as an equivalent quadratic-form maximization problem over the binary field.
For any data matrix $\mathbf X \in \mathbb R^{D \times N}$, the solution to ${\mathcal P}^{L_1}_3:{\bf r}_{L_1}=\operatorname*{arg\,max}_{{\bf r}\in{\mathbb R}^D,\left\|{\bf r}\right\|_2=1}\left\|{\bf X}^T{\bf r}\right\|_1$ is given by $${\bf r}_{L_1}=\frac{{\bf X}{\bf b}_\text{opt}}{\left\|{\bf X}{\bf b}_\text{opt}\right\|_2}
\label{eq:rL1}$$ where$${\bf b}_\text{opt}=\operatorname*{arg\,max}_{{\bf b}\in\{\pm1\}^N}\left\|{\bf X}{\bf b}\right\|_2=\operatorname*{arg\,max}_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf X}^T{\bf X}{\bf b}.
\label{eq:bopt}$$ In addition, $\left\|{\bf X}^T {\bf r}_{L_1}\right\|_1=\left\|{\bf X}{\bf b}_\text{{opt}}\right\|_2$. \[prop:quad\]
*Proof:* For any ${\bf z}\in{\mathbbm R}^N$, $\displaystyle\left\|{\bf z}\right\|_1=\text{sgn}\left({\bf z}\right)^T{\bf z}=\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf z}$. Therefore, we can rewrite the optimization problem as $$\max_{\left\|{\bf r}\right\|_2=1}\left\|{\bf X}^T{\bf r}\right\|_1=\max_{\left\|{\bf r}\right\|_2=1}\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf X}^T{\bf r}=\max_{{\bf b}\in\{\pm1\}^N}\max_{\left\|{\bf r}\right\|_2=1}{\bf r}^T{\bf X}{\bf b}.
\label{eq:rXb}$$ For any fixed vector ${\bf b}$, inner maximization in is solved by ${\bf r}=\frac{{\bf X}{\bf b}}{\left\|{\bf X}{\bf b}\right\|_2}$ and $$\max_{\left\|{\bf r}\right\|_2=1}{\bf r}^T{\bf X}{\bf b}=\left\|{\bf X}{\bf b}\right\|_2.
\label{eq:Xb2}$$ Combining (\[eq:rXb\]) and (\[eq:Xb2\]), we obtain $$\max_{\left\|{\bf r}\right\|_2=1}\left\|{\bf X}^T{\bf r}\right\|_1=\max_{{\bf b}\in\{\pm1\}^N}\left\|{\bf X}{\bf b}\right\|_2.$$ That is, $
\left\|{\bf X}^T{\bf r}_{L_1}\right\|_1=\left\|{\bf X}{\bf b}_\text{opt}\right\|_2$ where ${\bf b}_\text{opt}=\operatorname*{arg\,max}_{{\bf b}\in\{\pm1\}^N}\left\|{\bf X}{\bf b}\right\|_2$ and ${\bf r}_{L_1}=\frac{{\bf X}{\bf b}_\text{opt}}{\left\|{\bf X}{\bf b}_\text{opt}\right\|_2}$.
By Proposition \[prop:quad\], to find the principal $L_1$-norm component ${\bf r}_{L_1}$ we solve (\[eq:bopt\]) to obtain ${\bf b}_\text{opt}$ and then calculate $\frac{{\bf X}{\bf b}_\text{opt}}{\left\|{\bf X}{\bf b}_\text{opt}\right\|_2}$. The straightforward approach to solve (\[eq:bopt\]) is an exhaustive search among all $2^N$ binary vectors of length $N$. Therefore, with computational cost $2^N$, Proposition \[prop:quad\] identifies the $L_1$-optimal principal component of ${\bf X}$. As the data record size $N$ grows, calculation of the $L_1$ principal component by exhaustive search in (\[eq:bopt\]) becomes quickly infeasible. Proposition \[prop:NPhard\] below declares that, indeed, in its general form $\mathcal P_3^{L_1}$, $K=1$, is NP-hard for jointly asymptotically large $N, D$. McCoy and Tropp provide an alternative proof in [@McCoy2011], that is the earliest known to the authors.
The computation of the $L_1$ principal component of ${\bf X}\in{\mathbb R}^{D\times N}$ by maximum $L_1$-norm projection (Problem $\mathcal P_{3}^{L_1}$, $K=1$) is NP-hard in jointly asymptotic $N,D$. \[prop:NPhard\]
*Proof:* In (\[eq:rXb\]), for any fixed ${\bf r}\in{\mathbb R}^D$, ${\bf b}=\text{sgn}\left({\bf X}^T{\bf r}\right)$. Hence, $${\bf b}_\text{opt}=\text{sgn}\left({\bf X}^T{\bf r}_{L_1}\right).
\label{eq:brL1}$$ By (\[eq:rL1\]) and (\[eq:brL1\]), computation of the $L_1$ principal component of ${\bf X}$ is equivalent to computation of ${\bf b}_\text{opt}$ in (\[eq:bopt\]). Consider the special case of (\[eq:bopt\]) where ${\bf X}^T{\bf X}={\bf I}_N-{\bf a}{\bf a}^T$, $\mathbf a \in \mathbb R^{N}$, $\left\|{\bf a}\right\|_2=1$ (hence, $D=N-1$). Then, $$\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf X}^T{\bf X}{\bf b}=\max_{{\bf b}\in\{\pm1\}^N}\left\{\left\|{\bf b}\right\|_2^2-\left({\bf b}^T{\bf a}\right)^2\right\}=N-\min_{{\bf b}\in\{\pm1\}^N}\left({\bf b}^T{\bf a}\right)^2.$$ But $\min_{\mathbf b \in \{\pm 1 \}^{N}}(\mathbf b^T \mathbf a)^2$ is the NP-complete equal-partition problem [@Garey1979]. We conclude that computation of the $L_1$ principal component of ${\bf X}$ is NP-hard in jointly asymptotic $N,D$.
Existing Approaches in Literature
---------------------------------
Recently there has been a growing documented effort to calculate subspace components by $L_1$ projection maximization \[\]-\[\]. The work in [@Kwak2008] presented a suboptimal iterative algorithm for the computation of ${\bf r}_{L_1}$, which, following the formulation and notation of this present paper, initializes the solution to some arbitrary component $\mathbf r_{L_1}^{(0)}$ and executes $$\begin{aligned}
{\bf b}^{(i+1)}&=\text{sgn}\left({\bf X}^T{\bf r}_{L_1}^{(i)}\right),
\label{eq:Kwak1}\\
{\bf r}_{L_1}^{(i+1)}&=\frac{{\bf X}{\bf b}^{(i+1)}}{\left\|{\bf X}{\bf b}^{(i+1)}\right\|_2},
\label{eq:Kwak2}\end{aligned}$$ $i=0,1,2,\ldots$, until convergence. The work in [@Nie2011] presented an iterative algorithm for the joint computation of $K\geq1$ principal $L_1$-norm components. For the case where $K=1$, the iteration in [@Nie2011] simplifies to the iteration in [@Kwak2008] (that is, (\[eq:Kwak1\]), (\[eq:Kwak2\]) above). Therefore, for $K=1$, the algorithms in [@Kwak2008], [@Nie2011] are identical and can, in fact, be described by the simple single iteration $${\bf b}^{(i+1)}=\text{sgn}\left({\bf X}^T{\bf X}{\bf b}^{(i)}\right),\;\;\;i=1,2,\ldots,
\label{eq:Kwak3}$$ for the computation of ${\bf b}_\text{opt}$ in (\[eq:bopt\]). Equation , however, does not guarantee convergence to the $L_1$-optimal component solution (convergence to one of the many local maxima may be observed). In the following section, we present for the first time in the literature an optimal algorithm to calculate the $ L_1$ principal component of a data matrix with complexity polynomial in the sample size $N$ when the data dimension $D$ is fixed.
Exact Computation of the $L_1$ Principal Component in Polynomial Time
---------------------------------------------------------------------
Proposition \[prop:NPhard\] proves NP-hardness of the computation of the $L_1$ principal component ${\bf r}_{L_1}$ in $N,D$ (that is, when $N,D$ are jointly arbitrarily large). However, of engineering interest is the case of fixed data dimension $D$. In the following, we show for the first time in the literature that, if $D$ is fixed, then computation of ${\bf r}_{L_1}$ is no longer NP-hard (in $N$). We state our result in the form of Proposition \[prop:polynomial\] below.
For any fixed data dimension $D$, computation of the $L_1$ principal component of ${\bf X}\in{\mathbb R}^{D\times N}$ has complexity ${\mathcal O}\left(N^{\text{{rank}}({\bf X})}\right)$, $\text{{rank}}({\bf X})\leq D$. \[prop:polynomial\]
By Proposition \[prop:NPhard\], computation of the $ L_1$ principal component of $\mathbf X$ is equivalent to computation of $\mathbf b_{\text{opt}}$ in (\[eq:bopt\]). To prove Proposition \[prop:polynomial\], we will prove that ${\bf b}_\text{opt}$ can be computed with complexity ${\mathcal O}\left(N^{\text{rank}({\bf X})}\right)$. We begin our developments by defining $$d{\stackrel{\triangle}{=}}\text{rank}({\bf X})\leq D.
\label{eq:d}$$ Then, ${\bf X}^T{\bf X}$ also has rank $d$ and can be decomposed by $${\bf X}^T{\bf X}={\bf Q}{\bf Q}^T,\;\;\;{\bf Q}_{N\times d}=\left[{\bf q}_1\;{\bf q}_2\;\ldots\;{\bf q}_d\right],\;\;\;{\bf q}_i^T{\bf q}_j=0,\;i\neq j,
\label{eq:QQ}$$ where ${\bf q}_1$, ${\bf q}_2$, $\ldots$ , ${\bf q}_d$ are the $d$ eigenvalue-weighted eigenvectors of ${\bf X}^T{\bf X}$ with nonzero eigenvalue. By (\[eq:bopt\]), $${\bf b}_\text{opt}=\operatorname*{arg\,max}_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf Q}{\bf Q}^T{\bf b}=\operatorname*{arg\,max}_{{\bf b}\in\{\pm1\}^N}\left\|{\bf Q}^T{\bf b}\right\|_2.
\label{eq:Qb}$$
For the case $N<D$, the optimal binary vector ${\bf b}_\text{opt}$ can be obtained directly from (\[eq:bopt\]) by an exhaustive search among all $2^N$ binary vectors ${\bf b}\in\{\pm1\}^N$. Therefore, we can design the $L_1$-optimal principal component ${\bf r}_{L_1}$ with computational cost $2^N<2^D={\mathcal O}(1)$. For the case where the sample size exceeds the data dimension ($N\geq D$), we find it useful in terms of both theory and practice to present our developments separately for data rank $d=1$, $d=2$, and $2 < d \leq D$.\
*1) Case $d=1$:* If the data matrix has rank $d=1$, then ${\bf Q}={\bf q}_1$ and (\[eq:Qb\]) becomes $${\bf b}_\text{opt}=\operatorname*{arg\,max}_{{\bf b}\in\{\pm1\}^N}\left|{\bf q}_1^T{\bf b}\right|=\text{sgn}\left({\bf q}_1\right).$$ By (\[eq:rL1\]), the $L_1$-optimal principal component is $${\bf r}_{L_1}=\frac{{\bf X}\,\text{sgn}\left({\bf q}_1\right)}{\left\|{\bf X}\,\text{sgn}\left({\bf q}_1\right)\right\|_2}
\label{dequals1}$$ designed with complexity ${\mathcal O}\left(N\right)$. It is of notable practical importance to observe at this point that even when $\mathbf X$ is not of true rank one, presents us with a quality, trivially calculated approximation of the $L_1$ principal component of $\mathbf X$: Calculate the $L_2$ principal component $\mathbf q_1$ of the $N \times N$ matrix $\mathbf X^T \mathbf X$, quantize to $\text{sgn}(\mathbf q_1)$, and project and normalize to obtain $\mathbf r_{L_1} \simeq \mathbf X\,\text{sgn}(\mathbf q_1) / \| \mathbf X\,\text{sgn}(\mathbf q_1) \|_2$.\
*2) Case $d=2$:* If $d=2$, then ${\bf Q}=\left[{\bf q}_1\;\;{\bf q}_2\right]$ and (\[eq:Qb\]) becomes $${\bf b}_\text{opt}=\operatorname*{arg\,max}_{{\bf b}\in\{\pm1\}^N}\left\{\left({\bf q}_1^T{\bf b}\right)^2+\left({\bf q}_2^T{\bf b}\right)^2\right\}.
\label{eq:rank2}$$ The binary optimization problem was seen and solved in [@KP] by the auxiliary-angle method [@mackenthun], which was also used earlier in \[\],\[\]. Here, we define the $N\times1$ complex vector $${\bf z}{\stackrel{\triangle}{=}}{\bf q}_1+j{\bf q}_2$$ and rewrite (\[eq:rank2\]) as $${\bf b}_\text{opt}=\operatorname*{arg\,max}_{{\bf b}\in\{\pm1\}^N}\left|{\bf b}^T{\bf z}\right|.
\label{eq:bz}$$ We introduce the auxiliary angle $\phi\in\left[-\pi,\pi\right)$ and note that, for any complex scalar $w$, $$\text{Re}\left(we^{-j\phi}\right)\leq|w|$$ with equality if and only if $\phi=\text{angle}\left(w\right)$. That is, $$|w|=\max_{\phi\in\left[-\pi,\pi\right)}\text{Re}\left(we^{-j\phi}\right).$$ Therefore, the maximization in (\[eq:bz\]) can be rewritten as $$\begin{aligned}
\max_{{\bf b}\in\{\pm1\}^N}\left|{\bf b}^T{\bf z}\right|&=\max_{{\bf b}\in\{\pm1\}^N}\max_{\phi\in\left[-\pi,\pi\right)}\text{Re}\left({\bf b}^T{\bf z}e^{-j\phi}\right)=\max_{\phi\in\left[-\pi,\pi\right)}\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T\text{Re}\left({\bf z}e^{-j\phi}\right)
\label{eq:phib}\end{aligned}$$ where, for any given angle $\phi\in\left[-\pi,\pi\right)$, inner maximization is achieved by $${\bf b}\left(\phi\right)=\text{sgn}\left(\text{Re}\left({\bf z}e^{-j\phi}\right)\right).
\label{eq:bphi}$$ Then, the optimal vector ${\bf b}_\text{opt}$ in (\[eq:bz\]), i.e., the solution to (\[eq:bopt\]), is met if we scan the entire interval $\left[-\pi,\pi\right)$ and collect the locally optimal vector ${\bf b}\left(\phi\right)$ for any point $\phi\in\left[-\pi,\pi\right)$.
Interestingly, as we scan the interval $\left[-\pi,\pi\right)$, the locally optimal vector ${\bf b}\left(\phi\right)$ does not change unless the sign of $\text{Re}\left(z_ne^{-j\phi}\right)$ changes for some $n=1,2,\ldots,N$. Since the latter happens only at $\text{angle}\left(z_n\right)$ and $\text{angle}\left(z_n\right)+\pi$, we obtain $2N$ points in total at which ${\bf b}\left(\phi\right)$ changes. Next, we order the $2N$ points with complexity ${\mathcal O}\left(2N\log_22N\right)$ and create successively $2N$ binary vectors by changing each time the sign of $b_n$ if the $n$th element of ${\bf z}$ is the one that determines a sign change. It is observed that the $2N$ binary vectors that we obtain this way are pair-wise opposite (the vectors that are collected when $\phi\in\left[-\frac{\pi}{2},\frac{\pi}{2}\right)$ are opposite to the ones that are collected when $\phi\in\left[-\pi,-\frac{\pi}{2}\right)\cup\left[\frac{\pi}{2},\pi\right)$). Since opposite vectors result in the same metric value in (\[eq:bopt\]), we can restrict our search to $\left[-\frac{\pi}{2},\frac{\pi}{2}\right)$ and maintain optimality. Therefore, with overall complexity ${\mathcal O}\left(N\log_2N\right)$, we obtain a set of $N$ binary vectors that contains ${\bf b}_\text{opt}$. Then, we only have to evaluate the $N$ vectors against the metric of interest in (\[eq:bopt\]) to obtain ${\bf b}_\text{opt}$. We conclude that the $L_1$-optimal principal component of a rank-$2$ matrix ${\bf X}\in{\mathbb R}^{D\times N}$ is designed with complexity ${\mathcal O}\left(N\log_2N\right)$.
*3) Case $d>2$:* If $d>2$, we design the $L_1$-optimal principal component of ${\bf X}$ with complexity ${\mathcal O}\left(N^d\right)$ by considering the multiple-auxiliary-angle approach that was presented in [@KL] as a generalization of the work in [@KP].
Consider a unit vector ${\bf c}\in{\mathbbm R}^d$. By the Cauchy-Schwartz inequality, for any ${\bf a}\in{\mathbbm R}^d$, $${\bf a}^T{\bf c}\leq\left\|{\bf a}\right\|_2\left\|{\bf c}\right\|_2=\left\|{\bf a}\right\|_2$$ with equality if and only if ${\bf c}$ is codirectional with ${\bf a}$. Then, $$\max_{{\bf c}\in{\mathbbm R}^d,\;\left\|{\bf c}\right\|_2=1}{\bf a}^T{\bf c}=\left\|{\bf a}\right\|_2.
\label{eq:ca}$$ By (\[eq:ca\]), the optimization problem in (\[eq:Qb\]) becomes $$\max_{{\bf b}\in\{\pm1\}^N}\left\|{\bf Q}^T{\bf b}\right\|_2=\max_{{\bf b}\in\{\pm1\}^N}\max_{{\bf c}\in{\mathbbm R}^d,\; \left\|{\bf c}\right\|_2=1}{\bf b}^T{\bf Q}{\bf c}=\max_{{\bf c}\in{\mathbbm R}^d,\; \left\|{\bf c}\right\|_2=1}\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf Q}{\bf c}.
\label{eq:maxmax}$$ For every ${\bf c}\in{\mathbbm R}^d$, inner maximization in (\[eq:maxmax\]) is solved by the binary vector $${\bf b}({\bf c}) = \text{sgn}({\bf Q}{\bf c}),
\label{eq:bc}$$ which is obtained with complexity ${\mathcal O}(N)$. Then, by (\[eq:maxmax\]), the solution to the original problem in (\[eq:Qb\]) is met if we collect all binary vectors ${\bf b}({\bf c})$ returned as ${\bf c}$ scans the unit-radius $d$-dimensional hypersphere. That is, ${\bf b}_\text{opt}$ in (\[eq:Qb\]) is in[^4] $${\mathcal S}_1{\stackrel{\triangle}{=}}\hspace{-.5cm}\bigcup_{{\bf c}\in{\mathbbm R}^d,\,\left\|{\bf c}\right\|_2=1, \,c_d\geq0}\hspace{-.5cm}{\bf b}({\bf c}).
\label{eq:S1}$$ Two fundamental questions for the computational problem under consideration are what the size (cardinality) of set ${\mathcal S}_1$ is and how much computational effort is expended to form ${\mathcal S}_1$.
We address first the first question. We introduce the auxiliary-angle vector ${\boldsymbol\phi} =[\phi_1,\, \phi_2,\, \ldots, \,\phi_{d-1}]^T \in\Phi^{d-1}$, $\Phi{\stackrel{\triangle}{=}}\left[-\frac{\pi}{2},\frac{\pi}{2}\right)$, and parametrize ${\bf c}$ as follows, $${\bf c}(\boldsymbol{\phi}){\stackrel{\triangle}{=}}\begin{bmatrix}
\sin\phi_1\\
\cos\phi_1\sin\phi_2\\
\cos\phi_1\cos\phi_2\sin\phi_3\\
\vdots\\
\cos\phi_1\ldots\cos\phi_{d-2}\sin\phi_{d-1}\\
\cos\phi_1\ldots\cos\phi_{d-2}\cos\phi_{d-1}
\end{bmatrix}.$$ Then, we re-express the candidate set in (\[eq:S1\]) in the form $${\mathcal S}_1=\bigcup_{{\boldsymbol\phi}\in\Phi^{d-1}}{\bf b}({\boldsymbol\phi})
\label{eq:S1phi}$$ where, according to (\[eq:bc\]), $${\bf b}({\boldsymbol\phi}) =
\begin{bmatrix}
b_1({\boldsymbol\phi}), & b_2({\boldsymbol\phi}), & \ldots, & b_{N}({\boldsymbol\phi})
\end{bmatrix}^T
=\text{sgn}({\bf Q}{\bf c}({\boldsymbol\phi})).$$ We note that, for any point ${\boldsymbol\phi}$, each element $b_n({\boldsymbol\phi})$, $n=1,2, \ldots, N$, depends only on the corresponding row of ${\bf Q}$ and is determined by $b_n({\boldsymbol{\phi}})=\text{sgn}({\bf Q}_{n,:}\,{\bf c}({\boldsymbol{\phi}}))$. Hence, the value of the binary element $b_n({\boldsymbol{\phi}})$ changes only when $${\bf Q}_{n,:}\,{\bf c}({\boldsymbol{\phi}})=0.
\label{eq:Vnc}$$ To gain some insight into the process of introducing the auxiliary-angle vector $\boldsymbol{\phi}$, we notice that the points $\boldsymbol{\phi}$ that satisfy (\[eq:Vnc\]) determine a hypersurface (or $(d-2)$-manifold) in the $(d-1)$-dimensional space that partitions $\Phi^{d-1}$ into two regions. One region corresponds to $b_n=-1$ and the other corresponds to $b_n=+1$. A key observation in the algorithm is that, as ${\boldsymbol{\phi}}$ scans any of the two regions, the decision on $b_n$ does not change. Therefore, the $N$ rows of ${\bf Q}$ are associated with $N$ corresponding hypersurfaces that partition $\Phi^{d-1}$ into $P_1$ cells $C_1,C_2,\dots,C_{P_1}$ such that $\bigcup_{p=1}^{P_1}C_p=\Phi^{d-1}$, $C_p\cap C_q=\emptyset$ $\forall$ $p\ne q$, and each cell $C_p$ corresponds to a distinct vector ${\bf b}_p\in\{\pm1\}^N$. As a result, the candidate vector set is $S_1=\bigcup_{p=1}^P\{{\bf b}_p\}$.
In [@KL], it was shown that $P_1=\sum_{g=0}^{d-1}\binom{N-1}{g}$ if pairs of cells that correspond to opposite binary vectors (hence, equivalent vectors with respect to the metric of interest in (\[eq:Qb\])) are considered as one. Therefore, the candidate vector set ${\mathcal S}_1$ has cardinality $|{\mathcal S}_1|=\sum_{g=0}^{d-1}\binom{N-1}{g}={\mathcal O}\left(N^{d-1}\right)$. Fig. \[fig:surfaces\] presents a visualization of the algorithm/partition for the case of a data matrix $\mathbf X_{D \times N}$ of $N=8$ samples with rank $d=3\leq D\leq N$. Since $d=3$, the hypersurfaces (or $(d-2)$-manifolds) are, in fact, curves in the $2$-dimensional space that partition $\Phi^2$ into cells. The $P=\binom{7}{0} + \binom{7}{1} + \binom{7}{2} = 29 $ cells and associated binary candidate vectors are formed by the eight-row three-column eigenvector matrix ${\bf Q}$ of $\mathbf X^T \mathbf X$ and the scanning angle vector $\mathbf c ({\boldsymbol\phi})=[\sin \phi_1, \; \cos \phi_1 \sin \phi_2, \; \cos \phi_1 \cos \phi_2]^T$.
Regarding the cost of calculating $\mathcal S_1$, since each cell ${ C}$ contains at least one vertex (that is, intersection of $d-1$ hypersurfaces), see for example Fig. \[fig:surfaces\], it suffices to find all vertices in the partition and determine ${\bf b}$ for all neighboring cells. Consider $d-1$ arbitrary hypersurfaces; say, for example, ${\bf Q}_{1,:}{\bf c}({\boldsymbol\phi})=0$, ${\bf Q}_{2,:}{\bf c}({\boldsymbol\phi})=0$, $\ldots$, ${\bf Q}_{d-1,:}{\bf c}({\boldsymbol\phi})=0$. Their intersection satisfies ${\bf Q}_{1,:}{\bf c}({\boldsymbol\phi})={\bf Q}_{2,:}{\bf c}({\boldsymbol\phi})=\ldots={\bf Q}_{d-1,:}{\bf c}({\boldsymbol\phi})=0$ and is computed by solving the equation $${\bf Q}_{1:d-1,:}{\bf c}({\boldsymbol\phi})={\bf 0}.
\label{eq:Qc}$$ The solution to (\[eq:Qc\]) consists of the spherical coordinates of the unit vector in the null space of the $(d-1)\times d$ matrix ${\bf Q}_{1:d-1,:}$.[^5] Then, the binary vector ${\bf b}$ that corresponds to a neighboring cell is computed by $$\text{sgn}({\bf Q}\,{\bf c}({\boldsymbol\phi}))
\label{eq:sgnQc}$$ with complexity ${\mathcal O}(N)$. Note that (\[eq:sgnQc\]) presents ambiguity regarding the sign of the intersecting $d-1$ hypersurfaces. A straightforward way to resolve the ambiguity[^6] is to consider all $2^{d-1}$ sign combinations for the corresponding elements $b_1,b_2,\ldots,b_{d-1}$ and obtain the binary vectors of all $2^{d-1}$ neighboring cells. Finally, we repeat the above procedure for any combination of $d-1$ intersecting hypersurfaces among the $N$ ones. Therefore, the total number of binary candidates that we obtain (i.e., the cardinality of ${\mathcal S}_1$) is upper bounded by $2^{d-1}\binom{N}{d-1}={\mathcal O}(N^{d-1})$. Since complexity ${\mathcal O}(N)$ is required for each combination of $d-1$ rows of ${\bf Q}$ to solve (\[eq:sgnQc\]), the overall complexity of the construction of ${\mathcal S}_1$ is ${\mathcal O}(N^d)$ for any given matrix ${\bf Q}_{N\times d}$.
Our complete, new algorithm for the computation of the $L_1$-optimal principal component of a rank-$d$ matrix ${\bf X}\in{\mathbb R}^{D\times N}$ that has complexity ${\mathcal O}\left(N^d\right)$ is presented in detail in Fig. \[fig:algo\]. Computation of each element of $S_1$ (i.e., column of ${\bf B}$ in the algorithm) is performed independently of each other. Therefore, the proposed algorithm is fully parallelizable. The space complexity of the algorithm is $O(N)$, since after every computation of a new binary candidate the best binary candidate needs to be stored.
We note that the required optimal binary vector in (\[eq:Qb\]) can, alternatively, be computed through the algorithm in [@Allemand2001], [@Ferrez2005] with time complexity ${\mathcal O}\left(N^{d+1}\right)$ and space complexity at least ${\mathcal O}(N)$ based on the reverse search for cell enumeration in arrangements [@Avis1996] or with time complexity ${\mathcal O}(N^{d-1})$ but space complexity proportional to ${\mathcal O}\left(N^{d-1}\right)$ based on the incremental algorithm for cell enumeration in arrangements [@Edelsbrunner1986], [@Edelsbrunner1987]. Another algorithm that can solve (\[eq:Qb\]) with polynomial complexity is in [@BenAmeur2010]. Its time complexity is ${\mathcal O}(N^{d-1}\log N)$, while its space complexity is polynomially bounded by the output size (i.e., ${\mathcal O}(N^{d-1})$). In comparison to the above approaches, the algorithm in Fig. 2 is the fastest known with smallest (linear) space complexity. We conclude that the $L_1$-optimal principal component of a rank-$d$ data matrix ${\bf X}\in{\mathbb R}^{D\times N}$, $d \leq D \leq N$, is obtained with time complexity ${\mathcal O}\left(N^d\right)$ and space complexity ${\mathcal O}(N)$. That is, the time complexity is polynomial in the sample size with exponent equal to the rank of the data matrix, which is at most equal to the data dimension $D$. The space complexity is linear in the sample size.
Multiple $L_1$-norm Principal Components
========================================
In this section, we switch our interest to the joint design of $K>1$ principal $L_1$ components of a $D\times N$ data matrix ${\bf X}$. After we review suboptimal approaches from the recent literature, we generalize the result of the previous section and prove that, if the data dimension $D$ is fixed, then the $K$ principal $L_1$ components of ${\bf X}$ are computable in polynomial time ${\mathcal O}\left(N^{K \text{rank}(\mathbf X)-K+1}\right)$.
Exact Exhaustive-search Computation of Multiple $L_1$ Principal Components
--------------------------------------------------------------------------
For any $D\times K$ matrix ${\bf A}$, $$\max_{{\bf R}\in{\mathbbm R}^{D\times K},\,{\bf R}^T{\bf R}={\bf I}_K}\text{tr}\left({\bf R}^T{\bf A}\right)=\left\|{\bf A}\right\|_*
\label{eq:RAA}$$ where $\left\|{\bf A}\right\|_*$ denotes the nuclear norm (i.e., the sum of the singular values) of ${\bf A}$. Maximization in (\[eq:RAA\]) is achieved by ${\bf R}={\bf U}{\bf V}^T$ where ${\bf U}{\bf\Sigma}{\bf V}^T$ is the “compact” SVD of ${\bf A}$, ${\bf U}$ and ${\bf V}$ are $D\times d$ and $K\times d$, respectively, matrices with ${\bf U}^T{\bf U}={\bf V}^T{\bf V}={\bf I}_d$, ${\bf\Sigma}$ is a nonsingular diagonal $d\times d$ matrix, and $d$ is the rank of ${\bf A}$. This is due to the trace version of the Cauchy-Schwarz inequality [@Cauchy] according to which $$\begin{aligned}
\text{tr}\left({\bf R}^T{\bf A}\right)&=\text{tr}\left({\bf R}^T{\bf U}{\bf\Sigma}{\bf V}^T\right)=\text{tr}\left({\bf U}{\bf\Sigma}^\frac{1}{2}\cdot{\bf\Sigma}^\frac{1}{2}{\bf V}^T{\bf R}^T\right) \nonumber \\
&\leq\left\|{\bf U}{\bf\Sigma}^\frac{1}{2}\right\|_2\left\|{\bf\Sigma}^\frac{1}{2}{\bf V}^T{\bf R}^T\right\|_2=\left\|{\bf\Sigma}^\frac{1}{2}\right\|_2^2=\text{tr}\left({\bf\Sigma}\right)=\left\|{\bf A}\right\|_*
\label{eq:SA}\end{aligned}$$ with equality if $\left({\bf U}{\bf\Sigma}^\frac{1}{2}\right)^T={\bf\Sigma}^\frac{1}{2}{\bf V}^T{\bf R}^T$ which is satisfied by ${\bf R}={\bf U}{\bf V}^T$.
To identify the optimal $L_1$ subspace for any number of components $K$, we begin by presenting a property of ${\mathcal P}^{L_1}_3$ in the form of Proposition \[prop:nuclear\] below. Proposition \[prop:nuclear\] is a generalization of Proposition \[prop:quad\] and interprets ${\mathcal P}^{L_1}_3$ as an equivalent nuclear-norm maximization problem over the binary field.
For any data matrix $\mathbf X \in \mathbb R^{D \times N}$, the solution to ${\mathcal P}^{L_1}_3:\;\;\;\;{\bf R}_{L_1}=\operatorname*{arg\,max}_{{\bf R}\in{\mathbb R}^{D\times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf R}^T{\bf X}\right\|_1$ is given by $${\bf R}_{L_1}={\bf U}{\bf V}^T$$ where ${\bf U}$ and ${\bf V}$ are the $D\times K$ and $N\times K$ matrices that consist of the $K$ dominant-singular-value left and right, respectively, singular vectors of ${\bf X}{\bf B}_\text{{opt}}$ with $${\bf B}_\text{{opt}}=\operatorname*{arg\,max}_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf X}{\bf B}\right\|_*.
\label{eq:Bopt}$$ In addition, $\left\|{\bf R}_{L_1}^T{\bf X}\right\|_1=\left\|{\bf X}{\bf B}_\text{{opt}}\right\|_*$. \[prop:nuclear\]
*Proof:* We rewrite the optimization problem in (\[eq:RL1\]) as $$\begin{aligned}
\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}^T{\bf R}\right\|_1 &=\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\sum_{k=1}^K\left\|{\bf X}^T{\bf r}_k\right\|_1=\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\sum_{k=1}^K\max_{{\bf b}_k\in\{\pm1\}^N}{\bf b}_k^T{\bf X}^T{\bf r}_k \nonumber \\
&=\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\max_{{\bf B}\in\{\pm1\}^{N\times K}}\text{tr}\left({\bf B}^T{\bf X}^T{\bf R}\right) \nonumber \\
&=\max_{{\bf B}\in\{\pm1\}^{N\times K}}\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\text{tr}\left({\bf R}^T{\bf X}{\bf B}\right) = \max_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf X}{\bf B}\right\|_*.
\label{eq:RXB} \end{aligned}$$ That is, $\left\|{\bf R}_{L_1}^T{\bf X}\right\|_1=\left\|{\bf X}{\bf B}_\text{{opt}}\right\|_*$ where ${\bf B}_\text{opt}=\operatorname*{arg\,max}_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf X}{\bf B}\right\|_*$ and, by (\[eq:RAA\]) and (\[eq:SA\]), ${\bf R}_{L_1}={\bf U}{\bf V}^T$ where ${\bf U}{\bf\Sigma}{\bf V}^T$ is the “compact” SVD of ${\bf X}{\bf B}_\text{opt}$.
By Proposition \[prop:nuclear\], to find exactly the optimal $L_1$-norm projection operator ${\bf R}_{L_1}$ we can perform the following steps:
1. Solve (\[eq:Bopt\]) to obtain ${\bf B}_\text{opt}$.
2. Perform SVD on ${\bf X}{\bf B}_\text{opt}={\bf U}{\bf\Sigma}{\bf V}^T$.
3. Return ${\bf R}_{L_1}={\bf U}_{:,1:K}{\bf V}^T$.
Steps $1$ - $3$ offer for the first time a direct approach for the computation of the $K$ jointly-optimal $L_1$ principal components of ${\bf X}$. Step $1$ can be executed by an exhaustive search among all $2^{NK}$ binary matrices of size $N\times K$ followed by evaluation in the metric of interest in (\[eq:Bopt\]). That is, with computational cost $\mathcal O(2^{NK})$ we identify the $L_1$-optimal $K$ principal components of ${\bf X}$.
Existing Approaches in Literature
---------------------------------
For the case $K>1$, [@Kwak2008] proposed to design the first $L_1$ principal component ${\bf r}_{L_1}$ by the coupled iteration (\[eq:Kwak1\])-(\[eq:Kwak2\]) (which does not guarantee optimality) and then project the data onto the subspace that is orthogonal to ${\bf r}_{L_1}$; design the $L_1$ principal component of the projected data by the same coupled iteration; and continue similarly. To avoid the above suboptimal projection-greedy approach, [@Nie2011] presented an iterative algorithm for the computation of ${\bf R}_{L_1}$ altogether (that is the joint computation of the $K$ principal $L_1$ components). In the language of Proposition \[prop:nuclear\], the algorithm can be described as arbitrary initialization at some ${\bf R}_{L_1}^{(0)}$ followed by updates $$\begin{aligned}
{\bf B}^{(i+1)}&=\text{sgn}\left({\bf X}^T{\bf R}_{L_1}^{(i)}\right),\label{eq:UTA1}\\
\left({\bf U}^{(i+1)},{\bf\Sigma}^{(i+1)},{\bf V}^{(i+1)}\right)&=\text{SVD}\left({\bf X}{\bf B}^{(i+1)}\right),\\
{\bf R}_{L_1}^{(i+1)}&={\bf U}_{:,1:K}^{(i+1)}{{\bf V}^{(i+1)}}^T\label{eq:UTA3},\end{aligned}$$ for $i=0, 1,2,\ldots$, until convergence. Similar to the work in [@Kwak2008], the above iteration does not guarantee convergence to the $L_1$-optimal subspace.
Exact Computation of Multiple $L_1$ Principal Components in Polynomial Time
---------------------------------------------------------------------------
By the proof of Proposition \[prop:nuclear\], for any given ${\bf R}\in{\mathbb R}^{D\times K}$ the corresponding metric-maximizing binary matrix is ${\bf B}=\text{sgn}\left({\bf X}^T{\bf R}\right)$. Hence, $${\bf B}_\text{opt}=\text{sgn}\left({\bf X}^T{\bf R}_{L_1}\right).
\label{eq:BRL1}$$ By Proposition \[prop:nuclear\] and (\[eq:BRL1\]), computation of the $K$ principal $L_1$ components of ${\bf X}_{D \times N}$ is equivalent to computation of ${\bf B}_\text{opt}$ in (\[eq:Bopt\]), which indicates NP-hardness in $N,D$ (that is, when $N,D$ are arbitrarily large). As before, in this section we consider the case of engineering interest of fixed data dimension $D$. As in Section \[sec:OneComponent\], we show that, if $D$ is fixed, then computation of the $K$ principal $L_1$ components of ${\bf X}$ is no longer NP-hard (in $N$). We state our result in the form of the following proposition.
For any fixed data dimension $D$, optimal computation of the $K$ principal $L_1$ components of ${\bf X}\in{\mathbb R}^{D\times N}$ can be carried out with complexity ${\mathcal O}\left(N^{\text{{rank}}({\bf X})K-K+1}\right)$, $\text{{rank}}({\bf X})\leq D$. \[prop:polynomial\_multiple\]
To prove Proposition \[prop:polynomial\_multiple\], it suffices to prove that ${\bf B}_\text{opt}$ can be computed with complexity ${\mathcal O}\left(N^{\text{{rank}}({\bf X})K-K+1}\right)$. As in (\[eq:d\]), (\[eq:QQ\]), let $d$ denote the rank of ${\bf X}$ and ${\bf Q}{\bf Q}^T$ where ${\bf Q}\in{\mathbbm R}^{N\times d}$ is the eigen-decomposition matrix of ${\bf X}^T{\bf X}$. By (\[eq:Bopt\]), $${\bf B}_\text{{opt}}=\operatorname*{arg\,max}_{{\bf B}\in\{\pm1\}^{N\times K}}\sum_{k=1}^K\sqrt{\lambda_k\left[{\bf B}^T{\bf X}^T{\bf X}{\bf B}\right]}=\operatorname*{arg\,max}_{{\bf B}\in\{\pm1\}^{N\times K}}\sum_{k=1}^K\sqrt{\lambda_k\left[{\bf B}^T{\bf Q}{\bf Q}^T{\bf B}\right]}=\operatorname*{arg\,max}_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf Q}^T{\bf B}\right\|_*
\label{eq:QB}$$ where $\lambda_k [{\bf A}]$ denotes the $k$th eigenvalue of matrix ${\bf A}$, $k=1, \ldots, K$.
For the case $N<D$, the optimal binary matirx ${\bf B}_\text{opt}$ can be obtained directly from (\[eq:Bopt\]) by an exhaustive search among all $2^{NK}$ binary matrices ${\bf B}\in\{\pm1\}^{N\times K}$. Therefore, we can design the $L_1$-optimal $K$ principal components with computational cost $2^{NK}<2^{DK}={\mathcal O}(1)$.
For the (certainly more interesting) case where the sample size exceeds the data dimension, $N\geq D$, we present for the first time a generalized version of the approach in [@KP], [@KL] that introduces an orthonormal scanning matrix to maximize a rank-deficient nuclear norm. In particular, we observe by (\[eq:QB\]) that we need ${\bf B}_\text{{opt}}$ that solves $$\max_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf Q}^T{\bf B}\right\|_*\overset{(\ref{eq:RAA})}{=} \max_{{\bf B}\in\{\pm1\}^{N\times K}}\max_{{\bf C}\in{\mathbbm R}^{d\times K},\,{\bf C}^T{\bf C}={\bf I}_K}\text{tr}\left({\bf C}^T{\bf Q}^T{\bf B}\right)=\max_{{\bf C}\in{\mathbbm R}^{d\times K},\,{\bf C}^T{\bf C}={\bf I}_K}\max_{{\bf B}\in\{\pm1\}^{N\times K}}\text{tr}\left({\bf B}^T{\bf Q}{\bf C}\right).
\label{eq:BQC}$$ By interchanging the maximizations in (\[eq:BQC\]), for any fixed $d\times K$ matrix ${\bf C}$ the inner maximization with respect to $\mathbf B \in \{ \pm 1 \}^{N \times K}$ is solved by $${\bf B}({\bf C}) = \left[\text{sgn}({\bf Q}{\bf C}_{:,1}), \;\text{sgn}({\bf Q}{\bf C}_{:,2}), \;\ldots, \;\text{sgn}({\bf Q}{\bf C}_{:,K})\right],
\label{eq:BC}$$ which is obtained with complexity linear in $N$. Then, by (\[eq:BQC\]), the solution to our original problem in (\[eq:QB\]) is met if we collect all possible binary matrices ${\bf B}({\bf C})$ returned as the columns of ${\bf C}$ scan the unit-radius $d$-dimensional hypersphere while maintaining orthogonality among them. That is, ${\bf B}_\text{opt}$ in (\[eq:QB\]) is in[^7] $${\mathcal S}_K{\stackrel{\triangle}{=}}\hspace{-1cm}\bigcup_{\begin{smallmatrix}{\bf C}\in{\mathbbm R}^{d\times K},\,{\bf C}^T{\bf C}={\bf I}_K,\\C_{d,k}\geq0,\,k=1,2,\ldots,K\end{smallmatrix}}\hspace{-1cm}{\bf B}({\bf C}).$$ Then, by relaxing orthogonality among the columns of ${\bf C}$, $${\mathcal S}_K\subset\hspace{-1cm}\bigcup_{\begin{smallmatrix}{\bf C}\in{\mathbbm R}^{d\times K},\,\left[{\bf C}^T{\bf C}\right]_{k,k}=1,\\C_{d,k}\geq0,\,k=1,2,\ldots,K\end{smallmatrix}}\hspace{-1cm}{\bf B}({\bf C}){=}\Big(\bigcup_{\begin{smallmatrix}{\bf c}\in{\mathbbm R}^d,\,\left\|{\bf c}\right\|_2=1,\\c_d\geq0\end{smallmatrix}}\hspace{-.5cm}{\bf b}({\bf c})\Big)^K{=}{\mathcal S}_1\times{\mathcal S}_1\times\ldots\times{\mathcal S}_1={\mathcal S}_1^K,
\label{eq:S1K}$$ which implies that $$\left|{\mathcal S}_K\right|\leq\left|{\mathcal S}_1\right|^K=\left({\mathcal O}\left(N^{d-1}\right)\right)^K={\mathcal O}\left(N^{dK-K}\right).
\label{eq:SKO}$$ From (\[eq:SKO\]), we observe that the number of binary matrices that we collect as the columns of ${\bf C}$ scan the unit-radius $d$-dimensional hypersphere –with or without maintaining orthogonality– is polynomial in $N$. After ${\bf C}$ has finished scanning the hypersphere, all collected binary matrices in ${\mathcal S}_K$ are compared to each other against the metric of interest in (\[eq:QB\]) with complexity ${\mathcal O}(N)$ per matrix. Therefore, the complexity to solve (\[eq:Bopt\]) is determined by the complexity to build ${\mathcal S}_K$ or at most $\mathcal S_1^K$ since $\mathcal S_K \subset \mathcal S_1^K$ by .
Since ${\bf B}_\text{opt}\in{\mathcal S}_K{\subset}{\mathcal S}_1^K$, we already have a direct way to solve (\[eq:QB\]). First, we construct ${\mathcal S}_1$ with complexity ${\mathcal O}\left(N^d\right)$ as described in Section \[sec:OneComponent\]. We note that ${\mathcal S}_1$ contains ${\mathcal O}\left(N^{d-1}\right)$ binary vectors. Then, we construct ${\mathcal S}_1^K$ which consists of all selections of $K$ elements of ${\mathcal S}_1$ allowing repeated elements. The order of the elements in each selection can be disregarded, since the order of the columns of ${\bf B}$ does not affect the metric in (\[eq:QB\]). Hence, the total number of selections that we need to consider is the number of possible ways one can choose $K$ elements from a set of $\left|{\mathcal S}_1\right|$ elements disregarding order and allowing repetitions (i.e., the number of size-$K$ multisets of all ${\mathcal S}_1$), which equals [@stanley] $$P_K=\binom{\left|{\mathcal S}_1\right|+K-1}{K}={\mathcal O}\left(N^{dK-K}\right)$$ since $|{\mathcal S}_1|={\mathcal O}\left(N^{d-1}\right)$. For each one of the $P_K$ binary matrices, we evaluate the corresponding metric $\left\|{\bf Q}^T{\bf B}\right\|_*$ in (\[eq:QB\]) with complexity ${\mathcal O}(N)$. Then, we identify the optimal matrix ${\bf B}_\text{opt}$ by comparing the calculated metric values. Therefore, the overall complexity to solve (\[eq:Bopt\]) is ${\mathcal O}\left(N^{Kd-K}\right)\cdot{\mathcal O}\left(N\right)={\mathcal O}\left(N^{dK-K+1}\right)$.
The complete algorithm for the computation of the optimal $K$-dimensional ($K>1$) $L_1$-principal subspace of a rank-$d$ matrix ${\bf X}\in{\mathbb R}^{D\times N}$ with complexity ${\mathcal O}\left(N^{dK - K +1 }\right)$ is given in Fig. \[fig:algo\_multi\]. As a simple illustration of the practical computational cost of the presented algorithm, in Table I we show the average CPU time expended by an Intel$^{{\tiny \textregistered}}$ Core$^{{\tiny \texttrademark}}$ i5 Processor at 3.40 GHz running the algorithm of Fig. \[fig:algo\_multi\] in Matlab$^{{\tiny \textregistered}}$ R2012a to calculate the $K=2$ principal components of a $d \times N$ rank-$d$ data matrix for $d=3,4, 5, 6$ and $N = 4,6, \ldots, 14$ (we consider only the cases $N>d$). The presented CPU time for each $(d, N)$ case is the average over $100$ data matrix realizations created with independent zero-mean unit-variance Gaussian drawn entries. Importantly, per Figs. 2 and 3, both visiting the ${N \choose d-1}$ manifold-intersection points for constructing $\mathcal S_1$ (lines 2-8 of function *compute\_candidates* in Fig. \[fig:algo\]) and constructing $\mathcal S_K$ given $\mathcal S_1$ (line 4 of the $L_1$-principal subspace algorithm in Fig. \[fig:algo\_multi\]) are fully parallelizable actions that can be distributed over multiple processing units. Thus, the entire subspace calculation is fully parallelizable and the expended calculation time can be divided down by the number of available processors (plus necessary inter-processor communication overhead).
Experimental Studies
====================
In this section, we carry out a few experimental studies on $L_1$-subspace signal processing to motivate and illustrate the theoretical developments in the previous sections. Examples are drawn from the research fields of dimensionality reduction, data restoration, direction-of-arrival estimation, and image conditioning/reconstruction.
Experiment 1 - Data Dimensionality Reduction {#experiment-1---data-dimensionality-reduction .unnumbered}
--------------------------------------------
We generate a nominal data set $\mathbf X_{D \times N}$ of $N=50$ two-dimensional ($D=2$) observation points drawn from the Gaussian distribution $\mathcal N \left( \mathbf 0_2,
\begin{bmatrix}
15 & 13 \\ 13 & 26
\end{bmatrix}
\right)$ as seen in Fig. \[fig:dr1\]. We calculate and plot in Fig. \[fig:dr1\] the $L_2$ (by standard SVD) and $L_1$ (by Section III.C, Case $d=2$, complexity about $50 \log_2 50$) principal component of the data matrix $\mathbf X$.[^8] For reference purposes, we also plot the true nominal data maximum-variance direction, i.e., the dominant eigenvector of the autocorrelation matrix $\begin{bmatrix}
15 & 13 \\ 13 & 26
\end{bmatrix}$. Then, we assume that our data matrix is corrupted by four outlier measurements, $\mathbf o_1, \mathbf o_2, \mathbf o_3, \mathbf o_{4}$, shown in the bottom right corner of Fig. \[fig:dr2\]. We recalculate the $L_2$ and $L_1$ principal component of the corrupted data matrix $\mathbf X^{\text{CRPT}} = [\mathbf X, \mathbf o_1, \mathbf o_2 , \mathbf o_3, \mathbf o_{4}]$ and notice (Fig. \[fig:dr1\] versus Fig. \[fig:dr2\]) how strongly the $L_2$ component responds to the outliers compared to $L_1$. To quantify the impact of the outliers, in Fig. \[fig:dr3\] we generate $1000$ new independent evaluation data points from $\mathcal N \left( \mathbf 0_2,
\begin{bmatrix}
15 & 13 \\ 13 & 26
\end{bmatrix}
\right)$ and estimate the mean square-fit-error $\text{E} \left\{ \|\mathbf x - \mathbf r\mathbf r^T \mathbf x \|_2^2 \right\}$ when $\mathbf r=\mathbf r_{L_2}(\mathbf X^{\text{CRPT}})$ or $\mathbf r_{L_1}(\mathbf X^{\text{CRPT}})$. We find $\frac{1}{1000} \sum_{i=1}^{1000} \| {\mathbf x}_i - \mathbf r_{L_2}(\mathbf X^{\text{CRPT}})\mathbf r_{L_2}(\mathbf X^{\text{CRPT}})^T {\mathbf x}_i \|_2^2 = 34.417$ versus $\frac{1}{1000} \sum_{i=1}^{1000} \| {\mathbf x}_i - \mathbf r_{L_1}(\mathbf X^{\text{CRPT}})\mathbf r_{L_1}(\mathbf X^{\text{CRPT}})^T {\mathbf x}_i \|_2^2 = 11.555$. In contrast, when the principal component is calculated from the clean training set, $\mathbf r= \mathbf r_{L_2}(\mathbf X)$ or $ \mathbf r_{L_1}(\mathbf X)$, we find estimated mean square-fit-error $6.077$ and $ 6.080$, correspondingly. We conclude that dimensionality reduction by $L_1$ principal components may loose only minimally in mean-square fit compared to $L_2$ when the designs are from clean training sets, but can protect significantly when training is carried out in the presence of erroneous data.
Next, we will compare the dimensionality-reduction performance of the proposed $L_1$-principal subspace with that of other subspaces in the literature obtained by means of $L_1$-norm based methods. Specifically, alongside the $L_2$ (SVD) and $L_1$-principal component (proposed), we calculate the $R_1$-principal component [@Ding2006] as well as the direction obtained by means of $L_1$-factorization through alternating weighted median calculation [@Ke2003], [@Ke2005].[^9] All directions are calculated from an $(N=20)$-point corrupted data set $\mathbf X^{\text{CRPT}} \in \mathbb R^{2 \times 20}$ with $N_{\mathrm{out}}$ outliers drawn from $\mathcal N \left( \begin{bmatrix} 20 \\ -20 \end{bmatrix}, \begin{bmatrix} 5.73 & -4.494 \\ -4.494 & 5.27 \end{bmatrix} \right)$ and $N - N_{\mathrm{out}}$ nominal points drawn from $\mathcal N \left( \mathbf 0_{3}, \begin{bmatrix} 15 & 13 \\ 13 & 26 \end{bmatrix} \right)$. In Fig. \[fig:linefitting\_mse\], we plot the mean-squared-fit-error averaged over $10 000$ independent corrupted training data-set experiments as a function of the number of outlying points in the data set $N_{\mathrm{out}}$. We notice that, when designed on nominal data, all examined subspaces differ little, if any, from the $L_2$-principal subspace in mean-square fit error. However, when designed on outlier-corrupted data sets, the $L_1$-principal subspace exhibits notable robustness outperforming uniformly and significantly all other subspaces, especially in the $15\%$ - $40\%$ mid-range of corruption. Given that $L_1$ and $L_2$ start very near each other in mean-square-fit-error at $0\%$ corruption and meet again only at $100\%$ corruption, one is tempted to say that the $L_1$ subspaces are to be uniformly preferred over $L_2$ if the associated computational cost can be afforded.
Experiment 2 - Data Restoration {#experiment-2---data-restoration .unnumbered}
-------------------------------
As a toy numerical example, consider a hypothetical case where we collect from a sensor system eight samples of five-dimensional data. Due to the nature of the sensed source, the data are to lie in a lower-than-five dimensional space, say a plane. Say, then, the true data are given by the $\text{rank}$-$2$ data matrix below [$$\begin{aligned}
\mathbf X_{5 \times 8} =
\begin{bmatrix}
2.0724 & -1.2024 & 1.2956 & 2.8719 & 1.5637 & -2.9323 & -3.1792 & -1.4152 \\
-0.5233 & 0.2595 & -0.3298 & -0.7562 & -0.4087 & 0.7973 & 0.8235 & 0.4155 \\
0.0185 & -0.8158 & -0.0367 & -0.5406 & -0.2380 & 1.0108 & 0.3502 & 1.0487 \\
-0.6424 & 0.1476 & -0.4151 & -1.0486 & -0.5552 & 1.1989 & 1.0913 & 0.7355 \\
-2.1289 & 2.2734 & -1.2687 & -2.2200 & -1.2814 & 1.6751 & 2.7777 & 0.0851 \\
\end{bmatrix}. \nonumber\end{aligned}$$ ]{} Assume that due to sensor malfunction or data transfer error or data storage failure, we are presented instead with [$$\begin{aligned}
\mathbf X_{5 \times 8}^{\text{CRPT}} =
\begin{bmatrix}
2.0724 & {\color{red}\it{ 8.9538}} & 1.2956 & 2.8719 & {\color{red}\it{ 10.6817}} & -2.9323 & -3.1792 & -1.4152 \\
-0.5233 & {\color{red}\it{ 10.6187}} & -0.3298 & -0.7562 & {\color{red}\it{ 11.0235}} & 0.7973 & 0.8235 & 0.4155 \\
0.0185 & {\color{red}\it{ 11.3050}} & -0.0367 & -0.5406 & -0.2380 & 1.0108 & 0.3502 & 1.0487 \\
-0.6424 & 0.1476 & -0.4151 & -1.0486 & {\color{red}\it{ 7.8846}} & 1.1989 & 1.0913 & 0.7355 \\
-2.1289 & 2.2734 & -1.2687 & -2.2200 & -1.2814 & 1.6751 & 2.7777 & 0.0851 \\
\end{bmatrix}\nonumber\end{aligned}$$ ]{}\
where six of the original entries in two of the data points have been altered/overwritten and $\mathbf X^{\text{CRPT}} $ spans now a four-dimensional subspace of $\mathbb R^{5}$.
Our objective is to “restore” $\mathbf X^{\text{CRPT}} $ to $\mathbf X$ taking advantage of our knowledge (or assumption) of the rank of the original data. Along these lines, we project $\mathbf X^{\text{CRPT}}$ onto the span of its $K=2$ $L_2$- or $L_1$-principal components, $$\begin{aligned}
\hat{\mathbf X}=\mathbf R \mathbf R^T \mathbf X^{\text {CRPT}}\end{aligned}$$ where $\mathbf R_{5 \times 2}= [\mathbf r_{L_2}^{(1)}, \mathbf r_{L_2}^{(2)}]$ or $ [\mathbf r_{L_1}^{(1)}, \mathbf r_{L_1}^{(2)}]$. The resulting $L_2$- and $L_1$-derived representations of $\mathbf X$ are [$$\begin{aligned}
\hat{\mathbf X}_{L_2}=
\begin{bmatrix}
0.8029 & 8.2311 & 0.4919 & 0.9945 & 11.8445 & -0.9197 & -1.1528 & -0.3268\\
0.4839 & 11.0891 & 0.2888 & 0.5096 & 10.2500 & -0.3897 & -0.6347 & -0.0285\\
-0.5922 & 11.1679 & -0.3843 & -0.9862 & 0.0165 & 1.1412 & 1.0192 & 0.7148\\
0.6521 & 0.8969 & 0.4067 & 0.8926 & 6.6810 & -0.9024 & -0.9930 & -0.4245\\
-0.3868 & 2.8347 & -0.2455 & -0.5789 & -2.2540 & 0.6257 & 0.6220 & 0.3444
\end{bmatrix} \nonumber\end{aligned}$$ ]{} and [$$\begin{aligned}
\hat{\mathbf X}_{L_1}=
\begin{bmatrix}
2.0724 & -0.0303 & 1.2956 & 2.8719 & 2.9321 & -2.9323 & -3.1792 & -1.4152\\
-0.5233 & 0.1880 & -0.3298 & -0.7562 & -0.7283 & 0.7973 & 0.8235 & 0.4155\\
0.0185 & 3.2915 & -0.0367 & -0.5406 & 0.2476 & 1.0108 & 0.3502 & 1.0487\\
-0.6424 & 0.9300 & -0.4151 & -1.0486 & -0.8469 & 1.1989 & 1.0913 & 0.7355\\
-2.1289 & -4.2139 & -1.2687 & -2.2200 & -3.2976 & 1.6751 & 2.7777 & 0.0851
\end{bmatrix}, \nonumber\end{aligned}$$ ]{}
[respectively]{}. In Fig. \[fig:error\], we plot the element-by-element and per-measurement square-restoration error for the two projections. The relative superiority of $L_1$-subspace data representation is clearly captured and documented.
Experiment 3 - Direction-of-Arrival Estimation {#experiment-3---direction-of-arrival-estimation .unnumbered}
----------------------------------------------
We consider a uniform linear antenna array of $D=5$ elements that takes $N=10$ snapshots of two incoming signals with angles of arrival $\theta_1 = -30^\circ$ and $\theta_2 = 50^\circ$, $$\begin{aligned}
\mathbf x_n= A_1 \mathbf s_{\theta_1} + A_2 \mathbf s_{\theta_2} + \mathbf n_n, ~n=1, \ldots, 10, \label{cleanmes}\end{aligned}$$ where $A_1, A_2$ are the received-signal amplitudes with array response vectors $\mathbf s_{\theta_1}$ and $\mathbf s_{\theta_2}$, correspondingly, and $\mathbf n \sim \mathcal {CN}\left(\mathbf 0_{5}, \sigma^2 \mathbf I_{5} \right)$ is additive white complex Gaussian noise. We assume that the signal-to-noise ratio (SNR) of the two signals is $\text{SNR}_1=10 \log_{10}\frac{A_1^2}{\sigma^2}\text{dB}=2\text{dB}$ and $\text{SNR}_2=10 \log_{10}\frac{A_2^2}{\sigma^2}\text{dB}=3\text{dB}$. Next, we assume that one arbitrarily selected measurement out of the ten observations $\mathbf X_{5 \times 10}=[\mathbf x_1, \ldots, \mathbf x_{10}] \in \mathbb C^{5 \times 10}$ is corrupted by a jammer operating at angle $\theta_J= 20^\circ$ with amplitude $A_J = A_2$. We call the resulting corrupted observation set $\mathbf X^{\text{CRPT}} \in \mathbb C^{5 \times 10}$ and create the real-valued version $\tilde{\mathbf X}^{\text{CRPT}} = [ \text{Re} \{\mathbf X^{\text{CRPT}}\}^T, ~ \text{Im} \{\mathbf X^{\text{CRPT}}\}^T ]^T \in \mathbb R^{10 \times 10}$ by $\text{Re}\{ \cdot\}, \text{Im}\{ \cdot\} $ part concatenation. We calculate the $K=2$ $L_2$-principal components of $\tilde{\mathbf X}^{\text{CRPT}}$, $\mathbf R_{L_2} = [\mathbf r_{L_2}^{(1)}, \mathbf r_{L_2}^{(2)}] \in \mathbb R^{10 \times 2}$, and the $K=2$ $L_1$-principal components of $\tilde{\mathbf X}^{\text{CRPT}}$, $\mathbf R_{L_1} = [\mathbf r_{L_1}^{(1)}, \mathbf r_{L_1}^{(2)}] \in \mathbb R^{10 \times 2}$. In Fig. \[fig:DOA\], we plot the standard $L_2$ MUSIC spectrum [@schmidt] $$\begin{aligned}
P(\theta) \overset{\triangle}{=} \frac{1}{ \tilde{\mathbf s}_{\theta}^T (\mathbf I_{2D}- \mathbf R_{L_2} \mathbf R_{L_2}^T) \tilde{\mathbf s}_{\theta}}, ~\theta \in \left( -\frac{\pi}{2}, \frac{\pi}{2}\right),\end{aligned}$$ where $ \tilde{\mathbf s}_{\theta} = [ \text{Re}\{{\mathbf s}_{\theta}\}^T, ~ \text{Im}\{{\mathbf s}_{\theta}\}^T ]^T$, as well as what we may call “$L_1$ MUSIC spectrum” with $\mathbf R_{L_1}$ in place of $\mathbf R_{L_2}$. It is interesting to observe how $L_1$ MUSIC (in contrast to $L_2$ MUSIC) does not respond to the one-out-of-ten outlying jammer value in the data set and shows only the directions of the two actual nominal signals.
Experiment 4 - Image Reconstruction {#experiment-4---image-reconstruction .unnumbered}
-----------------------------------
Consider the “clean” $ 100 \times 64 $ gray-scale image $\mathbf A \in \{0, 1, \ldots, 255\}^{100 \times 64}$ of Fig. \[fig:alexis1\]. We assume that $\mathbf A$ is not available and instead we have a data set of $N=10$ corrupted/occluded versions of $\mathbf A$, say $\mathbf A_{1}, \mathbf A_{2}, \ldots, \mathbf A_{10}$. Each corrupted instance $\mathbf A_i$, $i=1, \ldots, 10$, is created by partitioning the original image $\mathbf A$ into sixteen tiles of size $25 \times 16$ and replacing three arbitrarily selected tiles by $25 \times 16$ grayscale-noise patches as seen, for example, in Fig. \[fig:alexis2\].
The $10$ corrupted instances are vectorized to form the data matrix $$\begin{aligned}
\mathbf M= [ \text{vec}(\mathbf A_1), \ldots, \text{vec}(\mathbf A_{10}) ] \in \{0, \ldots, 255\}^{6400 \times 10}.\end{aligned}$$ Next, we “condense” $\mathbf M$ to a rank-$2$ representation by both $L_2$- and $L_1$-subspace projection, $$\begin{aligned}
\hat{\mathbf M}_{L_{2/1}}= \mathbf R_{L_{2/1}} \mathbf R_{L_{2/1}}^T \mathbf M,\end{aligned}$$ where $\mathbf R_{L_{2/1}} \in \mathbb R^{6400 \times 2}$ consists of the $K=2$ $L_2$ or $L_1$, accordingly, principal components of $\mathbf M$. In Fig. \[fig:alexis3\] we show the projection of the corrupted image of Fig. \[fig:alexis2\] onto the $L_2$-derived rank-$2$ subspace (maximum-$L_2$-projection reconstruction). In Fig. \[fig:alexis4\], we show the projection of the same image onto the $L_1$-derived rank-$2$ subspace (maximum-$L_1$-projection reconstruction). Figs. \[fig:alexis3\] and (d) offer a perceptual (visual) interpretation of the difference between $L_2$ and $L_1$-subspace rank reduction. It is apparent that maximum-$L_1$-projection reconstruction offers a much clearer image representation of $\mathbf A$ than maximum-$L_2$-projection reconstruction. This is another result that highlights the resistance of $L_1$-principal subspaces against outlying data corruption.
Conclusions
===========
We presented for the first time in the literature optimal (exact) algorithms for the calculation of maximum-$L_1$-projection subspaces of data sets with complexity polynomial in the sample size (and exponent equal to the data dimension). It may be possible in the future to develop an $L_1$ principal-component-analysis (PCA) line of research that parallels the enormously rewarding $L_2$ PCA/feature-extraction developments. When $L_1$ subspaces are calculated on nominal “clean" training data, they differ little –arguably– from their $L_2$-subspace counterparts in least-squares fit. When, however, subspaces are calculated from data sets with possible erroneous, out-of-line, “outlier" entries, then $L_1$ subspace calculation offers significant robustness/resistance to the presence of inappropriate data values.\
\
**[ACKNOWLEDGEMENT]{}\
**
The authors would like to thank the Associate Editor and the four anonymous reviewers for their comments and suggestions that helped improve this manuscript significantly, both in presentation and content.
[99]{}
C. Eckart and G. Young, “The approximation of one matrix by another of lower rank,” [*Psychometrika*]{}, vol. 1, pp. 211-218, Sept. 1936.
G. H. Golub and C. F. Van Loan, [*Matrix Computations*]{}, 3rd Ed. Baltimore, MD: The Johns Hopkins Univ. Press, 1996.
Q. Ke and T. Kanade, “Robust subspace computation using L1 norm,” Internal Technical Report, Computer Science Dept., Carnegie Mellon Univ., CMU-CS-03-172, Aug. 2003.
Q. Ke and T. Kanade, “Robust $L_1$ norm factorization in the presence of outliers and missing data by alternative convex programming,” in [*Proc. IEEE Conf. Comput. Vision Pattern Recog. (CVPR)*]{}, San Diego, CA, June 2005, pp. 739-746.
A. Eriksson and A. van den Hengel, “Efficient computation of robust low-rank matrix approximations in the presence of missing data using the $L_1$ norm,” in [*Proc. IEEE Conf. Comput. Vision Pattern Recog. (CVPR)*]{}, San Francisco, CA, June 2010, pp. 771-778.
R. He, B.-G. Hu, W.-S. Zheng, and X.-W. Kong, “Robust principal component analysis based on maximum correntropy criterion,” [*IEEE Trans. Image Process.*]{}, vol. 20, pp. 1485-1494, June 2011.
L. Yu, M. Zhang, and C. Ding, “An efficient algorithm for L1-norm principal component analysis,” in [*Proc. IEEE Intern. Conf. Acoust. Speech and Signal Proc. (ICASSP)*]{}, Kyoto, Japan, Mar. 2012, pp. 1377-1380.
J. P. Brooks and J. H. Dulá, “The L1-norm best-fit hyperplane problem,” [*Appl. Math. Lett.*]{}, vol. 26, pp. 51-55, Jan. 2013.
J. P. Brooks, J. H. Dulá, and E. L. Boone, “A pure $L_1$-norm principal component analysis,” [*J. Comput. Stat. Data Anal.*]{}, vol. 61, pp. 83-98, May 2013.
J. S. Galpin and D. M. Hawkins, “Methods of $L_1$ estimation of a covariance matrix,” [*J. Comput. Stat. Data Anal.*]{}, vol. 5, pp. 305-319, 1987.
N. Kwak, “Principal component analysis based on L1-norm maximization,” [*IEEE Trans. Pattern Anal. Mach. Intell.*]{}, vol. 30, pp. 1672-1680, Sept. 2008.
N. Kwak and J. Oh, “Feature extraction for one-class classification problems: Enhancements to biased discriminant analysis,” [*Pattern Recog.*]{}, vol. 42, pp. 17-26, Jan. 2009.
X. Li, Y. Pang, and Y. Yuan, “L1-norm-based 2DPCA,” [*IEEE Trans. Syst., Man. Cybern., Part B: Cybern.*]{}, vol. 40, pp. 1170-1175, Aug. 2009.
Y. Pang, X. Li, and Y. Yuan, “Robust tensor analysis with L1-norm,” [*IEEE Trans. Circuits Syst. Video Tech.*]{}, vol. 20, pp. 172-178, Feb. 2010.
N. Funatsu and Y. Kuroki, “Fast parallel processing using GPU in computing L1-PCA bases,” in [*Proc. IEEE TENCON*]{}, Fukuoka, Japan, Nov. 2010, pp. 2087-2090.
M. McCoy and J. A. Tropp, “Two proposals for robust PCA using semidefinite programming,” [*Electron. J. Stat.*]{}, vol. 5, pp. 1123-1160, June 2011.
F. Nie, H. Huang, C. Ding, D. Luo, and H. Wang, “Robust principal component analysis with non-greedy $l_1$-norm maximization,” in [*Proc. Int. Joint Conf. Artif. Intell. (IJCAI)*]{}, Barcelona, Spain, July 2011, pp. 1433-1438.
D. Meng, Q. Zhao, and Z. Xu, “Improve robustness of sparse PCA by $L_1$-norm maximization,” [*Pattern Recogn.*]{}, vol. 45, pp. 487-497, Jan. 2012.
H. Wang, Q. Tang, and W. Zheng, “L1-norm-based common spatial patterns,” [*IEEE Trans. Biomed. Eng.*]{}, vol. 59, pp. 653-662, Mar. 2012.
H. Wang, “Block principal component analysis with L1-norm for image analysis,” [*Pattern Recogn. Lett.*]{}, vol. 33, pp. 537-542, Apr. 2012.
H. Q. Luong, B. Goossens, J. Aelterman, A. Pižurica, and W. Philips, “A primal-dual algorithm for joint demosaicking and deconvolution,” in [*Proc. IEEE Intern. Conf. Image Proc. (ICIP)*]{}, Orlando, FL, Oct. 2012, pp. 2801-2804.
Z. Gu, W. Lin, B.-S. Lee, and C. T. Lau, “Rotated orthogonal transform (ROT) for motion-compensation residual coding,” [*IEEE Trans. Image Process.*]{}, vol. 21, pp. 4770-4781, Dec. 2012.
C. Ding, D. Zhou, X. He, and H. Zha, “$R_1$-PCA: Rotational invariant $L_1$-norm principal component analysis for robust subspace factorization,” in [*Proc. Int. Conf. Mach. Learn.*]{}, Pittsburgh, PA, 2006, pp. 281-288.
X. Li, W. Hu, H. Wang, and Z. Zhang, “Linear discriminant analysis using rotational invariant $L_1$ norm,” [*Neurocomputing*]{}, vol. 73, pp. 2571-2579, Aug. 2010.
D. L. Donoho, “Compressed sensing,” *IEEE Trans. Inf. Theory*, vol. 52, pp. 1289-1306, Apr. 2006.
E. J. Candès, “Compressive sampling,” in *Proc. Intern. Cong. Math. (ICM)*, Madrid, Spain, Aug. 2006, pp. 1433-1452.
R. G. Baraniuk, “Compressive sensing,” *IEEE Signal Process. Mag.*, vol. 24, pp. 118-124, July 2007.
K. Gao, S. N. Batalama, D. A. Pados, and B. W. Suter, “Compressive sampling with generalized polygons,” *IEEE Trans. Signal Proc.*, vol. 59, pp. 4759-4766, Oct. 2011.
P. L. Combettes and J. Luo, “An adaptive level set method for nondifferentiable constrained image recovery,” [*IEEE Trans. Image Proc.*]{}, vol. 11, pp. 1295-1304, Nov. 2002.
Y. Nesterov and A. Nemirovski, “On first-order algorithms for $\ell_1$/nuclear norm minimization,” [*Acta Numerica*]{}, vol. 22, pp. 509-575, May 2013.
H. L. Van Trees, [*Detection, Estimation, and Modulation Theory (Part I)*]{}. Hoboken, NJ: Wiley, 2001.
M. R. Garey and D. S. Johnson, [*Computers and Intractability: A Guide to the Theory of NP-completeness*]{}. San Francisco, CA: W. H. Freeman and Company, 1979.
G. N. Karystinos and D. A. Pados, “Rank-2-optimal adaptive design of binary spreading codes,” [*IEEE Trans. Inf. Theory*]{}, vol. 53, pp. 3075-3080, Sept. 2007.
K. M. Mackenthun, Jr., “A fast algorithm for multiple-symbol differential detection of MPSK,” [*IEEE Trans. Commun.*]{}, vol. 42, pp. 1471-1474, Feb./Mar./Apr. 1994.
W. Sweldens, “Fast block noncoherent decoding,” [*IEEE Commun. Lett.*]{}, vol. 5, pp. 132-134, Apr. 2001.
I. Motedayen-Aval, A. Krishnamoorthy, and A. Anastasopoulos, “Optimal joint detection/estimation in fading channels with polynomial complexity,” [*IEEE Trans. Inf. Theory*]{}, vol. 53, pp. 209-223, Jan. 2007.
G. N. Karystinos and A. P. Liavas, “Efficient computation of the binary vector that maximizes a rank-deficient quadratic form,” [*IEEE Trans. Inf. Theory*]{}, vol. 56, pp. 3581-3593, July 2010.
K. Allemand, K. Fukuda, T. M. Liebling, and E. Steiner, “A polynomial case of unconstrained zero-one quadratic optimization,” [*Mathematical Programming*]{}, vol. A-91, pp. 49-52, Oct. 2001.
J.-A. Ferrez, K. Fukuda, and T. M. Liebling, “Solving the fixed rank convex quadratic maximization in binary variables by a parallel zonotope construction algorithm,” [*European Journal of Operational Research*]{}, vol. 166, pp. 35-50, 2005.
D. Avis and K. Fukuda, “Reverse search for enumeration,” [*Discrete Applied Mathematics*]{}, vol. 65, pp. 21-46, Mar. 1996.
H. Edelsbrunner, J. O’Rourke, and R. Seidel, “Constructing arrangements of lines and hyperplanes with applications,” [*SIAM J. Comput.*]{}, vol. 15, pp. 341-363, May 1986.
H. Edelsbrunner, [*Algorithms in Combinatorial Geometry*]{}. New York, NY: Springer-Verlag, 1987.
W. Ben-Ameur and J. Neto, “A polynomial-time recursive algorithm for some unconstrained quadratic optimization problems,” [*Discrete Applied Mathematics*]{}, vol. 159, pp. 1689-1698, Sept. 2011.
J. R. Magnus and H. Neudecker, [*Matrix Differential Calculus with Applications in Statistics and Econometrics*]{}, 2nd Ed. Chichester, UK: Wiley, 1999.
R. P. Stanley, [*Enumerative Combinatorics, Volume 1*]{}, 2nd Ed. New York, NY: Cambridge University Press, 2012.
R. O. Schmidt, “Multiple emitter location and signal parameter estimation,” *IEEE Trans. Antennas Propag.*, vol. AP-34, pp. 276-280, Mar. 1986.
{width=".8\textwidth"}
[0.6]{}
------------------------------------------------------------------------
------------------------------------------------------------------------
[**The Optimal $L_1$-Principal-Component Algorithm**]{}
------------------------------------------------------------------------
\
------------------------------------------------------
[**Input:**]{} $\mathbf X_{D \times N}$ data matrix
------------------------------------------------------
\
----- ----------------------------------------------------------------------------------------------------------------------------------
1: $\left( \mathbf U_{N \times d}, \mathbf \Sigma_{d \times d}, \mathbf V_{d\times d}\right) \leftarrow \mathrm{svd} (\mathbf X^T)$
2: $\mathbf Q_{N \times d} \leftarrow \mathbf {U \Sigma}$
3: $\mathbf B_{N \times P_1} \leftarrow \mathrm{compute\_candidates} (\mathbf Q)$, $\mathcal P \leftarrow \{1, 2, \ldots, P_1 \}$
4: $z_{\text{opt}} \leftarrow {\arg \max}_{z \in \mathcal P}~ \| \mathbf X \mathbf B_{:,z} \|_2$
5: $\mathbf b_{\text{opt}} \leftarrow \mathbf B_{:, z_{\text{opt}}}$
----- ----------------------------------------------------------------------------------------------------------------------------------
\
---------------------------------------------------------------------------------------------------------------------------
[**Output:**]{} $\mathbf r_{L_1} \leftarrow {\mathbf X \mathbf b_{\text{opt}}}/{\|\mathbf X \mathbf b_{\text{opt}}\|_2}$
---------------------------------------------------------------------------------------------------------------------------
\
------------------------------------------------------------------------
[Function *compute\_candidates*]{}
------------------------------------------------------------------------
\
------------------------------------------
[**Input:**]{} $\mathbf Q_{N \times m}$
------------------------------------------
\
----- -------------------------------------------------------------------------------------------------------------------------------------------------------------
1: if $m > 2$, $i \leftarrow 0$
2: for $\mathcal I \subset \{1, 2, \ldots, N \}$ s.t. $| \mathcal I|=m-1$, $i \leftarrow i+1$,
3: $\bar{\mathbf Q}_{(m-1) \times m} \leftarrow \mathbf Q_{\mathcal I, :}$
4: $\mathbf c_{m \times 1} \leftarrow \mathrm{null}(\bar{\mathbf Q})$, $\mathbf c \leftarrow \mathrm{sgn}(c_m) \mathbf c$
5: $\mathbf B_{:,i} \leftarrow \mathrm{sgn}(\mathbf Q \mathbf c)$
6: for $j =1 : m-1$,
7: $ \mathbf c_{(m-1) \times 1} \leftarrow \mathrm{null}(\bar{\mathbf Q}_{:/j,1:m-1}) $, $\mathbf c \leftarrow \mathrm{sgn}(c_{m-1}) \mathbf c$
8: $\mathbf B_{\mathcal I(j), i} \leftarrow \mathrm{sgn}(\bar{\mathbf Q}_{j,1:m-1} \mathbf c)$
9: $\mathbf B \leftarrow [\mathbf B, \mathrm{compute\_candidates}(\mathbf Q_{:,1:m-2})]$
10: elseif $m=2$,
11: for $i =1 : N$,
12: $\mathbf c_{2 \times 1} \leftarrow \mathrm{null}( \mathbf Q_{i, :})$, $\mathbf c \leftarrow \mathrm{sgn}(c_2) \mathbf c$
13: $\mathbf B_{:,i} \leftarrow \mathrm{sgn}(\mathbf Q \mathbf c)$, $\mathbf B_{i,i} \leftarrow \mathrm{sgn}({\mathbf Q}_{i,1})$
14: else, $\mathbf B \leftarrow \mathrm{sgn}(\mathbf Q)$
----- -------------------------------------------------------------------------------------------------------------------------------------------------------------
\
------------------------------
[**Output:**]{} $\mathbf B$
------------------------------
------------------------------------------------------------------------
------------------------------------------------------------------------
[0.6]{}
------------------------------------------------------------------------
------------------------------------------------------------------------
[**The Optimal $L_1$-Principal-Subspace Algorithm ($K>1$)**]{}
------------------------------------------------------------------------
\
-----------------------------------------------------------------------------------
[**Input:**]{} $\mathbf X_{D \times N}$ data matrix, subspace dimensionality $K$
-----------------------------------------------------------------------------------
\
----- -----------------------------------------------------------------------------------------------------------------------------------------------------------------
1: $\left( \mathbf U_{N \times d}, \mathbf \Sigma_{d \times d}, \mathbf V_{d\times d}\right) \leftarrow \mathrm{svd} (\mathbf X^T)$
2: $\mathbf Q_{N \times d} \leftarrow \mathbf {U \Sigma}$
3: $\mathbf B_{N \times P_1} \leftarrow \mathrm{compute\_candidates} (\mathbf Q)$, $\mathcal P \leftarrow \{1, 2, \ldots, P_1 \}$
4: $\mathbf z_{\text{opt}} \leftarrow {\arg \max}_{\mathbf z \in \mathcal P^{K}, z_1 \leq z_2 \leq \ldots \leq z_K }~ \| \mathbf X \mathbf B_{:, \mathbf z} \|_*$
5: $\mathbf B_{\text{opt}} \leftarrow \mathbf B_{:, \mathbf z_{\text{opt}}}$
6: $\left( \mathbf U_{D \times K}, \mathbf \Sigma_{K \times K}, \mathbf V_{K\times K}\right) \leftarrow \mathrm{svd} (\mathbf X \mathbf B_{\text{opt}})$
----- -----------------------------------------------------------------------------------------------------------------------------------------------------------------
\
---------------------------------------------------------------------
[**Output:**]{} $\mathbf R_{L_1} \leftarrow \mathbf U \mathbf V^T$
---------------------------------------------------------------------
------------------------------------------------------------------------
------------------------------------------------------------------------
[|c||\*[6]{}[c|]{}]{} & 4 & 6 & 8 & 10 & 12 & 14 \
3 & 0.0172 & 0.0406 & 0.0920 & 0.1966 & 0.3900 & 0.7160\
4 & - & 0.0624 & 0.3526 & 1.4212 & 4.5178 & 11.8686\
5 & - & 0.1014 & 0.8471 & 5.4944 & 26.3361 & 99.4600\
6 & - & - & 1.2308 & 12.2289 & 87.1546 & 471.2275\
\[tb:comp\]
[0.49]{} {width="\textwidth"}
\
[0.49]{} {width="\textwidth"}
[0.49]{} {width="\textwidth"}
{width="80.00000%"}
[.6]{} {width="\textwidth"}
\
[.6]{} {width="\textwidth"}
{width="80.00000%"}
[0.15]{}
[0.15]{} {width="80.00000%"}
[0.15]{} {width="80.00000%"}
[0.15]{}
[^1]: This paper was presented in part at the Tenth International Symposium on Wireless Communication Systems (ISWCS), Ilmenau, Germany, August 2013.
[^2]: Absolute-value errors put significantly less emphasis on extreme errors than squared-error expressions.
[^3]: A combined $L_1$/$L_2$-norm approach has been followed in \[\], \[\].
[^4]: The $d$th element of vector $\mathbf c$, $c_d$, can be set nonnegative without loss of optimality, because, for any given $\mathbf c$, $\| \mathbf c\|_2=1$, the binary vectors $\mathbf b(\mathbf c)$ and $\mathbf b(\text{sgn}(c_d)\mathbf c)$ result to the same metric value in .
[^5]: If ${\bf Q}_{1:d-1,:}$ is full-rank, then its null space has rank $1$ and ${\bf c}({\boldsymbol\phi})$ is uniquely determined (within a sign ambiguity which is resolved by $c_d\geq0$). If, instead, ${\bf Q}_{1:d-1,:}$ is rank-deficient, then the intersection of the $d-1$ hypersurfaces (i.e., the solution of (\[eq:Qc\])) is a $p$-manifold (with $p\geq1$) in the $(d-1)$-dimensional space and does not generate a new cell. Hence, linearly dependent combinations of $d-1$ rows of ${\bf Q}$ are ignored.
[^6]: An alternative way of resolving the sign ambiguities at the intersections of hypersurfaces was developed in [@KL] and led to the direct construction of a set ${\mathcal S}_1$ of size $\sum_{g=0}^{d-1}\binom{N-1}{g}={\mathcal O}(N^{d-1})$ with complexity ${\mathcal O}(N^d)$.
[^7]: Without loss of optimality, we set $C_{d,k} \geq 0$, $k =1, 2, \ldots, K$, since, for any given $\mathbf C$, $\mathbf C^T \mathbf C = \mathbf I_K$, the binary matrices $\mathbf B(\mathbf C)$ and $\mathbf B\big( \mathbf C \, \text{diag} ( \text{sgn}(\mathbf C_{d, :}) ) \big) $ result to the same metric value in .
[^8]: We note that without the presented algorithm, computation of the $L_1$ principal component of $\mathbf X_{2 \times 50}$ would have required complexity proportional to $2^{50}$ (by ), which is of course infeasible.
[^9]: Notice that for $R_1$-PCA [@Ding2006] and $L_1$-factorization [@Ke2003], [@Ke2005], no optimal solution exists in the literature so far.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the present work, we investigate the effects of long-range interactions on the phase transitions of two-dimensional ferromagnetic models with single-ion anisotropy at zero and finite temperatures. The Hamiltonian is given by $H=\sum_{i\neq j} J_{ij}(S_i^xS_j^x+S_i^yS_j^y+\lambda S_i^zS_j^z)+D\sum_{i}(S_i^z)^2$, where $J_{ij}=-J |r_j-r_i|^{-p}$ ($p\geq 3$) is a long-range ferromagnetic interaction ($J>0$) , $0\leq \lambda\leq 1$ is an anisotropic constant and $D$ is the single-ion anisotropic constant. It is well-known that the single-ion anisotropy $D$ creates a competition between an ordered state (favored by the exchange interaction) and a disordered state, even at zero temperature. For small values of $D$, the system has a spontaneous magnetization $m_z\neq 0$, while in the large-D phase $m_z=0$ because a state with $\langle S^z\rangle\neq 0$ is energetically unfavorable. Therefore, a phase transition due to quantum fluctuations occurs in some critical value $D_c$. For systems with short-range interaction $D_c\approx 6J$, depending of $\lambda$ constant, but in our model we have found larger values of $D$ due to the higher cost to flip a spin. Since low-dimensional magnetic systems with long range interaction can be ordered at finite temperature, we also have analyzed the thermal phase transitions (similar to the BKT transition). The model has been studied by using a Schwinger boson formalism as well as the Self-consistent Harmonic Approximation (SCHA) and both methods provide according results.'
author:
- 'A. R. Moura'
bibliography:
- 'manuscript.bib'
title: 'Phase transitions in the two-dimensional single-ion anisotropic ferromagnetic with long-range interactions'
---
Introduction
============
Usually, condensed matter physics considers short-range interactions between the nearest-neighbor spins as the main responsible for the properties of low-dimensional magnetic systems, existing an extensive number of works on this topic. However, longer-range interactions are also important to describe many other cases, where only short-range interactions are insufficient. For example, in antiferromagnetic models, second-neighbor interaction between spins are responsible for an increase in disorder due to lattice frustration once it is impossible a complete antiferromagnetic align. In addition, the properties of spin ice systems arise as consequence of the long-range dipolar interaction. It creates a power-law decaying interaction between the magnetic monopoles quasiparticles, similar to the electric charges.
It is well known that low-dimensional systems are more susceptible to quantum fluctuations which imply no continuous broken symmetry at finite temperatures for one and two-dimensional magnetic models with only short-range interactions (Mermin-Wagner theorem[@PRL17]). On the other hand, the long-range interactions are able to create a state with spontaneous magnetization at finite temperatures once the spin-waves do not have sufficient energy to flip spins. For a $d$-dimensional Heisenberg model with interaction between the sites $i$ and $j$ decaying as $r^p$, there is a state with spontaneous magnetization at finite low temperatures if $d<p<2d$ [@JP2; @CMP62] while for $p\geq 2d$ the long-range order ($LRO$) is lost at all finite temperatures [@JSP25; @CMP79]. The condition $p>d$ is applied in order to avoid an energy divergence per site in the ground state. If the symmetry is discrete instead continuous, as occurs in the Ising model, there is $LRO$ for $1<p\leq 2$ [@CMP118].
In this paper we have investigated the two-dimensional ($2d$) ferromagnetic model with long-range interactions and single-ion anisotropy. The Hamiltonian is given by: $$\label{eq:hamiltonian}
H=\sum_{i\neq j}J_{ij}\left(S_i^x S_j^x+S_i^y S_j^y + \lambda S_i^z S_j^z\right) + D\sum_i (S_i^z)^2,$$ where $J_{ij}=-J|\vec{r}_j-\vec{r}_i|^{-p}$ ($J>0$) is the long-range ferromagnetic interaction, $\lambda$ is an anisotropic constant and $D$ is the single-ion anisotropic coupling. The first sum is evaluated over all pairs of $i$ and $j$ sites while the second one is over all sites. For $\lambda=0$ one has the usual $XY$ model (considering ${\bf S}=(S^x,S^y,S^z)$) while $\lambda=1$ gives the isotropic Heisenberg model. The Hamiltonian makes sense only for spins larger than $1/2$ once the single-ion anisotropy is degenerate with respect the up and down states. Therefore, we have adopted $S=1$ for which there are two energy bands: $S_i^z=\pm1$ and $S_i^z=0$. Though the finite energy ground state requires $p>2$, we have considered only the decays with $p\geq 3$ because there is no physical models of interest with $2<p<3$. The one-dimensional case has been studied by Pires [@PLA202] and a first-order thermal phase transition has been found for $XY$ model and Planar Rotator Model (when ${\bf S}=({S^x,S^y)}$). The critical temperatures (the point where the spin-spin correlation vanishes) are consistent with Monte Carlo simulations [@NCD10] for $p=3/2$ and $p=2$. However, for $2d$ models, only few results are currently known, motivating us to develop this work.
In addition to the thermal phase transition, Hamiltonian (\[eq:hamiltonian\]) also presents a quantum phase transition ($QFT$) associated with the single-ion anisotropy. There is a competition between the exchange and the single-ion anisotropic energy. For small values of $D$, the quantum fluctuations are negligible and the spins tend to remain align with the $z$-axis, providing $\langle S^z \rangle\neq 0$. On the other hand, in the large-$D$ phase, the quantum fluctuations are larger and the energetic cost to keep a non-null $S^z$ spin component is too high. Therefore, the system goes to a gap phase with no spontaneous magnetization even at zero temperature. This transition occurs at the point $D=D_c$ and it is well documented for models with short-range interaction [@JPCM5; @PRB48; @PRB49; @PRB66; @PRB71; @JPCM20]. Below the critical point $D_c$, the system has a gapless energy spectrum and a quasi long-range order (quasi-$LRO$) with algebraic decay for the spin order-parameter correlation at low finite temperature. Above a critical transition temperature $T_c$, the order-parameter decays exponentially and there is no more quasi-$LRO$. Applying two different analytical methods, we have studied the quantum and thermal phase transitions as a function of the power-law exponent. In the next section we apply the bond operator (a Schwinger bosonic formalism) to determine the quantum phase transition at zero temperature. In section (\[scha\]), we have developed the Self-consistent Harmonic Approximation ($SCHA$) at zero temperature and in section (\[finiteT\]) the $SCHA$ is applied at finite temperatures. The conclusions are present in section (\[conclusion\]).
Bond operator formalism {#BO}
=======================
In this section, we use a $SU(3)$ Schwinger bosonic formalism, the so-called bond operator, to describe the $QFT$ at zero temperature in the large-$D$ phase. As commented earlier, for $D>D_c$ the system is disordered (vanishing spin-spin correlation) at zero temperature due to spin-waves fluctuations. To order to determine the critical point $D_c$, we have followed the procedures of references [@PRB56; @PRB71; @PLA360]. Considering a spin-$1$ model, we define three bosonic operators to represent the states of $S^z$: $|m_i=-1\rangle=a_{i,-1}^\dagger|0\rangle$, $|m_i=0\rangle=a_{i,0}^{\dagger}|0\rangle$ and $|m_i=1\rangle=a_{i,1}^\dagger|0\rangle$, where $|0\rangle$ is the vacuum state of the Fock space. The boson operator $a^\dagger_{im}$ creates a particle with $m_z=m$ on site $i$. The commutation relations $[S_i^+,S_j^-]=2S_i^z\delta_{ij}$ and $[S_i^z,S_j^{\pm}]=\pm S_i^{\pm}\delta_{ij}$ are valid and, to keep $S^2_i=S(S+1)$, we have to impose a local constraint $\sum_\mu a^\dagger_{i,\mu} a_{i,\mu}=S$ on each site. A condensation occurs in the $|m=0\rangle$ state once this is the smaller band energy (the $|m=\pm 1\rangle$ excited states are degenerate). Therefore, $N_0$, the number of particles in $|m=0\rangle$ state, obeys the condition $N_0=\langle a^\dagger_{i,0} a_{i,0}\rangle\gg 1$ and we can consider the approximation $[N_0,a^\dagger_{i,0}]=0$. Once $N_0$ and $a_{i,0}$ commute, we have treated the $a_{i,0}$ quantum operators as classical entities, i.e. real numbers. Thus we have adopted the mean-values $\langle a^\dagger_{i,0}\rangle=\langle a_{i,0}\rangle=\rho_0^{1/2}$ in the next equations ($\rho_0$ measures the density condensate in state $a_0^\dagger|0\rangle$). In the $SU(3)$ bosonic representation, the spin operators are written as: $$\begin{aligned}
S_i^+&=&\sqrt{2\rho_0}(u_i^\dagger+d_i),\nonumber\\
S_i^-&=&\sqrt{2\rho_0}(d_i^\dagger+u_i),\nonumber\\
S_i^z&=&u_i^\dagger u_i-d_i^\dagger d_i,\end{aligned}$$ on which we have properly replaced $a_{i,1}$ and $a_{i,-1}$ by $u_i$ and $d_i$, respectively. Therefore: $$\begin{aligned}
&&S_i^xS_j^x+S_i^yS_j^y+\lambda S_i^zS_j^z=\rho_0(d_id_j^\dagger+ d_iu_j + u_i^\dagger d_j^\dagger + \nonumber\\
&&+ u_i^\dagger u_j+d_i^\dagger d_j + d_i^\dagger u_j^\dagger + u_id_j+u_i u_{j}^\dagger)\nonumber\\
&& +\lambda(u_i^\dagger u_i-d_i^\dagger d_i)(u_j^\dagger u_j-d_j^\dagger d_j).\end{aligned}$$ The four operator terms are decoupled using a Hubbard-Stratonovich transform: $$\begin{aligned}
&&(u_i^\dagger u_i-d_i^\dagger d_i)(u_j^\dagger u_j-d_j^\dagger d_j)=2p^2-\frac{1}{2}(1-\rho_0)^2-\nonumber\\
&&-\frac{1}{2}m^2+\frac{1}{2}(1-\rho_0+m)(u_i^\dagger u_i+u_j^\dagger u_j)+\nonumber\\
&&+\frac{1}{2}(1-\rho_0-m)(d_i^\dagger d_i+d_j^\dagger d_j)-p(u_i d_j+d_iu_j+\nonumber\\
&&+d_j^\dagger u_j^\dagger+u_j^\dagger d_i^\dagger),\end{aligned}$$ where we have introduced the mean-field parameters $m=\langle u_i^\dagger u_i\rangle-\langle d_i^\dagger d_i\rangle$ and $p=\langle d_i^\dagger u_j^\dagger\rangle=\langle d_i u_j\rangle$. The anisotropic term is written as $(S_i^z)^2=u_i^\dagger u_i+d_i^\dagger d_i$ and after a Fourier transform, the Hamiltonian (\[eq:hamiltonian\]) is given by:
$$\begin{aligned}
\label{eq:hamiltonian_k}
H&=&\frac{1}{2}\sum_{\bf k}\left\{Z\left[\rho_0 J_{\bf k}-\lambda J_{\bf k} p\right](u_{-{\bf k}}d_{\bf k}+d_{-{\bf k}}u_{\bf k})+\left[\rho_0 ZJ_{\bf k}+\frac{ZJ_{\bf 0}}{2}(1-\rho_0+m)\lambda+D-\mu\right]u_{\bf k}^\dagger u_{\bf k}+\right.\nonumber\\
&&+\left.\left[\rho_0 ZJ_{\bf k}+\frac{ZJ_{\bf 0}}{2}(1-\rho_0 -m)\lambda+D-\mu\right]d_{\bf k}^\dagger d_{\bf k}+\textrm{H.c.}\right\}+H_0,\end{aligned}$$
with $H_0=\mu N (1-\rho_0)+\frac{\lambda}{4} NZ\left[4p^2-m^2-(1-\rho_0)^2\right]$, $Z=N-1$ and $N$ is the lattice sites number. In the last equation the constraint $\sum_\mu a^\dagger_{i,\mu} a_{i,\mu}=S$ was introduced as a Lagrange multiplier $\mu$ and we have defined $$\begin{aligned}
ZJ_{\bf k}=-JS_p({\bf k})=-J\sum_{\Delta{\bf r}\neq {\bf 0}} \frac{e^{-i {\bf k}\Delta{\bf r}}}{\Delta r^p}\end{aligned}$$ as the Fourier transform of the long-range ferromagnetic interaction. The function $S_p({\bf k})$ is calculated using the Ewald’s method, where the sum is divided in two rapidly convergent sums: the first over the space lattice and the second over the reciprocal lattice [@PR99; @JPC14]. After a straightforward calculation, we obtain for a square two-dimensional lattice [@PRB22]: $$\begin{aligned}
S_p({\bf k})&=&a^{-p}\frac{(\pi/4)^{p/2}}{\Gamma(p/2)}\left[\sum_{\bf l}e^{-ia{\bf k}{\bf l}}\Phi_{p/2-1}\left(\frac{\pi l^2}{4}\right)+\right.\nonumber\\
&&\left.+4\sum_{\bf l}\Phi_{-p/2}\left(\frac{|a{\bf k}-2\pi{\bf l}|^2}{\pi}\right)-\frac{2}{p}\right],\end{aligned}$$ where $\Delta{\bf r}=a{\bf l}=a(l_1\hat{x}+l_2\hat{y})$, $a$ being the lattice parameter and $l_1$ and $l_2$ integers numbers. The $\Phi$ functions are given by: $$\begin{aligned}
\Phi_m(x)=\int_1^\infty y^m e^{-y x}dy.\end{aligned}$$ The Hamiltonian (\[eq:hamiltonian\_k\]) is diagonalized by a Bogoliubov transform which provides: $$\begin{aligned}
H-E_g=\sum_{\bf k}\left[\omega_{\bf k}^{(+)}\alpha_{\bf k}^\dagger\alpha_{\bf k}+\omega_{\bf k}^{(-)}\beta_{\bf k}^\dagger\beta_{\bf k}\right],\end{aligned}$$ with $$\begin{aligned}
\omega_{\bf k}^{(\pm)}=\omega_{\bf k} \pm\frac{1}{2}Z\lambda m,\end{aligned}$$ $$\begin{aligned}
\omega_{\bf k}=\sqrt{\Lambda_{\bf k}^2-\Xi_{\bf k}^2},\end{aligned}$$ $$\begin{aligned}
\Lambda_{\bf k}=D-\mu +\frac{1}{2}ZJ_{\bf 0}\lambda(1-\rho_0)+\rho_0 ZJ_{\bf k}\end{aligned}$$ and $$\begin{aligned}
\Xi_{\bf k}=(\rho_0-\lambda p)ZJ_{\bf k}.\end{aligned}$$ The ground state energy is $E_g=\sum_{\bf k} \left(\omega_{\bf k}-\Lambda_{\bf k}\right) +H_0$. The new and old bosonic operators are related by: $$\begin{aligned}
d_{\bf k}=\xi_{\bf k}\beta_{\bf k}-\zeta_{\bf k}\alpha_{-{\bf k}}^\dagger\end{aligned}$$ and $$\begin{aligned}
u_{\bf k}=\xi_{\bf k}\alpha_{\bf k}-\zeta_{\bf k}\beta_{-{\bf k}}^\dagger\end{aligned}$$ where $\xi_{\bf k}^2=\frac{1}{2}(1+\frac{\Lambda_{\bf k}}{\omega_{\bf k}})$ and $\zeta_{\bf k}^2=\frac{1}{2}(-1+\frac{\Lambda_{\bf k}}{\omega_{\bf k}})$. The mean-field parameters $\rho_0$, $m$, $p$ and $\mu$ are determined by the the self-consistent equations obtained from the minimum of the Helmoltz free energy $F=-\beta^{-1}\ln\textrm{Tr}( e^{-\beta H})$. At zero temperature, the continuous limit of the self-consistent equations is given by: $$\begin{aligned}
\label{eq:rho}
\rho_0=2-\frac{1}{4\pi^2}\int \frac{\Lambda_{\bf k}}{\omega_{\bf k}}{\,\mathrm{d}}^2{\bf k},\end{aligned}$$ $$\begin{aligned}
\label{eq:p}
p=-\frac{1}{8\pi^2}\int\frac{\Xi_{\bf k}(J_{\bf k}/J_{\bf 0})}{\omega_{\bf k}}{\,\mathrm{d}}^2{\bf k},\end{aligned}$$ and $$\begin{aligned}
\label{eq:mu}
\mu=\frac{1}{4\pi^2}\int\frac{(\Lambda_{\bf k}-\Xi_{\bf k})ZJ_{\bf k}}{\omega_{\bf k}}{\,\mathrm{d}}^2{\bf k},\end{aligned}$$ where the integrals are evaluated over the first Brioullin Zone. The magnetization parameter $m$ is identically null at zero temperature (in the large-$D$ phase) and we can also approach $p\approx0$.
The energy $\omega_{\bf k}$ has a minimum at ${\bf k}^\ast=(0,0)$ and a finite gap $\Delta=\omega_{{\bf k}^\ast}$ exist for $D>D_c$. Even at zero temperature there is no spontaneous magnetization once the bosons condensate in the $|m=0\rangle$ state. Close to the critical point, $D\gtrsim D_c$, the gap decreases and when $D=D_c$, a quantum phase transition occurs. Below $D_c$, a gapless state takes place, which provides $\langle S^z\rangle\neq 0$, because the states $|m=1\rangle$ (or $|m=-1\rangle$) are filled. The bosonic formalism developed here is applied only for the large-$D$ phase. A similar method for $D<D_c$ phase can also be formulated, although, to our proposals, it is not necessary. Defining $y=2\rho_0\left[D-\mu+\frac{\lambda ZJ_{\bf 0}}{2}(1-\rho_0)\right]^{-1}$, equations (\[eq:rho\]) and (\[eq:mu\]) can be rewritten as: $$\begin{aligned}
2(2-\rho_0)=\left[I_1(y_m) + I_2(y_m)\right]\end{aligned}$$ and $$\begin{aligned}
y\mu=I_2(y_m)-I_1(y_m),\end{aligned}$$ which $I_1(y_m)=\frac{1}{4\pi^2}\int\frac{{\,\mathrm{d}}^2{\bf k}}{\sqrt{1+y ZJ_{\bf k}}}$ and $I_2(y_m)=\frac{1}{4\pi^2}\int\sqrt{1+yZJ_{\bf k}}{\,\mathrm{d}}^2{\bf k}$. In the gapless point ($D=D_c$) we have the minimum $y_m=-\frac{1}{ZJ_{\bf 0}}$ and using equations above, the critical point is found as: $$\begin{aligned}
D_c&=&\frac{\lambda ZJ_{\bf 0}}{4}\left[2-I_1(y_m)-I_2(y_m)\right]+\nonumber\\
&&+\frac{2}{y_m}\left[2-I_1(y_m)\right].\end{aligned}$$ The results for $D_c$ as function of the power-law exponent $p$ are show in Fig. (\[fig:dc\_bo\]). The $\lambda$ influence is small, mainly for $p<10$, and both $XY$ and isotropic model present approximately the same critical anisotropic constant $D_c$. For small $p$, the larger critical point ($D_c=17.72 J$ if $p=3$ and $D_c=12.86 J$ when $p=4$, both with $\lambda=1$) reflects the higher energetic cost to flip spins. Due the strong long-range interaction, it is easer to keep an ordered state which minimizes the exchange energy ($S_z\neq 0$) instead a state with $S_z=0$. To invert the dispute, making more favorable a state which minimizes the single-ion anisotropic energy, it is necessary a large anisotropic constant $D$. On the other hand, for $p\gg 1$ the $D_c$ tends to a constant value, $D_c=5.74 J$ if $\lambda=0$ and $D_c=6.10 J$ when $\lambda=1$. As expected, for long distances $\Delta r_{ij}$, the interaction is negligible and the system behaves like a model with short-range interactions. Therefore, the results for large $p$ agree with that obtained from models with nearest neighbors interactions. For two-dimensional antiferromagnetic models ($J=-1$) with only $SR$ interactions, we have $D_c=5.72 J$ for the $XY$ model [@PA388] and the set of values $D_c=5.77 J$ [@PLA360], $D_c=5.82 J$ [@PRB71] and $D_c=6.38 J$ [@PRB50] for the isotropic model. The same values are obtained also for a ferromagnetic coupling ($J=1$), once the condensation occurs at ${\bf k}=(0,0)$ instead ${\bf k}=(\pi,\pi)$ (for the antiferromagnetic case) and the Hamiltonian is invariant under the transformation ${\bf k}\to{\bf k}+{\pi}$ and $J\to -J$.
Close to the origin we can adopt the approximation: $S_p(k)\approx S_p(0)-\kappa_p f_p(k)$ where $S_p(0)=4^{1-n/2}\left[\zeta(\frac{p}{2},\frac{1}{4})-\zeta(\frac{p}{2},\frac{3}{4})\right]\zeta(\frac{p}{2})$, with $\zeta(s,q)$ being the Hurwitz zeta function and $\zeta(s)$ the zeta function. The constant $\kappa_p$ is numerically determined for each exponent $p$ and the function $f_p(k)$ is defined as [@PRB46; @PLA202]: $$\begin{aligned}
f_p(k)=\left\{
\begin{array}{lc}
k^{p-2}, & 2<p<4\\
k^2 \ln(k), & p=4\\
k^2, & p>4
\end{array}
\right.\end{aligned}$$ Thus, the excited state energy for $|{\bf k}-{\bf k}^\ast|\approx 0$ is $\omega_{\bf k}=\sqrt{\Delta^2+c^2 f_p^2(k)}$, where the gap energy and the long wavelength spin-wave velocity are given by: $$\begin{aligned}
\Delta=\left[D-\mu+\frac{\lambda ZJ_{\bf 0}}{2}(1-\rho_0)\right]\sqrt{1+ZJ_{\bf 0} y}\end{aligned}$$ and $$\begin{aligned}
c=\left[D-\mu+\frac{\lambda ZJ_{\bf 0}}{2}(1-\rho_0)\right]\sqrt{J \kappa_p y},\end{aligned}$$ respectively . For power-law decay with exponent $p>4$ we have relativistic dispersion relation, as occurs in the short-range interaction models, and the correlation length is expressed by: $$\begin{aligned}
\label{eq:cl}
\xi=\frac{c}{\Delta}=\sqrt{\frac{J\kappa_p y}{1+ZJ_{\bf 0} y}}.\end{aligned}$$ In the limit of low temperatures and long-distance $\Delta r_{ij}$, the equal time correlation function is given by: $$\begin{aligned}
&& \langle S_i^+ S_j^- \rangle =\frac{1}{4\pi^2}\int\langle S_k^+ S_k^- \rangle e^{i{\bf k}\cdot\Delta{\bf r}}{\,\mathrm{d}}^2{\bf k}\nonumber\\
&&=\frac{\rho_0}{\pi^2}\int\left(\frac{\Delta_{\bf k}-\Xi_{\bf k}}{\omega_{\bf k}}\right)\left(n_{\bf k}+\frac{1}{2}\right)e^{i{\bf k}\cdot\Delta{\bf r}}{\,\mathrm{d}}^2{\bf k}\nonumber\\
&&\approx \frac{T}{2J \kappa_p \pi^2}\int\frac{e^{i{\bf k}\cdot\Delta{\bf r}}}{k^2+\xi^{-2}}{\,\mathrm{d}}^2{\bf k} \propto \left(\frac{\Delta r}{\xi}\right)^{-\frac{1}{2}}e^{-\frac{\Delta r}{\xi}},\end{aligned}$$ where $\xi$ is finite in the large-$D$ phase. As expected, even at zero temperature, there is no $LRO$ when $D>D_c$ and the correlation function is similar to that obtained from the models with short-range interactions if $p>4$. When $D=D_c$ the system goes to a gapless state which provides an infinite correlation function and $LRO$ for $D<D_c$. In Fig. (\[fig:xi\_bo\]), we plot the correlation length at zero temperature for $p=2$, $4$ and $6$. In the large-$D$ phase, $\xi$ is finite but diverges when $D$ tends to $D_c$ (represented by the vertical lines).
The dynamical structure factor is calculated as: $$\begin{aligned}
S({\bf k},t)&=&\langle S_k^+(t)S_k^-(0)\rangle \nonumber\\
&=& \langle S_k^+S_k^-\rangle\left[e^{-i\omega_{\bf k}t}H(t)-e^{i\omega_{\bf k}t}H(-t)\right],\end{aligned}$$ where $H(t)$ is the Heaviside step function. Using $H(t)=\displaystyle\lim_{\epsilon\to 0^+}\frac{1}{2\pi i}\int_{-\infty}^{+\infty}\frac{e^{itx}}{x-i\epsilon}dx$, the time Fourier transform provides: $$\begin{aligned}
S({\bf k},\omega)=\frac{2c^2}{J\kappa_p(\omega^2-\omega_{\bf k}^2)}.\end{aligned}$$
Self-consistent Harmonic Approximation {#scha}
======================================
The Self-consistent Harmonic Approximation ($SCHA$) [@JPC7; @SSC104; @PA388; @PRB48; @PRB53; @PRB54; @PSS242; @JMMM357] is the most appropriate spin-wave method to treat the model at finite temperatures. The bosonic formalism used in the last sections provides reasonable results at zero temperature but, as argued by Yoshida [@JPSJ58], for $T>0$ the Schwinger formalism presents some divergences with another well-known results.
As demonstrated, due to the single-ion anisotropy, the analyzed system exhibits a quantum phase transition at $D=D_c$. In the large-$D$ phase, $D>D_c$, there is no spontaneous magnetization at finite and zero temperature. Below the critical point $D_c$, the $O(3)$ symmetry is spontaneously broken and $m\neq 0$ at zero temperature. If $2<p<4$, there is also $LRO$ at finite low temperatures (above a transition temperature the disordered state takes place). However, for $p\geq 4$ there is no spontaneous broken symmetry at any finite temperature and the system behaves like a short-range model for which the Mermin-Wagner theorem is valid. There is a thermal phase transition similar to the Berezinski-Kosterlitz-Thouless transition present in the $XY$ model. Below a critical temperature $T_c$, the model has a quasi long-range order with algebraic decay for the spin order-parameter correlation. For $T>T_c$ the decay becomes exponential and there is no more quasi $LRO$. In this section we have applied the $SCHA$ to determine the critical temperature $T_c$ beyond the $D_c$ constant. The main idea behind the $SCHA$ is replace the original Hamiltonian by another one with temperature dependent renormalized parameters. Using the Villain representation [@JP35], we written the spin operators as: $$\begin{aligned}
S_i^+=e^{i\phi_i}\sqrt{S(S+1)-S_i^z(S_i^z+1)}\end{aligned}$$ and $$\begin{aligned}
S_i^-=\sqrt{S(S+1)-S_i^z(S_i^z+1)}e^{-i\phi_i}.\end{aligned}$$ Therefore the Hamiltonian (\[eq:hamiltonian\]) is given by:
$$\begin{aligned}
\label{eq:hscha}
H=\frac{1}{2}\sum_{i\neq j}J_{ij}\left[\tilde{S}^2\sqrt{1-\left(\frac{S_i^z}{\tilde{S}}\right)^2}\sqrt{1-\left(\frac{S_j^z}{\tilde{S}}\right)^2}\cos(\phi_j-\phi_i)+
\lambda S_i^zS_j^z\right]+D\sum_i(S_i^z)^2,\end{aligned}$$
with $\tilde{S}=S(S+1)$. The mean-value $\langle \phi\rangle$ is not well defined to angles measured relative to a fixed axis, so we choose the angle operator $\phi_i$ relative to the direction of the instantaneous magnetization in order to avoid divergences. Considering $D<D_c$ and low temperatures, the spin field assumes a configuration with a small angular difference, $|\phi_j-\phi_i|\ll 1$. Thus, we expand the Hamiltonian in powers of $(S_i^z/\tilde{S})^2$ and $(\phi_j-\phi_i)^2$ which provides: $$\begin{aligned}
\label{eq:hscha0}
H&=&\frac{1}{2}\sum_{i\neq j}J_{ij}\left[-\frac{\rho_s \tilde{S}^2}{2}(\phi_j-\phi_i)^2-(S_i^z)^2+\lambda S_i^zS_j^z\right]+\nonumber\\
&&+D\sum_i (S_i^z)^2,\end{aligned}$$ where the spin stiffness: $$\begin{aligned}
\rho_s=\left\langle\sqrt{1-\left(\frac{S_i^z}{\tilde{S}}\right)^2}\sqrt{1-\left(\frac{S_j^z}{\tilde{S}}\right)^2}\cos(\phi_j-\phi_i)\right\rangle.\end{aligned}$$ takes into account contributions of anharmonic terms in the approximation. The average $\langle\cdots\rangle$ is taken with respect to the original Hamiltonian (\[eq:hscha\]). Once the fields $\phi_i$ and $S_i^z$ obey the fundamental Poisson bracket $\{\phi_i,S_i^z\}=\delta_{ij}$ for polar representation of a spin vector, the operators $\phi_i$ and $S_i^z$ are canonically conjugate, i.e. $[\phi_i,S_j^z]=i\delta_{ij}$ (adopting $\hbar=1$). After a Fourier transform, we have: $$\begin{aligned}
\label{eq:hschak}
H&=&\frac{1}{2}\sum_{\bf k}\left\{\left[ Z(\lambda J_{\bf k}-J_{\bf 0})+2D\right]S_{\bf k}^zS_{-{\bf k}}^z\right.+\nonumber\\
&&\left.+\rho_s\tilde{S}^2Z(J_{\bf k}-J_{\bf 0})\phi_{\bf k}\phi_{-{\bf k}}\right\},\end{aligned}$$ on which $J_{\bf k}$ is the same defined in previous section. The Hamiltonian is then diagonalized introducing a new bosonic operator $a_{\bf k}$ by the canonical transformation: $$\begin{aligned}
\phi_{\bf k}=\frac{1}{\sqrt{2}}\left[\frac{Z(\lambda J_{\bf k}-J_{\bf 0})+2D}{\rho_s \tilde{S}^2Z(J_{\bf k}-J_{\bf 0})}\right]^{\frac{1}{4}}(a_{\bf k}^\dagger+a_{-{\bf k}}),\end{aligned}$$ and $$\begin{aligned}
S_{\bf k}^z=\frac{i}{\sqrt{2}}\left[\frac{\rho_s \tilde{S}^2Z(J_{\bf k}-J_{\bf 0})}{Z(\lambda J_{\bf k}-J_{\bf 0})+2D}\right]^{\frac{1}{4}}(a_{\bf k}^\dagger-a_{-{\bf k}}).\end{aligned}$$ After a straightforward calculation, we obtain the harmonic diagonalized Hamiltonian: $$\begin{aligned}
H_0=\sum_{\bf k}\omega_{\bf k}\left(a_{\bf k}^\dagger a_{\bf k} +\frac{1}{2}\right),\end{aligned}$$ with the spin-wave energy $$\begin{aligned}
\omega_{\bf k}=\sqrt{\rho_s\tilde{S}^2Z(J_{\bf k}-J_{\bf 0})[Z(\lambda J_{\bf k}-J_{\bf 0})+2D]}.\end{aligned}$$ Through the diagonalized Hamiltonian, we have determined the mean-value fields: $$\begin{aligned}
\langle\phi_{\bf k}\phi_{-{\bf k}}\rangle_0=\frac{1}{2}\sqrt{\frac{Z(\lambda J_{\bf k}-J_{\bf 0})+2D}{\rho_s \tilde{S}^2Z(J_{\bf k}-J_{\bf 0})}}\coth\left(\frac{\beta\omega_{\bf k}}{2}\right)\end{aligned}$$ and $$\begin{aligned}
\langle S_{\bf k}^zS_{-{\bf k}}^z\rangle_0=\frac{1}{2}\sqrt{\frac{\rho_s\tilde{S}^2Z(J_{\bf k}-J_{\bf 0})}{Z(\lambda J_{\bf k}-J_{\bf 0})+2D}}\coth\left(\frac{\beta\omega_{\bf k}}{2}\right)\end{aligned}$$ To evaluate the spin stiffness, we calculate the average $\langle\cdots\rangle$ using the diagonalized Hamiltonian $H_0$. Since $\phi_i$ and $S_i^z$ are uncoupled operators and $\phi$ has a Gaussian distribution, $\rho_s$ is given by: $$\begin{aligned}
\label{eq:rho_scha}
\rho_s\approx\left[1-\left\langle\left(\frac{S_i^z}{\tilde{S}}\right)^2\right\rangle_0\right]e^{-\frac{1}{2}\langle(\phi_j-\phi_i)^2\rangle_0}.\end{aligned}$$ For small values of $D$, the spin-waves have small energy and the spin stiffness is finite; however, in the large-$D$ phase, the quantum fluctuations are sufficiently large, disordering the system. The point (at zero temperature) where $\rho_s$ abruptly vanishes is taken as the critical point $D_c$. Using Eq. (\[eq:rho\_scha\]), as well as the the mean-values for $\langle\phi_{\bf k}\phi_{-{\bf k}}\rangle_0$ and $\langle S_{\bf k}^z S_{-{\bf k}}^z\rangle_0$, we have determined the transition point $D_c$ as a function of the power-law exponent $p$. In Fig. (\[fig:dc\_scha\]) we plot the results from $SCHA$ analysis (for $\lambda=1$) and also those obtained in a previous section (using the bond operator formalism). As one can see, both methods provide according results. In the limit $p\gg 1$, in which the interactions behave like a short-range, the critical anisotropy is $D_c=6.72J$, slightly larger than those obtained from bond operator method ($D_c=6.10J$).
Finite temperature {#finiteT}
==================
As commented previously, beyond the quantum phase transition at $D=D_c$, there is a thermal transition which separates a phase with quasi-long-range order from another with null spin-spin correlation. Applying Eq. (\[eq:rho\_scha\]) to finite temperatures we calculate the critical point $T_c$ as the point where $\rho_s$ vanishes discontinuously. This temperature, as explained in next section, is a first approximation to a transition of $BKT$ kind. In the classical limit, we can use the approximation $\coth\left(\frac{\beta\omega_{\bf k}}{2}\right)\approx \frac{2T}{\omega_{\bf k}}$ (adopting $k_B=1$). Equation (\[eq:rho\_scha\]) is then written as: $$\begin{aligned}
\label{eq:rho_cl}
\rho_s=(1-I t)e^{-\frac{R t}{\rho_s}},\end{aligned}$$ in which $t=\frac{T}{J \tilde{S}^2}$ is the reduced temperature and the constants are given by: $$\begin{aligned}
R=\frac{1}{4\pi}\int \frac{J(1-\gamma_{\bf k})}{Z(J_{\bf k}-J_{\bf 0})}{\,\mathrm{d}}^2{\bf k}\end{aligned}$$ and $$\begin{aligned}
I=\frac{1}{4\pi}\int \frac{J}{Z(\lambda J_{\bf k}-J_{\bf 0})+2D}{\,\mathrm{d}}^2{\bf k}.\end{aligned}$$ The $I$ term measures the out-of-plane fluctuations and therefore $I=0$ for the planar rotator model since no $S^z$ components are allowed. For the other models $0<I\leq 1$. It is important to highlight that better results are reached for large $p$ since this case is closer to a short-range interaction model. For small power-law exponents $p$, the energy of the spin-waves is not sufficiently small and may be necessary to consider second order terms in the expansion of $\coth\left(\frac{\beta\omega_{\bf k}}{2}\right)$. We will use the classical approximation only for a qualitative analysis at low temperatures while the correct results will be directly determined by Eq. (\[eq:rho\_scha\]).
The critical temperature is obtained at the point where Eq. (\[eq:rho\_cl\]) admits as solution the trivial one. Considering only first order terms, $T_c$ is determined by $$\begin{aligned}
\label{eq:tc}
\frac{T_c}{J\tilde{S}^2}=\frac{1}{e R+I}\end{aligned}$$ with $e$ being the base of the natural logarithm. The classical result for $T_c$ shows an unexpected increasing behavior with respect to $D$. In the large-$D$ phase, there is no long-range order, even at zero temperature; however the classical model indicates a finite transition temperature. The problem is to consider low-energy spin-waves when quantum fluctuations are larger enough to disorder the system. Classically, the limit $D\to\infty$ provides the transition temperature for the planar rotator once there is no $S^z$ component. In Fig. (\[fig:tc\_classical\]) we plot $T_c$ as a function of $D$ using the classical approximation for $p=3$. The curve for the $XY$ model ($\lambda=0$), the isotropic model ($\lambda=1$) as well as the asymptotic limit of the planar rotator are plotted. The $\lambda$ influence is small, mainly for $D\gg 1$, and both curves tend to the planar rotator limit $T_c=3.60 J\tilde{S}^2$ when $D$ increases. A similar behavior is observed for all values of the power-law exponent $p$.
To include quantum fluctuations in the critical temperature, we have to evaluate the phase transition using the Eq. (\[eq:rho\_scha\]), disregarding any classical approximation. The results of $T_c$ as a function of the single-ion anisotropy $D$ for $p=$3, $4$ and $6$ are shown in Fig. (\[fig:Tc\_D\]). As expected, $T_c$ decreases as $D$ increases and close to the critical point $D_c$, there is a discontinuity associated with the $QFT$. Above the critical anisotropy there is no transition at finite temperature and the system is quantum disordered. The according results between classical and quantum calculations are recovered in the limit of large spin in which fluctuations are negligible.
The long-range interactions effects are shown in Fig. (\[fig:tc\_p\]). We plot the results for the $XY$ model (with $D=0$) taking into account the quantum fluctuations, through Eq. (\[eq:rho\_scha\]), as well as using the classical approximation from Eq. (\[eq:rho\_cl\]). For $p=3$ we have found $T_c=2.57 J\tilde{S}^2$ by using the classical approximation and $T_c=2.22 J\tilde{S}^2$ considering the spin-waves contributions. The large values for $T_c$, compared with the results for $p\gg 1$, reflect the large exchange spin energy and so the high thermal energy required for restoring the $O(3)$ symmetry. On the other hand, in the limit $p\gg 1$, the critical temperature tends to $T_c=1.08J\tilde{S}^2$ (classical) and $T_c=0.91J\tilde{S}^2$ (quantum). The result is consistent with the transition temperature $T_c=1.076 J\tilde{S}^2$ obtained for the classical $XY$ model with nearest-neighbor interactions [@PRB48-b]. Furthermore, the classical approximation overestimates the critical temperature which justifies the smaller temperature determined from Eq. (\[eq:rho\_scha\]).
It is well-known that the stiffness for the $XY$ model with short-range interactions should exhibit a universal jump at $T_{KT}$ associated with the Berezinskii-Kosterlitz-Thouless ($BKT$) transition. The $BKT$ transition is not associated with a spontaneous symmetry-breaking as occurs in the transitions described by the Mermin-Wagner theorem. Instead, it involves the emergence of topological configurations at finite temperature, the magnetic vortices. The angle field $\phi({\bf r})$ should be split in two parts $\phi({\bf r})=\phi_0({\bf r})+\psi({\bf r})$ where $\phi_0({\bf r})$ describes the small phase fluctuations of the order parameter while $\psi({\bf r})$ is the vortex field. The stiffness $\rho_s$ from Eq. (\[eq:rho\]) takes into account only the spin-wave contribution, neglecting any vortex effect. The vortex field can be implemented by the replacement $\rho_s\to\tilde{\rho}_s=\rho_s\eta$, in which $\eta$ is a renormalization factor. The $\eta$ term presents a discontinuity at $T_{KT}$ temperature given by $\lim_{T\to T_{KT}}J\tilde{S}^2\eta/T_{KT}=2/\pi$. Below $T_{KT}$, an ordered state occurs due the bounding of the vortex-antivortex pairs whilst above $T_{KT}$, the vortices are free guiding the system to a disordered state. The critical temperature $T_c$ obtained from Eq. (\[eq:tc\]) is only a first approximation to $T_{KT}$ while a better result is determined by the crossing between the curve $\rho_s$, from Eq. (\[eq:rho\_cl\]), and the line $\eta=2T/\pi J\tilde{S}^2$ [@EPJB44; @PRB48-b]. In general, the $T_{KT}$ temperature is smaller than $T_c$ and for $T<T_c$, the $SCHA$ provides good results at finite temperatures. For the two-dimensional $XY$ model, for example, $T_{KT}=0.83 J\tilde{S}^2$ [@PRB48-b]. Applying the same procedure to the $XY$ ferromagnet model with long-range interactions, we have found $T_{KT}=1.16 J\tilde{S}^2$ with $p=3$ and $T_{KT}=0.83 J\tilde{S}^2$ in the limit $p\gg 1$. Figure (\[fig:tkt\]) shows $T_{KT}$ as a function of $p$ for the $XY$ model.
Conclusion
==========
In this work we have investigated the two-dimensional single-ion anisotropic ferromagnetic with long-range interactions using a Schwinger bosonic formalism and the Self-consistent Harmonic Approximation. As in the short-range interaction case, the single-ion anisotropy $D$ is responsible for a quantum phase transition. The Hamiltonian is composed by two conflicting terms: the exchange and the single-ion anisotropic ones. Below a critical value $D_c$ the system has long-range order once an ordered state is energetic favorable. On the other hand, for the large-$D$ phase ($D>D_c$), there is a high cost to keep out-of-plane spin components and, therefore, the system prefers to keep a vanishing spontaneous magnetization state (even at zero temperature). The results obtained from both methods are similar and for large power-law exponent $p$ there is an excellent according with short-range interaction models. It is expected since for $p\gg 1$, the long-range interactions are very weak at long distances. Actually, in the approximation for long wavelength ($k\ll 2\pi/a$), the long-rage interaction model with $p>4$ has a relativistic energy spectrum, as occurs for a system with only short-range interactions. For small values of $p$, the critical anisotropy $D_c$ is larger than that of the short-range version, which reflects the high energetic cost to flip spins.
At finite low temperatures, there is a thermal phase transition for $D<D_c$. Using the $SCHA$, we have determined the critical temperature $T_c$ which separates a state with quasi long-range order from another one with vanishing short spin-spin correlation. In analogy with nearest-neighbor interaction models, there is a temperature $T_{KT}$ associated with the $BKT$ transition. The $SCHA$ does not consider vortex effects but they can be introduced by a renormalization factor in $\rho_s$. The critical $T_c$ gives only a first approximation to $T_{KT}$ and more precise results are obtained by the intersection of the stiffness $\rho_s$ curve with the line $\eta=2T/\pi J\tilde{S}^2$. Considering $D\approx 0$ and $p\gg 1$ we can use a classical approximation at sufficient low finite temperatures. However, for $D\approx D_c$ the quantum fluctuations are large and the approximation fails. While the classical result for $D\to\infty$ predicts a finite temperature $T_c$ (below which there is a ordered state), the quantum model indicates a disordered state even at zero temperature for any $D>D_c$. For large spins, the spin-wave fluctuations are negligible and we have according results for both limits, quantum and classical. As expected, the limit $p\gg 1$ always provides comparable results with the well-known short-range interaction models.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'By exhibiting a violation of a novel form of the Bell-CHSH inequality, Żukowski has recently established that the quantum correlations exploited in the standard perfect teleportation protocol cannot be recovered by any local hidden variables model. Allowing the quantum channel state in the protocol to be given by any density operator of two spin-1/2 particles, we show that a violation of a generalized form of Żukowski’s teleportation inequality can only occur if the channel state, considered by itself, violates a Bell-CHSH inequality. On the other hand, although it is sufficient for a teleportation process to have a nonclassical fidelity—defined as a fidelity exceeding $2/3$—that the channel state employed violate a Bell-CHSH inequality, we show that such a violation does *not* imply a violation of Żukowski’s teleportation inequality or any of its generalizations. The implication does hold, however, if the fidelity of the teleportation exceeds $2/3(1+1/2\sqrt{2})\approx .90$, suggesting the existence of a regime of nonclassical values of the fidelity, less than $.90$, for which the standard teleportation protocol can be modelled by local hidden variables.'
address:
- |
Departments of Philosophy and History and Philosophy of Science,\
10th floor, Cathedral of Learning, University of Pittsburgh,\
Pittsburgh, PA 15260, USA.\
email: [email protected]
- |
Department of Physics, University of Queensland\
Brisbane 4072, Queensland, Australia\
email: [email protected]
author:
- Rob Clifton
- Damian Pope
title: On the Nonlocality of the Quantum Channel in the Standard Teleportation Protocol
---
[2]{}
Introduction
============
It is well-known that the quantum correlations predicted by pure entangled states violate Bell-CHSH inequalities [@popescu_rohrlich] and, therefore, cannot be recovered by any local hidden variables model. Yet these same correlations cannot by themselves be used to transmit information between the locations of the entangled systems. This has sometimes been taken to suggest that quantum correlations are not inherently ‘nonlocal’, but simply ‘nonclassical’. On the other hand, it is a striking fact that quantum correlations *can* used, as in quantum teleportation, to increase the information carrying capacity of a classical channel. This new operational manifestation of quantum entanglement invites a deeper analysis of the extent to which the quantum teleportation process itself can be modelled classically, and what role, if any, nonlocality must play in explaining the success of teleportation.
The standard quantum teleportation protocol runs as follows [@6man]. Consider three qubit systems 1, 2, and 3, fix an arbitrary direction in space to define the $z$-direction for each qubit, and let $\left| \uparrow \right\rangle $ and $\left| \downarrow
\right\rangle $ denote the eigenstates of $\sigma_{z}$. We can imagine that all three qubits are initially in the possession of someone named Clare, who follows the instructions of Alice and Bob to prepare qubits $2+3$ in a quantum channel state given by some density operator $D$ of their choice, and prepares qubit $%
1 $ in a pure state $\left| \phi \right\rangle $, unknown to Alice or Bob. Clare then feeds qubits $2$ and $3$ to Alice and Bob, respectively, to use as their “quantum channel”, and Clare passes qubit $1$ to Alice so that she can teleport its unknown state to Bob. To execute the protocol, Alice measures the “Bell operator” on qubits $1+2$, which has eigenstates [@bmr]: $$\begin{aligned}
\left| \Phi _{\pm }\right\rangle &=&1/\sqrt{2}(\left| \uparrow
\right\rangle \left| \uparrow \right\rangle \pm \left| \downarrow
\right\rangle \left| \downarrow \right\rangle ), \\
\left| \Psi _{\pm }\right\rangle &=&1/\sqrt{2}(\left| \uparrow
\right\rangle \left| \downarrow \right\rangle \pm \left| \downarrow
\right\rangle \left| \uparrow \right\rangle ),\end{aligned}$$ and she obtains one of four results $n=1,...,4$ corresponding to the eigenstates above, with probabilities $p_{n}$. For each outcome $n$, Alice then sends classical information, via (say) a normal telephone call, instructing Bob to perform a corresponding unitary transformation $U_{n}$ on his qubit $3$. Having followed Alice’s instructions, Bob’s qubit will be left in one of four states $D_{n}$.
If the quantum channel state $D$ is a maximally entangled pure state, the unitaries $U_{n}$ can always be chosen so that no matter what Alice’s Bell operator measurement outcome, Bob will have prepared his particle by the completion of the protocol in exactly the same state $D_{n}=\left| \phi \right\rangle
\left\langle \phi \right| $ ($n=1,...,4$) as the state of qubit $1$ that Alice teleported. For example, when $D$ is the singlet state, qubit 3 has an equal probability of being in one of the four states $\left| \phi \right\rangle\left\langle \phi \right|$, $\sigma_{x}\left| \phi \right\rangle\left\langle \phi
\right|\sigma_{x}$, $\sigma_{y}\left| \phi \right\rangle\left\langle \phi
\right|\sigma_{y}$, $\sigma_{z}\left| \phi \right\rangle\left\langle \phi
\right|\sigma_{z}$, so that Bob’s performing, respectively, one of the four unitary transformations $$\begin{aligned}
\nonumber
U_{1}=I\ \ (\left| \Psi _{-}\right\rangle), & \ U_{2}=
\sigma_{x}\ \ (\left| \Phi _{-}\right\rangle),\\ \label{standard}U_{3}=\sigma_{y}\ \
(\left| \Phi _{+}\right\rangle),\ & U_{4}=\sigma_{z}\ \ (\left| \Psi _{+}\right\rangle),\end{aligned}$$ suffices for perfect teleportation.
For a mixed and/or non-maximally entangled quantum channel state $D$, perfect teleportation of an unknown state cannot be achieved (cf., e.g., [@werner]). However, one can introduce a natural measure of success for any fixed “strategy” $n\mapsto U_{n}$ of associating unitary operators $U_{n}$ with Alice’s Bell operator outcomes. This measure is the fidelity $\mbox{$\mathcal{F}$}_{\{U_{n}\}}(D)$ of transmission to Bob, given by the uniform average, over all possible unknown states $\left| \phi \right\rangle $, of the quantity $\sum_{n=1}^{4}p_{n}%
{Tr}(D_{n}\left| \phi \right\rangle \left\langle \phi
\right| )$ [@popescu],[@gisin]. Taking the supremum of $\mbox{$\mathcal{F}$}_{\{U_{n}\}}(D)$ over all possible strategies for associating unitary operators $%
U_{n}$ with Alice’s Bell operator outcomes, one obtains a number, $\mbox{$\mathcal{F}$}_{max}(D)$, representing the *maximum achievable* fidelity of teleportation for a given fixed quantum channel state $D$ [@horodeckis]. A density matrix $D$ is then regarded as useful for nonclassical teleportation just in case $\mbox{$\mathcal{F}$}_{max}%
(D)>2/3$, where $2/3$ is the maximum achievable fidelity without the quantum channel, or when a classical channel is substituted in its place [@popescu],[@massar].
Let us return, for the moment, to the case of perfect teleportation, where $%
D$ is a maximally entangled pure state and $\mbox{$\mathcal{F}$}_{max}(D)=1$. The striking thing is that Alice only directly communicates to Bob two classical bits of information when she conveys to him one of the four integers $%
n=1,...,4$. (We assume, as usual, that they have agreed in advance on a strategy $n\mapsto U_{n}$.) Indeed, one might think that Bob’s sure-fire reconstruction of the unknown state $\left|
\phi \right\rangle $ as a state of his qubit $3$ entails that he must (somehow) have actually received a *full qubit’s worth* of information (i.e., the information contained in specifying the two of infinitely many possible real numbers needed to fix the normalized expansion coefficients of the state $\left|
\phi \right\rangle $). A natural, though perhaps naive, explanation would be that the correlations inherent in a maximally entangled pure state carry the extra information to Bob instantaneously and nonlocally when that state “collapses” as a result of Alice’s Bell operator measurement. Indeed, the view that the quantum channel itself carries some share of the net information received by Bob appears to have been favoured by the inventors of teleportation ([@6man], p. 1896; cf. also [@popescu], p. 797), and is the springboard for a number of different explanations of teleportation from the point of view of particular interpretations of quantum theory ([@vaidman]–[@bohm]). There are sceptics, however, who have challenged the view that perfect teleportation requires nonlocal information transfer of any sort ([@hardy]–[@ari]). Moreover, one should bear in mind that the information about the unknown state $\left|
\phi \right\rangle $ “carried” by Bob’s qubit at the completion of a *single* run of the protocol is not actually *accessible* to him, since it is well-known that there is no way for him to discern the quantum-mechanical state of a single system.
In the case of explaining imperfect nonclassical teleportation, where $1>\mbox{$\mathcal{F}$}_{max}(D)>2/3$, the situation is more complicated, and appears to favour the sceptics. While $D$ must still be entangled (i.e., not a convex combination of product states of $2+3$), $D$ can still satisfy *all* Bell-CHSH inequalities [@popescu] $$|\left\langle\bar{\sigma}_{1}\otimes \bigl( \sigma_{1}
+ \sigma _{2} \bigr)
+ \bar{\sigma} _{2}\otimes \bigl( \sigma_{1}
- \sigma _{2} \bigr)\right\rangle_{D}|\leq 2,$$ where the $\bar{\sigma}$’s and $\sigma$’s denote arbitrary spin observables of qubits $2$ and $3$, respectively, and $\left\langle \cdot \right\rangle
_{D}$ denotes expectation value in the state $D$. More precisely, if we let $\beta (D)\in [2,2\sqrt{2}]$ be the maximum over all Bell-CHSH expressions of the above form, then $\mbox{$\mathcal{F}$}_{max}(D)>2/3$ does *not* imply $\beta (D)>2$, even though the converse implication does hold [@horodeckis]. So nonclassical teleportation is possible even in the absence of an independent argument for the nonlocality of the quantum channel state via Bell’s theorem. Notwithstanding this, one could adopt the view that the condition $\mbox{$\mathcal{F}$}_{max}(D)>2/3$ *itself* should be viewed as sufficient for thinking of the channel state $D$ as nonlocal (cf. [@popescu], p. 799). However, the sceptic is likely to point out that this gives further support to his view that nonclassical teleportation can occur without the involvement of nonlocality (cf. [@hardy], p. 6).
All parties to this discussion are agreed that good teleportation is made possible by the fact that the correlations of an entangled quantum channel state $D$ are nonclassical. They only diverge on the question of whether or not the standard teleportation protocol necessarily involves nonlocality. Recently, Żukowski [@z-guy] has made an interesting attempt to force this issue of nonlocality by asking an extremely relevant question: Can the correlations of $D$ that are *actually involved* in the standard teleportation protocol be recovered in a local hidden variables model? Żukowski attacks this question by making a novel use of a Bell-CHSH inequality, which we explain in the next section. The two sections following, Sections \[section\_three\] and \[section\_four\], investigate the connection between a generalized form of Żukowski’s teleportation inequality and the standard Bell-CHSH inequality for the channel state. In particular, we show that a violation of the former implies that the channel state satisfies $\beta(D) > 2$, but that $\beta(D) > 2$ does *not* imply a violation of even our generalized form of Żukowski’s teleportation inequality. However, in Section \[section\_five\] we show that such a violation occurs whenever the fidelity $\mbox{$\mathcal{F}$}_{st}(D)>2/3(1+1/2\sqrt{2})\approx .90$, where $\mbox{$\mathcal{F}$}_{st}$ is the fidelity associated with the *standard* choice for Bob’s unitary operators given in (\[standard\]). The implications of these results for the interpretation of quantum teleportation as nonlocal are discussed in Section \[section\_six\].
Żukowski’s Argument Simplified {#section_two}
==============================
In effect, Żukowski considers the following pair of $\pm 1$-valued functions of Alice’s Bell operator: $$\begin{aligned}
\nonumber
A_{1} & = & \left| \Psi _{-}\right\rangle \left\langle \Psi
_{-}\right| +\left| \Phi _{-}\right\rangle \left\langle \Phi
_{-}\right| \\ & & -\left| \Psi _{+}\right\rangle \left\langle \Psi
_{+}\right| -\left| \Phi _{+}\right\rangle \left\langle \Phi
_{+}\right| , \label{A} \\ \nonumber
A_{2} & = & \left| \Psi _{+}\right\rangle \left\langle \Psi
_{+}\right| +\left| \Phi _{-}\right\rangle \left\langle \Phi
_{-}\right| \\ & & -\left| \Psi _{-}\right\rangle \left\langle \Psi
_{-}\right| -\left| \Phi _{+}\right\rangle \left\langle \Phi
_{+}\right| . \label{B}\end{aligned}$$ The operator $A_{1}$ ($A_{2})$ coincides with the observable that represents the first (second) component of the *vector* observable $%
\overrightarrow{A}$ introduced by Żukowski ([@z-guy], Eqn. (12)). It is easy to see that the information obtained in a joint measurement of $%
A_{1}$ and $A_{2}$ on qubits $1+2$ is equivalent to the information one would obtain through a single Bell operator measurement. So we can equally well think of Alice as jointly measuring the observables $A_{j}$ ($j=1,2$) in her attempt to teleport the unknown state to Bob.
Now take $D$ to be the maximally entangled state $\left| \Phi
_{+}\right\rangle \left\langle \Phi _{+}\right| $, consider two alternative unknown states given by $$\begin{aligned}
\left| \phi _{1}\right\rangle &=&\frac{1}{\sqrt{2}}(\left| \uparrow
\right\rangle +\left| \downarrow \right\rangle ), \label{one} \\
\left| \phi _{2}\right\rangle &=&\frac{1}{\sqrt{2}}(\left| \uparrow
\right\rangle +i\left| \downarrow \right\rangle ),\; \label{two}\end{aligned}$$ and consider the following two alternative spin components of qubit $3$, $$\begin{aligned}
\sigma _{1} &=&e^{-i\pi /4}\left| \uparrow \right\rangle
\left\langle \downarrow \right| +e^{i\pi /4}\left| \downarrow
\right\rangle \left\langle \uparrow \right| , \label{three} \\
\sigma _{2} &=&e^{i\pi /4}\left| \uparrow \right\rangle \left\langle
\downarrow \right| +e^{-i\pi /4}\left| \downarrow \right\rangle
\left\langle \uparrow \right| . \label{four}\end{aligned}$$ (It is easily checked that the choices in (\[one\])–(\[four\]) correspond to the choices Żukowski ([@z-guy], p. 2) makes for the parameters he labels as $\beta ,\phi ,\beta',\phi'$.) Żukowski shows that the following Bell-CHSH-type inequality $$\left|
\begin{array}{c}
\left\langle A_{1}\otimes \sigma _{1}\right\rangle _{\left| \phi
_{1}\right\rangle \left\langle \phi _{1}\right| \otimes
D}+\left\langle A_{1}\otimes \sigma _{2}\right\rangle _{\left|
\phi _{1}\right\rangle \left\langle \phi _{1}\right| \otimes D}
\\
+\left\langle A_{2}\otimes \sigma _{1}\right\rangle _{\left| \phi
_{2}\right\rangle \left\langle \phi _{2}\right| \otimes
D}-\left\langle A_{2}\otimes \sigma _{2}\right\rangle _{\left|
\phi _{2}\right\rangle \left\langle \phi _{2}\right| \otimes D}
\end{array}
\right| \leq 2, \label{Z-expression}$$ which we henceforth for convenience call a “Bell teleportation inequality”, is violated by a factor of $\sqrt{2}$. Żukowski appears to interpret this violation as signifying that the quantum component of a perfect teleportation process involving the maximally entangled state $D=\left| \Phi
_{+}\right\rangle \left\langle \Phi _{+}\right| $ must be nonlocal.
This interpretation is not without plausibility. The idea is that to assess the nonlocality of $D$ *qua* the quantum component of Alice’s teleportation process, we should continue to have Alice measure the Bell operator but drop the classical channel, her phone call, from consideration. Although this frustrates any attempt to complete the process by Bob performing an appropriate unitary transformation on his qubit, it does not preclude asking whether what remains of the process can be modelled by a local hidden variables theory for any conceivable unknown state. One can imagine, then, that after preparing the quantum channel state, Clare randomly feeds to Alice one of two unknown states $\left| \phi
_{j}\right\rangle $, and makes a covert agreement with Bob that he should simply ignore all of Alice’s phone calls and randomly measure one of two spin components $\sigma _{j}$ on his qubit. After many measurement trials, Clare and Bob then reveal their charade to Alice, who grudgingly agrees to combine the information from her Bell operator outcomes with their information to see whether the Bell teleportation inequality (\[Z-expression\]) is violated. When they find that it *is* violated, the standard argument for Bell’s theorem (cf. [@z-guy], Eqns. (16),(17)) shows that no hidden variables model according to which: (i) Alice’s and Bob’s measurement outcomes are statistically independent at the level of the hidden variables; (ii) Alice’s outcomes are independent of the spin component Bob measures; and (iii) *Bob’s outcomes are independent of the unknown state Alice attempts to teleport*, can possibly account for the observed correlations.
Right after pointing out that inequality (\[Z-expression\]) is violated in the state $D=\left| \Phi
_{+}\right\rangle \left\langle \Phi _{+}\right| $, Żukowski remarks:
> It is an interesting fact, that needs further investigation, that the Bell inequality presented here is violated by the same factor $\sqrt{2}$ as the CHSH inequality for the usual Bell theorem involving a pair of particles in a maximally entangled state. This may imply that the quantum component of the teleportation process cannot be described in a local and realistic way, as long as the initial state of $B$ \[$2$\] and $C$ \[$3$\] admits no such model ([@z-guy], p. 3).
Żukowski’s conjecture appears to be that if the channel state violates [*any*]{} local hidden variables theory inequality, perhaps even collective measurement ones, then the correlations involved in the quantum component of teleportation will also violate a local hidden variables theory inequality. If this is indeed the conjecture, it would be difficult to decide, since we would need to take into account all possible local hidden variables theories inequalities for all possible measurement protocols on the channel particles alone and determine whether a violation of any such inequality suffices to prevent the recovery, within a local hidden variables model, of the correlations of the channel state that are exploited in the teleportation process. Consequently, we narrow our focus to the standard Bell-CHSH inequality for the channel state and a generalized form of (\[Z-expression\])’s Bell teleportation inequality for qubits 1–3 where the observables $A_{1}$ and $A_{2}$ are allowed to be *any* bivalent functions of Alice’s Bell operator.
Violation of a Bell teleportation inequality implies violation of a Bell-CHSH inequality by the channel state {#section_three}
=============================================================================================================
We first generalize Żukowski’s scheme to include a wider class of observables compatible with Alice’s Bell operator, by letting $A_{j}$ ($j=1,2$) be *any* pair of self-adjoint unitary ‘spin’ operators that are each bivalent functions of her Bell operator. Thus, $$\begin{aligned}
\nonumber
A_{j} & = & a_{j}\left| \Psi _{+}\right\rangle \left\langle \Psi
_{+}\right| +b_{j}\left| \Psi _{-}\right\rangle \left\langle
\Psi _{-}\right| \\ \label{define_A} & & +c_{j}\left| \Phi _{+}\right\rangle \left\langle \Phi
_{+}\right| +d_{j}\left| \Phi _{-}\right\rangle \left\langle
\Phi _{-}\right|,\ j=1,2,\end{aligned}$$ where $a_{j},b_{j},c_{j},d_{j}=\pm 1$. Now fix, once and for all, the channel state density operator $D$. By analogy with $\beta(D)$, define the number $\tau
(D)$ to be the maximum of the Bell teleportation expression $$\left|
\begin{array}{c}
\left\langle A_{1}\otimes \sigma _{1}\right\rangle _{\left| \phi
_{1}\right\rangle \left\langle \phi _{1}\right| \otimes
D}+\left\langle A_{1}\otimes \sigma _{2}\right\rangle _{\left|
\phi _{1}\right\rangle \left\langle \phi _{1}\right| \otimes D}
\\
+\left\langle A_{2}\otimes \sigma _{1}\right\rangle _{\left| \phi
_{2}\right\rangle \left\langle \phi _{2}\right| \otimes
D}-\left\langle A_{2}\otimes \sigma _{2}\right\rangle _{\left|
\phi _{2}\right\rangle \left\langle \phi _{2}\right| \otimes D}
\end{array}
\right| \label{eq:new}$$ over all choices of the unknown state vectors $\left| \phi
_{j}\right\rangle $ and all choices of the bivalent Bell operator functions $A_{j}$, i.e., all $\pm 1$ choices for the numbers $a_{j},b_{j},c_{j},d_{j}$.
Notice that, since there is a unitary operator $U$ on the state space of qubit $1$ mapping $\left| \phi_{2}\right\rangle$ to $\left|
\phi_{1}\right\rangle$, (\[eq:new\]) equals $$\label{gigo}
\left|\left\langle A_{1}\otimes(\sigma _{1}+
\sigma _{2})+
(UA_{2}U^{-1})\otimes(\sigma _{1}-\sigma _{2})
\right\rangle_{\left| \phi_{1}\right\rangle
\left\langle \phi_{1}\right|\otimes D}\right|,$$ which is just the absolute value of the expectation value of a Bell-CHSH operator on the bipartite system $(1+2)+3$. Thus, necessarily, $\tau (D)\in
[2,2\sqrt{2}]$. Note also that if the unknown states $\left|
\phi_{1}\right\rangle$ and $\left| \phi_{2}\right\rangle$ are compatible (i.e., either the same up to an irrelevant phase or orthogonal), then $U$ can always be chosen so that $UA_{2}U^{-1}$ is again a bivalent function of the Bell operator. This is most easily seen when $\left|
\phi_{1}\right\rangle=\left| \uparrow \right\rangle$ and $\left|
\phi_{2}\right\rangle=\left| \downarrow \right\rangle$, in which case the unitary that simply permutes these two $z$-eigenstates maps the four Bell operator eigenspaces into each other. (The general case can be reduced to this one by first applying a unitary transformation to the unknown states to bring them in line with the $z$-basis.) Thus, when the two alternative unknown states are compatible, $[A_{1},UA_{2}U^{-1}]=0$. In that case, it follows that expression (\[gigo\]), and thus (\[eq:new\]), cannot exceed 2, and the corresponding Bell teleportation inequality cannot be violated. This is to be expected, since it is known [@hi] that perfect teleportation of any collection of mutually compatible unknown states can be achieved with a separable channel state (all of whose correlations, including those involved in the teleportation, can be modelled by a local hidden variables theory).
It is not difficult to show that $\tau(D)=\beta
(D)=2\sqrt{2}$ if and only if $D$ is a maximally entangled pure state (by using the fact that any maximally entangled state is related to Żukowski’s choice $D=\left| \Phi
_{+}\right\rangle \left\langle \Phi _{+}\right| $ by a unitary operator on Bob’s qubit). We now concentrate on proving, quite generally, that $\beta (D)\geq \tau (D)$.
If $\tau (D)=2$ the claim is trivial, so we suppose $\tau
(D)>2$. Thus, there are unknown states $\left| \phi _{j}\right\rangle $ of qubit 1, bivalent Bell operator functions $A_{j}$, and spin components $\sigma
_{j}$ of qubit 3 such that $$\begin{aligned}
\nonumber
\left|
\begin{array}{c}
\left\langle A_{1}\otimes \sigma _{1}\right\rangle _{\left| \phi
_{1}\right\rangle \left\langle \phi _{1}\right| \otimes
D}+\left\langle A_{1}\otimes \sigma _{2}\right\rangle _{\left|
\phi _{1}\right\rangle \left\langle \phi _{1}\right| \otimes D}
\\
+\left\langle A_{2}\otimes \sigma _{1}\right\rangle _{\left| \phi
_{2}\right\rangle \left\langle \phi _{2}\right| \otimes
D}-\left\langle A_{2}\otimes \sigma _{2}\right\rangle _{\left|
\phi _{2}\right\rangle \left\langle \phi _{2}\right| \otimes D}
\end{array}
\right| \\ =\tau (D)>2. \label{mon}\end{aligned}$$ Consider the pair of self-adjoint operators $X_{j}$ on qubit 2’s state space defined via the following matrix elements in the $\sigma _{z}$ eigenbasis: $$\label{matrix}
\left(
\begin{array}{ll}
\left\langle \phi _{j}\right| \left\langle \uparrow \right|
A_{j}\left| \phi _{j}\right\rangle \left| \uparrow \right\rangle
& \left\langle \phi _{j}\right| \left\langle \uparrow \right|
A_{j}\left| \phi _{j}\right\rangle \left| \downarrow
\right\rangle \\
\left\langle \phi _{j}\right| \left\langle \downarrow \right|
A_{j}\left| \phi _{j}\right\rangle \left| \uparrow \right\rangle
& \left\langle \phi _{j}\right| \left\langle \uparrow \right|
A_{j}\left| \phi _{j}\right\rangle \left| \uparrow \right\rangle
\end{array}
\right) ,\ j=1,2. \label{friend}$$ Since each $A_{j}$ is self-adjoint, the matrices in (\[friend\]) are self-adjoint; thus each $X_{j}$ is self-adjoint as well. By linearity, (\[friend\]) implies: $$\left\langle v\right| X_{j}\left| w\right\rangle =\left\langle
\phi _{j}\right| \left\langle v\right| A_{j}\left|
\phi_{j}\right\rangle \left| w\right\rangle,\ j=1,2,
\label{googoogoo}$$ for all qubit 2 vectors $\left|
v\right\rangle ,\left| w\right\rangle$. In particular, when we set both $\left| v\right\rangle $ and $\left|
w\right\rangle $ equal to any unit eigenvector of $X_{j}$ with corresponding eigenvalue $\mu _{j}$, it follows that $\left| \mu _{j}\right|
\leq $ $\left\| A_{j}\right\| =1$. Thus each $X_{j}$ is a self-adjoint *contraction*, i.e., satisfies $\left\|
X_{j}\right\|
\leq 1$.
Next, multiplying both sides of (\[googoogoo\]) by $\left\langle
x\right| \sigma _{l}\left| y\right\rangle ,$ with $\left|
x\right\rangle ,\left| y\right\rangle$ arbitrary qubit 3 vectors, we obtain $$\begin{aligned}
\nonumber
& \left\langle v\right| \left\langle x\right| X_{j}\otimes \sigma
_{l}\left| w\right\rangle \left| y\right\rangle \\ = & \left\langle
\phi _{j}\right| \left\langle v\right| \left\langle x\right|
A_{j}\otimes \sigma _{l}\left| \phi _{j}\right\rangle \left|
w\right\rangle \left| y\right\rangle ,\ j,l=1,2. \label{yoyo}\end{aligned}$$ Now, since $D$ is Hermitian, $D=\sum_{k=1}^{4}\lambda _{k}\left|
e_{k}\right\rangle \left\langle e_{k}\right| $, where the $\{
e_{k}\}$ form an orthonormal basis for the 2+3 space. Since product vectors span this space, (\[yoyo\]) implies by linearity $$\begin{aligned}
\nonumber
\left\langle e_{k}\right| X_{j}\otimes \sigma _{l}\left|
e_{k}\right\rangle = \left\langle \phi _{j}\right| \left\langle
e_{k}\right| A_{j}\otimes \sigma _{l}\left| \phi
_{j}\right\rangle \left| e_{k}\right\rangle \; \\ k=1,...,4; j,l=1,2.
\label{googoo}\end{aligned}$$ Multiplying both sides of (\[googoo\]) by $\lambda _{k}$ and summing over $%
k$ results in: $$\left\langle X_{j}\otimes \sigma _{l}\right\rangle
_{D}=\left\langle A_{j}\otimes \sigma _{l}\right\rangle
_{\left| \phi _{j}\right\rangle \left\langle \phi _{j}\right|
\otimes D}\;j,l=1,2. \label{mouse}$$ Combining this with (\[mon\]) yields $$\left|
\begin{array}{c}
\left\langle X_{1}\otimes \sigma _{1}\right\rangle
_{D}+\left\langle X_{1}\otimes \sigma _{2}\right\rangle
_{D} \\
+\left\langle X_{2}\otimes \sigma _{1}\right\rangle
_{D}-\left\langle X_{2}\otimes \sigma _{2}\right\rangle
_{D}
\end{array}
\right| = \tau (D)>2. \label{yikes}$$ Finally note that, although the $X_{j}$ are in general only self-adjoint contractions, convexity arguments [@3man] show that if a bipartite state (here, our channel state $D$) violates a Bell-CHSH-type inequality with respect to self-adjoint contractions (as in (\[yikes\])), then that state also violates, *by at least the same amount*, a standard Bell-CHSH inequality with respect to observables [*all*]{} of which are self-adjoint unitary spin components. This completes the proof that $\beta
(D)\geq \tau (D)$.
An immediate corollary of this result is that there are channel states that permit nonclassical teleportation yet violate no Bell teleportation inequality. For example, consider the ‘Werner state’ [@werner_two] $$\label{w}
D_{W}=\bigl( 1-\frac{1}{\sqrt{2}} \bigr) \frac{1}{4} I \otimes I +
\frac{1}{\sqrt{2}} | \Psi_{-}\rangle \langle \Psi_{-}|.$$ Since spin operators, and hence Bell-CHSH operators, are traceless, $$\begin{aligned}
\nonumber
\beta(D_{W}) & = & \frac{1}{\sqrt{2}}\beta(| \Psi_{-}\rangle \langle
\Psi_{-}|) \\ & = & \frac{1}{\sqrt{2}}2\sqrt{2}=2\geq\tau(D_{W}),\end{aligned}$$ which forces $\tau (D_{W})=2$. On the other hand, the teleportation fidelity of $D_{W}$ relative to the standard choice of Bob’s unitary operators (which, in fact, gives the maximum achievable fidelity) is $$\begin{aligned}
\nonumber
{\cal F}_{st}(D_{W}) & = & (1-\frac{1}{\sqrt{2}}) \times \frac{1}{2}
+\frac{1}{\sqrt{2}} \times 1 \\ & = &
\frac{2}{3}(1+\frac{3\sqrt{2}-2}{8})>\frac{2}{3}.\end{aligned}$$
A channel state Bell-CHSH violation does not imply a violation of a Bell teleportation inequality. {#section_four}
==================================================================================================
We now exhibit a family of channel states $D$ for which the converse inequality $\tau
(D)\geq \beta (D)$ fails; indeed, for which $\beta (D)>2$ while $\tau (D)=2$. It follows that if one were to interpret [Ż]{}ukowski’s conjecture (discussed in Section \[section\_two\]) to be that a channel state Bell-CHSH violation implies a violation of his original inequality, then this particular conjecture is false.
By the reasoning in the previous section which led to Eqn. (\[yikes\]), we know that the maximum Bell teleportation violation given a channel state $D$ is attained by an expression of the form $|\langle Z\rangle_{D}|$ where $$\label{Z}
Z= X_{1}\otimes \sigma _{1}+
X_{1}\otimes \sigma_{2}+ X_{2}\otimes \sigma_{1}-
X_{2}\otimes \sigma_{2}, \label{yikes2}$$ and the self-adjoint contractions $X_{j}$ ($j=1,2$) have matrix elements given by (\[matrix\]). We may calculate these matrix elements explicitly by parameterizing the two alternative teleported states as $$\label{blob}
\left| \phi_{j}\right\rangle =\sin \theta _{j}\left| \uparrow
\right\rangle +\cos \theta _{j}e^{i\vartheta _{j}}\left| \downarrow
\right\rangle ,\;\theta _{j},\vartheta _{j}\in [0,2\pi ).$$ An elementary but tedious calculation then shows $$\begin{aligned}
\label{all_classes}
X_{j} &=&\frac{1}{4}(a_{j}+b_{j}+c_{j}+d_{j})I \nonumber \\
&&+\frac{1}{4}(a_{j}-b_{j}+c_{j}-d_{j})\sin 2\theta _{j}\cos \vartheta
_{j}\sigma _{x} \nonumber \\
&&+\frac{1}{4}(a_{j}+d_{j}-b_{j}-c_{j})\sin 2\theta _{j}\sin \vartheta
_{j}\sigma _{y} \\ \nonumber
&&+\frac{1}{4}(a_{j}+b_{j}-c_{j}-d_{j})\cos 2\theta _{j}\sigma _{z},\end{aligned}$$ where the numbers $a_{j},b_{j},c_{j},d_{j}=\pm 1$ correspond to the different choices that can be made for the bivalent functions $A_{j}$ of Alice’s Bell operator. The value of $\tau(D)$ is, therefore, given by $$\label{max}
\tau(D)=\max_{\theta_{j},\vartheta_{j}\in [0,2\pi
);a_{j},b_{j},c_{j},d_{j}=\pm 1;\sigma_{j}}
|\langle Z\rangle_{D}|.$$
Consider, now, the two parameter family of channel density operators given by $$D_{\lambda,\alpha}= (1-\lambda)\frac{1}{4} I\otimes I +\lambda
| \psi_{\alpha} \rangle \langle \psi_{\alpha} |,\
\lambda,\alpha\in [0,1],$$ where $$| \psi_{\alpha} \rangle = \alpha |+\rangle |+\rangle +
\sqrt{1-\alpha^{2}}|-\rangle
|-\rangle_,$$ $$| \pm \rangle = \frac{1 \pm \sqrt{3}}{\sqrt{2(3\pm\sqrt{3})}}
| \uparrow \rangle
+ \frac{1+i}{\sqrt{2(3\pm\sqrt{3})}}|\downarrow\rangle.$$ For later reference, note that $$\label{num1}
\langle\psi_{\alpha}|I\otimes
\sigma_{n}|\psi_{\alpha}\rangle=\frac{1}{\sqrt{3}}(2\alpha^{2}-1),\
n=x,y,z,$$ $$\begin{aligned}
\nonumber
\langle\psi_{\alpha}|\sigma_{m}\otimes
\sigma_{n}|\psi_{\alpha}\rangle=\frac{1}{3}(1+4\alpha\sqrt{1-\alpha^{2}}),
\\
(m,n)=(x,y),(y,x),(z,z), \label{num2}\end{aligned}$$ $$\begin{aligned}
\nonumber
\langle\psi_{\alpha}|\sigma_{m}\otimes
\sigma_{m}|\psi_{\alpha}\rangle & = &
\langle\psi_{\alpha}|\sigma_{m}\otimes
\sigma_{z}|\psi_{\alpha}\rangle \\ \label{num3}
& = &
\langle\psi_{\alpha}|\sigma_{z}\otimes
\sigma_{m}|\psi_{\alpha}\rangle \\ \nonumber & = & \frac{1}{3}
(1-2\alpha\sqrt{1-\alpha^{2}}),\
m=x,y.\end{aligned}$$ Since spin operators are traceless, the corresponding expectation values in the state $D_{\lambda,\alpha}$ are obtained by multiplying each of the above values by $\lambda$.
Again, because spin operators are traceless, the first maximally mixed component of $D_{\lambda,\alpha}$ does not contribute to any Bell-CHSH operator average; hence $\beta(D_{\lambda,\alpha})>2$ just in case $\lambda \beta(| \psi_{\alpha} \rangle \langle \psi_{\alpha} |) > 2$. It can be shown that [@popescu_rohrlich] $$\label{pop}
\beta(| \psi_{\alpha} \rangle \langle
\psi_{\alpha} |)=2\sqrt{1 + 4 \alpha^{2} (1-\alpha^{2})},$$ thus the condition for $\beta(D_{\lambda,\alpha})>2$ is simply $$\label{eq:bell}
\lambda \sqrt{1 + 4 \alpha^{2} (1-\alpha^{2})} > 1,\ \lambda,\alpha\in [0,1].$$ We now determine two additional conditions on $\lambda$ and $\alpha$, *consistent* with (\[eq:bell\]), which together are sufficient for $\tau(D_{\lambda,\alpha}) = 2$.
The set of 16 possible value assignments to the numbers $a_{j},b_{j},c_{j},d_{j}$ can be divided into three relevant classes. **Class I**: All assignments for which both $A_{1}$ and $A_{2}$ involve two value assignments of both signs (for example, the two value assignments $a_{j}=b_{j}=-c_{j}=-d_{j}=-1$, where $j=1,2$). **Class II**: All assignments for which exactly one of $A_{1}$ and $A_{2}$ involve two value assignments of both signs and the other involves three values assignments of one sign (for example, $a_{1}=b_{1}=-c_{1}=-d_{1}=-1$ and $a_{2}=b_{2}=c_{2}=-d_{2}=-1$). **Class III**: All assignments for which both $A_{1}$ and $A_{2}$ involve exactly three values assignments of one sign (for example, the value assignments $a_{j}=b_{j}=c_{j}=-d_{j}=-1$). Note that whenever $a_{j} = b_{j} = c_{j} = d_{j}$ for at least one value of $j$—i.e., whenever either $A_{1}$ or $A_{2}$ is multiple of the identity—no violation is possible, thus such assignments can be ignored.
**Class I**: Following [@popescu_rohrlich], we introduce the new spin operators $$\label{new_ops}
{\hat \sigma}_{1}= \frac{\sigma_{1}
+ \sigma_{2}}{2 \cos \xi},\
{\hat \sigma}_{2}= \frac{\sigma_{1} - \sigma_{2}}{2 \sin \xi},$$ so that (\[yikes2\]) becomes $$Z=2 (X_{1} \otimes {\hat \sigma}_{1})\cos \xi +
2 (X_{2} \otimes {\hat \sigma}_{2}) \sin \xi.$$ Using the fact that, whenever $M,N\in \mathbb{R}$, $M \cos \xi + N \sin \xi \leq \sqrt{M^{2} + N^{2}}$, we have $$\label{zee}
|\langle Z \rangle_{D_{\lambda,\alpha}}|
\leq 2 \sqrt{\langle X_{1} \otimes {\hat \sigma}_{1}
\rangle_{D_{\lambda,\alpha}}^{2}
+ \langle X_{2} \otimes {\hat \sigma}_{2} \rangle_{D_{\lambda,\alpha}}^{2}}.$$ Using similar reasoning, $$\langle X_{j} \otimes {\hat \sigma}_{j}\rangle_{D_{\lambda,\alpha}}
\leq \sqrt{ \sum_{n=x,y,z}\langle X_{j}\otimes \sigma_{n}
\rangle_{D_{\lambda,\alpha}}
^{2}} ,\ j=1,2,$$ therefore $$\label{temp}
|\langle Z\rangle_{D_{\lambda,\alpha}}| \leq 2\sqrt{ \sum_{n=x,y,z;j=1,2}\langle X_{j}\otimes \sigma_{n}
\rangle_{D_{\lambda,\alpha}}
^{2}}.$$ Assuming a **Class I** value assignment to the numbers $a_{j},b_{j},c_{j},d_{j}$, each self-adjoint contraction in (\[all\_classes\]) has the form $X_{j} = r_{j}\sigma_{m_{j}}$ where $r_{j}\in[-1,1]$ and $m_{j}=x,y,z$. Since $|\langle Z\rangle_{D_{\lambda,\alpha}}|$ cannot exceed the classical bound $2$ when $[X_{1},X_{2}]=0$, we need only consider the three cases where $m_{1}\not=m_{2}$, i.e., $$\begin{aligned}
\label{in}
X_{1}=r_{1} \sigma_{x}, & X_{2}=r_{2} \sigma_{y}, \\
X_{1}=r_{1} \sigma_{y}, & X_{2}=r_{2} \sigma_{z}, \\
X_{1}=r_{1} \sigma_{x}, & X_{2}=r_{2} \sigma_{z}, \end{aligned}$$ and the other three cases obtained by interchanging 1 and 2. For the first case, substitution into inequality (\[temp\]) and using the expectation values given in (\[num2\]) and (\[num3\]), we find $$\begin{aligned}
\nonumber
|\langle Z \rangle _{D_{\lambda,\alpha}} | & \leq &
\frac{2 \lambda}{3} \sqrt{(r_{1}^{2} + r_{2}^{2})
( 3+24\alpha^{2} (1-\alpha^{2}))} \\ \label{ineq}
& \leq & \frac{2 \sqrt{2} \lambda}{\sqrt{3}} \sqrt{1+8\alpha^{2} (1-\alpha^{2})}.\end{aligned}$$ The other two cases in (\[in\]) yield the same inequality (\[ineq\]), as do the cases obtained by interchanging 1 and 2 (which leaves the right-hand side of inequality (\[temp\]) unchanged). Thus, a sufficient condition for no **Class I** value assignment to yield a violation of a Bell teleportation inequality is $$\label{class_one}
\lambda \sqrt{2/3\bigl( 1 + 8\alpha^{2} (1-\alpha^{2})\bigr)} \leq 1.$$
**Class II**: In this case, we can see from (\[all\_classes\]) that exactly one of $X_{1}$ and $X_{2}$ takes the form $r \sigma_{m}$, where $m=x,y,z$ and $r\in[-1,1]$, while the other is of the form $\pm I/2 + \sigma/2$ where $\sigma$ is a spin-1/2 operator. Thus, $$\begin{aligned}
\label{third_case}
Z & = & r \sigma_{m} \otimes
\bigl( \sigma_{2} + {\bar \sigma}_{2} \bigr) +
1/2(\sigma \pm I)\otimes \bigl( \sigma_{1}- {\bar \sigma}_{2} \bigr) \\
\nonumber
& = & \pm I \otimes (\sigma_{1} - {\bar \sigma}_{2})/2
+ r \sigma_{m} \otimes (\sigma_{1} + {\bar \sigma}_{2})/2 + B'/2,\end{aligned}$$ where ${\bar \sigma}_{2}= \pm \sigma_{2}$ and $B'$ is a Bell-CHSH-type operator constructed out of self-adjoint contractions that are not necessarily unitary (unless $r=\pm 1$). Transforming the spin operators $(\sigma_{1},{\bar \sigma}_{2})\mapsto
(\hat{\sigma}_{1},\hat{\sigma}_{2})$ via Eqns. (\[new\_ops\]), and substituting into Eqn. (\[third\_case\]) yields $$\label{not}
Z= \pm (I \otimes {\hat \sigma}_{2})
\sin \xi
+ r (\sigma_{m} \otimes {\hat \sigma}_{1})
\cos \xi
+ B'/2.$$ Therefore, $$\begin{aligned}
\nonumber
& | \langle Z \rangle_{D_{\lambda,\alpha}} | \\ \nonumber
\leq & | \pm \langle I \otimes {\hat \sigma}_{2}\rangle_{D_{\lambda,\alpha}}
\sin \xi \\ \nonumber & +
r \langle \sigma_{m} \otimes {\hat \sigma}_{1} \rangle_{D_{\lambda,\alpha}}
\cos \xi |
+ 1/2 | \langle B' \rangle_{D_{\lambda,\alpha}} | \\ \label{partial}
\leq & \sqrt{ \langle I \otimes {\hat \sigma}_{2}
\rangle_{D_{\lambda,\alpha}}^{2} +
r^{2} \langle \sigma_{m} \otimes \hat{\sigma_{1}}
\rangle_{D_{\lambda,\alpha}}^{2} } \\ \nonumber &
+ 1/2 | \langle B' \rangle_{D_{\lambda,\alpha}} |. \end{aligned}$$ >From (\[pop\]) it follows that $$| \langle B' \rangle_{D_{\lambda,\alpha}} | \leq
2\lambda \sqrt{1 + 4 \alpha^{2} (1-\alpha^{2})}$$ and, clearly, $$r^{2}\langle \sigma_{m} \otimes {\hat \sigma}_{1} \rangle_{D_{\lambda,\alpha}}^{2}
= \lambda^{2} r^{2}\langle \sigma_{m} \otimes {\hat \sigma}_{1}
\rangle_{| \psi_{\alpha}\rangle\langle \psi_{\alpha}| }
\leq \lambda^{2}.$$ Moreover, $$\begin{aligned}
\nonumber
& \langle I \otimes \hat{\sigma}_{2}\rangle^{2}_{D_{\lambda,\alpha}}
\\ \leq
& \langle I \otimes \sigma_{x} \rangle_{D_{\lambda,\alpha}}^{2}
+ \langle I \otimes \sigma_{y}
\rangle_{D_{\lambda,\alpha}}^{2}
+ \langle I \otimes \sigma_{z} \rangle_{D_{\lambda,\alpha}}^{2}
\\ \nonumber = &\lambda^{2}
(2 \alpha^{2} - 1)^{2},\end{aligned}$$ using Eqn. (\[num1\]) for the last equality. Upon substitution of these last three inequalities back into inequality (\[partial\]), we obtain $$\begin{aligned}
\nonumber
|\langle Z \rangle_{D_{\lambda,\alpha}}| & \leq &
\lambda \bigl(\sqrt{2(1-2\alpha^{2} (1-\alpha^{2}))} \\ & & +
\sqrt{1 + 4\alpha^{2} (1-\alpha^{2})} \bigr) .\end{aligned}$$ Thus, a sufficient condition for no **Class II** value assignment to yield a violation of a Bell teleportation inequality is $$\label{class_three}
\lambda \bigl(\sqrt{2(1-2\alpha^{2} (1-\alpha^{2}))} +
\sqrt{1 + 4\alpha^{2} (1-\alpha^{2})} \bigr)\leq 2.$$
**Class III**: In this case, $Z$ takes the form $I \otimes \sigma + B/2$ where $\sigma$ is a spin-operator and $B$ a standard Bell-CHSH operator. Note that this is just a special case of the **Class II** expression for $Z$ in Eqn. (\[not\]), taking the $\pm$ sign to be $+$, $\hat{\sigma}_{2}=\sigma$, $\xi=\pi/2$, and $B' = B$. Thus inequality (\[class\_three\]) is also sufficient for no **Class III** value assignment to yield a violation of a Bell teleportation inequality.
Combining our results for each class, we see that $\beta (D_{\lambda,\alpha})>2$ while $\tau (D_{\lambda,\alpha})=2$ if and only if $\lambda$ and $\alpha$ satisfy conditions (\[eq:bell\]), (\[class\_one\]) and (\[class\_three\]). The family of channel density operators satisfying these conditions is indeed nonempty. For example, the conditions are satisfied by taking $\lambda = \sqrt{3/5}\approx .77$ and $\alpha = \sqrt{3}/2\approx .87$, and in that case we have $\beta(D_{\sqrt{3/5},\sqrt{3}/2})=2\sqrt{21/20}\approx 2.05$. Note, also, that choosing $\lambda=1$ does *not* satisfy the conditions, nor does choosing $\alpha^{2}
(1-\alpha^{2})=1/4$. Thus no channel density operator in the family is pure, and the pure entangled component $| \psi_{\alpha} \rangle \langle \psi_{\alpha} |$ of $D_{\lambda,\alpha}$ is never maximally entangled. Finally, since $\beta (D)>2$ implies ${\cal F}_{max}(D)>2/3$ [@horodeckis], each density operator in the family supplies another example of a state that permits nonclassical teleportation without violating any Bell teleportation inequality.
Fidelity values that imply a violation of a Bell teleportation inequality. {#section_five}
==========================================================================
In light of our previous results, it is natural to ask what values of the fidelity are sufficient for the quantum correlations in the standard teleportation protocol *not* to admit a local hidden variables explanation. We shall show in this section that ${\cal F}_{st}(D)>2/3(1+1/2\sqrt{2})\approx .90$ implies $\tau (D)>2$.
It is known [@horodecki] that for any Bell-CHSH operator $B$, $$\langle B \rangle_{D} = \mathbf{a}\cdot T(D)( \mathbf{b}+\mathbf{b'} )
+ \mathbf{a'}\cdot T(D)( \mathbf{b}-\mathbf{b'} )$$ where $\mathbf{a}, \mathbf{a'}, \mathbf{b}$ and $\mathbf{b'}$ are unit vectors in $\mbox{$\mathbb{R}$}^{3}$ defining the four spin operators that occur in $B$, and the $3 \times 3$ matrix $T(D)$ has components $$T_{mn}(D)=\mbox{Tr}[D(\sigma_{m}\otimes\sigma_{n})],\ m,n=x,y,z,$$ encoding the inter-particle correlations in state $D$. Setting $\mathbf{a}=\mathbf{x}$, $\mathbf{a'}=\mathbf{y}$, $\mathbf{b}=(\mathbf{x} + \mathbf{y})/\sqrt{2}$, and $\mathbf{b'}=(\mathbf{x} - \mathbf{y})/\sqrt{2}$, we obtain $$\begin{aligned}
\nonumber
\langle B \rangle_{D} & = &
\langle \sigma_{x}\otimes\sigma_{b}+\sigma_{x}\otimes\sigma_{b'}+
\sigma_{y}\otimes\sigma_{b}-\sigma_{y}\otimes\sigma_{b'} \rangle_{D}
\\ \label{T_bell}
& = & \sqrt{2} (T_{xx}(D) + T_{yy}(D)).\end{aligned}$$ Now choose the operators $A_{j}$ in Eqn. (\[define\_A\]) that correspond to $a_{1}=c_{1}=-b_{1}=-d_{1}=+1$ and $a_{2} = d_{2} = -b_{2} = -c_{2} = +1$, and the unknown states $\left| \phi_{j}\right\rangle$ in (\[blob\]) that correspond to $\theta_{1}=\pi/4$, $\vartheta_{1}=0$, $\theta_{2}=\pi/4$, and $\vartheta_{2}=\pi/4$. Substituting these values into (\[all\_classes\]), we obtain $X_{1}=\sigma_{x}$ and $X_{2}=\sigma_{y}$. Hence the Bell-CHSH operator $B$ in (\[T\_bell\]) has the form of a $Z$ operator as given in (\[Z\]), and $|\langle B \rangle_{D}|$ provides a lower bound for $\tau(D)$, i.e., $$\label{bog}
\tau(D)\geq \sqrt{2} |T_{xx}(D) + T_{yy}(D)|.$$ (As check on the correctness of this bound, note from (\[num3\]), that it falls below 2 when $D=D_{.77,.87}$, as it must.)
It can be shown quite generally that [@horodeckis] $$\label{fidelity}
{\cal F}_{st}(D) = \frac{1}{8} \sum_{n=1}^{4} (1 + \frac{1}{3}
\mbox{Tr}[T_{n}^{\dag} T(D) O_{n}])$$ where $$\begin{aligned}
T_{1} = \mbox{diag}(-1,-1,-1), &
T_{2} = \mbox{diag}(-1,1,1), \\ T_{3} = \mbox{diag}(1,-1,1), &
T_{4} = \mbox{diag}(1,1,-1),\end{aligned}$$ and the matrices $\{O_{n}\}$ are determined by the standard choice for Bob’s unitary operations in (\[standard\]) via the requirement that $U_{n}(\mathbf{b}\cdot\mathbf{\sigma})U_{n}^{-1}=(O_{n}^{\dag}\mathbf{b})\cdot
\mathbf{\sigma}$ for all $\mathbf{b}\in \mbox{$\mathbb{R}$}^{3}$. Calculating out these latter matrices explicitly yields $$\begin{aligned}
O_{1}=\mbox{diag}(1,1,1), & O_{2}=\mbox{diag}(1,-1,-1), \\
O_{3}=\mbox{diag}(-1,1,-1), & O_{4}=\mbox{diag}(-1,-1,1).\end{aligned}$$ (For example, for $n=2$, $U_{2}=\sigma_{x}$, and, using the fact that $\sigma_{x}^{2}=I$ and orthogonal spin components anti-commute, $$\begin{aligned}
\sigma_{x}(\mathbf{b}\cdot\mathbf{\sigma})\sigma_{x}
& = &
\sigma_{x}(b_{x}\sigma_{x}+b_{y}\sigma_{y}+b_{z}\sigma_{z})\sigma_{x} \\
& = & b_{x}\sigma_{x}-b_{y}\sigma_{y}-b_{z}\sigma_{z} \\ & = &
(O_{2}^{\dag}\mathbf{b})_{x}\sigma_{x}+(O_{2}^{\dag}\mathbf{b})_{y}\sigma_{y}+
(O_{2}^{\dag}\mathbf{b})_{z}\sigma_{z},\end{aligned}$$ which implies $O_{2}=\mbox{diag}(1,-1,-1)$.) Plugging all the matrices $\{T_{n}\}$ and $\{O_{n}\}$ back into (\[fidelity\]) gives $$\label{gog}
{\cal F}_{st}(D) =1/2 -1/6(T_{xx}(D) + T_{yy}(D) + T_{zz}(D)).$$
Finally, combining (\[bog\]) with Eqn. (\[gog\]), we have $$\tau(D)\geq \sqrt{2} |T_{zz}(D) + 6{\cal F}_{st}(D)-3|.$$ Assuming ${\cal F}_{st}(D)>2/3(1+1/2\sqrt{2})=2/3+\sqrt{2}/6$, and noting that $T_{zz}(D)\geq -1$, it follows that $$\begin{aligned}
\nonumber
& \sqrt{2}(T_{zz}(D) + 6{\cal F}_{st}(D)-3) \\ > & \sqrt{2}(-1+4+\sqrt{2}-3)=2,\end{aligned}$$ and therefore $\tau(D) > 2$, as claimed. Note that we have made no attempt to find the minimum value for ${\cal F}_{st}(D)$ that implies $\tau(D) > 2$. However, for the Werner state in (\[w\]), we know that ${\cal F}_{st}(D_{W})=
2/3(1+(3\sqrt{2}-2)/8)\approx .85$, yet $\tau(D_{W})=2$; so our bound of $.90$ cannot be decreased below $.85$.
Discussion {#section_six}
==========
We have compared a new class of “Bell teleportation inequalities”—our generalization of [Ż]{}ukowski’s inequality—to the well-known class of Bell-CHSH inequalities for the particle pair that makes up the quantum channel of a standard teleportation process. We found that while a Bell teleportation inequality cannot be violated unless the channel state already violates a Bell-CHSH inequality, the latter is no guarantee that the correlations of the channel state actually involved in the teleportation process will violate a Bell teleportation inequality. This suggests that it is generally *easier* for a local hidden variables theory to simulate teleportation than to simulate the results of a standard Bell correlation experiment on the channel particles alone. Moreover, if one were to interpret [Ż]{}ukowski’s conjecture (discussed in Section \[section\_two\]) to be that a channel state Bell-CHSH violation implies a violation of his original inequality, then we have shown this particular conjecture to be false.
Our results also contribute to the larger debate (discussed in our introduction) over the role that nonlocality plays in teleportation. We showed that a Bell teleportation inequality is always violated when the fidelity of transmission in the standard teleportation protocol exceeds the classical limit of $2/3$ by a factor of $1+1/2\sqrt{2}$. If one accepts that the violation of an inequality to which any local hidden variables theory is committed implies the presence of nonlocality, it follows that standard teleportation with a fidelity exceeding $2/3(1+1/2\sqrt{2})\approx .90$ cannot occur without the involvement of nonlocality. On the other hand, our results suggest the existence of a range $(.67,.90)$ of *nonclassical* values of the fidelity for which a local hidden variables theory of the teleportation process may be possible. Thus, the ability of an entangled channel state, such as $D_{W}$ or $D_{.77,.87}$, to permit nonclassical teleportation cannot by itself suffice for concluding that the channel state itself, or the teleportation it facilitates, is nonlocal; attention must also be paid to the magnitude of the fidelity achievable using the state.
Our tentative conclusion that local hidden variables models of quantum teleportation may exist for nonclassical fidelities up to $\approx 0.90$ is compatible with Gisin’s [@gisin] demonstration that the *end results* of quantum teleportation can be classically simulated up to a fidelity of $\approx 0.87$. However, this comparison should not be given too much weight because Gisin’s simulation has little to do with what actually goes on in the standard quantum teleportation protocol. In his simulation, Alice does not measure any Bell operator on qubits 1+2. Rather, it is assumed that she has full knowledge of the ‘unknown’ state’s expansion coefficients, and her task is simply to classically communicate to Bob as much information as she can about these coefficients by using only 2 bits. While it is certainly of interest to study how well teleportation can be achieved classically using various protocols different from the standard quantum protocol (see also [@gisin2]), we have confined ourselves in this paper to the possibility of local classical explanations of the latter only. While we have reached a negative conclusion for fidelities exceeding $\approx .90$, our results strongly suggest a positive answer for fidelities in the range $(.67,.90)$.
**Acknowledgements**
DTP acknowledges the support of a Graduate School Research Travelling Award from the University of Queensland and thanks the Department of Philosophy at the University of Pittsburgh for its hospitality. We would also like to thank Peter Drummond, Gerard Milburn, and William Munro for helpful discussion.
[99]{}
S. Popescu, D. Rohrlich, Phys. Lett. A 166 (1992) 293.
C. H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W. K. Wooters, Phys. Rev. Lett. 70 (1993) 1895.
S. L. Braunstein, A. Mann, M. Revzen, Phys. Rev. Lett. 68 (1992) 3259.
R. F. Werner, `quant-ph/0003070`.
S. Popescu, Phys. Rev. Lett. 72 (1994) 797.
N. Gisin, Phys. Lett. A 210 (1996) 157.
R. Horodecki, M. Horodecki, and P. Horodecki, ` quant-ph/9606027.`
S. Massar and S. Popescu, Phys. Rev. Lett. 74 (1995) 1259.
L. Vaidman, `quant-ph/9810089.`
J. Bub, Stud. Hist. Phil. Mod. Phys. 31 (2000) 74.
O. Maroney and B. J. Hiley, Found. Phys. 29 (1999) 1403.
L. Hardy, `quant-ph/9906123.`
D. Deutsch and P. Hayden, `quant-ph/9906007.`
A. Duwell, “Explaining Information Transfer in Quantum Teleportation”, Phil. Sci. (2001) forthcoming.
M. Żukowski, Phys. Rev. A 62 (2000) 032101.
S. Ghosh, G. Kar, A. Roy, and U. Sen, `quant-ph/0010012`.
R. Clifton, H. Halvorson, and A. Kent, Phys. Rev. A 61 (2000) 042101.
R. F Werner, Phys. Rev. A, 40 (1989) 4277.
R. Horodecki, P. Horodecki, and M. Horodecki, Phys. Lett. A 200 (1995) 340.
N. Cerf, N. Gisin, and S. Massar, Phys. Rev. Lett., 84 (2000) 2521.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Unbiased Learning-to-Rank framework [@Joachims:2017:ULB:3018661.3018699] has been recently introduced as a general approach to systematically remove biases, such as position bias, from learning-to-rank models. The method takes two steps - estimating click propensities and using them to train unbiased models. Most common methods proposed in the literature for estimating propensities involve some degree of intervention in the live search engine. An alternative approach proposed recently uses an Expectation Maximization (EM) algorithm to estimate propensities by using ranking features for estimating relevances [@Wang:2018:PBE:3159652.3159732]. In this work we propose a novel method to estimate propensities which does not use any intervention in live search or rely on any ranking features. Rather, we take advantage of the fact that the same query-document pair may naturally change ranks over time. This typically occurs for eCommerce search because of change of popularity of items[^1] over time, existence of time dependent ranking features, or addition or removal of items to the index (an item getting sold or a new item being listed). However, our method is general and can be applied to any search engine for which the rank of the same document may naturally change over time for the same query. We derive a simple likelihood function that depends on propensities only, and by maximizing the likelihood we are able to get estimates of the propensities. We apply this method to eBay search data to estimate click propensities for web and mobile search. We also use simulated data to show that the method gives reliable estimates of the “true” simulated propensities. Finally, we train a simple unbiased learning-to-rank model for eBay search using the estimated propensities and show that it outperforms the baseline model (which does not correct for position bias) on our offline evaluation metrics.'
author:
- Grigor Aslanyan
- Utkarsh Porwal
bibliography:
- 'bibliography.bib'
title: 'Direct Estimation of Position Bias for Unbiased Learning-to-Rank without Intervention'
---
Introduction
============
Modern search engines rely on machine learned methods for ranking the matching results for a given query. Training and evaluation of models for ranking is commonly known as **Learning-to-Rank (LTR)** [@Hang2011ASI]. There are two common approaches for collecting the data for LTR - **human judgements** and **implicit user feedback**. For human judgements samples of documents are gathered for a sample of queries and sent to human judges who analyze and label each document. The labels can be as simple as *relevant* vs. *not relevant* or can involve more levels of relevance. This labeled data is then used for training and/or evaluation of LTR models. Collecting human judged data can be expensive and time consuming and often infeasible. On the other hand, data from implicit user feedback, such as clicks, is essentially free and abundant. For that reason it is often the preferred method for collecting data for LTR. A major drawback of this method is that the data can be heavily biased. For example, users can only click on documents that have been shown to them (presentation bias) and are more likely to click on higher ranked documents (position bias). A lot of work in the LTR literature has focused on accounting for and removing these biases. In particular, the recent paper by Joachims et. al. [@Joachims:2017:ULB:3018661.3018699] has proposed a framework for systematically removing the biases from user feedback data. Following the title of the paper we will refer to this framework as **Unbiased Learning-to-Rank**. In particular, the authors have focused on removing the position bias by first estimating the click propensities and then using the inverse propensities as weights in the loss function. They have shown that this method results in an unbiased loss function and hence an unbiased model.
Unbiased Learning-to-Rank is a very appealing method for removing the inherent biases. However, to be able to successfully apply it one needs to first get a reliable estimate of click propensities. The method proposed in [@Joachims:2017:ULB:3018661.3018699] uses result randomization in the live search engine to estimate click propensities. Result randomization can negatively impact the quality of the search results, which will in turn result in poor user experience and potential loss of revenue for the company [@Wang:2018:PBE:3159652.3159732]. It also adds bookkeeping overhead.
Wang et. al. [@Wang:2018:PBE:3159652.3159732] have proposed an alternative method for estimating click propensities without result randomization. The method uses a regression-based Expectation Maximization (EM) algorithm to estimate the propensities. This method uses the ranking features to estimate relevances and has shown promising results for personal search. The only drawback of this method is the use of the ranking features for estimating the relevances. In the absence of good features this algorithm may not be able to produce good estimates of the propensity.
In this paper we propose a novel method for estimating click propensities without any intervention in the live search results page, such as result randomization. We use query-document pairs that appear more than once at different ranks to estimate click propensities. Note that each query-document pair does not have to appear a large number of times - it can be as few as twice. In comparison to the EM based algorithm in [@Wang:2018:PBE:3159652.3159732] our method does not rely on any ranking features to estimate the relevances. In fact, we completely eliminate the relevances from the likelihood function and directly estimate the propensities by maximizing a simple likelihood function. The main novelty of our work is that our method does not use any interventions in the live search engine and does not rely on indirect estimates of relevance through ranking features. Here we focus on eCommerce search as an example since it is natural for ranking to change over time for eCommerce platforms due to change of popularity of items over time, existence of time-dependent ranking features, and addition and removal of items to the index (items being sold or new items being listed). However, the method developed here is general and can be applied to any platform for which ranking can naturally change over time.
Our method can be further generalized to drop the requirement of the ranking to naturally change over time, which will make it applicable to practically any search engine. Our likelihood function can be used in a Bayesian setting for query-document pairs which did not necessarily appear at multiple different ranks. With appropriate prior functions one can use a Bayesian approach (for example, through Gibbs sampling) to estimate posteriors for the propensities. We plan to work on this Bayesian approach in a future study.
We use simulated data to test our method and get good results. We then apply our method on actual data from eBay search logs to estimate click propensities for both web and mobile search. Finally, we use our estimated propensities to train a simple unbiased learning-to-rank model for eBay search and compare it with a baseline “biased” model which does not account for position bias. Our results show that the unbiased model with our estimated propensities significantly outperforms the baseline on our offline evaluation metrics.
The rest of the paper is organized as follows. In Section \[section-related-work\] we discuss some of the related work in the literature. In Section \[section-unbiased-ltr\] we give a brief summary of the Unbiased Learning-to-Rank framework. In Section \[section-method\] we introduce our method for estimating click propensities. In Section \[sec-ebay\] we apply our method to eBay search logs and estimate propensities for web and mobile search. In Section \[section-sim\] we apply our method to simulated data and show that we are able to obtain very good estimates of the “true” simulated propensities. In Section \[section-models\] we train and evaluate simple proof-of-concept learning-to-rank models for eBay search using our estimated propensities and show that the unbiased model outperforms the baseline. Finally, we summarize our work in Section \[section-summary\] and discuss possible future directions for this research.
Related Work {#section-related-work}
============
Implicit feedback such as clicks are commonly used to train user facing machine learned systems such as ranking or recommender systems. Clicks are preferred over human judged labels as they are available in plenty, are available readily and are collected in a natural environment. However, such user behavior data can only be collected over the items shown to the users. This injects a presentation bias in the collected data. This affects the machine learned systems as they are trained on user feedback data as positives and negatives. It is not feasible to present many choices to the user and it affects the performance of these systems as we can not get an accurate estimate of positives and negatives for training with feedback available only on selective samples. This situation is aggravated by the fact that the feedback of the user not only depends on the presentation, it also depends on where the item was presented. This is a subclass of the presentation bias called position bias. Joachims et al. [@Joachims:2017:ULB:3018661.3018699] proved that if the collected user behavior data discounts the position bias accurately then the learned system will be the same as the one learned on true relevance signals.
Several approaches have been proposed to de-bias the collected user behavior data. One of the most common approaches is the use of click models. Click models are used to make hypotheses about the user behavior and then the true relevance is estimated by optimizing the likelihood of the collected clicks. There are several types of click models. One such model is a random click model (RCM) [@dupret2008user] where it is assumed that every document has the same probability of getting clicked and that probability is the model parameter. In a rank based click through rate model (RCTR) it is assumed that the probability of every document being clicked depends on its rank. Therefore, the total number of model parameters is the total number of ranks in the ranking system. Another model is the document based CTR model (DCTR) [@craswell2008experimental] where the click through rates are estimated for each query-document pair. In this model the total number of model parameters is the total number of query-document pairs. This model is prone to overfitting as the number of parameters grows with the training data size. Most commonly used click models are the position based model (PBM) [@craswell2008experimental; @Joachims2005] and the cascade model (CM) [@craswell2008experimental]. In PBM the hypothesis is that a document is only clicked if it is observed and the user found it attractive or relevant. In CM the hypothesis is that the user sequentially scans the whole document top to bottom and clicks when the document is found to be relevant. In this model the top document is always observed and consecutive documents are only observed if the previous ones were observed and were not deemed relevant. In our proposed method we make a similar hypothesis such as the position based method where the observation probability depends on the rank and the probability of relevance only depends on the query-document pair. However, our approach is to learn the click propensities instead of learning the true relevance by optimizing the likelihood of the collected clicks. More advanced click models, such as the user browsing model (UBM) [@dupret2008user], the dependent click model (DCM) [@guo2009efficient], the click chain model (CCM) [@guo2009click], and the dynamic Bayesian network model (DBN) [@chapelle2009dynamic] are also proposed. Chuklin et al. [@chuklin2015click] provides a comprehensive overview of click models.
Click models are trained on the collected user behavior data. Interleaving is another option that is deployed at the time of data collection. In interleaving different rank lists can be interleaved together and presented to the user. By comparing the clicks on the swapped results one can learn the unbiased user preference. Different methods for interleaving have been proposed. In the balanced interleave method [@Joachims03evaluatingretrieval] a new interleaved ranked list is generated for every query. The document constraint method [@he2009evaluation] accounts for the relation between documents. Hofmann et al. [@hofmann2011probabilistic] proposed a probabilistic interleaving method that addressed some of the drawbacks of the balanced interleave method and the document constraint method. One limitation of the interleaving method is that often the experimentation platform in eCommerce companies is not tied to just search. It supports A/B testing for all teams, such as checkout and advertisements. Therefore, the interleaving ranked list may not be supported as it is pertinent only for search ranking.
A more recent approach to address presentation bias is the unbiased learning-to-rank approach. In this click propensities are estimated and then the inverse propensities are used as weights in the loss function. Click propensities are estimated by presenting the same items at different ranks to account for click biases without explicitly estimating the query-document relevance. Click propensity estimation can either be done randomly or in a more principled manner. Radlinski et al. [@radlinski2006minimally] presented the FairPairs algorithm that randomly flips pairs of results in the ranking presented to the user. They called it randomization with minimal invasion. Carterette et al. [@Carterette2018] also presented a minimally invasive algorithm for offline evaluation. Joachims et al. [@Joachims:2017:ULB:3018661.3018699] proposed randomized intervention to estimate the propensity model. Radlinski et al. [@radlinski2008learning], on the other hand, proposed alteration in ranking in a more informed manner using Multi-Armed Bandits. The main drawback of randomization for propensity estimation is that it can cause bad user experience, book keeping overhead, and a potential loss in revenue. Wang et al. [@Wang:2018:PBE:3159652.3159732] proposed a method to estimate propensities without randomization using the EM algorithm. In most of the existing methods, propensity estimation is done first. Once the propensities are learned, an unbiased ranker is trained using the learned propensities. Recently Ai et al. [@ai2018unbiased] proposed a dual learning algorithm that learns an unbiased ranker and the propensities together.
Unbiased Learning-to-Rank {#section-unbiased-ltr}
=========================
In this section we introduce our notation and give a summary of the unbiased learning to rank methodology from [@Joachims:2017:ULB:3018661.3018699]. Note that our notation differs from that of [@Joachims:2017:ULB:3018661.3018699].
First, let us assume that we have unbiased data. The data consists of a sample $\mathbf{Q}$ of i.i.d. (independent and identically distributed) queries. Each query $q_i\in\mathbf{Q}$ comes with a set of documents $\mathbf{Y}_i$ with their known relevances. We denote the documents in $\mathbf{Y}_i$ by $y_{ij}$ and their corresponding relevances by $r_{ij}$. We would like to train a LTR model which is a function $F(q, y)$ that computes a score for a given query $q$ and document $y$. These scores can then be used to rank the documents (higher scored documents will be ranked higher). The model is trained to minimize a *loss function* for the training data. The loss function will have smaller values for ranking functions which give higher ranks to relevant documents and lower ranks to irrelevant documents in the training data. There are multiple approaches to choosing a model $F$ and a loss function.
There are three main classes of LTR models - **pointwise**, **pairwise**, and **listwise** [@Hang2011ASI; @Cao:2007:LRP:1273496.1273513]. Pointwise LTR models use document level loss functions, regardless of the query. This means that the model will try to produce scores for each document that is as close to its relevance as possible. Pairwise models, on the other hand, use pairs of documents with desired orders. For example, pairs of documents could contain one relevant and one irrelevant document for the query. The model tries to produce ranking scores that minimize the number of out-of-order pairs. In other words, it does not matter so much if the scores are close to relevances, as long as the ranking order for each pair of relevant and irrelevant documents is correct. Finally, listwise models take the whole query into account. Instead of using pairs the models will use the full list of documents for each query and will try to rank them as closely as possible to their desired order.
We assume that the loss function takes the form: $$\label{loss-general}
\Delta(F) = \sum_{y_{ij}}\delta(y_{ij}|F)$$ where the sum is taken over all documents $y_{ij}$ in the training set, and $\delta$ denotes a document level loss function. A simple example of a quadratic loss function would be $\delta_\mathrm{quad}(y_{ij}|F)=(r_{ij}-F(y_{ij}))^2$, where $r_{ij}$ is the relevance of document $y_{ij}$. Another popular choice is the cross-entropy loss function if the relevances are binary ($0$ or $1$): $\delta_{CE}(y_{ij}|F)=-r_{ij}\log F(y_{ij})-(1-r_{ij})\log(1-F(y_{ij}))$.
Note that the form (\[loss-general\]) does not restrict the loss function to be pointwise - it can also be pairwise or listwise. In fact, it has been shown in the literature that the unbiased learning-to-rank framework works better for pairwise/listwise models than pointwise ones [@Joachims:2017:ULB:3018661.3018699; @Wang:2018:PBE:3159652.3159732]. We will discuss the reasons behind this later in this section. A popular example of a listwise loss function (and one we will use later for our models) is the DCG (Discounted Cumulative Gain): $$\label{dcg}
\mathrm{DCG}=\sum_{i,j}\frac{\mathrm{rel}_{ij}}{\log_2(j+1)}$$ where $i$ is the index of the query, $j$ is the rank of a given document and $\mathrm{rel}_{ij}$ is its relevance, and the sum is taken over all queries and all documents for each query.
If we had data for a fair sample of queries and a fair sample of documents for these queries then minimizing the loss function (\[loss-general\]) above would result in an unbiased model (this is known in the literature as Empirical Risk Minimization [@Vapnik1998]). However, the data is often biased. For example, if we use click logs to determine relevances (a click on the document means the document is relevant, and no click means irrelevant) then we will only have data for documents that the users have actually seen. Using only that data for training will introduce a bias since some documents are more likely to be seen by a user than others. For example, documents that were ranked highly for a given query are more likely to be seen and receive clicks than documents ranked at lower positions (position bias). If we only use the data for documents for which the relevances have been revealed to us (e.g. the user has seen the document and decided to click or not click) we will end up with a biased model. On the other hand, we have no way of including the data for which the relevances have not been revealed. The Unbiased learning-to-rank methodology [@Joachims:2017:ULB:3018661.3018699] introduces a modification to the loss function (\[loss-general\]) such that it becomes an unbiased estimator for the true loss even if the data is biased. The requirement is that we should know the probabilities of observing the relevances for all of the documents in the data (in the context of this paper observing the relevance means that the user has examined the document and decided to click on it or not). In other words, for each document $y_{ij}$ in the training data we know the probability $p(y_{ij})$ of the relevance of that document being observed. This probability is commonly referred to as the **propensity**. For example, position bias would imply that $p(y_{ij})$ is larger for documents that had higher ranks when shown to the user compared to the ones at lower ranks. If the propensities are known then an unbiased estimator of the loss function (\[loss-general\]) is $$\label{loss-unbiased}
\hat{\Delta}(F) = \sum_{y_{ij}}\frac{o(y_{ij})\delta(y_{ij}|F)}{p(y_{ij})}=\sum_{y_{ij}:o(y_{ij})=1}\frac{\delta(y_{ij}|F)}{p(y_{ij})}\,.$$ Here $o(y_{ij})$ denotes if the document $y_{ij}$ has been *observed* ($1$ if it has been observed, $0$ otherwise). The document being observed is equivalent to the relevance being revealed to us. The equation above only includes data that has been observed, so it can be used in practice. To show that (\[loss-unbiased\]) is an unbiased estimator of (\[loss-general\]) we compute the expected value of (\[loss-unbiased\]): $$\begin{aligned}
\label{unbiased-proof}
\begin{split}
\mathbb{E}_{o(y_{ij})}[\hat{\Delta}(F)]&=\sum_{y_{ij}}\mathbb{E}_{o(y_{ij})}\left[\frac{o(y_{ij})\delta(y_{ij}|F)}{p(y_{ij})}\right]\\
&=\sum_{y_{ij}}\frac{p(y_{ij})\delta(y_{ij}|F)}{p(y_{ij})}\\
&=\sum_{y_{ij}}\delta(y_{ij}|F)\\
&=\Delta(F)\,.
\end{split}\end{aligned}$$
The unbiased loss function (\[loss-unbiased\]) can be used if we know all of the observed documents, as well as their propensities. However, in practice it might be hard to know all of the observed documents. We know for sure that clicked documents have been observed, but we may not have information about documents that have not been clicked. One approach is to assume that all of the documents have been observed up to the lowest ranked clicked document. However, this may be inaccurate since users sometimes skip certain documents and scroll down. Also, the users may observe more documents beyond the last clicked one without clicking. For this reason it is more desirable to have a loss function that includes only “relevant” documents, such as clicked ones. This is why pairwise/listwise models work better than pointwise ones for unbiased learning-to-rank [@Joachims:2017:ULB:3018661.3018699; @Wang:2018:PBE:3159652.3159732]. For example, the DCG loss function (\[dcg\]) only includes relevant documents, assuming $\mathrm{rel}_{ij}=0$ for irrelevant ones.
In this paper we will focus on position bias. We will assume that the probability of observing a document depends only on the original rank of the document (i.e. the rank at which the document was presented to the user at the time of data collection). We denote the propensity at rank $i$ by $p_i$ for $i\in[1,R_{\max}]$, where $R_{\max}$ is the maximum rank used in the dataset. If these propensities are known then the Unbiased learning-to-rank method described above can be used to remove the position bias from the LTR model.
Propensity Estimation Method {#section-method}
============================
The method proposed by Joachims et. al. [@Joachims:2017:ULB:3018661.3018699] for estimating click propensities is running an experimental intervention in the live search engine, where the documents at two selected ranks are swapped. By comparing the click through rates at these ranks before and after swapping one can easily estimate the ratios of propensities at these ranks. This process can be repeated for pairs of all different ranks to estimate the ratios of all of the propensities. For the loss function (\[loss-unbiased\]) it is sufficient to compute the propensities up to a constant multiplier. Therefore, estimating ratios of all of the propensities is enough for the purpose of training an unbiased LTR model.
Here we propose a novel methodology for estimating click propensities without any intervention. For some search engines, especially in eCommerce, the same query-document pair may naturally appear more than once at different ranks. Using the click data on such documents we can accurately estimate click propensities. It is not required that the same query-document pair should appear at different ranks a large number of times. It is sufficient to have only two data points for the same query-document pair, as long as they are at different ranks.
We model clicks by the following simple model (also used in [@Joachims:2017:ULB:3018661.3018699]) - *The probability of a click on a given document is the product of the probability of observing the document and the probability of clicking on the document for the given query assuming that it has been observed.* We assume that the probability of observing a document depends only on its rank and the probability of clicking on the document for a given query if it is observed depends only on the query and the document. Mathematically: $$\begin{aligned}
\label{prob-click}
\begin{split}
p(c=1|q,y)&=p(o=1|q,y)p(c=1|q,y,o=1)\\
&=p(o=1|rank(y))p(c=1|q,y,o=1)\\
&=p_{rank(y)}p(c=1|q,y,o=1)
\end{split}\end{aligned}$$ where $q$ denotes a query, $y$ denotes a document, $c$ denotes a click ($0$ or $1$), $o$ denotes observation ($0$ or $1$), and $p_i$ denotes the propensity at rank $i$.
Let us assume that our data $D$ consists of $N$ query-document pairs $x_j$ for $j\in[1,N]$. For a query-document pair $x_j$ we will denote the probability of clicking on the document after observing it by $z_j$. For each query-document pair $x_j$ we have a set of ranks $r_{jk}$ where the document has appeared for the query, and clicks $c_{jk}$ denoting if the document was clicked or not ($1$ or $0$) when it appeared at rank $r_{jk}$, for $k\in[1,m_j]$. Here we assume that the query-document pair $x_j$ has appeared $m_j$ separate times. For now we do not assume that $m_j$ must be greater than 1 - it can be any positive integer.
The probability of a click for query-document pair $x_j$ where the document appeared at rank $r_{jk}$ is, according to (\[prob-click\]): $$p(c=1)=p_{r_{jk}}z_j\,.$$
It follows that $$p(c=0)=1-p_{r_{jk}}z_j\,.$$
We can now introduce the following likelihood function: $$\label{likelihood}
\mathcal{L}(p_i,z_j|D)=\prod_{j=1}^N\prod_{k=1}^{m_j}\left[c_{jk}p_{r_{jk}}z_j+(1-c_{jk})(1-p_{r_{jk}}z_j)\right]\,.$$ Here the parameters are the propensities $p_i$ and the “relevances” $z_j$ (relevance here means probability of clicking for a given query-document pair assuming that the document has been observed). Theoretically, the parameters can be estimated by maximizing the likelihood function above. However, this can be challenging due to the large number of parameters $z_j$. In fact, we are not even interested in estimating the $z_j$ - we only need to estimate the propensities $p_i$, and the $z_j$ are nuisance parameters.
There are multiple approaches that one can take to estimate the propensities depending on the data itself. First of all, let us consider the query-document pairs that appeared only at one rank (either the query-document pair appeared only once or multiple times but always at the same rank). The parameters $p_i$ and $z_j$ appear only as a product of each other in the likelihood function above. If a query-document pair $x_j$ appeared only at rank $i$ then we will only get terms like $p_iz_j$. So these query-document pairs could be helpful in estimating the product of the propensity at the rank that they appeared at and the relevance $z_j$ but not each one individually. With $z_j$ unknown, this would not help to estimate the propensity at all. However, we should mention that in the presence of a reliable prior for $z_j$ and/or $p_i$ the likelihood function above can be used even for those query-document pairs that appeared only at one rank. In this case it would be more useful to take a Bayesian approach and estimate the posterior distribution for the propensities, for example using Gibbs sampling [@10.2307/2685208]. That study is beyond the scope of this work and will be explored by the authors in a future study.
From now on we will assume that the query-document pairs appear at least at two different ranks. Another extreme is the case when each query-document pair appears a large number of times at different ranks. In this case the estimation of the propensities (or more precisely their ratios) could be much simpler than maximizing the likelihood function above. For a pair of fixed ranks $i$ and $j$ one can take all query-document pairs that appeared at both of those ranks, then simply count the total number of clicks received at rank $i$ and rank $j$ and take their ratio. That would give a good estimate for the ratio $p_i/p_j$. We prove this in subsection \[sec-ratio\] below. Of course, for that estimate to be reliable one would need a large number of query-document pairs that appeared at both of these ranks. But if our data consists of query-document pairs that appeared a large number of times at different ranks then there should be enough data to estimate the ratios of propensities for the top ranks at least.
Let us now consider the case when the data consisits of a large number of query-document pairs that appeared a few times (can be as few as twice) at different ranks, but the query-document pairs do not appear a large enough number of times to be able to estimate the propensity ratios by simply counting clicks as described above. In this case we will actually need to maximize the likelihood above and somehow eliiminate the nuisance parameters $z_j$ to get estimates for the $p_i$. To the best of our knowledge this case has not been studied in the literature before and we will focus the rest of this work on this case. Also, the data we have collected from eBay search logs falls in this category, as will be discussed in Section \[sec-ebay\].
If a query-document pair appeared only a few times there is a good chance that it did not receive any clicks. These query-document pairs will not help in estimating the propensities by likelihood maximization because of the unknown parameter $z_j$. Specifically, for such query-document pairs we will have the terms $$\prod_{k=1}^{m_j}(1-p_{r_{jk}}z_j)\,.$$
If we use the maximum likelihood approach for estimating the parameters then the maximum will be reached by $z_j=0$ for which the terms above will be $1$. That means that these query-document pairs without any clicks will not change the maximum likelihood estimate of the propensities. For that reason we will only keep query-document pairs that received at least one click. However, we cannot simply drop the terms from the likelihood function for query-document pairs that did not receive any clicks. Doing so would bias the data towards query-document pairs with a higher likelihood of click. Instead, we will replace the likelihood function above by a conditional probability. Specifically, the likelihood function (\[likelihood\]) computes the probability of the click data $\{c_{jk}\}$ obtained for that query-document pair (i.e. the clicks the query-document pair received and did not receive at different ranks). If we are keeping only the query-document pairs that received at least one click then we need to replace that probability by a conditional probability - the probability of the click data $\{c_{jk}\}$ under the condition that there was at least one click received: $\sum_kc_{jk}>0$. The likelihood function for the query-document pair $x_j$ will take the form: $$\begin{aligned}
\label{likelihood_j}
\begin{split}
\mathcal{L}_j(p_i,z_j|D_j)&=P\left(D_j|\sum_k c_{jk}>0\right)\\
&=\frac{P(D_j\cap\sum_k c_{jk}>0)}{P(\sum_k c_{jk}>0)}\\
&=\frac{P(D_j)}{P(\sum_k c_{jk}>0)}\\
&=\frac{\prod_{k=1}^{m_j}\left[c_{jk}p_{r_{jk}}z_j+(1-c_{jk})(1-p_{r_{jk}}z_j)\right]}{1-\prod_{k=1}^{m_j}(1-p_{r_{jk}}z_j)}\,.
\end{split}\end{aligned}$$ Here $\mathcal{L}_j$ denotes the likelihood function for the query-document pair $x_j$, $D_j=\{c_{jk}\}$ denotes the click data for query-document pair $j$, and $P$ denotes probability. $\sum_k c_{jk} > 0$ simply means that there was at least one click. In the first line above we have replaced the probability of data $D_j$ by a conditional probability. The second line uses the formula for conditional probability. The probability of $D_j$ and at least one click just equals to probability of $D_j$ since we are only keeping query-document pairs that received at least one click. This is how the third line is derived. Finally, in the last line we have explicitly written out $P(D_j)$ in the numerator as above in (\[likelihood\]) and the probability of at least one click in the denominator (the probability of no click is $\prod_{k=1}^{m_j}(1-p_{r_{jk}}z_j)$ so the probability of at least one click is $1$ minus that).
The full likelihood is then the product of $\mathcal{L}_j$ for all query-document pairs: $$\label{likelihood-1}
\mathcal{L}(p_i,z_j|D)=\prod_{\substack{j=1\\\sum_k c_{jk} > 0}}^N\frac{\prod_{k=1}^{m_j}\left[c_{jk}p_{r_{jk}}z_j+(1-c_{jk})(1-p_{r_{jk}}z_j)\right]}{1-\prod_{k=1}^{m_j}(1-p_{r_{jk}}z_j)}\,.$$
From now on we will assume by default that our dataset contains only query-document pairs that received at least one click and will omit the subscript $\sum_k c_{jk} > 0$.
Our last step will be to simplify the likelihood function (\[likelihood-1\]). Typically the click probabilities $p_iz_j$ are not very large (i.e. not close to $1$). This is the probability that the query-document pair $j$ will get a click when displayed at rank $i$. To simplify the likelihood for each query-document pair we will only keep terms linear in $p_iz_j$ and drop higher order terms like $p_{i_1}z_{j_1}p_{i_2}z_{j_2}$. As discussed later in Section \[sec-ebay\], this assumption is valid for our data. In general, we expect this assumption to be valid for most search engines. It is certainly a valid assumption for lower ranks since click through rates are typically much smaller for lower ranks. One would simply need to verify this assumption for topmost ranks. Since we are dropping product terms the largest ones would be between ranks $1$ and $2$. For most search engines the click through rates at rank 2 are around 10% or below, which we believe is small enough to be able to safely ignore the product terms mentioned above (they would be at least $10$ times smaller than linear terms). We empirically show using simulations in Section \[section-sim\] that this assumption works very well for data similar to eBay data. If for other search engines the click through rates are much larger for topmost ranks we suggest keeping only those query-document pairs that appeared at least once at a lower enough rank. Also, using the methodology of simulations from Section \[section-sim\] one can verify how well this assumption works for their particular data.
Under the simplifying assumption we get for the denominator in (\[likelihood-1\]): $$\label{simplify-denom}
1-\prod_{k=1}^{m_j}(1-p_{r_{jk}}z_j)\simeq1-\left(1-\sum_{k=1}^{m_j}p_{r_{jk}}z_j\right)=z_j\sum_{k=1}^{m_j}p_{r_{jk}}\,.$$
Let us now simplify the numerator of (\[likelihood-1\]). Firstly, since the click probabilities are not large and each query-document pair appears only a few times we can assume there is only one click per query-document pair[^2]. We can assume $c_{jl_j}=1$ and $c_{jk}=0$ for $k\neq l_j$. The numerator then simplifies to $$\begin{aligned}
\label{simplify-numer}
\begin{split}
\prod_{k=1}^{m_j}\left[c_{jk}p_{r_{jk}}z_j+(1-c_{jk})(1-p_{r_{jk}}z_j)\right]&=p_{r_{jl_j}}z_j\prod_{\substack{k=1\\k\neq l_j}}^{m_j}(1-p_{r_{jk}}z_j)\\
&\simeq p_{r_{jl_j}}z_j\,.
\end{split}\end{aligned}$$
Using (\[simplify-denom\]) and (\[simplify-numer\]) the likelihood function (\[likelihood-1\]) simplifies to $$\label{likelihood-final}
\mathcal{L}(p_i,z_j|D)=\prod_{j=1}^N\frac{p_{r_{jl_j}}z_j}{z_j\sum_{k=1}^{m_j}p_{r_{jk}}}=\prod_{j=1}^N\frac{p_{r_{jl_j}}}{\sum_{k=1}^{m_j}p_{r_{jk}}}\,.$$
In the last step $z_j$ cancels out from the numerator and the denominator. Our assumption of small click probabilities, together with keeping only query-document pairs that received at least one click allowed us to simplify the likelihood function to be only a function of propensities. Now we can simply maximize the likelihood (\[likelihood-final\]) to estimate the propensities.
Eq. (\[likelihood-final\]) makes it clear why we need to include the requirement that each query-document pair should appear more than once at different ranks. If we have a query-document pair that appeared only once (or multiple times but always at the same rank) then the numerator and the denominator would cancel each other out in (\[likelihood-final\]). For that reason we will keep only query-document pairs that appeared at two different ranks at least.
It is numerically better to maximize the log-likelihood, which takes the form: $$\label{log-like}
\log\mathcal{L}(p_i|D)=\sum_{j=1}^N\left(\log(p_{r_{jl_j}})-\log\sum_{k=1}^{m_j}p_{r_{jk}}\right)\,.$$
Let us summarize our method of estimating propensities. Our data consists of query-document for which the document appeared at more than one rank and got at least one click. We then maximize the log-likelihood (\[log-like\]) to estimate the propensities. The log-likelihood is a sum over all query-document pairs in our dataset. For each query-document pair we take the log of the propensity for the rank at which the document got clicked (assuming there is only one) minus the log of the sum of the propensities for all the ranks where the document appeared. The only assumption we made in order to get the simple form of (\[log-like\]) is that the click probabilities are not large. As we will see in Section \[sec-ebay\] this is a reasonable assumption for eBay search. Our methodology of simulations in Secition \[section-sim\] can be used to verify the validity of the assumption. We have also discussed alternative approaches for estimating the click propensities for cases when the assumption might not work very well.
Propensity Ratio Estimation {#sec-ratio}
---------------------------
Here we consider the case when for two fixed ranks $i$ and $j$ a large number of query-document pairs appear at both of these ranks. As mentioned in Section \[section-method\] one can simply compute the number of clicks from those query-document pairs for each rank and take the ratio of those numbers to estimate the ratio $p_i/p_j$. We prove that statement below.
Let us assume that query-document pairs $\{x_k, k = 1,\dots,K\}$ appeared for both ranks $i$ and $j$. We will first assume that these query-document pairs appeared exactly once for each rank. We will later relax that assumption. The probability of a click at rank $i$ for query-document pair $x_k$ is $p_iz_k$ where $z_k$ is the “relevance” for query-document pair $x_k$ (i.e. the probability of a click under the assumption that it was observed). So the expected number of clicks $N_i$ for rank $i$ for all query-document pairs will be: $$\mathbb{E}[N_i]=\sum_{k=1}^Kp_iz_k=p_i\sum_{k=1}^Kz_k\,.$$ Similarly: $$\mathbb{E}[N_j]=\sum_{k=1}^Kp_jz_k=p_j\sum_{k=1}^Kz_k\,.$$
Taking the ratio[^3]: $$\frac{\mathbb{E}[N_i]}{\mathbb{E}[N_j]}=\frac{p_i\sum_{k=1}^Kz_k}{p_j\sum_{k=1}^Kz_k}=\frac{p_i}{p_j}\,.$$
Replacing the expected values of $N_i$ and $N_j$ above by their actual observed values we can get an estimator for $p_i/p_j$.
So far we have assumed that each query-document pair appears exactly once at each rank. The above proof can be easily extended to the case when the same query-document pair appears multiple times at each rank. In this case instead of counting clicks directly we will count clicks per impressions, i.e. for each query-document pair we will count the total number of clicks and divide by the total number of impressions. The expected value for clicks per impressions is still the click probability $p_iz_k$. Then we will take the sum for clicks per impressions for each rank and take the ratio. The derivation will remain exactly the same. The only difference is that $N_i$ and $N_j$ will now denote the sum of clicks per impressions instead of the total number of clicks.
Click Propensities for eBay Search {#sec-ebay}
==================================
![Click propensity estimated for eBay search for web data. The solid blue line is the direct estimation of propensities for each rank, and the red dashed line is the estimation using interpolation.[]{data-label="fig-prop-data"}](propensity_data)
We now apply the method developed above on eBay search data to estimate propensities. We collected a small sample (0.2%) of queries for four months of eBay search traffic. For each query we keep the top 500 items (as mentioned before, we use the terms “item” and “document” interchangeably). There are multiple sort types on eBay (such as Best Match, Price Low to High, Time Ending Soonest) and click propensities may differ for different sort types. In this paper we present our results on Best Match sort, and hence we keep only queries for that sort type. Furthermore, there are multiple different platforms for search (such as a web browser or a mobile app) which can have different propensities. We separate our dataset into two platforms - web and mobile, and estimate click propensities for each platform separately. For web queries the users can choose the number of items shown per page. The most common choice (and the default one) is 50 items per page. Click propensities might differ depending on the number of items per page. For web search we estimate the propensities for 50 items per page with list view, and filter our dataset accordingly.
Next, we must identify same query-document pairs and find cases where the document appeared at multiple different ranks. However, we would like to avoid the cases where some detail about the item (such as price) has changed between the queries, which could result in a different probability of a click on that item. eBay has inventory of both auction and fixed price items. Since the auction item properties (like current bid amount) change very frequently we exclude auction items from our dataset and keep only fixed price items. Furthermore, we check to make sure that the price of the item has not changed from query to query. We also keep the same query-item pairs from the same day only. This is a further caution to make sure that seasonality effects do not affect the popularity of the item. For the query side we identify two queries to be the same if they have the same keywords, as well as the same category and aspect (such as color, size) constraints. We then keep only those query-document pairs that appeared at two different ranks and got one click in one rank and no click in the other (in our data it is rare for the same query-document pair to appear at more than one rank and get more than one click).[^4]
We first check our assumption of small click probabilities. We compute the click through rates at ranks 1-500. As expected, rank 1 has the highest click through rate, but it is still under 10%. For lower ranks it is even smaller. This does justify our assumption of small click probabilities and we can apply our method above to estimate propensities.
![Click propensities for eBay search for mobile data. The solid blue line is the direct estimation at each rank and the red dashed line is the estimation using interpolation. For comparison, we also plot the propensities for web data using interpolation in solid green, which is the same as the red dashed line from Fig. \[fig-prop-data\].[]{data-label="fig-prop-data-mobile"}](propensity_data_mobile)
We first estimate propensities for web queries. Our dataset consists of about 40,000 query-item pairs, each of which appeared at two different ranks and received a click at one of the ranks. We use two methods for estimating propensities - direct and interpolation. In the direct method we treat the propensity at each rank as a separate parameter. We therefore get 500 different parameters to estimate. In the interpolation method we fix a few different ranks and use the propensities at those ranks as our parameters to estimate. The propensities for all the other ranks are computed as a linear interpolation in the log-log space, i.e. we approximate the log of the propensity as a linear function of the log of the rank. This results in the propensity being a power law of the rank. For the interpolation method our fixed ranks are 1, 2, 4, 8, 20, 50, 100, 200, 300, and 500. We choose a denser grid for higher ranks since there is more data and less noise for higher ranks, and the propensities can be estimated more accurately.
Our resulting propensity for web search is shown in Fig. \[fig-prop-data\]. The solid blue line shows the propensities estimated through the direct method, and the red dashed curve shows the propensities estimated through interpolation. Even though we estimate propensities up to rank 500, we plot them only up to rank 200 so that the higher ranks can be seen more clearly. The red dashed curve passes smoothly through the blue solid curve, which is reassuring. Note that the red dashed curve is not a fit to the blue one. The two are estimated directly from the data. For the blue curve the parameters are all of the propensities at each rank, whereas for the red dashed curve we only parametrize the propensities at select ranks and interpolate in between. We then maximize the likelihood for each case to estimate the parameters. The fact that the red dashed line appears to be a smooth fit to the solid blue shows that the interpolation method is useful in obtaining a smooth and less noisy propensity curve which is still very close to the direct estimation.
The propensities estimated from eBay mobile search data are shown in Fig. \[fig-prop-data-mobile\]. As in Fig. \[fig-prop-data\], the blue solid curve shows direct estimation, and the red dashed curve is estimation using interpolation. For comparison, we plot the propensities from web using interpolation in solid green. The blue solid curve shows a certain periodicity - the propensities seem to drop sharply near rank 25, then go back up at rank 40, drop again around rank 65, then back at rank 80, and so on. In fact, this reflects the way results are loaded in mobile search - 40 at a time. The blue curve seems to indicate that users observe the results at higher ranks with the usual decrease in interest, then they tend to scroll faster to the bottom skipping the results towards the bottom, then as the new batch is loaded they again gain more interest and it continues decreasing as they scroll down and the pattern repeats itself every 40 ranks. The red dashed curve matches the blue one reasonable well, but it fails to capture the periodic dips. This is due to our choice of knots for the linear spline. One can use the blue curve to choose new locations of the knots to be able to get a better interpolation for the propensities. The green solid curve matches fairly well with the blue one except for the dips. This means that the propensities for web and mobile are very similar, except for the periodic dips for mobile. The web results are shown 50 items per page, but we have not found any periodic dips for web search. Perhaps this indicates that for web search users do not tend to scroll quickly towards the end of the page and then regain interest as a new page is loaded. The smooth decline in propensities indicates that for web search users steadily lose interest as they scroll down, but the number of items per page does not affect their behavior. If they make it to the end of the page then they load the next page and continue.
Results on Simulations {#section-sim}
======================
![Propensity estimated from simulated data. The green solid curve shows the “true” propensity (\[prop-true\]). The blue solid curve is the estimated propensity using the direct estimation method. The red dashed curve is the estimation using interpolation.[]{data-label="fig-prop-simulation"}](propensity_simulation)
In this section we use simulated data to verify that the method of estimating propensities developed in Section \[section-method\] works well. For our simulations we choose the following propensity function as truth: $$\label{prop-true}
p_i^{\mathrm{sim}}=\min\left(\frac{1}{\log{i}},1\right)$$ which assigns propensity of $1$ for ranks $1$ and $2$, and then decreases as the inverse of the log of the rank.
Other than choosing our own version of propensities we simulate the data to be as similar to the eBay dataset as possible. We generate a large number of query-document pairs and randomly choose a mean rank $rank_{mean}$ for each query-document pair uniformly between 1 and 500. We randomly generate a click probability $z$ for that query-document pair depending on the mean rank $rank_{mean}$. We choose the distribution from which the click probabilities are drawn such that the click through rates at each rank match closely with the click through rates for real data, taking into account the “true” propensities (\[prop-true\]). We then generate two different ranks drawn from $\mathcal{N}(rank_{mean}, (rank_{mean} / 5)^2)$. For each rank $i$ we compute the probability of a click as $zp_i^{\mathrm{sim}}$. Then we keep only those query-document pairs which got at least one click and appeared at exactly two different ranks, in agreement with our method used for real eBay data. Finally we keep about 40,000 query-document pairs so that the simulated data is similar to the eBay web search data in size. This becomes the simulated data. Note that the real data contains many query-document pairs that appeared only once and some that appeared more than twice. We do not simulate such query-document pairs since they would be eventually all filtered out. In our simple method we keep only those query document pairs which appear twice exactly at two different ranks and receive at least one click. Note that we could have also kept query-document pairs that appeared more than twice. However, as mentioned in the previous Section, in actual eBay data we have much fewer pairs like that compared to the ones that appeared twice. So for the sake of simplicity we choose to only keep pairs that appeared exactly twice, without losing much of the data.
The estimated propensities on the simulated dataset are shown in Fig. \[fig-prop-simulation\]. The green solid curve shows the true propensity (\[prop-true\]), the blue solid curve shows the estimated propensity using the direct estimation method, and the red dashed curve is the estimated propensity using interpolation. As we can see, the estimations closely match with the truth. Furthermore, we can see that the interpolation method gives a better result by reducing the noise in the estimate. These results show that the propensity estimation method developed in this paper works well.
Unbiased Learning-to-Rank Models {#section-models}
================================
In this section we study the improvement in ranking models by using the estimated click propensities for eBay search data. Previous studies have consistently shown that unbiased learning-to-rank models significantly improve ranking metrics compared to their biased counterparts. Specifically, Joachims et. al. [@Joachims:2017:ULB:3018661.3018699] have shown that an unbiased learning-to-rank model significantly improves the average rank of relevant results for simulated data. Furthermore, they have performed an online interleaving experiment on a live search engine for scientific articles, which resulted in a significant improvement for the unbiased model. Wang et. al. [@Wang:2018:PBE:3159652.3159732] have shown an improvement in MRR (Mean Reciprocal Rank) for the unbiased learning-to-rank models for personal search.
Here we train simple proof-of concept models to check if unbiased ranking models show improvements over their biased counterparts. The training and evaluation of full production scale models is beyond the scope of this work and will be performed in the future. For our training data we collect a sample of about 40,000 queries which have received at least one click. The sample is collected from four days of search logs. We train listwise ranking models using the LambdaMART algorithm [@from-ranknet-to-lambdarank-to-lambdamart-an-overview]. We use the DCG metric (\[dcg\]) as our loss function. We define $\mathrm{rel}_{ij}$ to be $1$ if document $j$ was clicked, and $0$ otherwise. We train two models - baseline and unbiased. The baseline model uses (\[dcg\]) as its loss function without any correction for position bias and serves as our baseline, while the unbiased model uses inverse-propensity weighted relevances as in (\[loss-unbiased\]). We use the propensities estimated for eBay web search as shown in Fig. \[fig-prop-data\] red dashed curve. Our training and test data are also from web search (i.e. browser) only. We use $25$ features for both models, selected from our top ranking features. We also fix the model parameters: the number of trees is $100$ and the shrinkage is $0.1$. Since these are simple proof-of-concept models we have not done extensive feature selection and parameter tuning.
Our test data contains a sample of about 10,000 queries from four days of eBay search logs. Since the test data also has the same position bias as the training data we cannot rely on standard ranking metrics such as DCG, NDCG (Normalized Discounted Cumulative Gain), or MRR (Mean Reciprocal Rank). Another option would be to use inverse-propensity-weighted versions of these metrics to remove the presentation bias. However, the true propensities are unknown to us and we obviously cannot use estimated propensities for evaluation since part of the evaluation is checking if our estimate of propensities is a good one. For that reason we choose a different approach for evaluation. Namely, we fix the rank of items in the test data, i.e. we select items from different queries that appeared at a given fixed rank. By selecting the items from a fixed rank in the evaluation set we effectively eliminate position bias since all of the items will be affected by position bias the same way (the observation probability is the same for all the items since the rank is the same). Then we compare the two ranking models as classifiers for those items, which means that we evaluate how well the models can distinguish items that were clicked from ones that were not. We use AUC (Area Under the Receiver Operating Characteristic Curve) as our evaluation metric.
Rank AUC Baseline AUC Unbiased AUC Improvement
------ -------------- -------------- -----------------
1 0.484 0.499 3.1%
2 0.520 0.535 2.9%
4 0.527 0.549 4.2%
8 0.551 0.570 3.4%
16 0.603 0.644 6.8%
32 0.584 0.588 0.7%
: Performance of the unbiased ranking model compared to the baseline model. The unbiased model uses inverse propensity weighting with the propensities estimated in Section \[sec-ebay\], while the baseline model does not correct for position bias. The validation set contains documents from a fixed rank. The first column shows that fixed rank. The second and third columns show the AUC metric for the baseline and unbiased models, respectively. The last column shows the improvement in AUC for the unbiased model compared to the baseline.[]{data-label="eval-table"}
The results are presented in Table \[eval-table\], where we show results for fixed ranks $1$, $2$, $4$, $8$, $16$, and $32$. As we can see, for all ranks the unbiased model outperforms the baseline. The best improvement is for items at fixed rank $16$, with an AUC improvement of $6.8\%$. We also notice that for low ranks the AUC is close to $0.5$ for both models, meaning that the models are unable to distinguish between items that are likely to get clicked from ones that are not. This is due to the fact that the items have been ranked with the current production model and low ranked items are the ones that have a high likelihood to get clicked. Clearly our simple models are strongly correlated with the production model. This result simply means that after the production model selects the “best” items the new models do not have a strong distinguishing power. Nevertheless, the unbiased model has a higher distinguishing power than the baseline.
Summary and Future Work {#section-summary}
=======================
In this work we have introduced a new method for estimating click propensities for eCommerce search without randomizing the results during live search. Our method uses query-document pairs that appear more than once and at different ranks. Although we have used eCommerce search as our main example, the method is general and can be applied to any search engine for which ranking naturally changes over time. The clear advantage of our method over result randomization is that it does not affect live search results, which can have a negative impact on the engine as has been shown in the literature [@Wang:2018:PBE:3159652.3159732]. To the best of our knowledge there has been proposed only one other approach for estimating click propensities without result randomization by Wang et. al. [@Wang:2018:PBE:3159652.3159732]. Their method uses an EM (Expectation Maximization) algorithm to estimate the propensities and relies on ranking features for estimating and eliminating relevances. The advantage of our method over the EM method is that we estimate propensities directly and do not have to rely on ranking features. The results of the EM method strongly depend on the goodness of the features and their ability to predict the relevance of the document. For this reason we believe that our novel approach will find widespread use for unbiased learning-to-rank modeling.
We have used simulated data to show that our method can give accurate estimates of the true propensities. We have applied our method to eBay search results to separately estimate propensities for web and mobile search. We have also trained simple proof-of-concept ranking models and compared the performance of the unbiased model using the estimated propensities to the baseline model, which does not correct for position bias. Using a validation dataset of documents from a fixed rank we have shown that the unbiased model significantly outperforms the baseline in terms of the AUC metric.
This work focuses on estimating propensities from query-document pairs that appear at multiple different ranks. We have derived a simple method for estimating propensity ratios when each query-document pair appears a large number of times at different ranks. More importantly, we have addressed the case when the same query-document pair appears only a few number of times at different ranks (can be as few as twice), which is the case for our data. This method can be generalized to use query-document pairs that appeared at a single rank only by incorporating appropriate priors and using Gibbs sampling to estimate the posterior distribution for propensities. We plan to study this approach in a future work.
Previous studies have shown that the unbiased ranking models significantly outperform their “biased” counterparts both in terms of offline ranking metrics and A/B test results [@Joachims:2017:ULB:3018661.3018699; @Wang:2018:PBE:3159652.3159732]. This work has mostly focussed on propensity estimation. Because of space limitations we have only studied simple proof-of-concept learning-to-rank models. In the future we would like to further evaluate our estimated propensities and their use in ranking models. This can include training production level unbiased learning-to-rank models with appropriate feature selection and parameter tuning. The models can be evaluated offline on human judged data (which will remove the position bias from the validation set) and online through A/B tests.
Click propensities have been estimated in this work under the assumption that they are the same for all queries and all users. In the future it would be desirable to estimate and compare propensities for different classes of queries and/or users. For example, one could compare more frequent and less frequent queries, queries from different categories (such as electronics versus fashion), as well as different user demographics. Also, in this work we have only focused on the “Best Match” sort for eBay. Propensities for other sort types, such as “Price Low to High”, would also be of interest in a future study.
[^1]: Here an “item” is the same as a “document”. In this work we use the terms “item” and “document” interchangeably.
[^2]: This is true for our data as discussed later in Section \[sec-ebay\]. For the cases when most query-document pairs receive multiple clicks we suggest using a different method, such as computing the ratios of propensities by computing the ratios of numbers of clicks as discussed above
[^3]: Note that here we are taking the ratio of expected values rather than the expected value of the ratio. While in general they are not the same, to the first order approximation they are the same. We have empirically verified through simulations that taking the ratio of the observed values $N_i/N_j$ asymptotically approaches the ratio $p_i/p_j$ as the total number of query-document pairs is increased.
[^4]: Note that keeping only query-document pairs that appeared at two ranks exactly is in no way a requirement of our method. The method is general and can be used for query-document pairs that appeared more than twice. This is just intended to simplify our analysis without a significant loss in data, since it is rare for the same query-document pair to appear at more than two ranks.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We find that the requirement of model interpretations to be *faithful* is vague and incomplete. Indeed, recent work refers to interpretations as unfaithful despite adhering to the available definition. Similarly, we identify several critical failures with the notion of textual highlights as faithful interpretations, although they adhere to the faithfulness definition. With textual highlights as a case-study, and borrowing concepts from social science, we identify that the problem is a misalignment between the causal chain of decisions (causal attribution) and social attribution of human behavior to the interpretation. We re-formulate faithfulness as an accurate attribution of causality to the model, and introduce the concept of *aligned faithfulness*: faithful causal chains that are aligned with their expected social behavior. The two steps of causal attribution and social attribution *together* complete the process of explaining behavior, making the alignment of faithful interpretations a requirement. With this formalization, we characterize the observed failures of misaligned faithful highlight interpretations, and propose an alternative causal chain to remedy the issues. Finally, we the implement highlight explanations of proposed causal format using contrastive explanations.'
author:
- |
Alon Jacovi\
Bar Ilan University\
`[email protected]`\
Yoav Goldberg\
Bar Ilan University and Allen Institute for AI\
`[email protected]`
bibliography:
- 'tacl2018.bib'
title: Aligning Faithful Interpretations with their Social Attribution
---
=1
Introduction
============
In formalizing the desired properties of a quality interpretation, the NLP community has settled on the key property of *faithfulness* [@lipton2016mythos; @DBLP:journals/corr/abs-1711-07414; @wiegreffe2019attentionisnotnot], or how “accurately” the interpretation represents the true reasoning process of the model. Curiously, recent work in this area describes faithful interpretations of models as unfaithful when they fail to describe behavior that is “expected” of them [@sanjay2020], despite the interpretations seemingly complying with faithfulness definitions. We are interested in properly characterizing faithfulness, and exploring beyond faithfulness what is necessary of a quality explanation of a model decision to satisfy, in order to cope with this contradiction.
Several methods have been proposed to train and utilize models that faithfully “explain” their decisions with *highlights*—the guiding use-case in this work—otherwise known as an extractive rationale[^1], which aim to clarify what part of the input was important to the decision. The proposed models are modular in nature, composed of two stages of (1) *selecting* the highlighted text, and (2) *predicting* based on the highlighted text (*select-predict*, described in Section \[sec:background\]).
We take an extensive and critical look at the formalization of faithfulness and of explanations, with textual highlights as an example use-case. In particular, the *select-predict* models with faithful highlights provide us with more questions than they do answers, as we describe a variety of curious failure cases of such models in Section \[sec:limitations\], as well as experimentally validate that the failure cases are indeed possible and do occur in practice. Concretely, the behavior of the selector and predictor in these models do not necessarily line up with expectations of people viewing the highlight. Current literature in ML and NLP interpretability fails to provide a theoretical foundation to characterize these issues (Sections \[sec:limitations\], \[sec:plausibility\]).
In order to solve this problem, we turn to literature on the science of social explanations and how they are utilized and perceived by humans (Section \[sec:social\]): the social and cognitive sciences find that human explanations are composed of two, equally important parts: the attribution of a causal chain to the decision process (*causal attribution*), and the attribution of social or human-like behavior to the causal chain (*social attribution*) [@miller2017social].
We reformalize faithfulness as the (accurate) attribution of a causal chain of reasoning steps to the model decision process, which we find is the true nature behind the vague definition provided for this term until now. Fatally, the second key component of human explanations—the social attribution of behavior—has been missing from current formalization on the desiderata of artificial intelligence explanations in NLP. In Section \[sec:perceived-definition\] we define that a faithful interpretation—a causal chain of decisions—is *aligned* with human expectations if it is adequately constrained by the social behavior attributed to it by human observers.
Armed with this knowledge, we are able to characterize the mysterious failures of *select-predict* models described in Section \[sec:limitations\]: In Section \[sec:highlight-attribution\] we delve into the social attribution of highlights as explanation, outlining two possible attributions. We find that *select-predict* does not guarantee either. In Section \[sec:highlight-as-evidence\] we propose an alternative causal chain of the form of *predict-select-verify*. We note that the social attribution of this causal chain is of highlights to serve as *evidence* towards the predictor’s decision, and we formalize the constraints necessary to enforce this effect.
Finally, in Section \[sec:contrastive\] we discuss an implementation of *predict-select-verify*, i.e., designing the components in the roles predictor and selector. Designing the selector is non-trivial, as there are many possible options to select highlights that evidence the predictor’s decision, and we are only interested in selecting ones that are meaningful for the user to understand the decision. We leverage advancements from cognitive research on the internal structure of (human-given) explanations, dictating that explanations must be *contrastive* to hold tangible meaning to humans. We propose a classification *select-predict-verify* model which provides contrastive highlights—to our knowledge, a first in NLP—and qualitatively exemplify and showcase the solution.
#### Contributions.
We redefine “faithfulness” as the interpretation’s ability to represent the causal chain of model decisions, and formalize “aligned faithfulness” as the degree to which the causal chain is aligned with the social attribution of behavior that humans perceive from it. The new formalization allows us to identify issues with current models that derive faithful highlight interpretations, and design a new model pipeline that circumvents those issues. Finally, we propose an implementation of the new system with contrastive explanations, which are more intuitive and useful.
When designing interpretable models, we must formalize the social attribution of the interpretation—the *set of constraints* on model behavior to resemble human reasoning—in order to guarantee that the interpretation is aligned with expectations of human intent, in addition to being faithful.
Highlights as Faithful Interpretations {#sec:background}
======================================
Highlights, also known as extractive rationales, are binary masks over a given input which imply some kind of *behavioral interpretation* of a particular model’s decision process to arrive at a decision on the input.
Formally, given input $x \in R^n$ and model $m:~R^n~\longrightarrow~Y$, a highlight interpretation $h \in \mathbb{Z}_2^n$ is a binary mask over $x$ which attaches a meaning to $m(x)$, where the portion of $x$ highlighted by $h$ was important to the decision.
This functionality of $h$ was interpreted by @lei16 as an implication of a *behavioral process* of $m(x)$, where the decision process is a modular composition of two unfolding stages:
1. **Selector** component $m_s: R^n \longrightarrow \mathbb{Z}_2^n$ selects a binary highlight $h$ over $x$.
2. **Predictor** component $m_p: R^n \longrightarrow Y$ makes a prediction on the input $h \odot x$.
The final prediction of the system at inference is $m(x) = m_p(m_s(x) \odot x)$. We refer to $h$ as the **highlight** and $h \odot x$ as the **highlighted text**.
A highlight interpretation can be *faithful* or *unfaithful* to a model. Literature accepts a highlight interpretation as faithful if the highlight was the output of the selector component, and the input to the predictor component, which outputs the final prediction. These claims are, of course, verifiable only in select-predict models (notwithstanding advances in the understanding of other opaque models), and thus faithful highlight interpretations are limited to models of this structure.
#### Implementations.
Various methods have been proposed to build modular systems with the ability to derive faithful highlights in this way. @lei16 have proposed to train the selector and predictor end-to-end via the REINFORCE [@williams1992simple] algorithm. @bastings-etal-2019-interpretable have proposed to formalize the highlight as a latent variable sequence, replacing REINFORCE with the reparameterization trick [@kingma-vae]. Citing significant difficulty in training these solutions, @jain2020 have proposed to train the selector and predictor separately, by first training a separate model on the end task, and using an unfaithful highlight interpretation method on this model to select a highlight, then training the predictor on highlighted examples. We refer to this system by its given name as FRESH. Outside of NLP, @DBLP:conf/icml/ChenSWJ18 describe select-predict models for feature-vector inputs trained to maximize mutual information between the selected features and the response variable.
Utility of Highlights
=====================
We have described multiple possible approaches to faithful highlight interpretations. Is it enough, however, for highlights to be faithful—adhering to the select-predict composition—in order to be useful as indicators of the model’s decision process?
To answer these questions, we must first discuss potential uses of the technology. Throughout this work, we will refer to the following use-cases:
1. **Dispute**: A user may want to dispute a model’s decision, e.g. in a legal setting. A user can dispute either the selector or the predictor: a user may refer to words that were not selected in the highlight, and say: “the model ignored information *A*, but it shouldn’t have.” The user can also refer to the words that were fed to the predictor, and say: “based on this highlighted text, I would have expected a different outcome.”
2. **Debug**: Highlights allow a designation of model errors into two categories: did the model focus on the wrong part of the input, or did the model make the wrong prediction based on the correct part of the input? Based on the category, specific action may be taken to alleviate the specific error: influence the model to focus on the correct part of the input, or influence the model to make better judgements when the focus is correct.
3. **Advice**: When the user does not know the correct decision and is advised by a model, the user can be advised in two ways: (1) assuming that the user trusts the model, the highlight provides feedback on which part is important to make a decision; (2) if the user does not entirely trust the model, we assume that the user has a prior notion on attributes of the highlight selection process. If the model highlight is aligned or not aligned with this prior, then the user can opt to trust or not trust the model’s decision. For example, if the highlight is focused on punctuation and stop words, as the user believes that the highlight should include content words.
Limitations of Highlights as (Faithful) Interpretations {#sec:limitations}
=======================================================
We detail a variety of strange and surprising failures where select-predict models are uninformative to the above use-cases, as evidence of an apparent weakness in the formalization of faithfulness as a property of explanations. The failures are a possible risk in all current solutions of highlight interpretations.
Trojan Explanations
-------------------
Task information can manifest in the highlighted text in exceedingly unintuitive ways, where the highlight is faithful, but functionally useless to the intended use-cases. For example, consider the following case of a faithfully highlighted decision process:
1. The selector makes a prediction $y$, then selects an arbitrary highlight $h$ that encodes $y$.
2. The predictor then extracts $h$ from $h \odot x$ and decodes $y$ from $h$.
It is easy to see why the highlight becomes useless: the *meaning* that the highlight holds to the user is completely misaligned with the true role of the highlight in the decision process. E.g., in the *advice* use-case, the seemingly random highlighted text will cause the user to distrust the prediction of the model, despite the model quite realistically making informed and correct decisions based on well-generalizing reasoning.
Though this case may appear incredibly unnatural, it is nevertheless *not explicitly avoided* by faithful highlights of *select-predict* models or by the solutions available today. In other words, a highlight can still be regarded as perfectly faithful to a model that works in this way. As a result, there is no guarantee against highlights that encode task-relevant signal *by themselves*, without usage of the text itself. After all, we never demanded as such.
It should not be particularly surprising, then, that this is actually the case: the above “unintentional” exploit of the modular process is a perfectly valid trajectory in the training process of the highlight interpretation methods available today. We are able to verify this by attempting to predict the model’s output based on $h$ alone via another model (Table \[tab:trojan-results-h\]): although this experiment does not “prove” that the predictor utilizes this information, it shows that *there is no guarantee that it doesn’t*.
#### Definition.
The above is an example of a phenomenon we term the *Trojan explanation*: the explanation (in our case, $h \odot x$) carries information which is encoded in ways that are *unintuitive*, *difficult to discover* or otherwise *unintended* by the user which observes the interpretation as an explanation of model behavior. In the above case, this information is the prediction label, and the “unintuitive” encoding was manifested using $h$ alone, but of course the issue is not limited to those methods. The information encoded in $h \odot x$ can be anything which will be useful to predict $y$ and which the user will find hard to comprehend.
Below are cases of Trojans which are reasonably general to multiple tasks and use-cases:
1. **Highlight signal**: The label is encoded in the mask $h$ alone, requiring no information from the original text it is purported to focus on.
2. **Arbitrary token mapping**: The label is encoded via some mapping from highlighted tokens to labels which is considered arbitrary to the user, such as commas for a particular class, and periods for another.
3. **Arbitrary statistics mapping**: The label is encoded in arbitrary statistics of the highlighted text, such as quantity of capital letters, distance between commas, and so on.
4. **The default class**: In a classification case, a class can be predicted by precluding the ability to predict all other classes and selecting it by default. As a result, the selector may decide that the absence of class features in itself defines one of the classes.
In Table \[tab:trojan-results-statistics\] we report on an attempt to predict the decision (by training an MLP classifier) of select-predict models from quantities of seemingly irrelevant characters, such as commas and dashes, of the highlighted texts generated by the models. The results are compared against a baseline of attempting to predict the decisions based on the same statistics on the full text. Surprisingly, all models show an *increased* ability to predict their decisions on some level compared to the full text, showing that the highlights selected by these models carry *additional* information—which was not in the original input—that can be leveraged by the predictor to make decisions.
We stress that these Trojan explanations are not merely possible, but *just as reasonable* to the model as any other option, and difficult to counter explicitly, as we have not yet truly formalized the extent of what is regarded as a Trojan—or why it is undesirable at all[^2]—and possibilities of Trojans are conceptually limitless. Indeed, Trojan explanations have been observed in practice.[^3]
#### A note about FRESH.
The FRESH [@jain2020] solution to faithful highlights, via training the selector and predictor separately, may seem to be a natural solution to Trojan explanations, since the two models are unable to communicate during training. We note that even in the FRESH composition, *the selector was trained on the end task to make decisions before deriving a highlight*. As such, it is quite possible for the selector to disguise a Trojan “enemy” in the highlighted text.
The Dominant Selector
---------------------
The Trojan explanation example serves to illustrate that the select-predict process is not always representative of the expectation that a human will have when given a highlight interpretation of this process. Unfortunately, our troubles are not limited to Trojan explanations, for it is possible to represent unintuitive decision processes through faithful highlights in other ways.
To illustrate this point, consider the case of the *binary sentiment analysis* task, where the model predicts the polarity of a particular snippet of text. Assume that two fictional selector-predictor models attempt to classify a complex, mixed-polarity review of a product. Table \[fig:case2-ex\] describes two possible highlights faithfully attributed to these two models, on an example selected from the Amazon Reviews sentiment classification dataset.
Although the models made the same decision, the implied reasoning process is wildly different thanks to the different highlights chosen by the selector. In model (a), the selector clearly performed the entirety of the decision process. Had selector (a) chosen a negative-sentiment word such as “expensive”, the predictor would reasonably predict a negative sentiment for this example. In other words, for model (a), the selector dictates the entire decision process, even if the highlight is not a “Trojan” by any means.
Comparatively, the selector of model (b) performed a far simpler job, highlighting a selective summary of the review, with both positive and negative sentiments. The predictor then made a decision based on the information. How the predictor made the decision based on this highlighted text is unclear, but the division of roles between the selector and predictor is more intuitive to the observer than in the case of model (a).
Let us focus on the *dispute* use-case. Given a claim such as “the sturdiness of the product was not important to the final decision”, the claim appears safer in the case of model (b) than it is in the case of model (a), since in model (a), it is difficult to say how the selector arrived at the selected highlight—and thus, difficult to dispute a decision following the claim. There is an *implicit understanding* of the highlight of model (b) as a descriptor of a decision process that supports this claim.
#### To clarify,
in this failure case, the selector is *not* limited to reducing the input to a trivially solved snippet (such as selecting single words strongly associated with a class, per the example), but that **the selector is capable of dictating the final decision with power not intended to it**: *even an entirely random predictor can be manipulated to perform well by a crafty selector.* But *why* is this considered problematic, if it remains within the confines of faithful highlights?
Loss of Performance
-------------------
The selector-predictor format implies a loss of performance in comparison to models that classify the full available text in “one” opaque step—refer to Table \[tab:compromise\] for examples on the degree of decrease in performance on various classification datasets. In many cases, this decrease is severe. Although this phenomenon is treated as a reasonable and matter-of-fact necessity by literature on highlight interpretations and rationales, the intuition behind it is not trivial to us.
Naturally, humans are able to provide highlights of decisions without any loss of accuracy, even when the selector and predictor are separate people [@jain2020]. In addition, while interpretability may be prioritized over state-of-the-art performance in certain use-cases, there are also cases that will disallow the implementation of artificial models unless they are strong and interpretable, *simultaneously*.[^4]
Often, *sufficiency* is deemed to be a desiderata of highlights: the highlighted text should be sufficient to make an informed judgement [@yu2019comprehensiveness-sufficiency; @eraser2019] towards the correct label. Under the selector-predictor setup, the highlighted text is at least guaranteed to be sufficient to predict the same result as the model.
In this context, the following question is critical to the design of highlight models: why do models that provide highlight interpretations surrender performance to do so? Is this phenomenon necessary, or desirable? And *why*?
This question can be re-interpreted in the following way: in the *causal chain* of events behind the decision process for a given task, which comes first: the *prediction*, or the *highlight selection*? Is the answer constant, or contextual?
Consider the case of *agreement classification*, where the task is to classify whether a given snippet contains grammatically disagreeing words. In this case, most would agree that it is impossible to provide a sufficient highlight without making a decision first. However, for example, when deciding the polarity of a review snippet in a *sentiment classification* task, it may be natural to first make a selection of relevant phrases before finally making a decision based on them. Can we say that it is appropriate to surrender performance in order to provide highlight interpretations in either case, in both, or in neither?
Conclusion
----------
The described failure cases shed light on a missing step in the derivation of interpretations that are useful, despite adhering faithfully to a select-predict model. We conclude that **faithfulness is insufficient to formalize the desired properties of a behavioral explanation behind the decision of an artificial intelligence model.**
Plausibility is Not the Answer {#sec:plausibility}
==============================
Following the failure cases described in Section \[sec:limitations\], it may be theorized that plausibility is a desirable, or even necessary, condition for useful highlights and interpretations in general. After all, Trojan explanations are by default implausible. However, we argue that this is far from the case.
Plausibility [@wiegreffe2019attentionisnotnot] or persuasiveness [@DBLP:journals/corr/abs-1711-07414] is the property of how convincing the interpretation is towards the model prediction, regardless of whether the model was correct or not, and regardless of whether the interpretation is faithful or not. It is a property inspired by human-given explanations, which are post-hoc stories generated to plausibly justify our actions [@rudin2019stop]. Plausibility is generally quantified by the degree that the model’s highlights resemble gold-annotated highlights given by humans [@bastings-etal-2019-interpretable] or by querying for the feedback of humans directly [@jain2020].
Supposing that faithfulness has been “achieved” (unclear as that condition may be), plausibility is still without utility unless this plausibility is correlated with the likelihood of the model to be correct [@jacovi2020]. Although it is possible to quantify this correlation via Human-Computer Interaction (HCI) assignments (e.g., @10.1145/3301275.3302265), it remains irrelevant to other use-cases of interpretability, since model correctness is intractable. The failures discussed above stem not from how convincing the interpretation is, but from *how well the user understands the reasoning process of the model*. If the user is able to comprehend the steps that the model has taken to its decision, then the user will be the one to decide whether these steps are plausible or not, based on how closely these steps fit the user’s prior knowledge on whatever correct steps should be taken—whether the user knows the correct answer or not, and whether the model is correct or not.
For example, in Figure \[fig:case2-ex\], we are not interested in whether the highlighted text is plausible as an explanation to the decision—in which case, model (a) will likely be deemed superior—but rather, which highlight implies a more coherent decision process that the user can comprehend. It is the latter property, rather than the former, that will allow the model to be useful in any of the use-cases of highlight interpretations, such as dispute or debug.
This means that the important attribute of model (b)’s highlight is not that it may be a closer match to a gold-annotated human highlight; but that the roles of the selector and predictor resemble those of a human decision maker. The difference is that even if the model made a “wrong” decision, the latter attribute will stay valid, unlike the former.
On Faithfulness, Plausibility, and Explainability from the Science of Human Explanations {#sec:social}
========================================================================================
Clearly, the mathematical foundation of machine learning and natural language processing is insufficient to tackle the underlying issue behind the painful symptoms described in Section \[sec:limitations\]. In fact, formalizing the problem itself is difficult. What enables a faithful explanation to be understood as accurate to the model? And what causes an explanation to be perceived as a Trojan?
Although work by the community in this direction is well placed, it often neglects to draw upon the extremely vast library of work on the science of human explanation. As a result, some of the work re-invents the wheel, and we have yet arrived at a satisfactory formalization. We will attempt to better understand the problem by looking to the social sciences, and particularly philosophical research on human explanations of (human) behavior, to assist us.[^5]
The Composition of Explanations
-------------------------------
@miller2017social describes explanations of behavior as a social interaction of knowledge transfer between the explainer and the explainee, and thus *they are contextual*, and can be perceived differently depending on this context. Two central pillars of the explanation are *causal attribution*—the attribution of a causal chain[^6] of events to the behavior—and *social attribution*—the attribution of behavior to others [@heider58].
#### Causal attribution describes faithfulness.
We note a stark parallel between causal attribution and faithfulness, as it is understood by the NLP community: for example, the select-predict composition of modules defines an unfolding causal chain where the selector hides portions of the input, causing the predictor to make a decision based on the remaining portions.
#### Social attribution is missing.
@heider44 describe an experiment where participants attribute human concepts of emotion, intentionality and behavior to animated shapes. Clearly, the same phenomenon persists when humans attempt to understand the predictions of artificial models. What is the behavior attributed to the select-predict causal chain? And can models be constrained to adhere to this attribution?
Although informally, prior work on highlights has considered such factors before. @lei16 describe desiderata for highlights as being “short and consecutive”, and @jain2020 interpreted “short” as “around the same length as that of human-annotated highlights”. We assert that the true nature of these claims is an attempt to constrain highlights to the social behavior implicitly attributed to them by human observers in the select-predict paradigm. We discuss this further in the next section.
#### Plausibility as an incentive of the explainer, and not as a property of the explanation.
The utility of human explanations can be categorized across multiple axes [@miller2017social], among them are (1) *learning* a better internal model for future decisions and calculations [@lombrozo2006structure; @williams2013hazards]; (2) *examination* to verify the explainer has a correct internal prediction model; (3) *teaching*[^7] to modify the internal model of the explainer towards a more correct one (can be seen as the opposite end of (1)); (4) *assignment of blame* to a component of the internal model; and finally, (5) *justification* and *persuasion*.
Critically, the goal of justification and persuasion by the explainer may not necessarily be the goal of the explainee. Indeed, in the case of AI explainability, it is *not* a goal of the explainee to be persuaded that the decision is correct (even when it is), but to understand the decision process. If plausibility is a goal of the artificial model, this perspective outlines a game theoretic mismatch of incentives between the two players. And specifically in cases where the model is incorrect, it is interpreted as the difference between an innocent mistake and an intentional lie—of course, lying is considered more unethical. As a result, **modeling and pursuing plausibility in AI explanations is an ethical issue**.
Aligned Faithfulness {#sec:perceived-definition}
====================
We have covered the separation of causal attribution to the model from attribution of social behavior to it. In human explanations, this separation is formulated as a multi-step process in the transfer of knowledge from one person to the other. Unique to artificial explainers is a problem where there is a *misalignment* between the causal chain behind the decision, and the social attribution to it, as the (artificial) decision process does not necessarily resemble human behavior.
We claim that this problem is the root cause behind the symptoms described in Section \[sec:limitations\]. In this section we define the general desiderata of interpretations to satisfy to avoid this issue, separated from the narrative of highlight interpretations.
Definition
----------
By presenting to the user the causal pipeline of decisions in the model decision process as an interpretation of the decision process, the user naturally conjures social intent behind this pipeline. This intent is formalized as a set of constraints on the possible range of decisions at each step in the causal chain, which the model must adhere to in order to be considered as comprehensible to the user.
For an interpretation method to accurately describe the causal chain of decisions in a model decision, we say that it is *faithful*; and having accomplished that, for the method and model to adhere to the social attribution of behavior by humans, we say that the interpretation is *aligned*, short for *human-aligned*, or “aligned with human expectations”. Furthermore, we claim that this attribution of behavior is heavily contextual on the task and use-case of the model, and that it is incompatible with plausibility.
Social Attribution of Highlight Interpretations {#sec:highlight-attribution}
===============================================
As previously mentioned, unique to our setting in NLP and ML is the fact that the social attribution must lead the design of the causal chain, since we have control over one and not the other. In other words, we must first identify the behavior expected of the decision process, and constrain the decision process around it.
We arrive at two parallel attributions of human behavior to highlights, where each carries separate and distinct constraints on whether the highlight can describe behavior aligned with human intent.
#### Highlights as *summaries*.
The highlight serves as an extractive summary of the input text, filtering irrelevant information, making it easier to focus on the portions of the input most important.[^8] To illustrate, recall a student who is marking sentences and paragraphs in a book to make studying for a test easier. The highlight is merely considered a compressed version of the input, with sufficient information to make informed decisions in the future. It is not selected with an answer in mind, but in anticipation that an answer will be derived in the future, for a question that may not have even been asked yet.
#### Highlights as *evidence*.
Another name for highlight explanations in the NLP community is *extractive rationales*, or *rationalization*, due to this human characterization of the artificial explainer: the highlight serves as evidence towards a *prior* decision. The decision process behind the prior decision must consider the highlight as supporting evidence of the decision, whether the highlight is sufficient, comprehensive, or *neither*.
Issues with *Select-Predict*
----------------------------
Unfortunately, the select-predict causal chain supports neither attribution:
1. A summarizing selector is expected to filter out info which is irrelevant to making an informed decision, *without* making any decision in doing so. It is not guaranteed by black-box selectors which were explicitly trained on the end-task.
2. An evidencing or rationalizing selector should select the part of the input that supports the final decision *without influencing* this decision. This is clearly not the case in select-predict models.
The issues discussed in Section \[sec:limitations\] are direct results of the above conflation of interests. Trojan highlights and the dominant selector are a result of a selector that makes hidden and unintended decisions, so they serve as neither summary nor evidence towards the predictor’s decision. Loss of performance is due to the selector acting as an imperfect summarizer.
*Predict-Select-Verify* {#sec:highlight-as-evidence}
-----------------------
We focus on highlights as evidence: we propose the *predict-select-verify* causal chain as a solution that can be constrained to provide highlights as evidence. The decision pipeline is as followings:
1. The predictor $m_p$ makes a prediction $\hat{y}:=m_p(x)$ on the full text.
2. The selector $m_s$ selects $h := m_s(x)$ such that $m_p(h \odot x) = \hat{y}$.
In this framework, the selector provides evidence which is verified to be useful to the predictor towards a particular decision. Importantly, the final decision has been made on the full text, and the selector is constrained to provide a highlight that adheres to this exact decision. The selector does not purport to provide a highlight which is comprehensive of all evidence considered by the predictor, but it provides a *guarantee* that the highlighted text is sufficient for this prediction.
It is trivial to see that the predict-select-verify chain addresses all points of Section \[sec:limitations\]: Trojan highlights and dominant selectors are impossible, as the selector is constrained to only provide “retroactive” selections towards a specific priory-decided prediction. In other words, the selector has no power to influence the decision, since the decision was already made without its intervention. Loss of performance is impossible as the predictor is not constrained to make predictions on possibly insufficient subsets of the input (as it would be under a summarizing selector).
Constructing a *Predict-Select-Verify* Model with Contrastive Explanations {#sec:contrastive}
==========================================================================
In order to design a model adhering to the aforementioned constraints, we require solutions for the predictor and for the selector.
#### Predictor.
The predictor is constrained to be able to accept both full-text inputs and highlighted inputs. For this reason, we use masked language modeling (MLM) models, such as BERT, [@devlin2018bert] fine-tuned on the downstream task. The MLM pre-training is performed by recovering partially masked text, which conveniently suits our needs. We additionally provide randomly-highlighted inputs to the model during fine-tuning.
#### Selector.
The selector is constrained to select highlights for which the predictor made the same decision as it did on the full text. However, there are likely many possible choices that the selector may make under this constraints, as there are many possible highlights that all result in the same decision by the predictor. **We wish for the selector to select *meaningful* evidence to the predictor’s decision.** What is meaningful evidence? To answer this question, we again refer to cognitive science on necessary attributes of explanations that are easy to comprehend by humans.
Fact and Foil
-------------
An especially relevant finding of social science literature is of *contrastive explanations*, following the notion that the question “why $P$?” is may be followed by an addendum: “why $P$, rather than $Q$?” [@hilton1988logic]. We refer to $P$ as the *fact* and $Q$ as the *foil* [@lipton1990contrastive]. The concrete valuation in the community is that in the vast majority of cases, the cognitive burden of a “complete” explanation, i.e. where $Q$ is $\neg P$, is too great, and thus $Q$ is selected as a subset of all possible foils [@hilton1986knowledge; @Hesslow1988], and often not explicitly, but implicitly derived from context.
In classification tasks, the implication is that an interpretation of a prediction of a specific class is hard to understand, and should be contextualized by *the preference of the class over another*—and the selection of the foil (the non-predicted class) is non-trivial, and a subject of ongoing discussion even in human explanations literature.
Contrastive explanations have many implications for explanations in AI, and particularly for highlights, as a vehicle for explanations that are easy to understand. Although there is a modest body of work on contrastive explanations in machine learning [@Dhurandhar2018ExplanationsBO; @Chakraborti2019ContrastiveAF; @Chen2020TowardsTR], to our knowledge, there are none in NLP.
Contrastive Highlights
----------------------
An explanation in a classification setting should not only addresses the fact, but do so against a foil—where the fact is the class predicted by the model, and the foil is another class. Given two classes $c$ and $\hat{c}$,[^9] where $m_p(x) = c$ we are interested in deriving a contrast highlight explanation towards the question: “why did you choose $c$, rather than $\hat{c}$?”.
We assume a scenario where, having observed $c$, the user is aware of some highlight $h$ which should serve, they believe, as evidence for class $\hat{c}$.[^10] In other words, we assume the user believes a pipeline where $m_p(x) = \hat{c}$ and $m_s(x) = h$ is reasonable.
If $m_p(h \odot x) \neq \hat{c}$, then the user is made aware that the predictor disagrees that $h$ serves as evidence for $\hat{c}$.
Otherwise, $m_p(h \odot x) = \hat{c}$. We define: $$h_c := \operatorname*{argmin}_{ \substack{ h + \hat{h} \\
\text{s.\,t.}\, \, \, |\hat{h}| > 0 \\
\land \, m_p((h + \hat{h}) \odot x) = c
}
}
|h + \hat{h}|.$$
$h_c$ is the minimal highlight containing $h$ such that $m_p(h_c \odot x) = c$. Intuitively, the claim by the model is as such: “Because of $\hat{h}$, I consider $h_c$ as evidence towards $c$ despite $h$.”
We show examples of an implementation of this process in Table \[fig:topic-ex\] on examples from the AG News dataset. For illustration purposes, we selected incorrectly-classified examples, and selected the foil to be the true label of the example. The highlight for the foil was chosen by us. We stress that while the examples presented in these figures appear reasonable, the true goal of this method is not to provide highlights that seem justified, but to provide a framework which allows models to be meaningfully incorporated in use-cases of *dispute*, *debug*, and *advice*, with robust and proven guarantees of behavior.
For example, in each of the example use-cases:
1. **Dispute**: The user intends on verifying whether the model “correctly” considered a specific portion of the input when making the decision. I.e., the model made decision $\hat{c}$, where the user believes decision $c$ should have been made, and is supported by evidence $h \odot x$. If $m_p(h \odot x) \neq \hat{c}$, it is possible to dispute the claim that the model interpreted $h \odot x$ with “correct” evidence intent. Otherwise the dispute claim cannot be made, as the model provably considered $h$ as evidence for $c$, yet insufficiently so when combined with $\hat{h}$ as $h_c \odot x$.
2. **Debug**: Assuming $m_p(x)$ is incorrect, the user intends on performing error analysis by observing which portion of the input is sufficient to steer the predictor away from the correct decision $c$. This is provided by $\hat{h}$.
3. **Advice**: When the user is unaware of the answer, and is seeking perspective from a trustworthy model: then the user is given explicit feedback on which part of the input the model “believes” is sufficient to overturn the signal in $h$ towards $c$. Otherwise, if the model is not considered trustworthy, the user may gain or reduce trust by observing whether $m(h \odot x)$ and $\hat{h}$ align with user priors.
Discussion
==========
#### Causal attribution of heat-maps.
Recent work on the faithfulness of attention heat-maps [@Baan2019DoTA; @Pruthi2019LearningTD; @jain2019attentionisnot; @sofia-isattentioninterpretable; @wiegreffe2019attentionisnotnot] or saliency distributions [@alvarez2018robustness; @kindermans2017unreliability-of-saliency] cast doubt on their faithfulness as indicators to the significance of parts of the input (to a model decision). We argue that is a natural conclusion from the fact that as a community, we have not envisioned an appropriate causal chain that utilizes heat-maps in the decision process, reinforcing the claims in this work on the parallel between causal attribution and faithfulness.
#### Inter-disciplinary research.
Research on explanations in artificial intelligence will benefit from a deeper inter-disciplinary perspective on two axes: (1) literature on causality and causal attribution; and (2) literature on the social perception and attribution of human-like behavior to causal chains of decision or behavior.
Related Work
------------
Relevant to the issue of how the explanation is comprehended by people is *simulatability* [@kim2017interpretability], or the degree to which humans can simulate model decisions. @hase2020simulatability further refines simulatability evaluation by defining testing regimen where it can be properly assessed without false signal. While quantifying simulatability may be related to aligned faithfulness in some way, they are decidedly different, since e.g., simulatabiliy will not necessarily detect dominant selectors (Section \[sec:limitations\]), such as those which reduce the input to a trivial instance of a prior task prediction.
*Predict-select-verify* is reminiscent of iterative erasure as described by [@feng2018pathologies]. By iteratively removing “significant” tokens in the input, the authors discovered that a surprisingly small portion of the input could be interpreted as evidence for the model to make the prediction, leading to conclusions on the pathological nature of neural models and sensitivity to badly-structured inputs. This experiment retroactively serves as a successful use-case of *examination* and *debugging* using our described formulation.
@Kottur2017NaturalLD [@sanjay2020] describe failure cases of Trojans where communicating modules encode information unintuitively, despite an interpretable interface between them. @sanjay2020 additionally infer a particular social attribution on the causal chain of their proposed model for tasks of compositional reasoning with neural modular networks, with the motivation of overcoming the Trojan issue, though the work is not formalized as such.
The approach by @DBLP:conf/nips/ChangZYJ19 for class-wise highlights is reminiscent of contrastive highlights, but nevertheless distinct, since such highlights still explain a fact against all possible foils.
Conclusion
==========
Highlights are a popular format for explanations of decisions on textual inputs, relatively unique in that there are models available today with the ability to derive highlights faithfully. We analyze highlights as a guiding use-case in the pursuit of more rigorous formalization of what makes a quality explanation of an artificial intelligence model.
We redefine faithfulness as the degree to which the interpretation of causal events accurately represents the causal chain of decision making towards the decision. Next, we define aligned faithfulness as the degree to which the various decisions in the causal chain are constrained by the social attribution of behavior that humans expect from the interpretation. The two steps of causal attribution and social attribution *together* complete the process of explaining the decision process of the model to the human observer.
Using this formalization, we characterize various failures in faithful highlights that “seemed” strange, but could not be properly described previously, noting they are not properly constrained by their social attribution as summaries. We propose an alternative causal chain which can be constrained by the attribution of highlights as evidence. Finally, we illustrate a possible implementation of this model with practical utility of disputing, debugging or being advised by model decisions, by formalizing *contrastive explanations* in the highlight format.
Experimental Setup
==================
Datasets
--------
We elaborate on the experiments used to illustrate the points raised in Section \[sec:limitations\]. For the most part, these tasks are within precedence of relevant literature on highlights and rationales. The datasets use prior train/validation/test set splits when available. When only train/test were available, 20% of the train set was held out for validation. We elaborate below when no splits are available. Some of these datasets were not used by us explicitly, but we mention them here as we include performance reports on them (provided in previous papers).
#### Stanford Sentiment Treebank (SST-2/3/5) [@DBLP:conf/emnlp/SocherPWCMNP13]
Sentences or snippets of movie reviews with 2, 3 or 5-bucket polarity.
#### AG’s News Corpus (AG News) [@DBLP:conf/www/CorszoGR05]
News articles categorized into four topics of Science/Tech, Sports, Business, World.
#### Evidential Inference (Ev.Inf.) [@lehman-etal-2019-inferring]
Based on biomedical articles of controlled trials, and a given intervention, the task is to infer the outcome of the experiment as a result of the intervention (among *significantly increased*, *significantly decreased*, and *no significant change*), and give supporting portions of the text towards this conclusion. We use the article abstracts, and do not evaluate the extracted supporting snippet.
#### Movies [@DBLP:conf/emnlp/ZaidanE08]
Movie reviews labeled for binary polarity. Human-annotated highlights were collected for this dataset by @eraser2019.
#### Multi RC [@DBLP:conf/naacl/KhashabiCRUR18]
Binary classification of questions, answers and supporting snippets, into *true* and *false* labels. The binary setup was formulated by @eraser2019.
#### LGD [@linzen2016assessing]
A collection of template text sentences, each with two options for an agreeing or disagreeing pair of words, for evaluation on the syntactic ability of models. We converted this dataset into a binary classification dataset by requiring prediction on whether a given sentence contains a disagreement or whether all words are in agreement. This dataset is unique in that the choice of highlight is strict and punishing if not “correct”, providing a strong challenge to the selector in select-predict models. We have employed a train/dev/test split of 50%/25%/25% and enforced no overlap in sentences between the splits.
#### 20 Newsgroups (20 News) [@DBLP:conf/icml/Lang95; @DBLP:journals/ml/NigamMTM00]
A collection of newsgroup documents partitioned evenly across twenty categories.
#### Amazon Reviews [@DBLP:conf/recsys/McAuleyL13] and Elec [@DBLP:conf/naacl/Johnson015]
Amazon product reviews with binary polarity based on user star rating. Elec refers to electronic products only.
#### Beer [@DBLP:conf/icdm/McAuleyLJ12]
A collection of BeerAdvocate reviews with polarity scores on four aspects of appearance, smell, palate and taste.
[^1]: In the scope of this work, we use the term “highlights” for this type of explanation. Although the term “rationale” [@lei16] is more commonly used for this format in the NLP community, it is controversial, as it has been used ambiguously for different things in NLP literature historically (e.g., @DBLP:conf/naacl/ZaidanEP07 [@DBLP:conf/emnlp/BaoCYB18; @eraser2019]), it is a non-descriptive term and unintuitive without prior knowledge, and is seldom known or used outside of NLP disciplines. Most importantly, “rationalization” attributes human-like social behavior which is not necessarily compatible with the artificial explainer’s incentive, unless modeled explicitly. We elaborate on this justification further in Section \[sec:highlight-attribution\].
[^2]: It can be argued that for some tasks, it is reasonable for $h$ to encode label information, e.g., when particular labels are more often attributed by words in the beginning of the text, compared to others. Similar conclusions apply to other types of Trojans. How could we clarify when a property is or is not a Trojan for a given task?
[^3]: By us and others. Although presented in very different context, @sanjay2020 do in fact show cases of Trojan explanations in practice for a different class of compositional models.
[^4]: For example, consider the case of a doctor or patient seeking life-saving advice—it is difficult to quantify a trade-off between performance and explanation ability.
[^5]: Refer to @miller2017social for a substantial survey in this area, which was especially motivating to us.
[^6]: See @hilton05 for a breakdown of types of causal chains; we focus on unfolding chains in this work, but others may be relevant as well.
[^7]: Although (1) and (3) are considered one-and-the-same in the social sciences, we disentangle them as that is only the case when the explainer and explainee are both human.
[^8]: The definition of a summary and the utility of summarization in tasks [@yu2013-summaries] is beyond the scope of this work.
[^9]: Selecting the foil, or selecting what to explain, is a difficult and interesting problem even in philosophical literature [@Hesslow1988; @mcgill93; @chin2017contrastive]. In the classification setting, it is relatively simple, as we may request the foil (class) from the user, or provide separate contrastive explanations for each foil.
[^10]: In this work, we assume both $\hat{c}$ and $h$ are provided by the user. Additional strategies are possible, such as deriving candidates for $h$ (heuristically or otherwise) given $\hat{c}$, and allowing the user to choose and modify the $h$ candidates.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
, R. Fiore$^a$, M. Gravina$^a$ and A. Papa$^a$\
Dipartimento di Fisica, Università della Calabria,\
and Istituto Nazionale di Fisica Nucleare, Gruppo collegato di Cosenza\
I–87036 Arcavacata di Rende, Cosenza, Italy\
E-mail:
title: 'Screening masses in the SU(3) pure gauge theory and universality'
---
Introduction
============
In this work we compare the spectrum of the inverse decay lengths of Polyakov loop correlators in the (3+1)$d$ SU(3) gauge theory in the deconfined phase near the transition with the spectrum of massive excitations of the 3$d$ 3-state Potts model in the broken phase near the transition at zero magnetic field, which were determined in Ref. [@FFGP06]. The aim of the work is to verify if and to what extent the Svetitsky-Yaffe conjecture [@Svetitsky:1982gs] also holds for theories which undergo a [*weakly*]{} first order phase transition, using mass ratios as a probe. In particular we focus on the low-lying masses in two different sectors of parity and orbital angular momentum, 0$^+$ and 2$^+$. We expect that, if universality would apply in strict sense, these spectra should exhibit the same pattern, as suggested by several numerical determinations in the 3$d$ Ising class [@Caselle:1999tm; @Caselle:2001im; @Fiore:2002fj; @Fiore-Papa-Provero-2003].
We extend our numerical analysis to temperatures far away from the transition temperature $T_t$ in order to look for possible “scaling” of the fundamental masses with temperature. Moreover, we consider also the screening masses resulting from correlators of the real parts and of the imaginary parts of the Polyakov loop. These determinations can represent useful benchmarks for effective models of the high-temperature phase of SU(3), such as those based on mean-field theories of the Polyakov loop, suggested by R. Pisarski [@Pisarski0101168].
Screening masses from Polyakov loop correlators
===============================================
Screening masses are defined as the inverse decay lengths of the Yukawa-like potential between two static sources. They are generally determined through the correlation of suitable operators. In our case correlations are between operators with different spatial separation. The general large distance behavior for the correlation function $G(|z_1-z_2|)$, in an infinite lattice, is: $$G(|z_1-z_2|)= \sum_n a_n e^{-m_n |z_1-z_2|}\ ,
\label{corr-funct}$$ where $m_0$ is the fundamental mass, while $m_1$, $m_2$, ... are higher masses with the same angular momentum and parity quantum numbers of the fundamental mass. On a periodic lattice the above equation must be modified by the inclusion of the so called “echo” term: $$G(|z_1-z_2|)= \sum_n a_n\biggl[e^{-m_n |z_1-z_2|}+e^{-m_n (N_z-|z_1-z_2|)}\biggr]\ .
\label{corr-funct_echo}$$ The fundamental mass in a definite channel can be extracted from wall-wall correlators by looking for a plateau of the effective mass at large distances, $$m_{\mbox{\footnotesize eff}}(z)= \ln \frac{G(z-1)}{G(z)} \ .
\label{m_eff}$$
In the 0$^+$-channel, the connected wall-wall correlator in the $z$-direction is defined as $$G(|z_1-z_2|)=\mbox{Re}\langle \bar{P}(z_1) \bar{P}(z_2)^\dagger \rangle -|\langle P \rangle|^2\ ,
\label{corr0}$$ where $$\bar{P}(z)=\frac{1}{N_xN_y}\sum_{n_x=1}^{N_x} \sum_{n_y=1}^{N_y}P(n_x a,n_y a,z)\ ,
\label{wall0}$$ represents the Polyakov loop averaged over the $xy$-plane at a given $z$. [^1] The wall average implies the projection at zero momentum in the $xy$-plane.
For the 2$^+$-channel, we used the variational method [@Kronfeld; @Luscher-Wolff] (for more details, see [@FFGP07] and references therein.)
Our choice of wall-averaged operators in the 2$^+$-channel is inspired by Ref. [@Caselle1997] and reads $$\bar{P}_n(z)=\frac{1}{N_xN_y}\sum_{n_x=1}^{N_x} \sum_{n_y=1}^{N_y}P(n_x a,n_y a,z)
\biggl[P(n_x a+na,n_y a,z)-P(n_x a,n_y a+na,z)\biggr]\ .
\label{wall2}$$ In most cases we have taken 8 operators, corresponding to different values of $n$, with the largest $n$ almost reaching the spatial lattice size $N_x$.
We consider also correlators of the (wall-averaged) real and imaginary parts of the Polyakov loop, defined as $$\begin{aligned}
G_R(|z_1-z_2|) &=& \langle \mbox{Re} \bar P(z_1) \mbox{Re} \bar P(z_2)\rangle -
\langle \mbox{Re} \bar P(z_1)\rangle \langle \mbox{Re} \bar P(z_2)\rangle \;,
\label{Dumitru2:1}\\
G_I(|z_1-z_2|) &=& \langle \mbox{Im} \bar P(z_1) \mbox{Im} \bar P(z_2)\rangle\;.
\label{Dumitru2:2}\end{aligned}$$ The corresponding screening masses, $\hat m_R$ and $\hat m_I$, can be extracted in the same way as for the 0$^+$ mass. We have studied the ratio $m_I/m_R$ over a wide interval of temperatures above the transition temperature $T_t$ of (3+1)$d$ SU(3) and seen how it compares with the prediction from high-temperature perturbation theory, according to which it should be equal to 3/2 [@Nadkarni:1986cz; @Dumitru:2002cf], and with the prediction from the mean-field Polyakov loop model of Ref. [@Pisarski0110214], according to which it should be equal to 3 in the transition region. The interplay between the two regimes should delimit the region where mean-field Polyakov loop models should be effective.
Numerical results {#results}
=================
We used the Wilson lattice action and generated Monte Carlo configurations by a combination of the modified Metropolis algorithm [@Cabibbo-Marinari] with over-relaxation on SU(2) subgroups [@Adler]. The error analysis was performed by the jackknife method over bins at different blocking levels. We performed our simulations on a 16$^3\times$4 lattice, for which $\beta_t=5.6908(2)$ [@Boyd:1996bx], over an interval of $\beta$ values ranging from 5.69 to 9.0.
Screening masses are determined from the plateau of $m_{\mbox{\footnotesize eff}}(z)$ as a function of the wall separation $z$. In each case, the [*plateau mass*]{} is taken as the effective mass (with its error) belonging to the [*plateau*]{} and having the minimal uncertainty. We define [*plateau*]{} the largest set of consecutive data points, consistent with each other within 1$\sigma$. This procedure is more conservative than identifying the plateau mass and its error as the results of a fit with a constant on the effective masses $m_{\mbox{\footnotesize eff}}(z)$, for large enough $z$.
Just above the critical value $\beta_t$ we find a large correlation length, which is not of physical relevance. It is instead a genuine finite size effect [@Gavai-Karsch-Petersson] related to [*tunneling*]{} between degenerate vacua. This effect disappears by going to larger lattice volumes or moving away from $\beta_t$ in the deconfined phase. Tunneling can occur between the symmetric and the broken phase, and between the three degenerate vacua of the deconfined phase. When tunneling is active, the correlation function has the following expression [@Gavai-Karsch-Petersson]: $$G(|z_1-z_2|)\sim a_0e^{-m_0 |z_1-z_2|}+b_0 e^{-m_t |z_1-z_2|}\;,
\label{corr_funct_tunneling}$$ where $m_t$ is the inverse of the tunneling correlation length and is generally much smaller than the fundamental mass $m_0$ and therefore behaves as a constant additive term in the correlation function. [^2] The dependence on $m_t$ in the correlation function can be removed by extracting the effective mass by use of the combination $$m_{\mbox{\footnotesize eff}}(z)= \ln \frac{G(z)-G(z+1)}{G(z+1)-G(z+2)} \ .
\label{m_eff_1}$$ A typical example of the behavior of the effective mass with $z$ is shown in Fig. \[eff\_masses\_5.75\] for the 0$^+$ and the 2$^+$ channels. In Figs. \[masses\_vs\_beta\_1\] and \[masses\_vs\_beta\_2\] we show the behavior with $\beta$ of $\hat m_{0^+}$, $\hat m_{2^+}$, $\hat m_R$ and $\hat m_I$.
![Effective mass in the 0$^+$ (left) and the 2$^+$ (right) channel as a function of the separation between walls on the $(x,y)$ plane at $\beta=5.75$.[]{data-label="eff_masses_5.75"}](massa0.eps "fig:") ![Effective mass in the 0$^+$ (left) and the 2$^+$ (right) channel as a function of the separation between walls on the $(x,y)$ plane at $\beta=5.75$.[]{data-label="eff_masses_5.75"}](massa2.eps "fig:")
We observe from that $\hat m_{0^+}$ and $\hat m_R$ are consistent within statistical errors, this indicating that the Polyakov loop correlation is dominated by the correlation between the real parts. We can see that the fundamental mass in the 0$^+$ channel, as well as $\hat m_R$, becomes much smaller than 1 at $\beta_t$, as expected for a weakly first order phase transition. In the cases of $\hat m_{0^+}$ and of $\hat m_R$ we have made some determinations [*below*]{} $\beta_t$ (see Figs. \[masses\_vs\_beta\_1\] and \[masses\_vs\_beta\_2\]). It turns out that masses in lattice units take their minimum value just at $\beta_t$, where there is a “cusp” in the $\beta$-dependence. Such a behaviour was observed also by the authors of Ref. [@Datta:2002je], whose results, when the comparison is possible, agree with ours.
![Screening mass in the 0$^+$ channel (left) and in the 2$^+$ channel (right) [*vs.*]{} $\beta$.[]{data-label="masses_vs_beta_1"}](mass0_vs_beta.eps "fig:") ![Screening mass in the 0$^+$ channel (left) and in the 2$^+$ channel (right) [*vs.*]{} $\beta$.[]{data-label="masses_vs_beta_1"}](mass2_vs_beta.eps "fig:")
![Screening masses ${\hat m}_R$ (left) and ${\hat m}_I$ (right) [*vs.*]{} $\beta$.[]{data-label="masses_vs_beta_2"}](mass_r_vs_beta.eps "fig:") ![Screening masses ${\hat m}_R$ (left) and ${\hat m}_I$ (right) [*vs.*]{} $\beta$.[]{data-label="masses_vs_beta_2"}](mass_i_vs_beta.eps "fig:")
We have also looked for a scaling law for the fundamental mass in the $0^+$ channel, but with the understanding that any second-order-like scaling law, when applied to the region near a first order phase transition, should be taken as an [*effective*]{} description, which cannot hold too close to the transition point. With this spirit, we have compared our data with the scaling law $$\label{scal_rel}
\Bigg( \frac{\beta_1-\beta_t}{\beta_2-\beta_t} \Bigg)^{\nu} \sim
\frac{\hat m_{0^+}(\beta_1)}{\hat m_{0^+}(\beta_2)} \ ,$$ where $\hat m_{0^+}(\beta_1)$ and $\hat m_{0^+}(\beta_2)$ are the fundamental masses in the $0^+$ channel at $\beta_1$ and $\beta_2$, respectively. We have considered several choices of $\beta_1$ and found that for each of them there is a wide “window” of $\beta$ values above $\beta_t$ where the scaling law (\[scal\_rel\]) works, with a “dynamical” exponent $\nu$ (see Ref. [@FFGP07] for a details). For $\beta_1$=5.72 we have calculated also the $\chi^2$/d.o.f. when $\nu$ is put exactly equal to 1/3 (suggested in Ref. [@Fisher-Berker] to apply to the [*standard*]{} correlation function), getting $\chi^2$/d.o.f.=0.75 in the window from $\beta=5.715$ to $\beta=5.78$. In Fig. \[scaling\] we show, for this choice of $\beta_1$, the comparison between data and the “scaling” function with $\nu$ set equal to 1/3.\
![Comparison between the scaling function $[(\beta_1-\beta_t)/(\beta-\beta_t)]^{1/3}$ and the mass ratio $m_{0^+}(\beta_1)/m_{0^+}(\beta)$ for varying $\beta$, with $\beta_1$=5.72.[]{data-label="scaling"}](scaling_zoom.eps)
Then, we have considered the $\beta$-dependence of the ratio $m_{2^+}/m_{0^+}$, shown in Fig. \[ratio\_20\]. We have found that this ratio can be interpolated with a constant in the interval from $\beta_t$ to $\beta=5.77$. This constant turned out to be 3.172(65), with a $\chi^2$/d.o.f equal to 1.085. In the fit we excluded the point at $\beta$=5.695, for which the determination of $m_{2^+}$ is probably to be rejected. If the point at $\beta$=5.695 is included, the constant becomes 3.214(64) with $\chi^2$/d.o.f =2.21. The fact that the ratio $m_{2^+}/m_{0^+}$ is compatible with a constant in the mentioned interval suggests that $\hat m_{2^+}$ scales similarly to $\hat m_{0^+}$ near the transition. This constant turns out to be larger than the ratio between the lowest massive excitations in the same channels in the broken phase of the 3$d$ 3-state Potts model, which was determined in Ref. [@FFGP06] to be 2.43(10).
We have calculated the ratio $m_I/m_R$ for $\beta$ ranging from 5.695 up to 9.0. We observe from the right panel of Fig. \[ratio\_20\] that this ratio is compatible with 3/2 at the largest $\beta$ values considered, in agreement with the high-temperature perturbation theory. Then, when the temperature is lowered towards the transition, this ratio goes up to a value compatible with 3, in agreement with the Polyakov loop model of Ref. [@Pisarski0110214], which contains only quadratic, cubic and quartic powers of the Polyakov loop, i.e. the minimum number of terms required in order to be compatible with a first order phase transition. The same trend has been observed also in Ref. [@Datta:2002je].
![(Left) Ratio $m_{2^+}/m_{0^+}$ as a function of $\beta$ in the deconfined phase. The three upper horizontal lines represent the constant (with its error) which fits the data (see the text for details); the three lower horizontal lines represent the corresponding mass ratio (with its error) found in the 3$d$ 3-state Potts model [@FFGP06]. (Right) Ratio $m_I/m_R$ as a function of $\beta$ in the deconfined phase. The vertical line corresponds to the critical $\beta$ value.[]{data-label="ratio_20"}](ratio_20_new.eps "fig:")![(Left) Ratio $m_{2^+}/m_{0^+}$ as a function of $\beta$ in the deconfined phase. The three upper horizontal lines represent the constant (with its error) which fits the data (see the text for details); the three lower horizontal lines represent the corresponding mass ratio (with its error) found in the 3$d$ 3-state Potts model [@FFGP06]. (Right) Ratio $m_I/m_R$ as a function of $\beta$ in the deconfined phase. The vertical line corresponds to the critical $\beta$ value.[]{data-label="ratio_20"}](ratio_ri.eps "fig:")
Conclusions and outlook
=======================
In this work we have studied in the (3+1)$d$ SU(3) pure gauge theory above the deconfinement transition the lowest masses in the 0$^+$ and the 2$^+$ channels of angular momentum and parity and the screening masses resulting from the correlation between the real parts and between the imaginary parts of the Polyakov loop. The behavior of the ratio between the masses in the 0$^+$ and the 2$^+$ channels with the temperature suggests that they have a common scaling above the transition temperature. This ratio turns to be $\simeq$30% larger than the ratio of the lowest massive excitations in the same channels of the 3$d$ 3-state Potts model in the broken phase. This can be taken as an estimate of the level of approximation by which the Svetitsky-Yaffe conjecture, valid in strict sense only for continuous phase transitions, can play some role also for (3+1)$d$ SU(3) at finite temperature.
The dependence on the temperature of the ratio between the screening masses from the correlation between the real parts and between the imaginary parts of the Polyakov loop shows a nice interplay between the high-temperature regime, where perturbation theory should work, and the transition regime, where mean-field effective Polyakov loop models could apply.
[99]{}
R. Falcone, R. Fiore, M. Gravina and A. Papa, Nucl. Phys. B [**767**]{} (2007) 385 \[[hep-lat/0612016]{}\]. B. Svetitsky and L.G. Yaffe, Nucl. Phys. B [**210**]{} (1982) 423. M. Caselle, M. Hasenbusch and P. Provero, Nucl. Phys. B [**556**]{} (1999) 575 \[[hep-lat/9903011]{}\]. M. Caselle, M. Hasenbusch, P. Provero and K. Zarembo, Nucl. Phys. B [**623**]{} (2002) 474 \[[hep-th/0103130]{}\]. R. Fiore, A. Papa and P. Provero, Nucl. Phys. (Proc. Suppl.) [**119**]{} (2003) 490 \[[hep-lat/0208020]{}\]. R. Fiore, A. Papa and P. Provero, Phys. Rev. D [**67**]{} (2003) 114508 \[[hep-lat/0208020]{}\]. R.D. Pisarski, Phys. Rev. D [**62**]{} (2000) 111501 \[[hep-ph/0101168]{}\]. A.S. Kronfeld, Nucl. Phys. (Proc. Suppl.) [**17**]{} (1990) 313.
M. Lüscher and U. Wolff, Nucl. Phys. B [**339**]{} (1990) 222.
R. Falcone, R. Fiore, M. Gravina and A. Papa, Nucl. Phys. B. [**785**]{} (2007) 19 \[[arXiv:0704.3882 \[hep-lat\]]{}\].
V. Agostini, G. Carlino, M. Caselle and M. Hasenbusch, Nucl. Phys. B [**484**]{} (1997) 331 \[[hep-lat/9607029]{}\]. S. Nadkarni, Phys. Rev. D [**33**]{} (1986) 3738. A. Dumitru and R.D. Pisarski, Phys. Rev. D [**66**]{} (2002) 096003 \[[hep-ph/0204223]{}\]. A. Dumitru and R.D. Pisarski, Nucl. Phys. (Proc. Suppl.) [**106**]{} (2002) 483 \[[hep-ph/0110214]{}\].
N. Cabibbo and E. Marinari, Phys. Lett. B [**119**]{} (1982) 387. S.L. Adler, Phys. Rev. D [**23**]{} (1981) 2901. G. Boyd, J. Engels, F. Karsch, E. Laermann, C. Legeland, M. Lutgemeier and B. Petersson, Nucl. Phys. B [**469**]{} (1996) 419 \[[hep-lat/9602007]{}\]. R.V. Gavai, F. Karsch and B. Petersson, Nucl. Phys. B [**322**]{} (1989) 738 \[[cern-th-5221/88]{}\]. S. Datta and S. Gupta, Phys. Rev. D [**67**]{} (2003) 054503 \[[hep-lat/0208001]{}\]. M.E. Fisher and A.N. Berker, Phys. Rev. B [**26**]{} (1982) 2507.
[^1]: Here and in the following, $N_i$ ($i=x, y, z$) is the number of lattice sites in the $i$-direction.
[^2]: In (\[corr\_funct\_tunneling\]) we have taken into account only the lowest masses in the spectrum and, for brevity, omitted to write the “echo” terms.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Dmitry Ponomarev[^1]'
bibliography:
- 'lcsh.bib'
date:
-
-
title: |
Off-Shell Spinor-Helicity Amplitudes\
from Light-Cone Deformation Procedure
---
= 1mm
Introduction
============
Construction of consistent interacting field theories is an old and challenging problem. At the free level elementary fields in the Minkowski space can be identified with unitary irreducible representations of the Poincare group, which have been classified long time ago [@Wigner:1939cj], for a review see e.g. [@Bekaert:2006py]. For interacting theories constructed perturbatively the main consistency requirement is that the physical degrees of freedom defined at the free level interact without breaking Poincare invariance.
A natural way to construct Poincare invariant theories is to use Lorentz tensors. Having contracted tensor indices appropriately one automatically ensures Lorentz invariance of the action. If, moreover, the action does not depend on coordinates explicitly, then it is also translationally invariant. In this way, Lorentz tensors allow to make all Poincare symmetry of the theory manifest.
What makes the manifestly covariant approach much less trivial is that, unless special care is taken, it introduces unwanted degrees of freedom. For consistency these extra degrees of freedom should be removed from the theory, usually, either by constraints or as a result of gauge invariance. In particular, for massless fields with spin greater than one-half description in terms of Lorentz tensors requires these fields to be gauge ones. To avoid unwanted degrees of freedom at the interacting level gauge invariance should also be preserved. This leads to the manifestly Lorentz covariant deformation procedure, which amounts to a simultaneous deformation of the action and gauge transformations in a way that the action remains gauge invariant. For massless higher-spin fields this approach was used by many authors and leads to the conclusion that non-trivial local interactions of higher-spin fields in flat space cannot exist, see e.g. [@Aragone:1979hx; @Bekaert:2010hp; @Joung:2013nma][^2].
An alternative approach is to abandon Lorentz tensors and control Poincare symmetry explicitly by requiring that the Noether charges, deformed order by order, obey commutation relations of the Poincare algebra. A particular version of this procedure is the light-cone approach [@Bengtsson:1983pd; @Bengtsson:1983pg; @Bengtsson:1986kh], which we will use in this paper. A somewhat unattractive feature of the light-cone approach is that it requires manual control of all symmetries and as a result at first sight appears less economical than the manifestly Lorentz covariant one. On the other hand, it has an advantage of being completely general. It is quite remarkable that already at the cubic order by giving up manifest Lorentz covariance one can find additional consistent local interactions[^3]. These exotic vertices are known for a long time [@Bengtsson:1986kh; @Metsaev:1991mt; @Metsaev:1991nb; @Metsaev:1993ap], but only recently it was emphasised that they are missing in the manifestly Lorentz covariant classification [@Bengtsson:2014qza][^4]. Moreover, presence of the exotic vertices turns out to be crucial for consistency of higher-spin interactions at the quartic order [@Metsaev:1991mt; @Metsaev:1991nb]. In particular, they are present in the chiral higher-spin theory [@Ponomarev:2016lrm], which is a cubic theory consistent to all orders in interactions, see also [@Metsaev:1991mt; @Metsaev:1991nb].
In this paper we study the light-cone consistency conditions in four-dimensional flat space-time and show that they can be rephrased as the Ward identities for the off-shell light-cone amplitudes. This terminology deserves clarification. Firstly, a term ’Ward identity’ is usually used for constraints on the $S$-matrix, imposed by gauge invariance. In the light-cone approach gauge freedom is completely fixed and the constraints that we call ’Ward identities’ appear as a consequence of invariance of the $S$-matrix with respect to the global Poincare symmetry algebra. Nevertheless, we then show that these constraints are equivalent to the constraints imposed by gauge invariance in manifestly covariant approaches, thus justifying a name ’Ward identities’ that we use for them. Secondly, by an ’off-shell amplitude’ we, essentially, mean an amputated correlator, that is a correlator where propagators associated with external lines are removed, but external momenta are not put on-shell. In the following, these light-cone amplitudes will be built from the light-cone Hamiltonian according to the light-cone Feynman rules to be specified in the text.
Next, we solve these Ward identities and show that their general solution can be compactly written employing the spinor-helicity language as an arbitrary function of spinor contractions, obeying natural homogeneity constraints[^5]. The relation between the light-cone deformation procedure and the spinor-helicity approach was observed before. In [@Ananth:2012un; @Akshay:2014qea] it was found that the cubic light-cone interactions can be easily rewritten in the spinor-helicity form. The associated three-point amplitudes reproduce those found in [@Benincasa:2007xk]. Then, in [@Bengtsson:2016jfk; @Bengtsson:2016alt] the connection with the spinor-helicity approach was used to study the quartic sector of the light-cone consistency equations[^6].
In the present paper we show that the relation between the light-cone deformation procedure and the spinor-helicity representation is not accidental. Instead, spinor contractions and associated homogeneity constraints appear inevitably from the light-cone consistency conditions. Moreover, being essentially off-shell the light-cone approach provides a natural off-shell continuation of amplitudes written in the spinor-helicity representation. In this way, by combining benefits of the two approaches we obtain an attractive tool to study consistent theories of massless fields: on one hand, it is manifestly covariant, which makes it efficient and, on the other hand, it is completely general and hence captures all consistent local interactions.
The plan of the paper is as follows. In Section \[sec7oct1\] we review the basics of the light-cone deformation procedure. We start with the free theory and show how requiring closure of the Poincare algebra one derives kinematical and dynamical constraints. In Section \[section5oct1\] we show that the dynamical constraints can be equivalently rewritten as the Ward identities for the light-cone amplitude. To this end, in Section \[sec29oct1\] we first rewrite them in terms of the light-cone Hamiltonian only. In Section \[sect2oct1\] we collect some extra notations and identities necessary to streamline the following analysis. Next, in Section \[29oct3\] we show that after some manipulations the light-cone consistency conditions acquire a form of the Ward identities for the amplitude, that contains contact diagrams, exchanges with one internal line and extra terms bilinear in the light-cone Hamiltonian. After that, in Section \[sec29oct2\] we show that these extra terms produce contributions of exchanges with two internal lines as well as other terms, which are cubic in the light-cone Hamiltonian. Applying this procedure iteratively we recover the Ward identity for the complete amplitude in Section \[29oct4\]. This result is summarised in Proposition \[prop2oct1\]. Then, in Section \[sect1oct1\] we derive a general solution of the Ward identity. The solution is summarised in Proposition \[prop5oct1\]. Interpretation of these results in terms of the spinor-helicity approach is given in Section \[sect5oct3\]. We present our conclusions and discuss possible extension in Section \[sec7oct2\]. Appendix \[ap3oct1\] summarises our conventions.
Basics {#sec7oct1}
======
In this Section we review the basics of the light-cone deformation procedure for massless fields in four-dimensional flat space-time.
Free massless fields in the light-cone gauge
--------------------------------------------
The Poincare algebra commutation relations are given by $$\begin{aligned}
\notag
[P^a,P^b] &\,= 0,\\
\notag
[J^{ab},P^c] &\,= P^a \eta^{bc}- P^b \eta^{ac},\\
[J^{ab}, J^{cd}] &\, = J^{ad} \eta^{bc} -J^{bd} \eta^{ac} -J^{ac} \eta^{bd}+ J^{bc} \eta^{ad},\end{aligned}$$ where $x^a=\{x^-,x^+,x,\bar x\}$. For more details on our conventions see Appendix \[ap3oct1\]. It admits helicity-$\lambda$ representations $$\begin{aligned}
\notag
P^a \cdot \Phi^{\lambda}&\,\equiv \partial^a \Phi^{\lambda},\\
\label{29sep2}
J^{ab}\cdot \Phi^{\lambda}&\,\equiv (x^a\partial^b - x^b \partial^a+S^{ab} )\Phi^\lambda,\end{aligned}$$ where $S^{ab}$ is the spin part of the angular momentum. In the light-cone gauge $$\label{29sep4}
S^{+a}\cdot \Phi^{\lambda} =0,\qquad S^{ab} \partial_a \cdot \Phi^{\lambda}=0.$$ The first condition in (\[29sep4\]) implies that the only non-vanishing components of $S^{ab}$ are $S^{x\bar x}$, $S^{x-}$ and $S^{\bar x -}$. The second condition allows to express all of them in terms of $S^{x\bar x}$ $$\label{21sep3}
S^{x-}\cdot \Phi^{\lambda}= -S^{x\bar x}\cdot \frac{\partial}{\partial^+}\Phi^{\lambda}, \qquad
S^{\bar x -}\cdot \Phi^{\lambda}= S^{x\bar x}\cdot \frac{\bar\partial}{\partial^+}\Phi^{\lambda}.$$ Thus, the helicity representation is specified by the action of $S^{x\bar x}$ generating the Wigner little group. It is conventional to define $$\label{29sep3}
S^{x\bar x} \cdot \Phi^{\lambda} = -\lambda \Phi^{\lambda},$$ where $\lambda$ is the helicity.
#### Free action and canonical analysis.
The canonically normalised action for the set of free massless fields is given by $$\label{19sep6}
S_2\equiv \int d^4 x L_2, \qquad L_2 = -\frac{1}{2}\sum_{\lambda} \partial_a\Phi^{-\lambda} \partial^a \Phi^\lambda.$$ Here we do not impose any restrictions on the spectrum of values of $\lambda$ except that opposite helicities should enter together. For example, for the spin $s$ field one has $\lambda = \{s,-s\}$ and $$L_2= -\partial_a\Phi^{-s}\partial^a \Phi^s.$$ In the scalar case $\lambda=0$ and $$L_2 = -\frac{1}{2} \partial_a\Phi^{0} \partial^a \Phi^0.$$
In the light-cone coordinates $\partial^-$ is the time derivative, so the canonical momentum is $$\Pi^\lambda \equiv \frac{\delta L_2}{\delta( \partial^- \Phi^\lambda)}= -\partial^+ \Phi^{-\lambda}.$$ Then, the Poisson bracket is $$[\partial^+ \Phi^\lambda(x^\perp,x^+),\Phi^{\mu}(y^\perp,x^+)]_P= \delta^{\lambda+\mu,0}\delta^{3}(x^\perp,y^\perp),$$ where we use $x^{\perp}\equiv \{x,\bar x,x^- \}$. The canonical Hamiltonian is given by $$\label{21sep8}
H_2\equiv \sum_{\lambda } \int d^{3}x^\perp(\Pi^\lambda \partial^-\Phi^{\lambda} - L_2) =
\sum_{\lambda} \int d^{3}x^\perp \partial \Phi^{-\lambda} \bar\partial \Phi^\lambda,$$ where one integrates over equal-time hypersurfaces.
It is not hard to see that due to the fact that the Lagrangian (\[19sep6\]) is first order in time derivatives, the theory features constraints. Their analysis has been discussed by many authors, for a review see e.g. [@Heinzl:2000ht]. As a result, to account for the constraints one should replace the Poisson bracket by the Dirac one $$\label{21sep6}
[\partial^+ \Phi^\lambda(x^\perp,x^+),\Phi^{\mu}(y^\perp,x^+)]= \frac{1}{2}\delta^{\lambda+\mu,0}\delta^{3}(x^\perp,y^\perp),$$ or, equivalently, $$\label{21sep7}
[\Phi^\lambda(x^\perp,x^+),\Phi^{\mu}(y^\perp,x^+)]= \frac{1}{\partial^+_x - \partial^+_y}\delta^{\lambda+\mu,0}\delta^{3}(x^\perp,y^\perp).$$
The canonical Hamiltonian (\[21sep8\]) and the Dirac bracket (\[21sep7\]) define the time evolution $$\label{19sep11}
\partial^- F(\Phi) = [F(\Phi), H_2].$$ In particular, one can verify that $$\partial^-\Phi^\lambda(x) = [\Phi^\lambda(x), H_2] = -\frac{\partial \bar \partial}{\partial^+} \Phi^\lambda(x),$$ which is consistent with the variation of (\[19sep6\]).
#### Noether currents and charges.
The action (\[19sep6\]) is invariant with respect to transformations (\[29sep2\]). The associated Noether currents are well known $$\begin{aligned}
\notag
P^i \quad \to \;\quad T^{i,j} &\,= \sum_{\lambda}\frac{\delta L_2}{\delta (\partial_j \Phi^\lambda)}\partial^i \Phi^\lambda - \eta^{ij} L_2,\\
J^{ij} \quad \to \quad L^{ij,k} &\,= x^i T^{j,k}-x^j T^{i,k}+R^{ij,k},
\label{21sep10}
\end{aligned}$$ where $R^{ij,k}$ is the spin current $$\label{21sep12}
R^{ij,k} = \sum_{\lambda}\frac{\delta L_2}{\delta( \partial_k \Phi^\lambda)} S^{ij}\cdot \Phi^\lambda$$ and $S^{ij}$ was given in (\[21sep3\]), (\[29sep3\]).
Accordingly, we define the Noether charges $$\label{21sep13}
P^i_2 = \int d^3x^\perp T^{i,+}, \qquad J^{ij}_2 = \int d^3x^\perp L^{ij,+},$$ which with a slight abuse of notations we denoted by the same symbols as the algebra generators themselves. For simplicity we choose the integration hypersurface to be $x^+=0$. Explicitly the charges (\[21sep13\]) read $$\label{21sep14}
P^i_2 = -\sum_{\lambda}\int d^3x^\perp \partial^+\Phi^{-\lambda} p_2^i \Phi^\lambda, \qquad J_2^{ij} = -\sum_{\lambda}\int d^3x^\perp \partial^+\Phi^{-\lambda} j_2^{ij} \Phi^\lambda,$$ where $$\begin{aligned}
\notag
p_2^+ &\,=\partial^+, & p_2^- &= -\frac{\partial \bar\partial}{\partial^+}, \qquad\qquad \quad\; p_2 = \partial, \qquad\qquad\quad\; \bar p_2=
\bar\partial,\\
\notag
j_2^{+-} &\,= - x^-\partial^+,& j_2^{x\bar x} &= x\bar\partial - \bar x \partial -\lambda, \\
\notag
j_2^{x+}&\, = x\partial^+, & j_2^{x-} &= -x\frac{\partial \bar\partial}{\partial^+} - x^- \partial+\lambda\frac{\partial}{\partial^+}, \\
j_2^{\bar x+}&\, = \bar x\partial^+,
& j_2^{\bar x-} &= -\bar x\frac{\partial \bar\partial}{\partial^+} - x^- \bar\partial
-\lambda\frac{\bar\partial}{\partial^+}.
\label{21sep15}\end{aligned}$$ This representation coincides with the original one (\[29sep2\]) up to terms that vanish on $x^+=0$ and on the mass-shell. As expected, the charges (\[21sep14\]) generate the algebra action via the commutator $$\label{21sep16}
[\Phi^\lambda, P_2^i] = p_2^i \Phi^\lambda, \qquad [\Phi^\lambda, J_2^{ij}] = j_2^{ij} \Phi^\lambda.$$ Moreover, the charge $P^-_2$ associated with the light-cone time translation is the canonical Hamiltonian $H_2$ (\[21sep8\]).
#### Fourier transform.
It is convenient to make the Fourier transform with respect to spatial coordinates $x^-$, $x$ and $\bar x$ $$\begin{aligned}
\notag
\Phi(x,x^+) &\,= (2\pi)^{-\frac{3}{2}} \int{e^{+i (x^- p^+ + \bar x p +x\bar p)}\Phi(p,x^+)d^{3}p^\perp},\\
\label{21sep17}
\Phi(p,x^+) &\,= (2\pi)^{-\frac{3}{2}} \int{e^{-i (x^- p^+ + \bar x p +x\bar p)}\Phi(x,x^+)d^{3}x^\perp},
\end{aligned}$$ followed by a change of variables $p=iq$. This allows to avoid complex factors and effectively amounts to the substitution $$\label{21sep19}
\frac{\partial}{\partial x^i} \to q_i, \qquad x^i \to -\frac{\partial}{\partial q_i}.$$ We will also use $q^\perp \equiv \{q,\bar q,q^+\}$ and $\beta \equiv q^+ $.
In these terms the canonical commutator reads $$\label{29sep5}
[\Phi^{\lambda_1}(q^\perp_1,x^+),\Phi^{\lambda_2}(q^\perp_2,x^+)] = \frac{\delta^{\lambda_1+\lambda_2,0}\delta^3(q^\perp_1+q^\perp_2)}{\beta_1-\beta_2}$$ and the Noether charges are $$\begin{aligned}
\notag
P_2^i &\,= \sum_{\lambda}\int d^3q^\perp_1 d^3q^\perp_2 \delta^3(q^\perp_1+q^\perp_2)\beta_1 \Phi^{-\lambda}(q^\perp_1,x^+) p^i_2(q_2,\partial_2) \Phi^\lambda(q^\perp_2,x^+),\\
\label{29sep6}
J_2^{ij} &\,= \sum_{\lambda}\int d^3q^\perp_1 d^3q^\perp_2 \delta^3(q^\perp_1+q^\perp_2)\beta_1 \Phi^{-\lambda}(q^\perp_1,x^+) j^{ij}_2(q_2,\partial_2) \Phi^\lambda(q^\perp_2,x^+),\end{aligned}$$ where $$\begin{aligned}
\notag
p_2^+ &\,=q^+, & p_2^- & = -\frac{q \bar q}{\beta}, \qquad\qquad \quad\; p_2 = q, \qquad\qquad \quad\; \bar p_2=
\bar q,\\
\notag
j_2^{+-} &\,= \frac{\partial}{\partial \beta}\beta, & j_2^{x\bar x} &= N_q- N_{\bar q} -\lambda,\\
\notag
j_2^{x+}&\, = -\beta \frac{\partial}{\partial \bar q}, & j_2^{x-} &= \frac{\partial}{\partial \bar q}\frac{q \bar q}{\beta} + q \frac{\partial}{\partial \beta}+\lambda\frac{q}{\beta},\\
j_2^{\bar x+}&\, = -\beta \frac{\partial}{\partial q},
& j_2^{\bar x-} &= \frac{\partial}{\partial q}\frac{q \bar q}{\beta} + \bar q \frac{\partial}{\partial \beta}
-\lambda\frac{\bar q}{\beta}
\label{29sep7}\end{aligned}$$ and $$N_q \equiv q\frac{\partial}{\partial q}, \qquad N_{\bar q} \equiv \bar q\frac{\partial}{\partial \bar q}.$$
Introducing interactions consistently
-------------------------------------
At the interacting level the action receives non-linear corrections and so do the charges (\[29sep6\]). The only consistency requirement that one imposes is that they still generate the Poincare algebra. The standard lore of the light-cone approach is that it is sufficient to deform only the generators, that are transversal to the light-cone $x^+=0$ [@Dirac:1949cp]. These are $$H\equiv P^- ,\qquad J\equiv J^{x-}, \qquad \bar J\equiv J^{\bar x-}$$ and they are called dynamical generators[^7]. The remaining generators are called kinematical. Let us collectively denote by $D$ and $K$ the charges of the dynamical and the kinematical generators respectively. Then, at the non-linear level $$D =D_2 + \delta D, \qquad K = K_2.$$
Accordingly, one can break the Poincare algebra commutation relations into classes depending on the types of generators they feature. The simplest type of commutators is $$[K,K]=K,$$ as it is automatically satisfied at the non-linear level. Other two groups of commutators $$\begin{aligned}
[K,D] = K \qquad &\Rightarrow \qquad [K,\delta D] = 0,\\
[K,D] = D \qquad &\Rightarrow \qquad [K,\delta D] = \delta D\end{aligned}$$ are very similar to each other. They both result in linear differential equations on deformations $\delta D$, which will be called kinematical constraints. These constraints can be solved easily for $\delta D$ at any order of deformation. The last type of commutators is $$[D,D]=0$$ and it presents the main difficulty of the light-cone deformation procedure. We will now consider these issues one by one in more details.
#### Deformation.
A general ansatz for the dynamical generators at the non-linear level is $$H=H_2+\sum_n H_n ,\quad J=J_2+\sum_n J_n, \quad \bar J=\bar J_2+\sum_n \bar J_n,$$ where $$\begin{aligned}
\notag
H_n =&\; \frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp \delta^3 (\sum_{i=1}^n q^\perp_i) h^{\lambda_1 \dots \lambda_n}_n \prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i),\\
\notag
J_n =&\; \frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp \delta^3 (\sum_{i=1}^n q^\perp_i) \Big[j_n^{\lambda_1 \dots \lambda_n} -\frac{1}{n}
h_n^{\lambda_1\dots \lambda_n}\big( \sum_j \frac{\partial}{\partial \bar q_j}\big)\Big]\prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i),\\
\label{22sep1}
\bar J_n =&\; \frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp \delta^3 (\sum_{i=1}^n q^\perp_i) \Big[\bar j_n^{\lambda_1 \dots \lambda_n} -\frac{1}{n}
h_n^{\lambda_1\dots \lambda_n}\big( \sum_j \frac{\partial}{\partial q_j}\big)\Big]\prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i).\end{aligned}$$ Here $h_n$, $j_n$ and $\bar j_n$ generalise operators appearing in (\[29sep7\]). They all depend on transversal momenta $q^\perp_i$, but not on $q_i^-$. On the one hand, $q_i^-$-independence of interaction vertices can always be achieved by field redefinitions, using the fact that on-shell $$q_i^- \approx h_2(q^\perp_i)\equiv -\frac{q_i\bar q_i}{\beta_i}.$$ On the other hand, interactions that are free of time derivatives are convenient, as they do not deform the canonical bracket.
In (\[22sep1\]) the momentum delta-functions ensure translation invariance along spatial directions, which implies that the kinematical constraints arising from commutators with $P_2$, $\bar P_2$ and $P^+_2$ are automatically satisfied. The $h$-dependent corrections in the ansatzes for $J_n$ and $\bar J_n$ in (\[22sep1\]) is just a standard trick which slightly simplifies the remaining kinematical constraints for $j$ and $\bar j$.
#### Kinematical constraints.
To evaluate the remaining kinematical constraints it is convenient to use $$\begin{aligned}
\label{30sep1}
[F(\Phi),J_2^{ij}] &\,= [\Phi,J_2^{ij}] \,\frac{\delta F(\Phi)}{\delta \Phi} = j_2^{ij} \Phi \,\frac{\delta F(\Phi)}{\delta \Phi},\end{aligned}$$ which is a result of consecutive application of (\[21sep16\]) to each $\Phi$ entering $F(\Phi)$. Employing (\[30sep1\]) to evaluate commutators with $H_n$, we find $$\begin{aligned}
\label{30sep2}
[J_2^{x+},H_n]=0\qquad \Rightarrow &\qquad \delta^3 (\sum_{i=1}^n q^\perp_i)\sum_{i=1}^n \beta_i \frac{\partial}{\partial \bar q_i} h_n^{\lambda_1\dots \lambda_n}=0,\\
\label{30sep3}
[J_2^{\bar x+},H_n]=0\qquad \Rightarrow &\qquad \delta^3 (\sum_{i=1}^n q^\perp_i)\sum_{i=1}^n \beta_i \frac{\partial}{\partial q_i} h_n^{\lambda_1\dots \lambda_n}=0,\\
\label{30sep4}
[J_2^{x\bar x},H_n]=0\qquad \Rightarrow &\qquad \delta^3 (\sum_{i=1}^n q^\perp_i)\sum_{i=1}^n (N_{q_i}- N_{\bar{q}_i}+\lambda_i)h_n^{\lambda_1\dots \lambda_n}=0,\\
\label{30sep5}
[J_2^{+-},H_n]+H_n=0\qquad \Rightarrow &\qquad \delta^3 (\sum_{i=1}^n q^\perp_i) \sum_{i=1}^n \beta_i \frac{\partial}{\partial \beta_i} h_n^{\lambda_1\dots \lambda_n}=0.\end{aligned}$$ Constraints for $j_n$ and $\bar j_n$ are analogous and can be found, e.g. in [@Ponomarev:2016lrm].
The first two conditions (\[30sep2\]), (\[30sep3\]) imply that $h_n$ can depend on $q_i$ and $\bar q_i$ only through their particular combinations with $\beta_i$ $$\bar{\mathbb{P}}_{ij} \equiv \bar q_i\beta_j -\bar q_j \beta_i , \qquad \mathbb{P}_{ij} \equiv q_i\beta_j -q_j \beta_i.$$ The remaining two conditions (\[30sep4\]), (\[30sep5\]) simply specify the homogeneity degrees of $h_n$ on its arguments.
#### Dynamical commutators.
Let us now consider the dynamical equation $$\label{22sep5}
[H,J]=0 \qquad \Rightarrow \qquad [H_2, J_n] + [H_3, J_{n-1}] + \dots + [H_{n-1}, J_3]+ [H_n,J_2]=0.$$ The charges $H_2$ and $J_2$ are already known, so the first and the last commutators can be readily computed. Employing (\[30sep1\]) for $H_2$ we find $$\begin{aligned}
[H_2, J_n]
= -\frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp \delta (\sum_{i=1}^n q^\perp_i)
{\cal H}^{\{i\}}
\Big[j_n^{\lambda_1 \dots \lambda_n} +\frac{1}{n}
\big( \sum_j \frac{\partial}{\partial \bar q_j}\big) h_n^{\lambda_1\dots \lambda_n}\Big]
\prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i),
\label{22sep6}
\end{aligned}$$ where $${\cal H}^{\{i\}}\equiv \sum_{i=1}^n h_2(q^\perp_i).$$
Analogously, the last commutator in (\[22sep5\]) gives $$\begin{aligned}
\label{22sep7}
[H_n,J_2] = \frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp \delta (\sum_{i=1}^n q^\perp_i){\cal J}^{\{i\}} h_n^{\lambda_1 \dots \lambda_n}
\prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i),
\end{aligned}$$ where $$\label{22sep8}
{\cal J}^{\{i\}} =\sum_{i=1}^n \Big(-\frac{q_i\bar q_i}{\beta_i}\frac{\partial}{\partial \bar q_i}-q_i \frac{\partial}{\partial \beta_i}+\lambda_i \frac{q_i}{\beta_i}\Big).$$
Eventually, (\[22sep5\]) becomes $$\begin{aligned}
\notag
&\,-\frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp \delta (\sum_{i=1}^n q^\perp_i)
{\cal H}^{\{i\}}
\Big[j_n^{\lambda_1 \dots \lambda_n} +\frac{1}{n}
\big( \sum_j \frac{\partial}{\partial \bar q_j}\big) h_n^{\lambda_1\dots \lambda_n}\Big]
\prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i)\\
\notag
&\, \qquad \qquad\qquad\qquad\qquad \;+
\frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp \delta (\sum_{i=1}^n q^\perp_i){\cal J}^{\{i\}} h_n^{\lambda_1 \dots \lambda_n}
\prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i)\\
\label{22sep9}
&\,\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad \qquad \quad+ [H_3, J_{n-1}] + \dots + [H_{n-1}, J_3]=0.
\end{aligned}$$
The equation $[H,\bar J]=0$ is analogous. Moreover, $[H,J] = 0$ and $[H,\bar J]=0$ together imply that the last consistency condition $[J,\bar J]=0$ is also satisfied [@Ponomarev:2016lrm]. Hence, in order to proceed it is enough to learn how to solve (\[22sep9\]) efficiently.
Towards the Ward identity {#section5oct1}
=========================
It is hard not to notice that the light-cone consistency condition (\[22sep9\]) is reminiscent of some constraint imposed on the total amplitude made of $H$. Indeed, along with a contribution from the contact n-point interaction $h_n$, it contains terms $[H_m,J_{n+2-m}]$, which are naturally associated with the exchanges involving $m$- and $(n+2-m)$-point vertices. On the other hand, it is not at all obvious how to make this relation precise. Firstly, (\[22sep9\]) contains $J$ and $H$ on equal footing. While $H$ is trivially related to the vertices the way they appear in the action, for $J$ this relation is less obvious. Secondly, it is not immediately clear how (\[22sep9\]) produces contributions associated with exchanges involving two or more internal lines.
In the amplitude language, the main consistency requirement any interacting field theory should satisfy is that the $S$-matrix is Poincare invariant. Given that the $S$-matrix is essentially the transition amplitude between the on-shell states of the free theory, the Poincare algebra acts on the $S$-matrix by the free theory generators. Of course, one expects that consistency conditions in different approaches are related to each other. Hence, the light-cone consistency condition (\[22sep9\]) should be related to Poincare invariance of the $S$-matrix with respect to the free theory transformations. Our goal in this Section is to clarify this relation. This will enable us to rewrite the light-cone consistency conditions in the $S$-matrix-like form and then solve them in Section \[sect1oct1\].
More precisely, we will show that $$\begin{aligned}
\notag
[H,J] = 0 \qquad \Leftrightarrow\qquad [A,J_2]=0,\\
\label{15oct1}
[H,\bar J] = 0 \qquad \Leftrightarrow \qquad [A,\bar J_2]=0,\end{aligned}$$ where $A$ will be specified later. At this point we just note that $A$, similarly to $H$, is given by a space-time integral, where as a kernel instead of $h$ one has a certain off-shell continuation of the amplitude built of $h$. Moreover, one can then trivially show that $$\label{22oct1}
[H,K]=0 \qquad \Leftrightarrow \qquad [{A},K]=0,$$ were $K$ are kinetic generators as well as $$\label{22oct2}
[{A},H] = {\cal H}\, {A}\approx 0.$$ Combining (\[15oct1\])-(\[22oct2\]) together, we find that the light-cone consistency conditions imply Poincare invariance of the $S$-matrix, as expected. However, we would like to emphasise, that (\[15oct1\])-(\[22oct2\]) hold off-shell. Also, let us stress that despite these formulas will be derived for massless particles in flat four-dimensional space, they are, clearly, completely general and should be valid for any number of dimensions, types of particles and the value of the cosmological constant.
Eliminating $J$ {#sec29oct1}
---------------
What complicates the analysis of (\[22sep9\]) is that it has to be solved for two unknowns $h_n$ and $j_n$. At the same time, conceptually, it is clear that once $h_n$ is known one can find the action, which, if Poincare-invariant, defines $j_n$. So, our first goal is to eliminate $j_n$ in favour of $h_n$.
An important observation, which will be used extensively throughout the paper is that the operator, that acts on $h_n$ in (\[22sep9\]) has the following property $$\label{1oct3}
\big({\cal J}^{\{i\}}- {\cal H}^{\{i\}}\frac{1}{n}
\sum_{i=1}^n \frac{\partial}{\partial \bar q_i}\big) (\sum_{i=1}^n q^{\perp}_i)\propto (\sum_{i=1}^n q^{\perp}_i).$$ This allows to use momentum conservation inside $h_n$ without changing its contribution to (\[22sep9\]). On the other hand, since $$\big[\frac{1}{n}
\sum_{i=1}^n \frac{\partial}{\partial \bar q_i}, \sum_{j=1}^n \bar q_j\big]=1,$$ one can always add to $h_n$ terms proportional to the total momentum $$\label{23oct1}
h_n^{\lambda_1\dots \lambda_n} \quad \to \quad \tilde h_n^{\lambda_1\dots \lambda_n}=
h_n^{\lambda_1\dots \lambda_n} + \alpha \sum_{j=1}^n \bar q_j,$$ so that $$\label{22sep10}
j_n^{\lambda_1 \dots \lambda_n} +\frac{1}{n}
\big( \sum_j \frac{\partial}{\partial \bar q_j}\big) \tilde h_n^{\lambda_1\dots \lambda_n}=0$$ is satisfied. In other words, once a solution $h^n$ of (\[22sep9\]) is found, one can replace it with $\tilde h_n$, which additionally satisfies (\[22sep10\]). Moreover, using the fact that $h^n$ enters $H^n$ multiplied by the momentum conserving delta function, the replacement of $h^n$ with $\tilde h^n$ leaves commutators $[H_n,J_m]$ intact as well. In the following we will omit the tilde and write just $h_n$. Clearly, a similar argument works for $\bar j_n$.
The condition (\[22sep10\]) allows to solve for $j_n$ in terms of $h_n$. This leads to $$\begin{aligned}
\label{22sep12}
J_n
=-\frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp d\bar \varepsilon \dot\delta(\bar\varepsilon) \delta(\sum_{i=1}^n q^\perp_i+ \bar\varepsilon)
h_n^{\lambda_1 \dots \lambda_n} \prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i),
\end{aligned}$$ where we found convenient to introduce an extra variable $\bar\varepsilon$ to write a derivative of the momentum conserving delta-function in a more concise and symmetric form. Comparing it with (\[22sep1\]) we find that this derivative is the only difference between $H_n$ and $J_n$.
This has the following simple interpretation. Given that at the non-linear level from all $P^i$ one deforms only $P^-$, using (\[21sep10\]) we find $$\label{23oct2}
\delta L^{x-,+}=x \delta T^{-,+} + \delta R^{x-,+}, \qquad \delta L^{\bar x -,+}= \bar x \delta T^{-,+}+\delta R^{\bar x-,+}.$$ Integrating it over $x^+=0$ and making the Fourier transform, one can express $\delta J^{x-}$ and $\delta J^{\bar x-}$ in terms of $\delta P^-$. Taking into account definitions (\[22sep1\]) one can see that (\[22sep10\]), (\[22sep12\]) just mean that in (\[23oct2\]) the spin current remains undeformed, $\delta R=0$.
With (\[22sep10\]) imposed, the consistency condition (\[22sep9\]) simplifies to $$\begin{aligned}
\label{22sep11}
[H_3, J_{n-1}^{z-}] + \dots + [H_{n-1}, J_3^{z-}]+
\frac{1}{n!}\sum_{\lambda_i}\int d^{3n}q^\perp \delta (\sum_{i=1}^n q^\perp_i){\cal J}^{\{i\}} h_n^{\lambda_1 \dots \lambda_n}
\prod_{i=1}^n\Phi^{\lambda_i}(q^\perp_i)=0.
\end{aligned}$$ To evaluate commutators we use $$[F(\Phi),G(\Phi)] = [\Phi^i,\Phi^j] \;\frac{\delta F(\Phi)}{\delta \Phi^i} \frac{\delta G(\Phi)}{\delta \Phi^j}.$$ A straightforward computation gives $$\begin{aligned}
\notag
& [H_n,J_m]+[H_m,J_n] = \frac{1}{(n-1)!}\frac{1}{(m-1)!}\sum_{\lambda_i,\lambda_j} \int d^{3n} q^\perp_i d^{3m}q^\perp_j \delta(\sum_{i=1}^n q^\perp_i)\delta(\sum_{j=n+1}^{n+m} q^\perp_j)
\\
& \frac{\delta(q^\perp_{1}+q^\perp_{n+1})\delta^{\lambda_{1}+\lambda_{n+1},0}}{\beta_{1}-\beta_{n+1}}
\Big( \frac{\partial}{\partial \bar q_{1}}-\frac{\partial}{\partial \bar q_{n+1}} \Big)h_n^{\lambda_{1}\dots \lambda_{n}}
h_m^{\lambda_{n+1}\dots \lambda_{n+m}}\prod_{i=2}^n\Phi^{\lambda_i}(q^\perp_i)\prod_{j=n+2}^{n+m}\Phi^{\lambda_j}(q^\perp_j).
\label{22sep13}
\end{aligned}$$ Note that when $m=n$ one has $[H_n,J_n]$ only once and thus obtains only a half of the right hand side of (\[22sep13\]).
Identities and notations {#sect2oct1}
------------------------
So far we were quite explicit with the variables that $h$ depends on, momentum conserving delta-functions, contractions with fields, etc. To remove unnecessary information that just repeats form line to line, we introduce shortcut notations. Let us illustrate them by the example of (\[22sep13\]) which we will write as $$\label{1oct1}
[H_n,J_m]+[H_m,J_n] = \frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{\beta_{i_j}-\beta_{j_i}}\Big( \frac{\partial}{\partial \bar q_{i_j}}-\frac{\partial}{\partial \bar q_{j_i}} \Big)
h_n^{\{i\}}(q^\perp_{i_j})h_m^{\{j\}}(q^\perp_{j_i}).$$ Here $\{i\}$ refers to the set of indices carried by the variables $h_n^{\{i\}}$ depends on. The set $\{i\}$ has a special element $i_j$, which is associated with a field, that was removed by the commutator. The same holds for $h_m^{\{j\}}$. Dependence of $h$ on special momenta will be important, so it is written explicitly. The momentum conserving delta-functions in our new notations impose $$\sum_{\{ i\}} q^\perp_i = 0 ,\qquad \sum_{\{ j\}} q^\perp_j = 0 ,\qquad q^\perp_{i_j}+ q^\perp_{j_i}=0.$$ In the following they will be implicit. We will also use $$\label{1oct7}
\{i\}_j\equiv \{i\}/i_j, \qquad \{ i - j\}\equiv \{i\}_j \cup \{j\}_i.$$
Below (\[1oct1\]) will be related to an exchange involving vertices $h_n$ and $h_m$ with $\{i\}$ and $\{j\}$ labelling fields entering the first and the second vertices respectively. Moreover, $\{i\}_j$ and $\{j\}_i$ label external legs of the diagram, while $i_j$ and $j_i$ are labels for the exchanged field, see Figure \[fig2vert\].
Next, one has $$\begin{aligned}
\notag
(\sum_{\{i\}_j} \vec q_i)^2 &\; \equiv 2 (\sum_{\{i\}_j} q_i)(\sum_{\{i\}_j} \bar q_i)+2(\sum_{\{i\}_j} \beta_i)(\sum_{\{i\}_j} q^-_i)\\
&\;\approx -2 (-\sum_{\{i\}_j} \beta_i)\Big( h_2 (-\sum_{\{i\}_j} q^\perp_i)+\sum_{\{i\}_j} h_2(q^\perp_i)\Big)=-2\beta_{i_j}{\cal H}^{\{i\}},
\label{23sep3}
\end{aligned}$$ where we used $(\vec q)^2\equiv 2q^-q^++2q\bar q$ for the momentum squared. Then $s_i$ defined by $$\label{1oct2}
s_i \equiv - 2 \beta_{i_j}{\cal H}^{\{i\}}$$ can be interpreted as the Mandelstam variable. Note that in (\[23sep3\]) we employed $q_i^- \approx h_2(q_i^\perp)$ for external particles, which holds only when they are on-shell. So, (\[23sep3\]) should be taken just as a motivation for definition (\[1oct2\]), which is understood off-shell. We also introduce the symmetric Mandelstam variable $$\label{8oct2}
s_{i,j} \equiv \frac{1}{2}(s_i +s_j)= -\frac{1}{2}(\beta_{i_j}-\beta_{j_i})({\cal H}^{\{i\}}-{\cal H}^{\{j\}}).$$
![This figure illustrates our conventions on labelling external and internal lines in the case of two vertices.[]{data-label="fig2vert"}](2vertices)
Below we will use a slightly modified version of (\[1oct3\]) $$\label{1oct4}
\big({\cal J}^{\{i\}}- {\cal H}^{\{i\}}
\frac{\partial}{\partial \bar q_{i_j}}\big) (\sum_{\{i\}} q^{\perp}_i)\propto (\sum_{\{i\}} q^{\perp}_i)$$ as well as $$\label{1oct5}
({\cal J}^{\{i\}}+{\cal J}^{\{j\}}) (q_{i_j}^\perp+ q_{j_i}^\perp) \propto (q_{i_j}^\perp+ q_{j_i}^\perp).$$ Finally, we introduce a notation for $q^\perp_{i_j}$ when it is expressed in terms of external momenta $$\label{1oct8}
Q^\perp_i \equiv -\sum_{\{i\}_j}q^\perp_i, \qquad Q^\perp_{i,j}=\frac{1}{2}(Q^\perp_i - Q^\perp_j).$$
Single exchanges {#29oct3}
----------------
First we note that $$\begin{aligned}
\notag
&\frac{1}{2}({\cal H}^{\{i\}} - {\cal H}^{\{j\}}) \big( \frac{\partial}{\partial \bar q_{i_j}}- \frac{\partial}{\partial \bar q_{j_i}} \big) h_n^{\{i\}}(q^\perp_{i_j})
h_m^{\{j\}}(q^\perp_{j_i})\\
\label{25sep8}
& \qquad \qquad \qquad \qquad =\big( {\cal H}^{\{i\}}\frac{\partial}{\partial \bar q_{i_j}}
+{\cal H}^{\{j\}}\frac{\partial}{\partial \bar q_{j_i}}\big)
h_n^{\{i\}} \big(\tfrac{1}{2}(q^\perp_{i_j}-q^\perp_{j_i})\big) h_m^{\{j\}} \big(\tfrac{1}{2}(q^\perp_{j_i}-q^\perp_{i_j})\big),
\end{aligned}$$ where the arguments $q^\perp_{i_j}$ and $q^\perp_{j_i}$ should be replaced before evaluating derivatives as indicated. Then the right hand side of (\[1oct1\]) can be written as $$\begin{aligned}
\notag
& [H^n,J^m]+[H^m, J^n]\\
&\;\;=\frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}}\big( -{\cal H}^{\{i\}}\frac{\partial}{\partial \bar q_{i_j}}
-{\cal H}^{\{j\}}\frac{\partial}{\partial \bar q_{j_i}}\big)
h_n^{\{i\}} \big( \tfrac{1}{2}(q^\perp_{i_j}-q^\perp_{j_i})\big) h_m^{\{j\}} \big(\tfrac{1}{2}(q^\perp_{j_i}-q^\perp_{i_j})\big),
\end{aligned}$$ where $s_{i,j}$ was defined in (\[8oct2\]).
Next we proceed by adding and subtracting ${\cal J}^{\{i\}}+ {\cal J}^{\{j\}}$, so as to produce operators that commute with the total momentum as in (\[1oct4\]) $$\begin{aligned}
\notag
& [H_n,J_m]+[H_m, J_n]\\
\notag
& \qquad =\frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}}\big({\cal J}^{\{i\}}+ {\cal J}^{\{j\}}- {\cal H}^{\{i\}}\frac{\partial}{\partial \bar q_{i_j}}
-{\cal H}^{\{j\}}\frac{\partial}{\partial \bar q_{j_i}}\big) \\
\notag
&\qquad\qquad\qquad\qquad\qquad\qquad \qquad\qquad \qquad \;
h_n^{\{i\}} \big( \tfrac{1}{2}(q^\perp_{i_j}-q^\perp_{j_i})\big) h_m^{\{j\}} \big( \tfrac{1}{2}(q^\perp_{j_i}-q^\perp_{i_j})\big)\\
\label{1oct6}
&\qquad -\frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}} \big({\cal J}^{\{i\}}+ {\cal J}^{\{j\}}\big) h_n^{\{i\}} \big( \tfrac{1}{2}(q^\perp_{i_j}-q^\perp_{j_i})\big) h_m^{\{j\}} \big( \tfrac{1}{2}(q^\perp_{j_i}-q^\perp_{i_j})\big).
\end{aligned}$$ The operator that appears in the first term permits us to use momentum conservation inside $h_m$ and $h_n$, so we can eliminate $q^\perp_{i_j}$ and $q^\perp_{j_i}$ in favour of external momenta. After that $h_n$ and $h_m$ no longer depend on exchanged momenta explicitly, hence differential operators acting on them can be dropped. As a result, for the first term in (\[1oct6\]) we obtain $$\begin{aligned}
\notag
&\frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}}\big({\cal J}^{\{i\}}+ {\cal J}^{\{j\}}- {\cal H}^{\{i\}}\frac{\partial}{\partial \bar q_{i_j}}
-{\cal H}^{\{j\}}\frac{\partial}{\partial \bar q_{j_i}}\big) \\
\notag
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad \qquad \qquad
h_n^{\{i\}} \big(\tfrac{1}{2}(q^\perp_{i_j}-q^\perp_{j_i})\big) h_m^{\{j\}} \big( \tfrac{1}{2}(q^\perp_{j_i}-q^\perp_{i_j})\big)\\
&\qquad\qquad\qquad\qquad\qquad\qquad\quad =\frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}}{\cal J}^{\{i-j\}}
h_n^{\{i\}} ( Q^\perp_{i,j}) h_m^{\{j\}} ( Q^\perp_{j,i}),
\end{aligned}$$ where $\{i-j\}$ and $Q^\perp_{i,j}$ were defined in (\[1oct7\]) and (\[1oct8\]).
Employing (\[1oct5\]), the second term reads $$\begin{aligned}
\notag
& -\frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}} \big({\cal J}^{\{i\}}+ {\cal J}^{\{j\}}\big) h_n^{\{i\}} \big( \tfrac{1}{2}(q^\perp_{i_j}-q^\perp_{j_i})\big) h_m^{\{j\}} \big( \tfrac{1}{2}(q^\perp_{j_i}-q^\perp_{i_j})\big)\\
&\qquad\qquad\qquad\qquad\qquad=- \frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}}\big({\cal J}^{\{i\}}+{\cal J}^{\{j\}}\big) h_n^{\{i\}}(q^\perp_{i_j})
h_m^{\{j\}}(q^\perp_{j_i}).
\end{aligned}$$
Combining both contributions, we find that $$\begin{aligned}
\notag
[H_n,J_m]+[H_m, J_n]
&= \frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}}{\cal J}^{\{i-j\}}
h_n^{\{i\}} (Q^\perp_{i,j}) h_m^{\{j\}} (Q^\perp_{j,i})\\
\label{25sep9}
& - \frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}}\big({\cal J}^{\{i\}}+{\cal J}^{\{j\}}\big) h_n^{\{i\}}(q^\perp_{i_j})h_m^{\{j\}}(q^\perp_{j_i}).
\end{aligned}$$
To interpret this result, first note that in the first term ${\cal J}^{\{ i-j\}}$ acts only on external momenta. Let us denote it just by ${\cal J}$. In Section \[sect1oct1\] it will be shown that $s_{i,j}$ commutes with ${\cal J}$. Using these facts one can rewrite the consistency condition (\[22sep11\]) as $$\begin{aligned}
\notag
&\sum_{k=0}^{n-2}[H_{n-k},J_{k+2}]={\cal J}\Big( \frac{1}{n!} h_n^{\{i\}} + \sum_{k=1}^{n-3} \frac{1}{(n-k-1)!}\frac{1}{(k+1)!}
h^{\{i\}}_{n-k} ( Q^\perp_{i,j}) \frac{1}{s_{i,j}} h_{k+2}^{\{j\}} ( Q^\perp_{j,i})\Big)\\
\label{1oct9}
& \qquad\qquad\qquad
-\sum_{k=1}^{n-3} \frac{1}{(n-k-1)!}\frac{1}{(k+1)!}\frac{1}{s_{i,j}}\big({\cal J}^{\{i\}}+{\cal J}^{\{j\}}\big)h_{n-k}^{\{i\}}(q^\perp_{i_j})
h_{k+2}^{\{j\}}(q^\perp_{j_i})=0.\end{aligned}$$ After multiplying the first line by $n!$ in brackets we recover the sum of the contact $n$-point diagram and of all exchanges involving $n$ external particles and a single propagator. The combinatorial factors $$\frac{n!}{(n-k-1)!(k+1)!},$$ that appear in front of exchanges count all possible channels that each given exchange can have. As it was noted below (\[22sep13\]), for $n-k = k+2$ the exchange will get an extra factor of $1/2$, which is the standard symmetry factor associated with a symmetry that interchanges identical vertices.
In other words, up to contributions from diagrams involving more than one propagator, the consistency condition (\[1oct9\]) looks exactly as the Ward identity for the $n$-point amplitude, where ${\cal J}$ is the operator that verifies gauge invariance. At this point the term Ward identity may sound misleading, because in the light-cone approach gauge invariance is completely fixed. We will justify this terminology in Section \[sect5oct3\], where we will make a connection between the light-cone approach and the spinor-helicity formalism.
Double exchanges {#sec29oct2}
----------------
The second term on the right hand side of (\[25sep9\]) is responsible for contributions of multiple exchanges to the Ward identity. To prevent possible confusions, let us clarify, that by double and multiple exchanges we mean tree-level diagrams that involve two or more internal lines.
In this Section we will show how to reproduce contributions from diagrams with two exchanges. To this end, let us focus on a particular diagram, which consists of two vertices $h^{\{k\}}_p$ and $h^{\{j\}}_m$ connected by a pair of exchanges and a vertex $h^{\{l\}}_q$, see Figure \[fig3vert\]. The remaining contributions will be omitted, which will be indicated by $\to$ instead of $=$. To have the same number of external fields as in (\[25sep9\]) we have to demand $p+q-2 = n$. One finds $$\begin{aligned}
\notag
& [H_n,J_m]+[H_m,J_n] \to - \frac{1}{(n-1)!}\frac{1}{(m-1)!}\frac{1}{s_{i,j}}{\cal J}^{\{i\}} h_n^{\{i\}}(q^\perp_{i_j})h_m^{\{j\}}(q^\perp_{j_i})\\
\label{2oct1}
& \to \frac{1}{(p-1)!}\frac{1}{(q-1)!}\frac{n}{(m-1)!}\frac{1}{s_{kl,j}} \frac{1}{\beta_{k_l}-\beta_{l_k}}
h^{\{j\}}_m(q^\perp_{j_i}) \big(\frac{\partial}{ \partial \bar q_{k_l}}-\frac{\partial}{ \partial \bar q_{l_k}} \big)h^{\{k\}}_p(q^\perp_{k_l})h^{\{l\}}_q(q^\perp_{l_k}).
\end{aligned}$$ To obtain the last line we used the consistency condition that relates $h_n^{\{i\}}$ to commutators involving the Hamiltonians of lower degrees $h^{\{k\}}_p$ and $h^{\{l\}}_q$. We also renamed $s_{i,j}\to s_{kl,j}$ to be consistent with the fact that the legs of the diagram, that before using the consistency condition for $h_n^{\{i\}}$ were labelled by $\{i\}$, after that belong to $h^{\{k\}}_p$ and $h^{\{l\}}_q$ and hence are labelled by the sets $\{k\}$ and $\{l\}$.
![This figure illustrates how after using the consistency condition for $h_n^{\{i\}}$ its external lines previously labelled by $\{i\}_j$ split into two groups labelled by $\{k\}_l$ and $\{l\}_{kj}$. Accordingly, we rename $i_j\to l_j$, $j_i\to j_l$ and $s_{i,j}\to s_{kl,j}$.[]{data-label="fig3vert"}](3vertices)
We would like to remind the reader, that in the first line of (\[2oct1\]) there is an implicit delta-function $\delta(q^\perp_{i_j}+q^\perp_{j_i})$, which relates momenta on different sides of the exchange. When we go to the second line, the set $\{ i\}$ is split into $\{k\}$ and $\{l\}$ and the special index $i_j$ of the set $\{i\}$ can either belong to $\{k\}$ or to $\{l\}$. Here we keep only the terms, relevant to $\{ k-l-j\}$ exchange that is those with $i_j \in \{l\}$. This produces an extra combinatorial factor $(q-1)/n$. As a result $$\begin{aligned}
\notag
& [H_n,J_m]+[H_m,J_n] \\
\label{2oct02}
&\to \frac{1}{(p-1)!}\frac{1}{(q-2)!}\frac{1}{(m-1)!}\frac{1}{s_{kl,j}} \frac{1}{\beta_{k_l}-\beta_{l_k}}
\big(\frac{\partial}{ \partial \bar q_{k_l}}-\frac{\partial}{ \partial \bar q_{l_k}} \big)h^{\{k\}}_p(q^\perp_{k_l})h^{\{l\}}_q(q^\perp_{l_k},q^\perp_{l_j})
h^{\{j\}}_m(q^\perp_{j_l}),
\end{aligned}$$ where we also relabelled $i_j$ to $l_j$ and $j_i$ to $j_l$.
Leaving aside the combinatorial factor for a moment, we proceed with the rest analogously to a single exchange case $$\begin{aligned}
\notag
W^{\{k\text{-}l\text{-}j\}}_1&\,\equiv \frac{1}{s_{kl,j}}\frac{1}{\beta_{k_l}-\beta_{l_k}}\big( \frac{\partial}{ \partial \bar q_{k_l}}-
\frac{\partial}{\partial \bar q_{l_k}}\big) h^{\{k\}}_p(q^\perp_{k_l})h^{\{l\}}_q(q^\perp_{l_k},q^\perp_{l_j})
h^{\{j\}}_m(q^\perp_{j_l})\\
\notag
&\,=-\frac{1}{s_{kl,j}} \frac{1}{s_{k,lj}}({\cal H}^{\{k\}} - {\cal H}^{\{l\}} - {\cal H}^{\{j\}})
\big( \frac{\partial}{ \partial \bar q_{k_l}}-
\frac{\partial}{\partial \bar q_{l_k}}\big) h^{\{k\}}_p(q^\perp_{k_l})h^{\{l\}}_q(q^\perp_{l_k},q^\perp_{l_j})
h^{\{j\}}_m(q^\perp_{j_l})
\\
\notag
&\,=-\frac{1}{s_{kl,j}} \frac{1}{s_{k,lj}} \big( {\cal H}^{\{k\}} \frac{\partial}{\partial \bar q_{k_l}}+
({\cal H}^{\{l\}} + {\cal H}^{\{j\}})\frac{\partial}{\partial \bar q_{l_k}}\big)
h_p^{\{k\}}\big( \tfrac{1}{2}(q^\perp_{k_l}-q^\perp_{l_k})\big)\\
\label{25sep12}
&\,\qquad \qquad\qquad \qquad\qquad \quad h_q^{\{l\}}\big(\tfrac{1}{2}(q^\perp_{l_k}-q^\perp_{k_l}),\tfrac{1}{2}(q^\perp_{l_j}-q^\perp_{j_l})\big)
h_m^{\{j\}}\big( \tfrac{1}{2}(q^\perp_{j_l}-q^\perp_{l_j})\big).
\end{aligned}$$ In a similar way, from $[H_p,J_{q+m-2}]+[H_{q+m-2},J_p]$ we find another contribution $$\begin{aligned}
\notag
W^{\{k\text{-}l\text{-}j\}}_2&\,\equiv -\frac{1}{s_{kl,j}} \frac{1}{s_{k,lj}} \big( ({\cal H}^{\{k\}} + {\cal H}^{\{l\}})\frac{\partial}{\partial \bar q_{l_j}} +{\cal H}^{\{j\}} \frac{\partial}{\partial \bar q_{j_l}}\big)
h_p^{\{k\}}\big( \tfrac{1}{2}(q^\perp_{k_l}-q^\perp_{l_k})\big)\\
\label{2oct2}
&\,\qquad \qquad\qquad \qquad\qquad \quad h_q^{\{l\}}\big(\tfrac{1}{2}(q^\perp_{l_k}-q^\perp_{k_l}),\tfrac{1}{2}(q^\perp_{l_j}-q^\perp_{j_l})\big)
h_m^{\{j\}}\big( \tfrac{1}{2}(q^\perp_{j_l}-q^\perp_{l_j})\big).
\end{aligned}$$
For the sum of two terms we proceed as for the single exchange case in (\[1oct6\]): we add and subtract ${\cal J}^{\{k\}}+{\cal J}^{\{l\}}+{\cal J}^{\{j\}}$ producing an operator that allows to use momentum conservation inside the diagram. Namely, $$\begin{aligned}
\notag
W^{\{k\text{-}l\text{-}j\}}_1&+W^{\{k\text{-}l\text{-}j\}}_2 =\frac{1}{s_{kl,j}} \frac{1}{s_{k,lj}}\big({\cal J}^{\{k\}}+{\cal J}^{\{l\}}+{\cal J}^{\{j\}}\\
\notag
& \qquad \qquad \qquad - {\cal H}^{\{k\}} \frac{\partial}{\partial \bar q_{k_l}}-
({\cal H}^{\{l\}} + {\cal H}^{\{j\}})\frac{\partial}{\partial \bar q_{l_k}} -({\cal H}^{\{k\}} + {\cal H}^{\{l\}})\frac{\partial}{\partial \bar q_{l_j}} -{\cal H}^{\{j\}} \frac{\partial}{\partial \bar q_{j_l}}\big)\\
\notag
& \quad \qquad\qquad\qquad h_p^{\{k\}}\big( \tfrac{1}{2}(q^\perp_{k_l}-q^\perp_{l_k})\big)
h_q^{\{l\}}\big(\tfrac{1}{2}(q^\perp_{l_k}-q^\perp_{k_l}),\tfrac{1}{2}(q^\perp_{l_j}-q^\perp_{j_l})\big)
h_m^{\{j\}}\big( \tfrac{1}{2}(q^\perp_{j_l}-q^\perp_{l_j})\big)\\
& \qquad \qquad -\frac{1}{s_{kl,j}} \frac{1}{s_{k,lj}}\big({\cal J}^{\{k\}}+{\cal J}^{\{l\}}+{\cal J}^{\{j\}}\big)h^{\{k\}}_p(q^\perp_{k_l})h^{\{l\}}_q(q^\perp_{l_k},q^\perp_{l_j})
h^{\{j\}}_m(q^\perp_{j_l}).
\end{aligned}$$ Employing (\[1oct4\]) and (\[1oct5\]), it is not hard to see that the operator in the first term allows to use $$\begin{aligned}
\notag
\sum_{\{k\}} q^\perp_k =0, \quad \sum_{\{l\}_j} q^\perp_l+ \sum_{\{j\}_l} q^\perp_j =0,\\
\sum_{\{j\}} q^\perp_j =0, \quad \sum_{\{l\}_k} q^\perp_l+ \sum_{\{k\}_l} q^\perp_k =0.
\end{aligned}$$ These momentum conservation conditions permit us to eliminate $q^\perp_{k_l}$, $q^\perp_{l_k}$, $q^\perp_{l_j}$ and $q^\perp_{l_j}$ expressing them in terms of external momenta. After that vertices no longer depend on exchanged momenta, so differential operators acting on them can be dropped. Thus, we find $$\begin{aligned}
\notag
W^{\{k\text{-}l\text{-}j\}}_1+W^{\{k\text{-}l\text{-}j\}}_2
&=\frac{1}{s_{kl,j}} \frac{1}{s_{k,lj}} {\cal J}^{\{ k-l-j\}}
h_p^{\{k\}}( Q^\perp_{k,lj})
h_q^{\{l\}}(Q^\perp_{lj,k},Q^\perp_{kl,j})
h_m^{\{j\}}( Q^\perp_{j,kl})\\
\label{2oct3}
& -\frac{1}{s_{kl,j}} \frac{1}{s_{k,lj}}\big({\cal J}^{\{k\}}+{\cal J}^{\{l\}}+{\cal J}^{\{j\}}\big)h^{\{k\}}_p(q^\perp_{k_l})h^{\{l\}}_q(q^\perp_{l_k},q^\perp_{l_j})
h^{\{j\}}_m(q^\perp_{j_l}).\end{aligned}$$
The second line is responsible for contributions of diagrams involving at least three propagators. Reinstating the combinatorial factor from (\[2oct02\]), for the first line of (\[2oct3\]) we find $$\begin{aligned}
\notag
&[H_n,J_m]+[H_m,J_n] +[H_p,J_{q+m-2}]+[H_{q+m-2},J_p]\\
\label{2oct4}
& \quad \to \frac{1}{(p-1)!}\frac{1}{(q-2)!}\frac{1}{(m-1)!} {\cal J}
h_p^{\{k\}}( Q^\perp_{k,lj})\frac{1}{s_{kl,j}}
h_q^{\{l\}}(Q^\perp_{lj,k},Q^\perp_{kl,j})\frac{1}{s_{k,lj}}
h_m^{\{j\}}( Q^\perp_{j,kl}).
\end{aligned}$$ This gives one of the double exchange contributions to the Ward identity involving a contact vertex of degree ${p+q+m-4}$. This contact vertex in our conventions goes with a prefactor $1/(p+q+m-4)!$. Normalising the contribution of the contact diagram to unity, we find that the double exchange from (\[2oct4\]) appears with a factor $$\frac{(p+q+m-4)!}{(p-1)!(q-2)!(m-1)!},$$ which just gives the total number of channels. This is exactly what we should expect from the Feynman rules if we symmetrise them over external fields, as in our approach.
General case {#29oct4}
------------
In the previous two Sections we showed that the consistency condition for the non-linear deformation in the light-cone approach (\[22sep5\]) can be rewritten as the Ward identity involving contributions from contact diagrams, single exchanges and some extra terms bilinear in vertices of degrees lower than that of a contact contribution (\[1oct9\]). Employing the consistency condition for these extra terms, they were shown to produce contributions of double exchanges plus extra terms, trilinear in vertices of yet lower degrees (\[2oct3\]). This procedure should be repeated recursively until it terminates due to ${\cal J}h_3 =0$, see (\[22sep11\]), thus reproducing the Ward identity for the complete amplitude. The equation $[H,\bar J]=0$ is analogous. The results of this recursive procedure can be summarised as
\[prop2oct1\] The light-cone consistency condition can be equivalently rewritten as a set of Ward identities for all $n$-point amplitudes $$\label{2oct5}
{\cal J}^{\{i\}} {\cal A}_n (q^\perp_i)=0, \qquad \bar {\cal J}^{\{i\}} {\cal A}_n (q^\perp_i)=0,$$ where $$\begin{aligned}
\notag
{\cal J}^{\{i\}}&\, =\sum_{i=1}^n \Big(-\frac{q_i\bar q_i}{\beta_i}\frac{\partial}{\partial \bar q_i}-q_i \frac{\partial}{\partial \beta_i}+\lambda_i \frac{q_i}{\beta_i}\Big),\\
\bar{\cal J}^{\{i\}}&\, =\sum_{i=1}^n \Big(-\frac{q_i\bar q_i}{\beta_i}\frac{\partial}{\partial q_i}-\bar q_i \frac{\partial}{\partial \beta_i}+\lambda_i \frac{\bar q_i}{\beta_i}\Big)
\end{aligned}$$ and ${\cal A}_n (q^\perp_i)$ is the off-shell light-cone amplitude constructed according to the following Feynman rules:
- the propagator, splitting the external fields into sets labelled by $\{k\}_l$ and $\{l\}_k$, is given by $$\label{8nov1}
\frac{1}{s_{k,l}} = -\frac{2}{(\beta_{k_l}-\beta_{l_k})({\cal H}^{\{k\}}-{\cal H}^{\{l\}})},$$ where $$\begin{aligned}
\notag
\beta_{k_l} = -\sum_{\{k\}_l} \beta_{k},& \qquad {\cal H}^{\{k\}} = \sum_{\{k\}_l} h_2(q_k^\perp) + h_2\big(-\sum_{\{k\}_l} q_k^\perp\big)\\
\beta_{l_k} = -\sum_{\{l\}_k} \beta_{l},& \qquad {\cal H}^{\{l\}} = \sum_{\{l\}_k} h_2(q_l^\perp) + h_2\big(-\sum_{\{l\}_k} q_l^\perp\big).
\end{aligned}$$ More explanations on our conventions are given in Section \[sect2oct1\].
- Inside vertices, the momenta of exchanged fields $q_{k_l}^\perp$ and $q_{l_k}^\perp$ should be expressed in terms of momenta of external fields as $$\begin{aligned}
\notag
&q^\perp_{k_l} \to Q_{k,l} = \frac{1}{2}(-\sum_{\{k\}_l} q^\perp_{k}+\sum_{\{l\}_k} q^\perp_l),\\
&q^\perp_{l_k} \to Q_{l,k} = \frac{1}{2}(-\sum_{\{l\}_k} q^\perp_{l}+\sum_{\{k\}_l} q^\perp_k).
\label{2oct6}
\end{aligned}$$
Alternatively, (\[2oct5\]) can be written as $$\begin{aligned}
[A,J_2]=0, \qquad
\label{9oct2}
[A,\bar J_2]=0,\end{aligned}$$ where $A$ is made of the amplitude ${\cal A}(q^\perp)$ by contracting it with fields.
Let us now make several comments. First, as it was noted above, the amplitude ${\cal A}(q^\perp)$ is an off-shell object. Indeed, in the light-cone approach neither the free Hamiltonian nor interaction terms depend on $q_i^-$. For this simple reason ${\cal A}(q^\perp)$ defined above makes perfect sense for any $q_i^-$, not only for on-shell values $q_i^- \approx h_2(q_i^\perp)$. Moreover, ${\cal A}({q^\perp})$ is related unambiguously to the Hamiltonian, so it defines the action.
It is worth to note the unusual form of the propagator (\[8nov1\]), which we employed to construct ${\cal A}(q^\perp)$. It follows from (\[23sep3\]) that this propagator coincides with the one that comes from the covariant Feynman rules when external particles are on-shell. They are, nevertheless, different for off-shell external momental. This difference can be explained by the fact that the framework we are dealing with is the light-cone Hamiltonian perturbation theory, which has some peculiar features, see e.g. [@Kogut:1969xa].
We recall that to derive (\[2oct5\]) we fixed the freedom of integration by parts in interaction terms by (\[22sep10\]). Of course, this condition should not be important for consistency. In particular, it should be possible to use conservation of the total momentum inside ${\cal A}_n(q^\perp)$. The resulting amplitude will no longer be annihilated by ${\cal J}^{\{i\}}$, but, as it is not hard to see, will produce terms proportional to ${\cal H}^{\{i\}}$. Returning back to the consistency conditions in the original form (\[22sep9\]), one can see that such extra terms are harmless, as they can be absorbed by the appropriate redefinition of $j_n$. Thus, one can reformulate Proposition \[prop2oct1\] above in a form, where integration by parts is allowed, but (\[2oct5\]) holds only up to terms proportional to ${\cal H}^{\{i\}}$. One can then use the freedom of integration by parts to simplify slightly the Feynman rules presented above. In particular, it is not necessary to express momenta of internal lines in a symmetric way in terms of external momenta on both sides of a given propagator as in (\[2oct6\]).
Note, that for four external lines (\[2oct5\]) was found in [@Metsaev:1991mt; @Metsaev:1991nb]. More precisely, it was observed that one can solve the light-cone consistency condition at the four-point level by taking the quartic vertex to be minus the sum of exchanges. Our analysis gives a simple derivation of this fact and extends it to all orders.
Finally, let us point out that even though the form (\[2oct5\]) of the consistency condition looks much more natural from the perturbative field theory perspective, it is sometimes instructive to keep in mind the original form of the consistency condition as well. In particular, it is clear from (\[22sep11\]) that once cubic interactions are consistent by themselves, that is $[H_3,J_3]=0$, then by setting all higher vertices to zero we obtain a consistent theory. At the same time, using the language of (\[2oct5\]) this translates into the statement, that once we verified that cubic vertices result into consistent four-point exchanges, then all higher-point exchanges involving only cubic vertices are also consistent. The latter statement is not so easily seen only from (\[2oct5\]). The origin of this non-trivial translation between the two languages is the iterative procedure that we needed to undertake to derive (\[2oct5\]) from (\[22sep11\]). It would be interesting to find more non-trivial examples of this phenomenon.
Solution of the Ward identity {#sect1oct1}
=============================
Previous analysis emphasises the central role played by operators ${\cal J}$ and $\bar{\cal J}$. It was shown that consistency of the non-linear deformation implies that these operators annihilate the total amplitude. In this Section we find a general solution of these equations.
External scalars {#sect5oct2}
----------------
To start we find a general solution of (\[2oct5\]) in the case when all external fields are scalars. The first equation explicitly reads $$\label{3oct2}
{\cal J}^{\{i\}} {\cal A}_n(q_i^\perp) = 0, \qquad {\cal J}^{\{i\}} = \sum_{i=1}^n\left(-\frac{q_i\bar q_i}{\beta_i}\frac{\partial}{\partial \bar q_i}-q_i \frac{\partial}{\partial \beta_i}\right).$$
We remind the reader that the Hamiltonian should satisfy kinematical constraints (\[30sep2\])-(\[30sep5\]). By inspecting the Feynman rules given in the previous Section, it is not hard to see that the same kinematical constraints carry over to the total amplitude.
As it was mentioned above, the first two conditions (\[30sep2\]), (\[30sep3\]) imply that the amplitude is a general function of the following varaibles $$\label{3oct1}
\bar{\mathbb{P}}_{ij} \equiv \bar q_i\beta_j -\bar q_j \beta_i , \qquad \mathbb{P}_{ij} \equiv q_i\beta_j -q_j \beta_i, \qquad \beta_i.$$ This can be seen, for instance, by considering characteristic vector fields associated with differential equations (\[30sep2\]), (\[30sep3\]). Variables (\[3oct1\]) provide an overcomplete set of invariants of the associated characteristic flows.
When we add an extra differential equation (\[3oct2\]), clearly, the set of invariants can only be reduced. Upon proper dressing with $\beta$ dependence, both $\bar{\mathbb{P}}_{ij}$ and $\mathbb{P}_{ij}$ can be promoted to solutions of (\[3oct2\]) $$\bar{\mathbb{P}}_{ij} \to \bar\Pi_{ij} \equiv \frac{\bar{\mathbb{P}}_{ij}}{\beta_i\beta_j}: \qquad \qquad
{\cal J}^{\{i\}} \bar\Pi_{ij}=0,\qquad {\cal J}^{\{i\}} \mathbb{P}_{ij}=0.$$ It is not hard to see that there are no solutions of (\[3oct2\]) that depend on $\beta_i$ alone, so ${\cal A}_n(q_i^\perp)$ is an arbitrary function of $\bar\Pi_{ij}$ and $\mathbb{P}_{ij}$. It is convenient to phrase this conclusion as $$\label{3oct4}
{\cal A}_n(q_i^\perp) = \phi(\bar{\mathbb{P}}_{ij},\mathbb{P}_{ij},\beta_i), \qquad -N_{\beta_i} = N_{\langle i|},$$ where $$N_{\langle i|}\equiv \sum_{j=1}^n N_{\mathbb{P}_{ij}}$$ counts the total homogeneity degree of all $\mathbb{P}_{ij}$’s with fixed $i$ and any $j$.
Similarly, to solve $$\label{3oct3}
\bar{\cal J}^{\{i\}} {\cal A}_n(q_i^\perp) = 0, \qquad \bar{\cal J}^{\{i\}} = \sum_{i=1}^n\left(-\frac{q_i\bar q_i}{\beta_i}\frac{\partial}{\partial q_i}-\bar q_i \frac{\partial}{\partial \beta_i}\right)$$ we introduce new variables $\Pi_{ij}$ $${\mathbb{P}}_{ij} \to \Pi_{ij} \equiv \frac{{\mathbb{P}}_{ij}}{\beta_i\beta_j}: \qquad \qquad
\bar{\cal J}^{\{i\}} \bar{\mathbb{P}}_{ij}=0,\qquad \bar{\cal J}^{\{i\}} \Pi_{ij}=0$$ and find that ${\cal A}_n(q_i^\perp)$ is some function of $\bar{\mathbb{P}}_{ij}$ and $\Pi_{ij}$, which is equivalent to $$\label{3oct5}
{\cal A}_n(q_i^\perp) = \phi(\bar{\mathbb{P}}_{ij},\mathbb{P}_{ij},\beta_i), \qquad -N_{\beta_i} = N_{|i]},$$ where $$N_{|i]}\equiv \sum_{j=1}^n N_{\bar{\mathbb{P}}_{ij}}$$ counts the total homogeneity degree of all $\bar{\mathbb{P}}_{ij}$’s with fixed $i$ and any $j$.
Combining (\[3oct4\]) and (\[3oct5\]) we find that $$\label{3oct6}
N_{|i]} = N_{\langle i|}.$$ Provided (\[3oct6\]) is satisfied, we can rewrite the remaining homogeneity constraint as $$\label{8oct1}
-N_{\beta_i} = \frac{ N_{|i]} +N_{\langle i|}}{2}.$$ It can be solved as $$\label{3oct7}
{\cal A}_n(q_i^\perp) = \chi ([ij],\langle ij\rangle),$$ where $$\label{27sep14}
[ij] \equiv \frac{\sqrt{2}}{\sqrt{\beta_i\beta_j}}\bar{\mathbb{P}}_{ij}, \qquad \langle ij \rangle \equiv -\frac{\sqrt{2}}{\sqrt{\beta_i\beta_j}}\mathbb{P}_{ij}.$$ Here the numerical coefficients have been chosen so as to agree with the standard conventions of the spinor-helicity approach reviewed in Appendix \[ap3oct1\].
In the next Section it will be used that $[ij]$ and $\langle ij\rangle$ can be factorised in terms of helicity spinors $$\label{5oct1}
| i]_a =
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_i}}\left(\begin{array}{c}
\bar q_i\\
-\beta_i
\end{array}\right),
\qquad
\langle i |_{\dot b} = \frac{2^{\frac{1}{4}}}{\sqrt{\beta_i}}
\left(\begin{array}{ccc}
q_i&& -\beta_i
\end{array}\right)$$ as $$\begin{aligned}
[ij] &\,= \varepsilon^{ab} |j]_a |i]_b=
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_j}}
\left(\begin{array}{ccc}
\bar q_j&& -\beta_j
\end{array}\right)
\left(
\begin{array}{ccc}
0 && 1\\
-1 && 0
\end{array}
\right)
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_i}}\left(\begin{array}{c}
\bar q_i\\
-\beta_i
\end{array}\right),
\\
\langle ij\rangle&\, = \varepsilon^{\dot a\dot b} \langle i|_{\dot a}\langle j|_{\dot b}=
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_i}}
\left(\begin{array}{ccc}
q_i&& -\beta_i
\end{array}\right)
\left(
\begin{array}{ccc}
0 && 1\\
-1 && 0
\end{array}
\right)
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_j}}\left(\begin{array}{c}
q_j\\
-\beta_j
\end{array}\right).\end{aligned}$$ In these terms (\[3oct6\]) reads $$\label{3oct8}
N_{| i]} - N_{\langle i |}=0, \qquad N_{| i]} = | i ]\frac{\partial}{\partial | i]}, \qquad N_{\langle i|} = \langle i |\frac{\partial}{\partial \langle i |}.$$ Equation (\[3oct7\]) supplemented with a homogeneity constraint (\[3oct8\]) gives a general solution of the Ward identities (\[3oct2\]), (\[3oct3\]) in the case of external scalars.
General case {#sect9oct1}
------------
In this Section we prove
\[prop5oct1\] A general solution of the Ward identities $$\begin{aligned}
\notag
{\cal J}^{\{i\}} {\cal A}_n(q_i^\perp) &\,= 0, \qquad {\cal J}^{\{i\}} = \sum_{i=1}^n\left(-\frac{q_i\bar q_i}{\beta_i}\frac{\partial}{\partial \bar q_i}-q_i \frac{\partial}{\partial \beta_i}+\lambda_i\frac{q_i}{\beta_i}\right),
\\
\label{3oct08}
\bar{\cal J}^{\{i\}} {\cal A}_n(q_i^\perp) &\,= 0, \qquad \bar{\cal J}^{\{i\}} = \sum_{i=1}^n\left(-\frac{q_i\bar q_i}{\beta_i}\frac{\partial}{\partial q_i}-\bar q_i \frac{\partial}{\partial \beta_i}-\lambda_i\frac{\bar q_i}{\beta_i}\right) \end{aligned}$$ is given by $$\label{3oct9}
{\cal A}_n(q_i^\perp) = \chi ([ij],\langle ij\rangle),$$ where $\chi$ satisfies a homogeneity condition $$\label{3oct10}
(-N_{| i]} + N_{\langle i|} +2\lambda_i) \chi ([ij],\langle ij\rangle)=0$$ for every $i$. Spinor products are defined in (\[27sep14\]).
To show this, let us first verify that (\[3oct9\]), (\[3oct10\]) is indeed a solution. To evaluate how the differential part of ${\cal J}^{\{i\}}$ acts on $\chi$ we use $$\begin{aligned}
\notag
-\sum_{i=1}^n\frac{q_i\bar q_i}{\beta_i}\frac{\partial}{\partial \bar q_i}|j] &\,= -\sum_{i=1}^n\frac{q_i}{\beta_i} N_{\bar q_i}
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_j}}\left(\begin{array}{c}
\bar q_j\\
-\beta_j
\end{array}\right)
=-\frac{q_j}{\beta_j}
\left(\begin{array}{cccc}
1 &&& 0\\
0 && &0
\end{array}\right)|j],
\\
\notag
-\sum_{i=1}^n q_i\frac{\partial}{\partial \beta_i} |j] &\,= -\sum_{i=1}^n\frac{q_i}{\beta_i} N_{ \beta_i}
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_j}}\left(\begin{array}{c}
\bar q_j\\
-\beta_j
\end{array}\right)=
\frac{1}{2}\frac{q_j}{\beta_j}
\left(\begin{array}{ccc}
1 && 0\\
0 && -1
\end{array}\right)|j],
\\
-\sum_{i=1}^n q_i\frac{\partial}{\partial \beta_i} \langle j| &\,=
-\sum_{i=1}^n\frac{q_i}{\beta_i} N_{ \beta_i}
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_j}}\left(\begin{array}{ccc}
q_j&& -\beta_j
\end{array}\right)
=
\frac{1}{2}\frac{q_j}{\beta_j}
\langle j|
\left(\begin{array}{ccc}
1 && 0\\
0 && -1
\end{array}\right)
\label{3oct11}\end{aligned}$$ and the Leibniz rule. We find $$\begin{aligned}
\notag
{\cal J}^{\{i\}} \chi ([ij],\langle ij\rangle) =&\, \sum_{i=1}^n \frac{q_i}{\beta_i}\Big( -\frac{1}{2} |i] \frac{\partial}{\partial |i]}
+
\frac{1}{2}\langle i |\left(\begin{array}{ccc}
1 && 0\\
0 && -1
\end{array}\right) \frac{\partial}{\partial \langle i|}
+\lambda_i
\Big)\chi ([ij],\langle ij\rangle) \\
&\qquad \qquad \qquad \qquad \qquad=-\sum_{i=1}^n \frac{q_i}{\beta_i} \langle i |
\left(\begin{array}{cccc}
0 &&& 0\\
0 && &1
\end{array}\right) \frac{\partial}{\partial \langle i|}
\chi ([ij],\langle ij\rangle) ,\end{aligned}$$ where to get to the last line we used the homogeneity condition (\[3oct10\]). Now we take into account that in $\chi$ spinors can only appear in the form of spinor contractions $${\cal J}^{\{i\}} \chi ([ij],\langle ij\rangle) =-
\sum_{i,k,l=1}^n \frac{q_i}{\beta_i} \langle i |
\left(\begin{array}{cccc}
0 &&& 0\\
0 && &1
\end{array}\right) \frac{\partial\langle kl \rangle}{\partial \langle i|}
\frac{\partial}{\partial\langle kl\rangle}\chi ([ij],\langle ij\rangle).$$ Using that $$\begin{aligned}
\notag
&-
\sum_{i=1}^n \frac{q_i}{\beta_i} \langle i |
\left(\begin{array}{cccc}
0 &&& 0\\
0 && &1
\end{array}\right) \frac{\partial}{\partial \langle i|}\langle kl \rangle =
-\frac{q_k}{\beta_k}\frac{\sqrt{2}}{\sqrt{\beta_k\beta_l}}
\left(\begin{array}{ccc}
q_k&& -\beta_k
\end{array}\right)
\left(\begin{array}{cccc}
0 &&& 0\\
0 && &1
\end{array}\right)
\left(
\begin{array}{ccc}
0 && 1\\
-1 && 0
\end{array}
\right) \left(\begin{array}{c}
q_l\\
-\beta_l
\end{array}\right)\\
& \qquad \qquad\qquad\qquad \qquad+
\frac{q_l}{\beta_l}\frac{\sqrt{2}}{\sqrt{\beta_k\beta_l}}
\left(\begin{array}{ccc}
q_l&& -\beta_l
\end{array}\right)
\left(\begin{array}{cccc}
0 &&& 0\\
0 && &1
\end{array}\right)
\left(
\begin{array}{ccc}
0 && 1\\
-1 && 0
\end{array}
\right) \left(\begin{array}{c}
q_k\\
-\beta_k
\end{array}\right)=0,\end{aligned}$$ we find $${\cal J}^{\{i\}} \chi ([ij],\langle ij\rangle) =0.$$ Analogously one can show that $\chi$ is annihilated by $\bar {\cal J}^{\{i\}}$, so we conclude that (\[3oct9\]), (\[3oct10\]) indeed solves (\[3oct08\]).
Finally, let us assume that there is another solution $\chi'$ of (\[3oct08\]), which cannot be written in the form (\[3oct9\]), (\[3oct10\]). Then, as it is not hard to see, $\chi'/\chi$ satisfies the scalar Ward identities (\[3oct2\]), (\[3oct3\]) and cannot be written in the form (\[3oct7\]), (\[3oct8\]). This contradicts the results of the previous Section, hence, $\chi'$ does not exist.
In the four-point case the light-cone Ward identity (\[3oct08\]) was solved previously in [@Metsaev:1991mt; @Metsaev:1991nb; @Bengtsson:2016hss]. Our solution extends these result to any number of external points. Moreover, unlike previous results, solution (\[3oct9\]), (\[3oct10\]) is not limited to the class of polynomials in transverse momenta. This extension is, in fact, important, as the primary meaning of this solution is to give all possible consistent total amplitudes, which are typically non-local due to contributions from exchanges.
#### Back to the Mandelstam variables.
Now we can easily resolve a loose end left from Section \[section5oct1\] and prove that the Mandelstam variables can be pulled through ${\cal J}$ to form exchanges. The algebraic part of ${\cal J}$ clearly commutes with $s_{i,j}$, so it remains to prove that $s_{i,j}$ is annihilated by the differential part of ${\cal J}$.
It is straightforward to compute $$(\sum_{\{i\}_j} \vec q_i)^2 = \sum_{k,l\in \{i\}_j} \vec q_k\cdot \vec q_l \approx \sum_{k,l\in \{i\}_j}
\big( q_k\bar q_l + \bar q_k q_l - \frac{q_k\bar q_k}{\beta_k}\beta_l- \beta_k\frac{q_l\bar q_l}{\beta_l} \big)
= \frac{1}{2} \sum_{k,l\in \{i\}_j}{[kl]\langle kl \rangle}.$$ Comparing this with (\[23sep3\])-(\[8oct2\]) we find $$\label{5oct2}
s_i = \frac{1}{2}\sum_{k,l\in \{i\}_j}{[kl]\langle kl \rangle}, \qquad s_{i,j} = \frac{1}{4}\sum_{k,l\in \{i\}_j}{[kl]\langle kl \rangle}+ \frac{1}{4}\sum_{k,l\in \{j\}_i}{[kl]\langle kl \rangle}.$$ Note that these formulas hold off-shell. They imply that $s_{i,j}$ can be presented in the form (\[3oct7\]), (\[3oct8\]) and consequently commutes with ${\cal J}$, as it was argued. The same applies to $\bar{\cal J}$.
Interpretation {#sect5oct3}
==============
In preceding Sections we first found that the light-cone consistency condition can be rewritten in the form of the Ward identity for the amplitude, constructed form the light-cone Hamiltonian. Then we showed that a general solution of the Ward identity can be conveniently presented in terms of spinor products. These spinor products is a basic building block of the spinor-helicity approach, which is effectively used for computations of amplitudes in theories of massless particles. In particular, the spinor-helicity approach is to large extent responsible for the existence of extremely compact representations of tree and loop partial amplitudes in QCD. In this Section, we would like to draw a link between the outcome of our light-cone analysis and the spinor-helicity approach aiming to interpret the results we found, use the ideas from the spinor-helicity approach to plot a strategy for construction of massless higher-spin interactions and to see whether the light-cone analysis has something new to offer compared to the spinor-helicity approach. We start by briefly reviewing the spinor-helicity approach. Our review is not meant to be self-contained. We refer the reader to [@Dixon:1996wi; @Bern:2007dw; @Elvang:2013cua] for general reviews and to [@Benincasa:2007xk; @Conde:2016izb] for discussions more focused on the higher-spin case.
In four space-time dimension, any null vector can be represented as $$\label{5oct3}
q_{a\dot b} \equiv q_\mu (\sigma^\mu)_{ a \dot b}
=\sqrt{2}\left(
\begin{array}{ccc}
q^- && \bar q\\
q & &-q^+
\end{array}\right)
\approx
\sqrt{2}
\left(\begin{array}{ccc}
-\frac{q\bar q}{\beta} && \bar q\\
q && -\beta
\end{array}\right)= - | q]_a \langle q|_{\dot b},$$ where $| q]_a$ and $\langle q|_{\dot b}$ are simply on-shell Weyl spinors with momentum $q$ $$\label{9oct1}
| q]_a =
\frac{2^{\frac{1}{4}}}{\sqrt{\beta}}\left(\begin{array}{c}
\bar q\\
-\beta
\end{array}\right),
\qquad
\langle q |_{\dot b} = \frac{2^{\frac{1}{4}}}{\sqrt{\beta}}
\left(\begin{array}{ccc}
q&& -\beta
\end{array}\right).$$ In the spinor-helicity approach representation (\[5oct3\]) is used to define momenta of massless particles of any spin. We will use the standard notation $|i] \equiv | q_i]$ and $\langle i | \equiv \langle q_i |$. More conventions are given in Appendix \[ap3oct1\].
To encode polarisations one uses auxiliary massless vectors called reference momenta. For example, the polarisation vector $\epsilon^\mu_1(q;k)$ of a helicity one boson of momentum $q$ is defined as $$\label{5oct4}
\epsilon_1^\mu(q;k)=-\frac{\langle k| \gamma^\mu | q]}{\sqrt{2}\langle k q\rangle},$$ where $k$ is the reference momentum. It is easy to see that the polarisation vector defined above is transversal to $q$, that is $q_\mu \epsilon_1^\mu(q;k)=0$, as required by the Ward identity. The arbitrariness in the choice of the reference momentum is just a manifestation of gauge redundancy. For each external field one is free to choose a reference momentum independently. However, this choice should be consistent for all diagrams relevant to the process. Then, as a consequence of gauge invariance, auxiliary reference vectors drop out from the final answer. In other words, a consistent amplitude should be a function of spinor products $[ij]$ and $\langle ij \rangle$ only.
It is clear from (\[5oct3\]) that $q_\mu$ is invariant with respect to the following scaling transformations $$\label{23oct3}
\langle q| \to t\langle q|,
\qquad | q] \to t^{-1} | q].$$ For $|\lambda|=1$ this generates the action of the Wigner little group on the helicity spinors. If we compute amplitudes using the Feynman rules, then we can see that the only way this scaling contributes to the amplitude is through polarisation vectors. For a helicity one boson the polarisation vector (\[5oct4\]) scales as $$\epsilon_1^\mu(q;k) \to t^{-2} \epsilon_1^\mu(q;k).$$ More generally, the Wigner little group acts on the helicity-$\lambda$ polarisation vector as $$\label{9oct3}
\epsilon_1^\mu(q;k) \to t^{-2\lambda} \epsilon_1^\mu(q;k).$$ Each external field can be subjected to (\[23oct3\]) independently. For an amplitude expressed in terms of spinor contractions to reproduce this scaling behaviour it should satisfy $$\label{9oct4}
( -N_{| i]} + N_{\langle i |} +2\lambda_i ) {\cal A}_n([jk],\langle jk\rangle ) = 0.$$
We find that despite a slightly different motivation, the spinor-helicity approach produces the same constraints that we found previously from the Poincare algebra closure in the light-cone deformation procedure. As we just reviewed, in the spinor-helicity approach the only two constraints are gauge invariance, which requires the amplitude to be expressed in terms of spinor products, and invariance with respect to the action of the Wigner little group, which fixes the homogeneity degrees of spinors. In the light-cone approach the only constraint is invariance with respect to the fully non-linear action of the Poincare algebra, which according to Proposition \[prop2oct1\] can be replaced by invariance of the amplitude with respect to the linear action of the algebra generators. Solving these constraints, we indeed find that the light-cone amplitude can be expressed in terms of spinor products only, which through the spinor-helicity approach relates (\[3oct08\]) to gauge invariance and, thus, justifies the term the Ward identity that we used[^8]. Moreover, (\[3oct08\]) fixes the homogeneity degrees of spinors, which from the spinor-helicity perspective is related to the Wigner little group invariance. This relation can be easily seen from the light-cone approach itself. Namely, by acting along the lines of Section \[sect9oct1\], one can show that $[{\cal A}_n,J_2^{x\bar x}]=0$ implies (\[3oct10\]).
So far we could see that on-shell amplitudes from the spinor-helicity approach appear to be very similar to light-cone amplitudes ${\cal A}_n(q^\perp)$, which we introduced above. However, it is important to remember that ${\cal A}_n(q^\perp)$ in the light-cone approach is well defined for off-shell momenta. In particular, ${\cal A}_n(q^\perp)$ can be used to define the light-cone Hamiltonian and then the action via the Legendre transform. This difference originates from the way one defines spinors $|i]$ and $\langle i|$ in the two approaches. In the spinor-helicity approach spinors $|i]$ and $\langle i|$ are only defined for null momenta, that is for $q^-\approx h_2(q^\perp)$, see (\[5oct3\]), (\[9oct1\]). In the light-cone approach one defines spinors by the very same formula (\[9oct1\]), but the momentum is not required to be massless and $q^-$ can be arbitrary. Clearly, factorisation formula (\[5oct3\]) does not work off-shell, but in the light-cone approach it is never used[^9]. It is quite remarkable that this rather trivial off-shell extension of spinor products and, hence, of the amplitude turns out to be consistent without any further constraints. This can be regarded as a simple consequence of $q^-$-independence of the Feynman rules presented in Proposition \[prop2oct1\].
An immediate benefit from the off-shell extension of helicity spinors is that they can be used for momenta on internal lines. So, in the light-cone approach exchanges can also be written in terms of spinor products. More details on how they are constructed can be found in Proposition \[prop2oct1\]. Generically, when internal momenta are expressed in terms of momenta on external lines, exchanges cannot be written in terms of spinor products any more. This implies that they do not satisfy the Ward identity. Consequently, individual contact interactions, in general, violate the Ward identity either.
This suggests to reconsider the light-cone deformation procedure and focus on seeking the amplitude instead of the Hamiltonian. The benefit of this strategy is concisely summarised by (\[9oct2\]): the amplitude satisfies a simple linear differential equation, while the equation for the Hamiltonian is quadratic in deformations. Having found a general solution for the amplitude we, thus, found a general solution of the light-cone deformation procedure. Note, however, that these solutions generically are associated with non-local Hamiltonians. The problem of finding consistent amplitudes that result in local Hamiltonians deserves a separate thorough analysis. Let us, nevertheless, make few comments on this point.
#### Locality.
Typically, in the light-cone approach one defines locality as a requirement that the Hamiltonian is polynomial in transverse momenta or, equivalently, in $\bar{\mathbb{P}}_{ij}$ and ${\mathbb{P}}_{kl}$. The same Hamiltonian can brought to many different forms using momentum conservation. Clearly, the Hamiltonian is local if there exist at least one of its forms where it is polynomial in transverse momenta. Due to the possibility to use momentum conservation locality may not be manifest. For example, the antiholomorphic part of the Yang-Mills cubic vertex can be written as $$\label{10dec1}
h_{3}= \frac{[12]^3}{[23][31]}=\sqrt{2}\frac{\bar{\mathbb{P}}^3_{12}\beta_3}{\bar{\mathbb{P}}_{23}\bar{\mathbb{P}}_{31}\beta_1\beta_2}.$$ This vertex is superficially non-local as it contains powers of transverse momenta in the denominator. However, it is easy to see that momentum conservation implies $$\label{10dec2}
\bar{\mathbb{P}}_{12}=\bar{\mathbb{P}}_{23}=\bar{\mathbb{P}}_{31},$$ hence, non-locality of (\[10dec1\]) is spurious. This example illustrates that to make locality manifest it may be required to break the spinor-helicity representation. At the cubic level this subtlety does not result in any difficulties, because modulo momentum conservation there is only one possible Lorentz-invariant variable that depends on the antiholomorphic momentum (\[10dec2\]) and, similarly, only one that depends on the holomorphic one. It would be interesting to clarify how this phenomenon extends to higher-point amplitudes.
Locality of the Hamiltonian is naturally translated into the language of amplitudes using the framework of on-shell methods [@Britto:2004ap; @Britto:2005fq]. Namely, it is required that amplitudes have no other singularities than those produced by exchanges. In many cases one can also justify that amplitudes vanish in certain directions at complex infinity. This allows to reconstruct them unambiguously from their singularities. In this respect we would like to clarify that the light-cone deformation procedure alone does not impose any constraints on the behaviour of amplitudes at infinity[^10]. Let us also note that the solution (\[3oct9\]), (\[3oct10\]) is not limited to polynomials in transverse momenta, so it is applicable to amplitudes, which are typically non-local due to contributions from exchanges.
It is worth to remark separately that in higher-spin theories imposing locality as described above, most likely, would rule out any interactions at all. It is then suggestive to replace locality in a strict sense by a milder requirement that coefficients of higher derivative terms decrease fast enough, so that the amplitudes associated with contact interactions do not contain singularities. This weaker version of locality was discussed at length in [@Bekaert:2015tva]. Let us also note that it is this weaker version of locality, that is effectively implemented by the on-shell methods, as they just require contact interactions to be free of singularities.
Finally, we remark that contrary to the way one usually defines locality for the light-cone deformation procedure, the on-shell methods require locality of interactions only on-shell. It would be interesting to clarify whether this difference can play any role.
Conclusion {#sec7oct2}
==========
This paper contains two main results. First, we show that the light-cone consistency conditions can be equivalently rephrased as a set of Ward identities for the light-cone off-shell amplitudes. Then we give a general solution to these Ward identities. This solution acquires an extremely simple form when written using the spinor-helicity language. More precisely, the general solution is just any function of spinor products that satisfies a well-known constraint relating homogeneity degrees of spinors with helicities of external fields. These results are summarised in Proposition \[prop2oct1\] and Proposition \[prop5oct1\].
Our primary goal is to employ the light-cone analysis to construct interactions of massless higher-spin fields. In this respect, our results provide a general solution to this problem in the case when locality of interactions is not required. Of course, this way of solving the consistent interaction problem is to large extent trivial, see, e.g., [@Barnich:1993vg]. Nevertheless, this gives a good starting point to address the problem of local interactions.
To construct local interactions, it is suggestive to proceed in the spirit of the on-shell methods [@Britto:2004ap; @Britto:2005fq], that is by reconstructing amplitudes from singularities associated with exchanges. In other words, it seems more reasonable to seek not the individual vertices, but the total amplitude. Indeed, the total amplitude satisfies the Ward identities, which have been solved in the present paper in complete generality. At the same time, constraints on individual contact interactions are much more complicated and from particular examples we know that individual vertices can be quite cumbersome [@Ponomarev:2016lrm]. More generally, amplitudes, being physically observable quantities, are much more constrained than individual vertices which are, moreover, prone to ambiguities of a gauge choice, field redefinitions, different sets of auxiliary fields etc. One should not expect a simple form of individual vertices unless these ambiguities are fixed wisely. On the contrary, as a result of strong constraints put on them, amplitudes admit a concise spinor-helicity representation.
Having written the light-cone consistency condition in the spinor-helicity form, we are able to clarify that the light-cone analysis does not impose any constraints on the behaviour of the amplitude at infinity. Hence, the no-go conclusions based on BCFW [@Benincasa:2007xk; @Fotopoulos:2010ay; @Benincasa:2011pg; @McGady:2013sga; @Ponomarev:2016jqk; @Bengtsson:2016alt], in principle, can be circumvented, see also [@Bengtsson:2016hss]. Moreover, the explicit analysis of the quartic self-interaction sector [@Ponomarev:2016lrm] shows that the relevant consistent interaction does exist.
It is worth to emphasise that the light-cone approach leads to a natural off-shell continuation of the spinor-helicity representation, which is usually defined on-shell. Once continued off-shell the light-cone amplitudes unambiguously define the action of the theory. It would be interesting to see whether the light-cone off-shell continuation can be used to promote on-shell results, such as colour-kinematics duality [@Bern:2008qj; @Bern:2010ue], to the off-shell level. Other interesting directions include extensions of the spinor-helicity approach to massive particles[^11], as well as to AdS.
Acknowledgements {#acknowledgements .unnumbered}
================
I am grateful to E. Skvortsov for many stimulating discussions. I would also like to thank R. Metsaev and A. Tseytlin for useful comments on the draft. I am grateful to A. Ochirov for explanations on the spinor-helicity approach and to W. Siegel for helpful correspondence. I acknowledge a kind hospitality at the program “Higher Spin Theory and Duality" MIAPP, Munich (May 2-27, 2016) organized by the Munich Institute for Astro- and Particle Physics (MIAPP). This work was supported by the ERC Advanced grant No.290456.
Notations {#ap3oct1}
=========
Here we collected various notations used throughout the paper.
#### Light-cone coordinates.
We work with the 4d Minkowski space endowed with the mostly plus metric $$ds^2 = -(dx^0)^2+ (dx^1)^2+(dx^2)^2+(dx^3)^2.$$ In the light-cone coordinates $$\begin{aligned}
\notag
x^+ &\,= \frac{1}{\sqrt{2}}(x^3+x^0), & x^-&\, = \frac{1}{\sqrt{2}}(x^3-x^0),\\
\label{29sep1}
x &\,=\frac{1}{\sqrt{2}}(x^1-ix^2), & \bar x &\,= \frac{1}{\sqrt{2}}(x^1+ix^2),\end{aligned}$$ it becomes $$ds^2 = 2dx^+ dx^- + 2 dx d\bar x.$$ Accordingly, we denote $$\begin{aligned}
\notag
\partial^- &\,= \frac{1}{\sqrt{2}}(\partial^3-\partial^0), & \partial^+ &\, = \frac{1}{\sqrt{2}}(\partial^3 + \partial^0),\\
\bar\partial &\,= \frac{1}{\sqrt{2}}(\partial^1-i\partial^2), & \partial &\, =\frac{1}{\sqrt{2}}(\partial^1 + i \partial^2),\end{aligned}$$ which implies $$\partial^+ x^- = \partial^- x^+ = \bar\partial x = \partial\bar x = 1.$$ In the light-cone approach $x^+$ is taken to be the time variable and $\partial^-$ is the time derivative. Moreover, one assumes that $\partial^+$ is non-zero and can always be inverted.
#### Spinor-helicity.
For spinor-helicity conventions we follow [@Elvang:2013cua]. We choose the Pauli matrices as $$\sigma^0 =
\left(\begin{array}{cccc}
1 &&& 0\\
0 && &1
\end{array}\right), \quad \sigma^1 =
\left(\begin{array}{cccc}
0 & && 1\\
1& && 0
\end{array}\right),
\quad
\sigma^2 =
\left(\begin{array}{ccc}
0 && -i\\
i&& 0
\end{array}\right), \quad
\sigma^3 =
\left(\begin{array}{ccc}
1 && 0\\
0&& -1
\end{array}\right).$$ Then $$\label{27sep10}
q_{a\dot b} \equiv q_\mu (\sigma^\mu)_{ a \dot b}
=\sqrt{2}\left(
\begin{array}{ccc}
q^- && \bar q\\
q & &-q^+
\end{array}\right)
\approx
\sqrt{2}
\left(\begin{array}{ccc}
-\frac{q\bar q}{\beta} && \bar q\\
q && -\beta
\end{array}\right)= - | q]_a \langle q|_{\dot b},$$ where $$\label{27sep12}
| q]_a =
\frac{2^{\frac{1}{4}}}{\sqrt{\beta}}\left(\begin{array}{c}
\bar q\\
-\beta
\end{array}\right),
\qquad
\langle q |_{\dot b} = \frac{2^{\frac{1}{4}}}{\sqrt{\beta}}
\left(\begin{array}{ccc}
q&& -\beta
\end{array}\right).$$ In these terms $$\label{27sep13}
[pq] = [ p |^a |q]_a = \varepsilon^{ab} |q]_a |p]_b, \qquad \langle p q\rangle = \langle p|_{\dot a} |q\rangle^{\dot a} = \varepsilon^{\dot a\dot b} \langle p|_{\dot a}\langle q|_{\dot b},$$ where $$\varepsilon^{ab} = \varepsilon^{\dot a\dot b} =
\left(
\begin{array}{ccc}
0 && 1\\
-1 && 0
\end{array}
\right) = - \varepsilon_{ab}= -\varepsilon_{\dot a\dot b}.$$ Rewriting spinor contractions as matrix products we find $$\begin{aligned}
[ij]&\, =
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_j}}
\left(\begin{array}{ccc}
\bar q_j&& -\beta_j
\end{array}\right)
\left(
\begin{array}{ccc}
0 && 1\\
-1 && 0
\end{array}
\right)
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_i}}\left(\begin{array}{c}
\bar q_i\\
-\beta_i
\end{array}\right)=\frac{\sqrt{2}}{\sqrt{\beta_i\beta_j}}\bar{\mathbb{P}}_{ij},
\\
\langle ij\rangle &\,=
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_i}}
\left(\begin{array}{ccc}
q_i&& -\beta_i
\end{array}\right)
\left(
\begin{array}{ccc}
0 && 1\\
-1 && 0
\end{array}
\right)
\frac{2^{\frac{1}{4}}}{\sqrt{\beta_j}}\left(\begin{array}{c}
q_j\\
-\beta_j
\end{array}\right)=-\frac{\sqrt{2}}{\sqrt{\beta_i\beta_j}}{\mathbb{P}}_{ij}.\end{aligned}$$
[^1]: [email protected]
[^2]: There are also other types of no-go arguments, e.g. [@Weinberg:1964ew; @Coleman:1967ad]. For a comprehensive review on no-go theorems for massless higher-spin interactions and how they can be circumvented, see [@Bekaert:2010hw].
[^3]: At least formally, the exotic vertices can be written in the Lorentz covariant form, but this requires non-localities [@Conde:2016izb; @Sleight:2016xqq].
[^4]: Note that within the manifestly covariant framework there exist some additional lower-derivative deformations of the gauge algebra and gauge transformations, that cannot be promoted to the action level, see e.g. [@Boulanger:2006gr; @Bekaert:2010hp]. It would be interesting to see whether they are related to the exotic vertices from the light-cone approach.
[^5]: Following the higher-spin literature we use ’spinor-helicity’ just to term a representation for the total amplitude when it is given as a function of spinor products. In contrast, originally it rather means a set of tools employed to evaluate QCD Feynman diagrams, which include: the spinor-helicity representation for the polarisation vectors, prescriptions for a convenient choice of reference spinors, factorisation of external momenta in terms of spinors, etc.
[^6]: In a slightly different context the relation between the light-cone gauge and the spinor-helicity approach was found in [@Chalmers:1998jb]. Namely, the spinor-helicity approach was identified as fixing the space-cone gauge, which is closely related to the light-cone one.
[^7]: In fact, one can even succeed by deforming only $P^-$, but then this deformation rather plays a role of the amplitude than of the Hamiltonian, as it will be discussed later, i.e. see Proposition \[prop2oct1\].
[^8]: Note that in the spinor-helicity terms the light-cone gauge (\[29sep4\]) can be viewed as a particular choice of the reference vector along $x^-$ direction.
[^9]: This off-shell continuation appeared in [@Bardeen:1995gk; @Cangemi:1996rx]. Later it was used in [@Cachazo:2004kj] to extend Yang-Mills MHV amplitudes off-shell, which were then treated as vertices in the action, see also [@Gorsky:2005sf; @Mansfield:2005yd].
[^10]: See [@Benincasa:2011kn; @Benincasa:2011pg; @McGady:2013sga] for extensions of the on-shell methods, which do not require constraints on amplitudes at infinite momenta.
[^11]: One natural way to do that is to represent massive momenta by pairs of massless ones [@Conde:2016vxs; @Conde:2016izb].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Vincent Humilière$^1$
title: 'Hofer’s distance on diameters and the Maslov index'
---
\[section\] \[theo\][Proposition]{} \[theo\][Lemma]{} \[theo\][Definition]{} \[theo\][Corollary]{} \[theo\][Conjecture]{} \[theo\][Question]{}
Introduction and result
=======================
This note is inspired by Khanevsky’s article [@khanevsky]. We will adopt his settings throughout the paper. Let us recall them briefly. Let $D\subset{\mathbb{R}}^2=\{(x,y)\,|\,x\in{\mathbb{R}},y\in{\mathbb{R}}\}$ be the open unit 2-disk, endowed with the symplectic structure $\omega=\frac1\pi dx\wedge dy$ (so that the total area of $D$ is 1). We call *standard diameter* and denote by $L_0$ the intersection of $D$ with the $x$-axis ${\mathbb{R}}\times\{0\}$. Given a time-dependent Hamiltonian $H$ on $D$, we denote by $X_H$ its Hamiltonian vector field defined by $\omega(X_H,\cdot)=dH$, and by $\phi_H^t$ the Hamiltonian isotopy generated by $H$. All Hamiltonian functions will be assumed compactly supported. By definition, a *diameter* is a curve which is isotopic to the standard diameter via a compactly supported Hamiltonian isotopy. We denote by $\mathcal{E}$ the set of all diameters.
The set $\mathcal{E}$ can be endowed with the so-called *Hofer distance* which is defined as follows: $$d(L_1,L_2)=\inf\left\{\int_0^1\left(\max_{x\in D} H_t-\min_{x\in D} H_t\right)dt\right\},$$ where the infimum is taken over all smooth and compactly supported time dependent Hamiltonians $H_t$ such that $\phi_H^1(L_1)=L_2$. The distance $d$ may be defined on more general symplectic manifolds. Its non-degeneracy is a deep result (see [@chekanov]). The value of $\int_0^1(\max {H_t}-\min {H_t})dt$ is called the *Hofer energy* of the isotopy generated by $H_t$.
An interesting fact pointed out in Khanevsky’s paper is that Hofer’s distance descends to a metric on the *reduced diameter space* which is defined as $\hat{\mathcal{E}}= \mathcal{E}/\sim$, where $L\sim L'$ iff $L=\phi(L')$ with $\phi\in\mathcal{S}$, $\mathcal{S}$ being the group of Hamiltonian diffeomorphisms which globally fix the standard diameter $L_0$. Several other invariants of $\mathcal{E}$ descend to $\hat{\mathcal{E}}$. Khanevsky considers two of them: the number of intersection points between diameters and (much more involved) an invariant $r_A$ constructed from the Entov-Polterovich quasimorphism and the Calabi morphism. Roughly speaking, he proves that $r_A$ is dominated by $d$ (this implies that $\hat{\mathcal{E}}$ is unbounded) and that $d$ is dominated by the number of intersection points. In the present note we are interested in a third invariant which is the Maslov index.
The Maslov index of a transverse intersection point of two given Lagrangians is an integer which is invariant under Hamiltonian transformations. Unless we choose some convention, this index is defined up to an additive constant but the difference of index between two points (the “index gap”) is well defined. It can be constructed in a natural way by an abstract construction [@viterbo]. However, in our case, it can be constructed in a naive way as follows.
First, applying a Hamiltonian diffeomorphism, we may suppose that one of our two diameters is the standard one. Then, consider a diameter $L$ intersecting $L_0$ transversely. Note that in our context, every diameter coincides with the standard diameter in the complement of a compact set. Therefore, we say that $L$ and $L_0$ are transverse if we can write $L\cap L_0=(-1,-1+{\varepsilon}]\cup\{x_i\}_{i\in \{1,\ldots,N\}}\cup[1-{\varepsilon}',1)$, where ${\varepsilon},{\varepsilon}'>0$ and the points $x_1,\ldots,x_N$ are transverse intersections (in the usual sense). We suppose that the points $x_1,\ldots,x_N$ are ordered by their position on $L$ (not $L_0$). We choose a convention for the index of $x_1$, namely we set $\mu(x_1)=0$. Then, we construct the Maslov index $\mu$ of the other intersection points inductively as follows.
We denote by $D^+$ the upper half-disk and $D^-$ the lower half-disk. For $i\in\{1,\cdots,N\}$, we set $\mu(x_{i+1})=\mu(x_i)+\delta$, where $$\delta=\left\lbrace
\begin{array}{cl}
1, & \text{if }L|_{[x_i,x_{i+1}]}\subset D^-\text{ and }x_i < x_{i+1}\text{ on } L_0\\
-1, & \text{if }L|_{[x_i,x_{i+1}]}\subset D^-\text{ and }x_i > x_{i+1}\text{ on } L_0\\
-1, & \text{if }L|_{[x_i,x_{i+1}]}\subset D^+\text{ and }x_i < x_{i+1}\text{ on } L_0\\
1, & \text{if }L|_{[x_i,x_{i+1}]}\subset D^+\text{ and }x_i > x_{i+1}\text{ on } L_0
\end{array}
\right.$$
\[remarque indice\]
1. Intuitively, $\mu(x_i)$ measures how much $L$ “twists” in the positive direction before reaching $x_i$.
2. The index $\mu$ descends to the reduced diameter space $\hat{\mathcal{E}}$.
3. Note that by definition, the index gap between two intersection points that are consecutive on $L$ is $1$ or $-1$. The abstract definition of the Maslov index implies that we can invert the roles of $L$ and $L_0$ and therefore that the index gap between two intersection points that are consecutive on $L_0$ is also either $1$ or $-1$.
To avoid any confusion, we will sometimes denote $\mu(x,L)$ instead of $\mu(x)$. We then set $$\mu_{\max}(L)=\max_{i=1,\ldots,N}\mu(x_i,L)\quad\text{ and }\quad\mu_{\min}(L)=\min_{i=1,\ldots,N}\mu(x_i,L).$$ The maximal index gap $\mu_{\max}(L)-\mu_{\min}(L)$ descends to the reduced diameter space $\hat{\mathcal{E}}$. It is therefore natural to try to compare it with the other invariants considered by Khanevsky. The following inequality is a consequence of Remark \[remarque indice\].4. $$2(\mu_{\max}(L)-\mu_{\min}(L)){\leqslant}\sharp(L\cap L_0)+1.$$ Our main result is then the following theorem.
\[theoreme hofer-maslov\] For any diameter $L$ transverse to $L_0$, with at least two transverse intersection points with $L_0$, $$d(L,L_0){\leqslant}{\mu_{\max}}(L)-{\mu_{\min}}(L)-\frac12.$$
The maximal index gap ${\mu_{\max}}(L)-{\mu_{\min}}(L)$ vanishes if and only if there is only one transverse intersection point. In that case, $d(L,L_0){\leqslant}\frac12$.
This implies $d(L,L_0){\leqslant}\frac12\cdot\sharp(L\cap L_0)$. Therefore our result implies a linear estimate as in [@khanevsky] but with $\frac12$ as multiplicative constant, while Khanevsky’s constant is $\frac18$. Nevertheless, in many cases, our upper bound is much better. Moreover, after I completed this note, Michael Khanevsky explained to me an argument based on his paper’s settings (in particular “diameter trees”) showing that it might be possible to improve the multiplicative constant 1 in Theorem \[theoreme hofer-maslov\] to the constant $\frac14$.
As far as I know, the analogous problem in higher dimension is completely open.
Let us give an idea of the proof of Theorem \[theoreme hofer-maslov\]. Suppose that ${\mu_{\min}}(L)<0$. We first concentrate on intersection points of minimal index. Given such a point, it is possible to construct a very simple Hamiltonian isotopy with several properties including the fact that it removes the intersection point (see Section 3). We then show that such simple transformations can be performed in an appropriate order so that all intersection points of minimal index are removed and so that the total energy needed for these transformation is less than 1. We get a new diameter $L'$ satisfying ${\mu_{\min}}(L'){\leqslant}{\mu_{\min}}(L)+1$ satisfying $d(L',L)=1$. We then proceed by induction until ${\mu_{\min}}=0$, and work similarly for points of positive index (see Section 4).
A consequence of our proof is that for any diameter $L$ with ${\mu_{\min}}(L)=0$, there exists a non-negative Hamiltonian function $H$ such that $d(\phi_H^1(L_0),L){\leqslant}1$.
Aknowlegments {#aknowlegments .unnumbered}
-------------
I am very grateful to Frédéric Le Roux for useful discussions and for listenning to me patiently. I also thank Leonid Polterovich for drawing my attention on Michael Khanevsky’s paper. I thank the three of them for their comments and suggestions on the first version of the paper. I am also grateful to Hélène Eynard for her careful reading. Finally I thank an anonymous referee for pointing out some mistakes and for his suggestions to improve the paper.
Two lemmas on the Maslov index
==============================
Let $L$ be a fixed diameter. We first recall the abstract definition of the Maslov index gap between two transverse intersection points, as introduced by Viterbo in [@viterbo]. Let $x,y$ be two transverse intersection points. Denote by $\gamma:[0,1]\to L$ a curve on $L$ with $\gamma(0)=x$ and $\gamma(1)=y$. The derivative $\dot{\gamma}$ determines a path $[\dot{\gamma}]$ in the projective space ${\mathbb{R}}P^1={\mathbb{R}}/{\mathbb{Z}}$. This path can be turned into a loop after concatenation with a path $\tau:[0,1]\to{\mathbb{R}}P^1$, such that $\tau(0)=[\dot{\gamma}](1)$, $\tau(1)=[\dot{\gamma}](0)$ and such that for any $s$, $\tau(s)$ is transverse to $L_0$. Let $\tilde{\gamma}:[0,1]\to{\mathbb{R}}$ be the lift of this loop satisfying $\tilde{\gamma}(0)=0$. The index gap $\mu(y,L)-\mu(x,L)$ is then the Maslov index of the loop, i.e, the integer $\tilde{\gamma}(1)$.
We will need the two following lemmas.
\[lemme indices preserves\] Let $x,y$ be two transverse intersection points between $L$ and $L_0$. Let $(h^t)_{t\in[0,1]}$ be a compactly supported Hamiltonian isotopy of the disk. Suppose that for any $t\in[0,1]$, the points $x$ and $y$ remain intersection points of $L_0$ and $h^t(L)$. Suppose moreover that the tangent line to $h^t(L)$ at $x$ (resp. $y$) remains constant for $t\in[0,1]$. Then the index gap remains constant: $$\forall t\in[0,1],\ \mu(y,h^t(L))-\mu(x,h^t(L))=\mu(x,L)-\mu(y,L).$$
Let $\gamma$ be a curve joining $x$ to $y$ on $L$, as above. Let $\alpha_t$ be a continuous family of curves joining $x$ to $h^t(x)$ on $h^t(L)$ and $\beta_t$ a continuous family of curves joining $h^t(y)$ to $y$ on $h^t(L)$. Then consider the family of curves $\gamma_t$ obtained as a smooth reparametrization of the concatenation of three paths: $$\gamma_t=\alpha_t\ast (h^t\circ\gamma)\ast\beta_t.$$ Our assumptions imply that the family of loops $[\dot{\gamma_t}]\ast\tau$, where $\tau$ is as above, is a homotopy with fixed end points in ${\mathbb{R}}P^1$. Thus, $\tilde{\gamma_t}(1)=\tilde{\gamma}(1)$ for any $t$, and the lemma follows.
As in the introduction, we write $L\cap L_0=(-1,-1+{\varepsilon}]\cup\{x_i\}_{i\in \{1,\ldots,N\}}\cup[1-{\varepsilon}',1)$, where ${\varepsilon},{\varepsilon}'>0$ and the points $x_1,\ldots,x_N$ are transverse intersections ordered by their position on $L$.
\[lemme x1xN\] We suppose that both arcs $L|_{[0,x_1]}$ and $L|_{[x_N,1]}$ are included in $D^+$. Then, $\mu(x_N)-\mu(x_1)=1$.
![Proof of Lemma \[lemme x1xN\][]{data-label="dessin x1xN"}](dessin_x1xN.eps){width="7cm"}
Let $\psi$ be a Hamiltonian diffeomorphism such that $L=\psi(L_0)$. Let $\gamma$ be a smooth path on $D$ which is the concatenation of two paths: a path $\alpha$ that parametrizes the arc $L_{[x_1,x_N]}$ and a path $\beta$ from $x_N$ to $x_1$ disjoint from $L_0$ and $L$ (see Figure \[dessin x1xN\]). The path $\gamma$ is a simple curve which is the boundary of an embeded disk. Therefore, the Maslov index of $[\dot{\gamma}]$ equals 2. Since the contribution of $\beta$ to the Maslov index is obviously 1, it follows that the contribution of $\alpha$ is also 1. Thus, $\mu(x_N)-\mu(x_1)=1$.
Removing points of minimal index
================================
Let $L$ be a diameter transverse to $L_0$. In this section we describe simple Hamiltonian transformations that allow to remove an intersection point with extremal index. We describe it only for points of minimal index. Points of maximal index can be treated in a similar way.
![Large and small regions of a minimal index point[]{data-label="dessin petites_grandes_regions"}](dessin_petites_grandes_regions.eps){width="6cm"}
We start with the following remark on the shape of $L$ near a point of minimal index. Let $x_0$ be a point with minimal Maslov index. Consider the two transverse intersection points of $L$ and $L_0$ which are consecutive to $x_0$ on $L$. Here, we make the assumption that those two points exist, i.e., that $x_0\neq x_1$ and $x_0\neq x_N$. The minimality of $\mu(x)$ imposes that one of them (call it $y$) is on the left of $x_0$ on $L_0$ and the other (call it $z$) is on its right. It also forces the arc $\alpha=L|_{[y,x_0]}$ to lie in $D^+$ and the arc $\beta=L|_{[x_0,z]}$ to lie in $D^-$ (see figure \[dessin petites\_grandes\_regions\]).
We assume that one of the two closed regions delimited respectively by $\alpha$ and $L_0$ on one side, $\beta$ and $L_0$ on the other side has a smaller area. This region will be called the *small region* of $x_0$ and denoted by $r(x_0)$. Its area will be called the *weight* of $x_0$ and denoted by $w(x_0)$. The largest region will be called the *large region* of $x_0$ and denoted $R(x_0)$. By *half-disk* we will mean any intersection of a smooth closed disk with one of the two standard half disks $D^+$ and $D^-$ (figure \[dessin petites\_grandes\_regions\]).
\[lemme technique\] Let ${\varepsilon}>0$, $U$ a neighbourhood of $r(x_0)\cup R(x_0)$, and $A$ any half disk included in the interior of $R(x_0)$, of area less than $w(x_0)$ and such that any connected component of $A\cap L$ is an arc with at least one extremity on $L_0$. Then there exists a Hamiltonian diffeomorphism $\phi$ supported in $U$ with $\|\phi\|{\leqslant}w(x_0)+{\varepsilon}$, such that
- $\phi$ removes at least the intersection points $x_0$ and $z$ and does not create new intersection points, i.e., $\phi(L)\cap L_0\subset L\cap L_0-\{x_0,z\}$,
- $\phi$ does not change the index gap between any two remaining intersection points,
- $\phi$ maps the small region $r(x_0)$ into the opposite half disk. In particular it removes all intersection points contained in $r(x_0)$.
- $\phi$ maps $A$ into the opposite half disk.
![Illustration of Lemma \[lemme technique\] (the dotted arrows indicate roughly how $\phi$ acts)[]{data-label="dessin lemme technique"}](dessin_lemme_technique.eps){width="13cm"}
We will use several times the following fact and therefore state it in a separate lemma (see Figure \[dessin hamiltonien simple\]).
\[lemme hamiltonien precis\] Let $R=[a,b]\times[c,d]$, $R'=[a',b']\times[c',d']$ be two disjoint closed rectangles in ${\mathbb{R}}^2$ with same area $\alpha$. We suppose moreover that $a'<b'<a<b$. Let $\gamma$ (resp. $\gamma'$) be a curve that does not meet $R\cup R'$ and joins a point of $[a,b]\times \{c\}$ (resp. $[a,b]\times \{d\}$) to a point of $[a',b']\times \{c'\}$ (resp. $[a',b']\times \{d'\}$). We also suppose that both curves are disjoint. Let ${\varepsilon}>0$ and let $V$ be an open neighbourhood of $R\cup R'\cup\gamma\cup\gamma'$. Then there exists a smooth autonomous Hamiltonian function $H$ satisfying
- $H=0$ on the unbounded component of ${\mathbb{R}}^2-V$ and $H=\alpha+2{\varepsilon}$ on the bounded component of ${\mathbb{R}}^2-V$.
- $\min H =0$ and $\max H=\alpha+2{\varepsilon}$
- On each rectangle $R$ and $R'$, $H$ is an affine function of $x$. More precisely, $H(x,y)=(c-d+{\varepsilon})(x-b)+{\varepsilon}$ on $R$ and $H(x,y)=(d'-c'+{\varepsilon})(x-a')+{\varepsilon}$ on $R'$.
- $\phi_H^1(R)=R'$ and $\phi_H^1(R')=R$.
Note that the second condition on $H$ implies that the Hofer energy of the time one map $\phi_H^1$ is less or equal to $\alpha+2{\varepsilon}$. Note also that the third condition means that the flow lines restricted to both rectangles are vertical straight lines.
![Illustration of Lemma \[lemme hamiltonien precis\].[]{data-label="dessin hamiltonien simple"}](dessin_hamiltonien_simple.eps){width="6cm"}
![Image of Figure \[dessin hamiltonien simple\] by some area preserving diffeomorphism.[]{data-label="dessin anneau"}](dessin_anneau.eps){width="6cm"}
Under some area preserving map, the situation is conjugated to that where $V$ is a standard annulus of the form $\{(x,y)\in{\mathbb{R}}^2\,|\,r_{\min}^2<x^2+y^2<r_{\max}^2\}$. The transformation can also be chosen so that the rectangles become angular sectors of a sub-annulus. In polar coordinates $(\rho,\theta)$, we may suppose that $R$ is now $$\{(\rho,\theta)\in(0,+\infty)\times(-\pi,\pi]\,|\,\rho_{\min}{\leqslant}\rho {\leqslant}\rho_{\max}\text{ and } \eta{\leqslant}\theta{\leqslant}\pi-\eta\}$$ and that $R'$ is $$\{(\rho,\theta)\in(0,+\infty)\times(-\pi,\pi]\,|\,\rho_{\min}{\leqslant}\rho {\leqslant}\rho_{\max}\text{ and } -\pi+\eta{\leqslant}\theta{\leqslant}-\eta\},$$ for some real numbers $\rho_{\min}$, $\rho_{\max}$ and $\eta$ such that $r_{\min}<\rho_{\min}<\rho_{\max}<r_{\max}$, and $0<\eta$. The straight vertical lines in the rectangles become pieces of circles centered at the origin. See figure \[dessin anneau\].
Therefore, we only need to find $H$ in this situation, which is easy: take $H$ in the form $H(\rho,\theta)=h(\rho)$, where $h$ is some non-increasing function which satisfies the following properties.
- if $\rho{\leqslant}r_{\min}$ then $h(\rho)=\alpha+2{\varepsilon}$,
- if $\rho_{\min}{\geqslant}\rho{\geqslant}\rho_{\max}$ then $h(\rho)=\frac12(\rho_{\max}^2-\rho^2)+{\varepsilon}$,
- if $\rho{\geqslant}r_{\max}$ then $h(\rho)=0$.
The associated flow acts as a linear rotation on the sub-annulus delimited by $\rho_{\min}$ and $\rho_{\max}$. The time one map makes half a turn and hence sends $R$ onto $R'$ and $R'$ onto $R$ as wished.
First, recall that at any moment of the proof we can apply a small Hamiltonian perturbation and that all objects may thus be supposed to be in generic position. Let ${\varepsilon}>0$ sufficiently small and $U$ an open neighbourhood of $r(x_0)\cup R(x_0)$. The first step of the proof is to note that after conjugation by some area preserving map of the plane, we may suppose that $D^- $ and $D^+$ are rectangles, that $r(x_0)$ is included in a rectangle $R$ af area less than $w(x_0)+{\varepsilon}$ and that this rectangle $R$ is included in $U\cap D^-$. We may also suppose that $A$ lies inside a rectangle $R'\subset D^+$ included in $R(x_0)$ and with the same area as $R$, with $R\cap R'=\emptyset$. The arcs which are the connected components of $L\cap R'$ either have both extremities on $L_0$ or may be assumed to be pieces of straight vertical lines with one extremity on $L_0$.
In this situation, let $\gamma'$ be a smooth curve included in $D^+\cap U$ joining (as in Lemma \[lemme hamiltonien precis\]) the top sides of $R$ and $R'$, and $\gamma$ be a smooth curve in $D^-\cap U$ joining the bottom sides of $R$ and $R'$. Let $V$ be a neighborhood of $R\cup R'\cup\gamma\cup\gamma'$ included in $U$ and sufficiently narrow to avoid meeting $L_0$ except in the neighbourhood of $R'$ and $R$. Now, we may apply Lemma \[lemme hamiltonien precis\] which provides us with a Hamiltonian function $H$. Then set $\phi=\phi_H^{\tau}$ where $\tau<1$ will be appropriately chosen in the sequel.
Since $\phi_H^1$ sends $R$ onto $R'$, $r(x_0)$ is sent entirely into $D^+$ by $\phi$ provided $1>\tau>\inf\{t\,|\,\phi_H^t(r(x_0))\subset D^+\}$. In particular the intersection points $x_0$ and $z$ disappear. In the same way $A$ is sent into $D^-$. We have to make sure that no new intersection point appears. These new points would necessarily come from the arcs of $L$ meeting $V$. If $V$ is sufficiently narrow, such an arc must meet $R$, $R'$ or $\gamma\cup\gamma'$. First remark that such an arc cannot meet $R$ except if it is included in $r(x_0)$ ($R$ is supposed sufficiently narrow). But then, the whole arc is sent to $D^+$ and no new intersection point appears. Then, note that the assumption made on the intersection $R'\cap L$ implies that along the isotopy, the intersection points either disappear or do not move at all and no new point appears. Finally, the arcs which correspond to intersections between $L$ and $\gamma\cup\gamma'$ do not create new intersection points either provided $\tau$ is sufficiently close to $\inf\{t\,|\,\phi_H^t(r(x_0))\subset D^+\}$.
The intersection points that remain after applying $\phi$ are either those not contained in the support of $\phi$ or some of those included in $R'$. In both cases these points satisfy the assumptions of Lemma \[lemme indices preserves\]. Thus, for any pair of remaining points $u$, $v$, Lemma \[lemme indices preserves\] gives: $\mu(v,\phi(L))-\mu(u,\phi(L))=\mu(v,L)-\mu(u,L)$.
Finally, according to Lemma \[lemme hamiltonien precis\], our diffeomorphism $\phi$ has energy less than approximately the area of the rectangle $R$. But this area is also approximately the area of $r(x_0)$, that is $w(x_0)$.
![Proof of Lemma \[lemme technique\].[]{data-label="dessin preuve lemme technique"}](dessin_preuve_lemme_technique.eps){width="9cm"}
\[remark partial removing\] Note that if we take the isotopy $\phi_H^t$ of the proof of Lemma \[lemme technique\] at some smaller time, we can construct for any half-disk $A\subset R(x_0)$ of area less than $w(x_0)$ a diffeomorphism similar to that in Lemma \[lemme technique\] that does not remove the intersection point $x_0$, but decreases $r(x_0)$ by approximately the area of $A$. The energy needed is approximately the area of $A$. We will call this procedure a *partial removing* of $x_0$ associated to $A$.
\[remark region du bas\] We consider again an isotopy like the one appearing in the proof of Lemma \[lemme technique\]. Let $B$ be a half disk included in $D^-$ and containing $r(x_0)$. It will be useful to figure out how the area of $B$ behaves along the isotopy $\phi_H^t$ in the following two cases.
First, if $A\cap L_0\subset B\cap L_0$, then it is possible to choose the support of the isotopy so that it does not meet $\partial B\backslash L_0$. In this case, the area of $B$ remains constant along the isotopy (see Figure \[dessin remarques\] (a)).
Second, if $A$ is entirely on the left of $B$ (i.e., $\forall a\in A\cap L_0,\forall b\in B\cap L_0, a{\leqslant}b$). Then, for any other half disk $A'$ containing $A$ and on the left of $B$, we can choose the path $\gamma$ (resp. $\gamma'$) so that once it has entered $B$ (resp. $A'$) it does not exit $B$ (resp. $A'$) anymore (i.e., $\gamma\cap B$ is connected). This choice made, we see that along our (area preserving) isotopy, the respective areas of $A$, $A'$ and $B$ decrease at the same speed exactly (see Figure \[dessin remarques\] (b)).
![Illustration of Remark \[remark region du bas\][]{data-label="dessin remarques"}](dessin_remarques.eps){width="10cm"}
With the help of these remarks, we can prove the following refinement of Lemma \[lemme technique\], illustrated by Figure \[dessin lemme technique 2\].
![Settings of Lemma \[lemme technique 2\][]{data-label="dessin lemme technique 2"}](dessin_lemme_technique2.eps){width="9cm"}
\[lemme technique 2\] Let ${\varepsilon}>0$ and $U$ be an open neighbourhood of $r(x_0)\cup R(x_0)$. Let $A_1,\ldots, A_p=A'_1\ldots A'_n, B_1,\ldots, B_n$ be closed half disks such that
- the $A_i$’s are included in $R(x_0)$, are pairwise disjoint and are ordered by their position on $L_0$ from the right to the left,
- $\sum_{i=1}^{p-1}\textrm{area}(A_i)<w(x_0)$,
- $A_p=A'_1\subset \cdots\subset A'_n\subset R(x_0)\subset D_+$,
- $r(x_0)\subset B_n\subset\cdots\subset B_1=B\subset D_-$,
- for any index $i$, $A_i'$ is on the left of $B_i$ and intersects at one point exactly,
- for any indices $i,j$, either $A_i\subset A_j'$ or $A_i\cap A_j'=\emptyset$.
Then, there exists a Hamiltonian diffeomorphism $\phi$ supported in $U$ with $\|\phi\|{\leqslant}w(x_0)+{\varepsilon}$, such that
- $\phi(L)\cap L_0\subset L\cap L_0-\{x_0,z\}$,
- $\phi$ does not change the index gap between two remaining intersection points,
- $\phi$ maps the small region $r(x_0)$ into the opposite half disk,
- $\phi$ maps each $A_i$, $1{\leqslant}i{\leqslant}p-1$ into the opposite half disk,
- for any index $j\in\{1,\ldots, n\}$, the map $\phi$ decreases the areas of $A'_j$ and $B_j$ by exactly the same amount, except in the special case $j=1$ and $w_0>\sum_{i=1}^{p}\textrm{area}(A_i)$, for which $A_1'=A_p$ is entirely mapped to the opposite half disk.
We first apply successively partial removings of $x_0$ associated to each $A_i$ with $1{\leqslant}i{\leqslant}p-1$ (see Remark \[remark partial removing\]). We get diffeomorphisms $\phi_1$,...,$\phi_{p-1}$. We can choose them so that the support of each $\phi_i$ meets neither the $A_j$’s for $j>i$, nor the $A'_j$’s that do not contain $A_i$. Moreover, according to Remark \[remark region du bas\], this can be done so that for any index $j$, if $A'_j$ contains $A_i$, then the areas of $A'_j$ and $B_j$ are reduced by the same amount exactly, and if $A'_j$ does not contain $A_i$, then $A'_j$ and $B_j$ remain globally unchanged.
After the action of the composition $\phi_{p-1}\circ\cdots\circ\phi_{1}$, the half-disks $A_1,\cdots,A_{p-1}$ have been moved to $D_-$, the small region $r(x_0)$ has now area $c=w(x_0)-\sum_{i=1}^{p-1}\textrm{area}(A_i)$ and moreover $A_p=A'_1$ and $B_1$ remain unchanged.
Then, two cases may occur. If $c>\textrm{area}(A_p)$, then we apply Lemma \[lemme technique\] for a half-disk $A$ containing $A_p$ and of area approximately $c$. If $c<\textrm{area}(A_p)$, then we apply Lemma \[lemme technique\] for a half-disk $A$ included in $A_p$ and of area approximately $c$. In the first case, this gives a diffeomorphism $\phi_p$ which maps $A_p$ to $D_-$. In the second case, this gives a diffeomorphism $\phi_p$ which, according to Remark \[remark region du bas\], can be chosen so that both areas of $A_p=A'_1$ and $B_1$ are decreased by $c$. Finally we see that the diffeomorphism $\phi=\phi_{p}\circ\cdots\circ\phi_{1}$ suits our needs.
Proof of the theorem
====================
We will prove Theorem \[theoreme hofer-maslov\] by induction on the maximal index gap ${\mu_{\max}}(L)-{\mu_{\min}}(L)$. Before to start, we apply an arbitrarily small perturbation on $L$ so that we may assume that $\mu(x_1)=\mu(x_N)=0$. This can be achieved by creating new transverse intersection points near the boundary to be in the situation of Lemma \[lemme x1xN\] and then add one more intersection point to get $\mu(x_1)=\mu(x_N)$.
First, suppose that ${\mu_{\max}}(L)-{\mu_{\min}}(L)=1$. In this case, the intersection points of $L$ and $L_0$ are in a very simple configuration: the indices are successively $0$, $\pm 1$, 0, $\pm1$... and there exists some area preserving map of the plane, that preserves the $x$-axis and sends $L$ to the graph of the differential of a compactly supported function $S:{\mathbb{R}}\to{\mathbb{R}}$. In this case, the Hofer distance is less than $\max S-\min S$. This value can be interpreted as the difference between the areas of some domains delimited by $L$ and $L_0$ respectively in $D_+$ and $D_-$. This leads to the estimate $d(L,L_0){\leqslant}\frac12$, which proves the theorem in the case ${\mu_{\max}}(L)-{\mu_{\min}}(L)=1$.
Suppose now that ${\mu_{\max}}(L)-{\mu_{\min}}(L)>1$. Then, at least one of the two integers ${\mu_{\min}}(L)$ and ${\mu_{\max}}(L)$ does not vanish. Suppose ${\mu_{\min}}(L)< 0$. We are going to construct a Hamiltonian diffeomorphism $h$, with energy $\|h\|{\leqslant}1$, such that ${\mu_{\min}}(h(L))={\mu_{\min}}(L)+1$. If ${\mu_{\max}}{(L)}>0$, an analogous construction could also be performed to reduce ${\mu_{\max}}{(L)}$ by 1. In any case, we are able to reduce the maximal index gap ${\mu_{\max}}-{\mu_{\min}}$ by 1. It follows by induction that there exists a Hamiltonian diffeomorphism $g$ that sends $L$ to a diameter $L'$ with ${\mu_{\max}}(L')-{\mu_{\min}}(L')=1$ and $\|g\|{\leqslant}{\mu_{\max}}(L)-{\mu_{\min}}(L)-1$. This will conclude the proof of Theorem \[theoreme hofer-maslov\] since $$d(L,L_0){\leqslant}d(L,L')+d(L',L_0){\leqslant}\|g\|+\frac12.$$
Let us now suppose ${\mu_{\min}}(L)<0$ and construct a Hamiltonian diffeomorphism $h$ with energy less than 1 and such that ${\mu_{\min}}(h(L))={\mu_{\min}}(L)+1$.
Let $M(L)$ be the (finite) set of transverse intersection points of $L$ with $L_0$ whose index is ${\mu_{\min}}(L)$. Since ${\mu_{\min}}(L)< 0$, the points $x_1$ and $x_N$ do not belong to $M(L)$. Let $M^-(L)$ (resp. $M^+(L)$) be the set of all $x\in M(L)$ whose small region is contained in the lower (resp. upper) half-disk (Figure \[dessin mplus mmoins\]). The points of $M(L)$ are ordered by their position on $L_0$. We will denote by $\preceq$ this order relation and by $\prec$ the associated strict order. Namely, $x\prec y$ means that $x$ is strictly on the left of $y$. In a similar way, a region delimited by $L$, $L_0$ and two intersection points $y,z$ (i.e., a Whitney disk) will be said to be *on the left* of an intersection point $x$ if $y$ and $z$ are on the left of $x$.
![Intersection points belonging to $M^-$ or $M^+$.[]{data-label="dessin mplus mmoins"}](dessin_m_plusmoins.eps){width="9cm"}
When we consider a minimum point in $(M^-(L), \preceq)$, we can refine Lemma \[lemme technique\] as follows.
\[lemme comb\] Let $x_0$ be the minimum in $(M^-(L),\preceq)$. Then there exists a Hamiltonian diffeomorphism $\phi$ meeting the conditions of Lemma \[lemme technique\] and satisfying moreover the following:
- $M(\phi(L))\varsubsetneq M(L)$ and $\forall q\in M^-(\phi(L)), q\succ x_0$.
- For any $q\in M(\phi(L))$, consider the region $\mathcal{R}$ of $q$ that is included in $D^-$ where $q$ is seen as a point of $M(L)$. If $\mathcal{R}$ is on the right of $r(x_0)$, then $\mathcal{R}$ does not meet the support of $\phi$.
- $M^+(\phi(L))\subset M^+(L)$.
Let us postpone the proof of this lemma to the end of the section and achieve the proof of Theorem \[theoreme hofer-maslov\].
Applying Lemma \[lemme comb\] to the minimum point $x_1$ in $(M^-(L),\preceq)$, we get a Hamiltonian diffeomorphism $\phi_1$ meeting several conditions. Then, we repeat inductively the operation on the successive minimum points $x_1, x_2, x_3,\ldots$, constructing diffeomorphisms $\phi_1,\phi_2,\phi_3,\ldots$ until $M^-$ is empty. The process stops at some point since according to property (i) in Lemma \[lemme comb\] the (finite) cardinal of $M(L)$ decreases at each step. We have constructed a finite family of Hamiltonian diffeomorphisms $\phi_1,\ldots,\phi_p$ such that if we write $\psi=\phi_p\circ\ldots\circ\phi_1$ then $M^-(\psi(L))=\emptyset$.
Moreover, the energy of each $\phi_i$ is approximately the area $w(x_i)$ of the corresponding small region $r(x_i)$. But from conditions (i) and (ii) in Lemma \[lemme comb\], the small regions of the points involved in the process are pairwise disjoint. Indeed, these conditions imply that each $r(x_i)$, which is a priori a region of $x_i$ seen as a point in $M(\phi_{i-1}\ldots\phi_1(L))$, is actually already a region of $x_i$ seen as a point in $M(L)$ (the region is invariant under each successive diffeomorphism). But, two such regions are either disjoint or nested. This last possibility cannot happen for two small regions involded in our process: the largest of those two small regions would be associated to the $\preceq$-smallest intersection point (Figure \[dessin inclusion petites regions\]), thus we would apply Lemma \[lemme comb\] first to this larger small region during the process and this would remove the other intersection point. As a consequence, the energy of $\psi$ is bounded above as follows: $$\|\psi\|{\leqslant}\sum_{i=1}^p\|\phi_i\|{\leqslant}\sum_{i=1}^p(w(x_i)+{\varepsilon}){\leqslant}\text{area}(D^-)+p{\varepsilon}=\frac12+p{\varepsilon}.$$ Since we can do it for any small enough ${\varepsilon}>0$, we get $\|\psi\|{\leqslant}\frac12$.
![If $r(x)\subset r(x')$, then $x'\preceq x$.[]{data-label="dessin inclusion petites regions"}](dessin_inclusion_petites_regions.eps){width="5cm"}
Now we apply the same process to the points of $M^+(L)$: we exchange the roles of $M^+(L)$ and $M^-(L)$, which means in particular that we have to start from the right. We get a Hamiltonian diffeomorphism $\psi'$ with Hofer energy less than $\frac12$ and such that $M^+(\psi'\circ\psi(L))=\emptyset$. Moreover, it follows from Lemma \[lemme comb\] (iii) that $M^-(\psi'\circ\psi(L))\subset M^-(\psi(L))=\emptyset$. Hence $M(\psi'\circ\psi(L))=\emptyset$.
As a conclusion, we can move with energy at most 1 any diameter with given ${\mu_{\min}}$ to a diameter with minimal index ${\mu_{\min}}+1$. As explained in the beginning of the section, this concludes the proof of Theorem \[theoreme hofer-maslov\]. $\quad\Box$
\(i) According to Lemma \[lemme technique\], $\phi$ removes some intersection points letting the remaining intersection points at the same place and with the same index. Thus, $M(\phi(L))\subset M(L)$ is obvious.
Let $q\in M^-(\phi(L))$. First, $q\neq x_0$ because the intersection $x_0$ is removed by $\phi$. If $q\in M^-(L)$ then $q\succ x_0$ by definition of $x_0$. Let us show that $\phi$ may be chosen so that $q\in M^+(L)$ implies $q\succ x_0$. The two conditions $q\in M^+(L)$ and $q\in M^-(\phi(L))$ mean that either the area of $r(q)$ is expanded by $\phi$ or that of $R(q)$ is reduced by $\phi$ (maybe both). We will study these two cases separately. But we must first remark that the vector field generating $\phi$ is oriented downward for points of $L_0$ located on the left of (a neighbourhood of) $x_0$ and is oriented upward for points of $L_0$ located on the right of $x_0$ (see Figures \[dessin preuve lemme technique\] and \[dessin hamiltonien simple\]).
Suppose now that $r(q)$ is expanded by $\phi$. Then, since it is included in $D^+$, it cannot be entirely on the left of $x_0$ because otherwise, $\phi$ would decrease it. Since $q$ is on the right of $r(q)$, it follows that $q$ is on the right of $x_0$.
Suppose now that $R(q)$ is reduced by $\phi$. Then a similar argument implies that $R(q)$ is not entirely on the left of $x_0$ and two cases could happen: either $R(q)$ is entirely on the right of $x_0$, or $x_0$ belongs $R(q)$. The first case implies $q\succ x_0$ as wished. In the second case, there are again two possiblities (Figure \[dessin dernieres possibilites\]): (a) $q\notin R(x_0)$, (b) $q\in R(x_0)$. We want to avoid these to cases. To avoid the case (a) we only have to choose the map $\phi$ to have a support sufficiently n so that $R(q)$ remains invariant. Avoiding the case (b) requires a bit more work and we will need to apply the refined Lemma \[lemme technique 2\] to appropriate regions $A_i$, $A'_j$ and $B_j$, $1{\leqslant}i{\leqslant}p$, $0{\leqslant}j{\leqslant}n$.
![Two possibilities for $x_0\in R(q)$: (a) $q\notin R(x_0)$, (b) $q\in R(x_0)$.[]{data-label="dessin dernieres possibilites"}](dessin_dernieres_possibilites.eps){width="10cm"}
Suppose that the set $C$ of points $q\in R(x_0)\cap M^+(L)$ such that $r(x_0)\subset R(q)$ is not empty. We want to choose $\phi$ so that $C$ does not meet $M^-(\phi(L))$. To do so, we classify the points in $C$ into a partition of four sets $C_1$, $C_2$, $C_3$ and $C_4$, constructed by the following algorithm.
We initialize our sets as $C_1=\emptyset$, $C_2=\emptyset$, $C_3=\emptyset$ and $C_4=C$. We also initialize some auxiliary variables: a set $E=C$, a real variable $\alpha=0$ and a point $q=\max(E)$. Then we iterate the following operations until $E=\emptyset$.
If $E\neq \emptyset$ and $\alpha+\omega(q)<w(x_0)$, then modify all the variables by the rules (the symbole $":="$ means “is replaced by”):
$C_1:= C_1\cup\{q\}$
$C_2:= C_2$
$C_3:=C_3\cup \{q'\in E\,|\,q'\in r(q), q'\neq q\}$
$C_4:=C_4\backslash \{q'\in E\,|\,q'\in r(q)\}$
$\alpha:=\alpha+w(q)$
$E:=E\backslash\{q'\in E\,|\,q'\in r(q)\}$.
If $E\neq \emptyset$ and $\alpha+\omega(q)>w(x_0)$, then modify the variables by the rules:
$C_1:= C_1$
$C_2:= C_2\cup\{q\}$
$C_3:=C_3$
$C_4:=C_4\backslash \{q\}$
$\alpha:=\alpha$
$E:=\{q'\in E\,|\,r(q')\subset r(q), q'\neq q\}$.
The result of the algorithm is illustrated by Figure \[dessin algorithm\].
![Small regions of points in $C$ are represented. The small regions are colored according to their belonging to $C_1$, $C_2$, $C_3$ or $C_4$ at the end of the algorithm.[]{data-label="dessin algorithm"}](dessin_algorithme.eps){width="10cm"}
Once it has been performed, we decide to call $A_1,A_2,...$ the small regions of the points of $C_1$, ordered from the right to the left. We denote by $A'_1, A'_2, \ldots$ the small regions of the points of $C_2$, ordered by inclusion from the smallest to the largest. We also denote by $B_1, B_2,\ldots$ the large regions associated to $A'_1, A'_2, \ldots$. This family of regions satisfies the settings of Lemma \[lemme technique 2\] which provides us with a diffeomorphism $\phi$ with good properties. Let us analyse how it acts on the set $C$. First, points in $C_1$ are removed. Then, points in $C_2$ remain but the area of their small region decreases by exactly the same amount as their large region. Points in $C_3$ have their small regions included in small regions of points in $C_1$ and thus are removed. Finally points in $C_4$ have their small and large regions disjoint from the support of $\phi$ (they are on the left of its support). In any cases, we see that no points of $C$ belong to $M_-(\phi(L))$, which is exactly what we wished for.
\(ii) This condition is obviously satisfied provided the support of $\phi$ is chosen sufficiently narrow.
\(iii) Let $x\in M^+(\phi(L))$. Suppose $r(x)\subset D^-$ where $x$ is seen as an intersection point of $L$ and $L_0$ (i.e. $x\in M^-(L)$). Then either the region $r(x)$ is increased by $\phi$ or $R(x)$ is decreased (it may be both). But this imposes $x\prec x_0$. Contradiction.
[10]{}
Yuri Chekanov. Invariant [F]{}insler metrics on the space of [L]{}agrangian embeddings. , 234:605–619, 2000.
Mickael Khanevsky. Hofer’s metric on the space of diameters. , 1(4):407–416, 2009.
Claude Viterbo. Intersections de sous-variétés Lagrangiennes, fonctionnelles d’action et indice des systèmes hamiltoniens. , 115:361–390, 1987.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the present paper we study the long wavelength and slow time scale behavior of a coasting beam in a resonator adopting a broad-band impedance model. Based on the renormalization group approach we derive a set of coupled evolution equations for the beam envelope distribution function and the resonator voltage amplitude. The equation for the resonator voltage amplitude is further transformed into a generalized Ginzburg-Landau equation.'
address: |
[*Fermi National Accelerator Laboratory* ]{}\
[*P. O. Box 500, Batavia, IL 60510, USA*]{}\
[*E. mail: [email protected]*]{}
author:
- ' Stephan I. Tzenov'
title: 'Fermilab-Pub-98/275 Formation of Patterns and Coherent Structures in Charged Particle Beams'
---
Introduction.
=============
So far nonlinear wave phenomena have received scant attention in the study of collective effects in charged particle beams. Considerable experimental and simulation data however exists suggesting that these phenomena should be included into the entire physical picture of beam propagation in accelerators and storage rings.
A vast literature in the field of plasma physics is dedicated to the study of nonlinear wave-particle processes due to space charge interparticle forces. In high energy particle accelerators, where space charge forces are negligibly small, of particular interest is the coherent state of the beam under the influence of wakefields, or in the frequency domain, machine impedance. This state is highly nonlinear and depends on the interaction of nonlinear waves, involving some weak dissipative mechanisms that balance beam fluctuations driven be the wakefields.
In a previous work [@tzenov], [@colestock] we studied nonlinear behavior of a coasting beam under the influence of a resonator impedance. Starting from the gas-dynamic equations for longitudinal motion and using a renormalization group (RG) approach [@goldenfeld], [@kunihiro] we found a set of coupled nonlinear equations for the beam density and resonator voltage. However, as is well-known, the hydrodynamic approximation is valid when the beam is close to a local equilibrium, which in a number of practically important cases may well be far from reality.
The present paper, providing a complete kinetic description of the processes involved, is aimed to overcome the above mentioned difficulties. In what follows we study the longitudinal dynamics of a coasting beam in a resonator adopting a broad-band impedance model. We are interested in describing slow motion of beam patterns (droplets) neglecting fast oscillations of beam density and voltage on the resonator at a frequency close to the resonant frequency. We employ the RG method to derive amplitude equations governing the dynamics of slow processes. In Section II we obtain the desired equations for the longitudinal envelope distribution function and the amplitude of the resonator voltage. In Section III we proceed to transform the equation for the voltage amplitude into a generalized Ginzburg-Landau equation by solving explicitly the Vlasov equation for the envelope distribution function. Finally in Section IV we draw some conclusions resulting from the work performed.
The Amplitude Equations.
========================
The starting point for the subsequent analysis is the system of equations:
$$\frac{\partial f}{\partial T}+v\frac{\partial f}{\partial \theta }+\lambda V%
\frac{\partial f}{\partial v}=0,$$
$$\frac{\partial ^2V}{\partial T^2}+2\gamma \frac{\partial V}{\partial T}%
+\omega ^2V=\frac{\partial I}{\partial T}, \label{kinetic}$$
$$I\left( \theta ;T\right) =\int dvvf\left( \theta ,v;T\right) ,$$
for the longitudinal distribution function $f\left( \theta
,v;T\right) $ of an unbunched beam and the variation per turn of the voltage $V\left( \theta ;T\right) $ on a resonator. All dependent and independent variables, as well as free parameters in equations (\[kinetic\]) are dimensionless and have been rescaled according to the relations:
$$T=\omega _st\quad ;\quad v=\frac 1{\omega _s}\frac{d\theta }{dt}=1+\frac{%
k_o\Delta E}{\omega _s}\quad ;\quad \omega =\frac{\omega _R}{\omega _s}\quad
;\quad \gamma =\frac \omega {2Q},$$
$$\lambda =\frac{e^2R\gamma k_o\rho _o}\pi .$$
Here $\omega _s$ is the angular revolution frequency of the synchronous particle, $\Delta E$ is the energy error, $\omega _R$ is the resonant frequency, $Q$ is the quality factor of the resonator, $R$ is the resonator shunt impedance and $\rho _o$ is the uniform beam density distribution at the thermodynamic limit. Furthermore
$$k_o=-\frac{\eta \omega _s}{\beta _s^2E_s}$$
is the proportionality constant between the frequency deviation and the energy deviation of a non synchronous particle with respect to the synchronous one, while $\eta =\alpha _M-\gamma _s^{-2}$ ($\alpha _M$ - momentum compaction factor) is the phase slip coefficient. The voltage variation per turn $V,$ the beam current $I$ and the longitudinal distribution function $f$ entering equations (\[kinetic\]) have been rescaled as well from their actual values $V_a$, $I_a$ and $f_a$ as follows:
$$V_a=2e\omega _s\rho _o\gamma RV\quad ;\quad I_a=e\omega _s\rho _oI\quad
;\quad f_a=\rho _of$$
Let us introduce the Radon transform [@horvath], [@klim] of the distribution function $f\left( \theta ,v;T\right) $
$$f\left( \theta ,v;T\right) =\int d\xi F\left( \theta ,\xi ;T\right) \delta
\left[ v-U\left( \theta ,\xi ;T\right) \right] . \label{radon}$$
In the definition (\[radon\]) $\xi $ can be viewed as a Lagrange variable which is usually determined from the condition that the distribution function $f\left( \theta ,v;T\right) $ be equal to a specified distribution, say the equilibrium distribution for instance:
$$f\left( \theta ,v;T\right) =f_0\left( \xi \right) \qquad \Rightarrow \qquad
v=U\left( \theta ,\xi ;T\right) .$$
Substitution of eq. (\[radon\]), into the system (\[kinetic\]) yields:
$$\frac{\partial F}{\partial T}+\varepsilon \frac \partial {\partial \theta
}\left( FU\right) =0,$$
$$\frac{\partial U}{\partial T}+\varepsilon U\frac{\partial U}{\partial \theta
}=\lambda V, \label{hydrodyn}$$
$$\frac{\partial ^2V}{\partial T^2}+2\gamma \frac{\partial V}{\partial T}%
+\omega ^2V=\int d\xi \frac \partial {\partial T}\left( FU\right) ,$$
where the fact that the azimuth $\theta $ is a slow variable (the dependence of $F,$ $U$ and $V$ on $\theta $ is through a stretched variable $%
\zeta =\varepsilon \theta $) has been taken into account. Note that the system (\[hydrodyn\]) resembles the set of gas-dynamic equations, governing the longitudinal motion of the beam. It bears however, additional information about the velocity distribution, embedded in the dependence on the Lagrange variable $\xi $, and takes into account its overall effect through the integral on the right hand side of the third equation.
We next examine the solution of the system of equations (\[hydrodyn\]) order by order in the formal small parameter $\varepsilon $ by carrying out a naive perturbation expansion. The zero order solution (stationary solution) is readily found to be:
$$F_0=F_0\left( \xi \right) \quad ;\quad U_0=U_0\left( \xi \right) \quad
;\quad V_0\equiv 0$$
and in particular one can choose
$$F_0\left( \xi \right) =f_0\left( \xi \right) \qquad ;\qquad U_0\left( \xi
\right) =1+\xi .$$
Combining the first order equations
$$\frac{\partial F_1}{\partial T}=0\qquad ;\qquad \frac{\partial U_1}{\partial
T}=\lambda V_1,$$
$$\frac{\partial ^2V_1}{\partial T^2}+2\gamma \frac{\partial V_1}{\partial T}%
+\omega ^2V_1=\int d\xi F_0\frac{\partial U_1}{\partial T},$$
yields trivially a unique equation for $V_1$:
$$\frac{\partial ^2V_1}{\partial T^2}+2\gamma \frac{\partial V_1}{\partial T}%
+\omega _o^2V_1=0\qquad ;\qquad \omega _o^2=\omega ^2-\lambda .$$
Solving the first order equations one easily obtains:
$$V_1\left( \theta ;T\right) =E\left( \theta ;T_0\right) e^{i\omega _1\Delta
T}+c.c.$$
$$U_1\left( \theta ,\xi ;T\right) =u_o\left( \theta ,\xi ;T_0\right) +\lambda
\frac{E\left( \theta ;T_0\right) }{i\omega _1}e^{i\omega _1\Delta T}+c.c.
\label{first}$$
$$F_1\left( \theta ,\xi ;T\right) =R_o\left( \theta ,\xi ;T_0\right) ,$$
where
$$\omega _1=\omega _q+i\gamma \quad ;\quad \omega _q^2=\omega _o^2-\gamma
^2\quad ;\quad \Delta T=T-T_0, \label{const}$$
and the amplitudes $E\left( \theta ;T_0\right) ,$ $u_o\left(
\theta ,\xi ;T_0\right) ,$ $R_o\left( \theta ,\xi ;T_0\right) $ are yet unknown functions of $\theta $, $\xi $ and the initial instant of time $T_0.$ Proceeding further with the expansion in the formal parameter $\varepsilon $ we write down the second order equations
$$\frac{\partial F_2}{\partial T}+F_0\frac{\partial U_1}{\partial \theta }+U_0%
\frac{\partial R_o}{\partial \theta }=0,$$
$$\frac{\partial U_2}{\partial T}+U_0\frac{\partial U_1}{\partial \theta }%
=\lambda V_2,$$
$$\frac{\partial ^2V_2}{\partial T^2}+2\gamma \frac{\partial V_2}{\partial T}%
+\omega ^2V_2=\int d\xi \left( F_0\frac{\partial U_2}{\partial T}+\lambda
R_oV_1+U_0\frac{\partial F_2}{\partial T}\right)$$
and by elimination of $U_2$ and $F_2$ from the third equation we obtain:
$$\frac{\partial ^2V_2}{\partial T^2}+2\gamma \frac{\partial V_2}{\partial T}%
+\omega _o^2V_2=\lambda V_1\int d\xi R_o-2\int d\xi U_0F_0\frac{\partial
U_1}{\partial \theta }-\int d\xi U_0^2\frac{\partial R_o}{\partial \theta }%
.$$
Solving the above equation and subsequently the two other equations for $U_2$ and $F_2$ we find the second order solution as follows:
$$V_2\left( \theta ;T\right) =-\frac 1{\omega _o^2}\int d\xi \left( U_0^2%
\frac{\partial R_o}{\partial \theta }+2F_0U_0\frac{\partial u_o}{\partial
\theta }\right) +$$
$$+\frac{\lambda \Delta T}{2i\omega _q}\left( E\int d\xi R_o+\frac{2i}{\omega
_1}\frac{\partial E}{\partial \theta }\int d\xi F_0U_0\right) e^{i\omega
_1\Delta T}+c.c.$$
$$U_2\left( \theta ,\xi ;T\right) =-\Delta TU_0\frac{\partial u_o}{\partial
\theta }-\frac{\lambda \Delta T}{\omega _o^2}\int d\xi \left( U_0^2\frac{%
\partial R_o}{\partial \theta }+2F_0U_0\frac{\partial u_o}{\partial \theta }%
\right) +$$
$$+\frac{\lambda U_0}{\omega _1^2}\frac{\partial E}{\partial \theta }%
e^{i\omega _1\Delta T}-\frac{\lambda ^2\Delta T}{2\omega _1\omega _q}\left(
E\int d\xi R_o+\frac{2i}{\omega _1}\frac{\partial E}{\partial \theta }\int
d\xi F_0U_0\right) e^{i\omega _1\Delta T}+ \label{second}$$
$$+\frac{\lambda ^2}{2i\omega _q\omega _1^2}\left( E\int d\xi R_o+\frac{2i}{%
\omega _1}\frac{\partial E}{\partial \theta }\int d\xi F_0U_0\right)
e^{i\omega _1\Delta T}+c.c.$$
$$F_2\left( \theta ,\xi ;T\right) =-\left( U_0\frac{\partial R_o}{\partial
\theta }+F_0\frac{\partial u_o}{\partial \theta }\right) \Delta T+\frac{%
\lambda F_0}{\omega _1^2}\frac{\partial E}{\partial \theta }e^{i\omega
_1\Delta T}+c.c.$$
In a way similar to the above we write the third order equations as
$$\frac{\partial F_3}{\partial T}+F_0\frac{\partial U_2}{\partial \theta }%
+\frac \partial {\partial \theta }\left( F_1U_1\right) +U_0\frac{\partial F_2%
}{\partial \theta }=0,$$
$$\frac{\partial U_3}{\partial T}+U_1\frac{\partial U_1}{\partial \theta }+U_0%
\frac{\partial U_2}{\partial \theta }=\lambda V_3,$$
$$\frac{\partial ^2V_3}{\partial T^2}+2\gamma \frac{\partial V_3}{\partial T}%
+\omega ^2V_3=\int d\xi \left[ F_0\frac{\partial U_3}{\partial T}+R_o\frac{%
\partial U_2}{\partial T}+\frac{\partial \left( F_2U_1\right) }{\partial T}%
+U_0\frac{\partial F_3}{\partial T}\right] .$$
Solving the equation for $V_3$
$$\frac{\partial ^2V_3}{\partial T^2}+2\gamma \frac{\partial V_3}{\partial T}%
+\omega _o^2V_3=$$
$$=\int d\xi \left[ -2F_0U_1\frac{\partial U_1}{\partial \theta }-2F_0U_0%
\frac{\partial U_2}{\partial \theta }-2U_0\frac{\partial \left(
R_oU_1\right) }{\partial \theta }\right] +$$
$$+\int d\xi \left( \lambda R_oV_2+\lambda V_1F_2-U_0^2\frac{\partial F_2}{%
\partial \theta }\right)$$
that can be obtained by combining the third order equations, and subsequently solving the two other equations for $U_3$ and $F_3$ we obtain the third order solution:
$$V_3\left( \theta ;T\right) =\frac{\lambda \Delta T}{2i\omega _q}\left\{
\frac{2i}{\omega _1}\frac \partial {\partial \theta }\left[ E\int d\xi
\left( u_oF_0+U_0R_o\right) \right] -\right.$$
$$-\frac 3{\omega _1^2}\left( \int d\xi F_0U_0^2\right) \frac{\partial ^2E}{%
\partial \theta ^2}+\frac 1{2i\omega _q}\left( 1-i\omega _q\Delta T\right)
\left[ \int d\xi \left( U_0\frac{\partial R_o}{\partial \theta }+F_0\frac{%
\partial u_o}{\partial \theta }\right) \right] E+$$
$$\left. +\frac \lambda {4\omega _q^2}\left( 1-i\omega _q\Delta T\right)
\left( \int d\xi R_o\right) \left[ E\int d\xi R_o+\frac{2i}{\omega _1}%
\left( \int d\xi F_0U_0\right) \frac{\partial E}{\partial \theta }\right]
\right\} e^{i\omega _1\Delta T}+$$
$$+c.c.+\mathrm{oscillating\ terms} \label{third}$$
$$U_3\left( \theta ,\xi ;T\right) =-u_o\frac{\partial u_o}{\partial \theta }%
\Delta T+\frac{\lambda ^2}{2\gamma \omega _o^2}\frac{\partial \left|
E\right| ^2}{\partial \theta }e^{-2\gamma \Delta T}+\mathrm{oscillating\
terms}$$
$$F_3\left( \theta ,\xi ;T\right) =-\frac{\partial \left( R_ou_o\right) }{%
\partial \theta }\Delta T+\mathrm{oscillating\ terms}$$
Next we collect the secular terms that would contribute to the amplitude equations when applying the RG procedure. Setting now $\varepsilon
=1$ we write down the part of the solution of the system (\[hydrodyn\]) that has to be renormalized
$$F_{RG}\left( \theta ,\xi ;T,T_0\right) =\widetilde{F}\left( \theta ,\xi
;T_0\right) -\Delta T\frac \partial {\partial \theta }\left[ \widetilde{F}%
\left( \theta ,\xi ;T_0\right) \widetilde{U}\left( \theta ,\xi ;T_0\right)
\right] ,$$
$$U_{RG}\left( \theta ,\xi ;T,T_0\right) =\widetilde{U}\left( \theta ,\xi
;T_0\right) -\Delta T\widetilde{U}\left( \theta ,\xi ;T_0\right) \frac
\partial {\partial \theta }\widetilde{U}\left( \theta ,\xi ;T_0\right) +$$
$$+\frac{\lambda ^2}{2\gamma \omega _o^2}\frac{\partial \left| E\left( \theta
;T_0\right) \right| ^2}{\partial \theta }e^{-2\gamma \Delta T},$$
$$V_{RG}\left( \theta ;T,T_0\right) =\left\{ E+\frac{\lambda \Delta T}{%
2i\omega _q}\left[ E\int d\xi R_o+\frac{2i}{\omega _1}\frac \partial
{\partial \theta }\left( E\int d\xi \widetilde{F}\widetilde{U}\right)
-\right. \right. \label{rgperturb}$$
$$-\frac 3{\omega _1^2}\left( \int d\xi F_0U_0^2\right) \frac{\partial ^2E}{%
\partial \theta ^2}-\frac{iE}{2\omega _q}\frac \partial {\partial \theta
}\left( \int d\xi \widetilde{F}\widetilde{U}\right) +$$
$$\left. \left. +\frac \lambda {4\omega _q^2}\left( \int d\xi R_o\right)
\left( E\int d\xi R_o+\frac{2i}{\omega _1}\frac{\partial E}{\partial \theta
}\int d\xi F_0U_0\right) \right] \right\} e^{i\omega _1\Delta T}+c.c.$$
where
$$\widetilde{F}\left( \theta ,\xi ;T_0\right) =F_0\left( \xi \right)
+R_o\left( \theta ,\xi ;T_0\right) \quad ;\quad \widetilde{U}\left( \theta
,\xi ;T_0\right) =U_0\left( \xi \right) +u_o\left( \theta ,\xi ;T_0\right) .$$
Following Kunihiro [@kunihiro] we represent the solution (\[rgperturb\]) as a family of trajectories or curves $\left\{ \Re
_{T_0}\right\} =\left[ F_{RG}\left( T_0\right) ,U_{RG}\left( T_0\right)
,V_{RG}\left( T_0\right) \right] $, being parameterized with $T_0$. The RG equations are defined as the envelope equations for the one-parameter family $\left\{ \Re _{T_0}\right\} $:
$$\left. \left( \frac{\partial F_{RG}}{\partial T_0},\frac{\partial U_{RG}}{%
\partial T_0},\frac{\partial V_{RG}}{\partial T_0}\right) \right| _{T_0=T}=0.
\label{envequation}$$
From the above definition (\[envequation\]) it is straightforward to obtain the desired RG equations:
$$\frac{\partial \widetilde{F}}{\partial T}+\frac \partial {\partial \theta
}\left( \widetilde{F}\widetilde{U}\right) =0,$$
$$\frac{\partial \widetilde{U}}{\partial T}+\widetilde{U}\frac{\partial
\widetilde{U}}{\partial \theta }=-\frac{\lambda ^2}{\omega _o^2}\frac{%
\partial \left| \widetilde{E}\right| ^2}{\partial \theta },$$
$$\frac{2i\omega _q}\lambda \left( \frac \partial {\partial T}+\frac \partial
{\partial \theta }+ \gamma \right) \widetilde{E}=\widetilde{E}\int d\xi
\left( \widetilde{F}-F_0\right) +\frac{2i}{\omega _1}\frac \partial
{\partial \theta }\left( \widetilde{E}\int d\xi \widetilde{F}\widetilde{U}%
\right) - \label{rgequations}$$
$$-\frac 3{\omega _1^2}\left( \int d\xi F_0U_0^2\right) \frac{\partial ^2%
\widetilde{E}}{\partial \theta ^2}-\frac i{2\omega _q}\widetilde{E}\frac
\partial {\partial \theta }\left( \int d\xi \widetilde{F}\widetilde{U}%
\right) +$$
$$+\frac \lambda {4\omega _q^2}\left[ \int d\xi \left( \widetilde{F}%
-F_0\right) \right] \left[ \widetilde{E}\int d\xi \left( \widetilde{F}%
-F_0\right) +\frac{2i}{\omega _1}\frac \partial {\partial \theta }\left(
\widetilde{E}\int d\xi \widetilde{F}\widetilde{U}\right) \right] ,$$
where
$$\widetilde{E}\left( \theta ;T\right) =E\left( \theta ;T\right)
e^{- \gamma T}.$$
The final step consists in defining the envelope distribution function $%
G\left( \theta ,v;T\right) $ by the Radon transform
$$G\left( \theta ,v;T\right) =\int d\xi \widetilde{F}\left( \theta ,\xi
;T\right) \delta \left[ v-\widetilde{U}\left( \theta ,\xi ;T\right) \right] .
\label{radonenvel}$$
By virtue of (\[radonenvel\]) the system of RG equations (\[rgequations\]) is equivalent to the following system of equations for the envelope distribution function $G\left( \theta ,v;T\right) $ and the resonator voltage amplitude $\widetilde{E}\left( \theta ;T\right) $:
$$\frac{\partial G}{\partial T}+v\frac{\partial G}{\partial \theta }-\frac{%
\lambda ^2}{\omega _o^2}\frac{\partial \left| \widetilde{E}\right| ^2}{%
\partial \theta }\frac{\partial G}{\partial v}=0, \label{vlasov}$$
$$\frac{2i\omega _q}\lambda \left( \frac \partial {\partial T}+\frac \partial
{\partial \theta }+ \gamma \right) \widetilde{E}=\widetilde{E}\int dv\left(
G-f_0\right) +\frac{2i}{\omega _1}\frac \partial {\partial \theta }\left(
\widetilde{E}\int dvvG\right) -$$
$$-\frac 3{\omega _1^2}\left[ \int dvv^2f_0\left( v\right) \right] \frac{%
\partial ^2\widetilde{E}}{\partial \theta ^2}-\frac i{2\omega _q}\widetilde{E%
}\frac \partial {\partial \theta }\left( \int dvvG\right) +$$
$$+\frac \lambda {4\omega _q^2}\left[ \int dv\left( G-f_0\right) \right]
\left[ \widetilde{E}\int dv\left( G-f_0\right) +\frac{2i}{\omega _1}\frac
\partial {\partial \theta }\left( \widetilde{E}\int dvvG\right) \right] .
\label{resonator}$$
The system of equations (\[vlasov\]) and (\[resonator\]) provides a complete description of nonlinear particle-wave interaction. It governs slow processes of beam pattern dynamics through the evolution of the amplitude functions. In (\[vlasov\]) one can immediately recognize the Vlasov equation for the envelope distribution function $G\left( \theta ,v;T\right) $ with the ponderomotive force, due to fast oscillations at frequency close to the resonant frequency. It may be worth noting that the system (\[vlasov\]) and (\[resonator\]) intrinsically contains the nonlinear Landau damping mechanism, a fact that will become apparent from the treatment in the next Section.
Derivation of the Generalized Ginzburg-Landau Equation.
=======================================================
In order to solve equation (\[vlasov\]) we perform a Fourier transform and obtain:
$$\left( \Omega -kv\right) G\left( \chi \right) =$$
$$-\frac{\lambda ^2}{\left( 2\pi \right) ^4\omega _o^2}\int d\chi _1d\chi
_2d\chi _3\delta \left( \chi -\chi _1-\chi _2-\chi _3\right) \left(
k-k_1\right) \frac{\partial G\left( \chi _1\right) }{\partial v}\widetilde{E}%
\left( \chi _2\right) \widetilde{E}^{*}\left( \chi _3\right) ,
\label{fourier}$$
where
$$\chi =\left( k,\Omega \right) \qquad ;\qquad \delta \left( \chi \right)
=\delta \left( k\right) \delta \left( \Omega \right)$$
and the Fourier transform of a generic function $g\left( \theta
;T\right) $ is defined as
$$g\left( \theta ;T\right) =\frac 1{\left( 2\pi \right) ^2}\int d\Omega
dkg\left( k;\Omega \right) e^{i\left( k\theta -\Omega T\right) },$$
$$g\left( k;\Omega \right) =\int d\theta dTg\left( \theta ;T\right)
e^{-i\left( k\theta -\Omega T\right) },$$
$$g^{*}\left( k;\Omega \right) =\left[ g\left( -k;-\Omega \right) \right]
^{*}.$$
Solving equation (\[fourier\]) perturbatively we represent its solution in the form:
$$G\left( \chi \right) =\left( 2\pi \right) ^2f_0\left( v\right) \delta \left(
\chi \right) +\widetilde{G}\left( \chi \right) \quad ;\quad \widetilde{G}%
\left( \chi \right) =\sum\limits_{n=1}^\infty G_n\left( \chi \right) ,
\label{gensol}$$
where
$$G_1\left( \chi \right) =-\frac{\lambda ^2}{\left( 2\pi \right) ^2\omega _o^2}%
\frac k{\Omega -kv}\frac{\partial f_0}{\partial v}\int d\chi _1d\chi
_2\delta \left( \chi -\chi _1-\chi _2\right) \widetilde{E}\left( \chi
_1\right) \widetilde{E}^{*}\left( \chi _2\right) , \label{sol1}$$
$$G_n\left( \chi \right) =-\frac{\lambda ^2}{\left( 2\pi \right) ^4\omega _o^2}%
\frac 1{\Omega -kv}*$$
$$\ast \int d\chi _1d\chi _2d\chi _3\delta \left( \chi -\chi _1-\chi _2-\chi
_3\right) \left( k-k_1\right) \frac{\partial G_{n-1}\left( \chi _1\right) }{%
\partial v}\widetilde{E}\left( \chi _2\right) \widetilde{E}^{*}\left( \chi
_3\right) . \label{soln}$$
The Fourier transform of equation (\[resonator\]) yields the linear dispersion relation
$$\Omega =-i\gamma +k-\frac{\lambda \left\langle v\right\rangle _0}{\omega
_q\omega _1}k+\frac{3\lambda \left\langle v^2\right\rangle _0}{2\omega
_q\omega _1}k^2=$$
$$=k-\frac{\lambda \left\langle v\right\rangle _0}{\omega _o^2}k+\frac{%
3\lambda \left\langle v^2\right\rangle _0}{2\omega _q\omega _o^2}k^2-i\gamma
\left( 1-\frac{\lambda \left\langle v\right\rangle _0}{\omega _q\omega _o^2}%
k+\frac{3\lambda \left\langle v^2\right\rangle _0}{\omega _o^4}k^2\right) .
\label{dispersion}$$
The integrals over $v$ of the envelope distribution function $%
G\left( \theta ,v;T\right) $
$$I_0\left( \theta ;T\right) =\int dvG\left( \theta ,v;T\right) =\frac
1{\left( 2\pi \right) ^2}\int dvd\Omega dkG\left( k,v;\Omega \right)
e^{i\left( k\theta -\Omega T\right) },$$
$$I_1\left( \theta ;T\right) =\int dvvG\left( \theta ,v;T\right) =\frac
1{\left( 2\pi \right) ^2}\int dvd\Omega dkvG\left( k,v;\Omega \right)
e^{i\left( k\theta -\Omega T\right) }$$
entering equation (\[resonator\]) can be computed in a straightforward manner. Substituting the solution (\[gensol\])-(\[soln\]) into the above equations with the linear dispersion relation (\[dispersion\]) in hand, up to second order in $G\left( k,v;\Omega \right) $, we find
$$I_0\left( \theta ;T\right) =1-W\left[ \left| \widetilde{E}\left( \theta
;T\right) \right| \right] +..., \label{i0}$$
$$I_1\left( \theta ;T\right) =1-\left( 1-\frac \lambda {\omega _o^2}\right)
W\left[ \left| \widetilde{E}\left( \theta ;T\right) \right| \right] +...,
\label{i1}$$
where the function $W$ is defined as
$$W\left( z\right) =\frac{\lambda ^2}{\omega _o^2\sigma _v^2}\left( 1+i\gamma
_L\right) \left( 1-\frac{\lambda ^2}{2\omega _o^2\sigma _v^2}z^2\right) z^2,$$
and
$$\gamma _L=\frac \lambda {\omega _o^2\sigma _v}\sqrt{\frac \pi 2}\exp \left( -%
\frac{\lambda ^2}{2\omega _o^4\sigma _v^2}\right)$$
is the Landau damping factor. In the above calculations the equilibrium distribution function has been taken to be the Gaussian one
$$f_0\left( v\right) =\frac 1{\sigma _v\sqrt{2\pi }}\exp \left[ -\frac{\left(
v-1\right) ^2}{2\sigma _v^2}\right] ,$$
where
$$\sigma _v=\frac{\left| k_o\right| \sigma _E}{\omega _s}$$
and $\sigma _E$ is the r.m.s. of the energy error, proportional to the longitudinal beam temperature. By substitution of the expressions (\[i0\]) and (\[i1\]) into equation (\[resonator\]) we arrive at the generalized Ginzburg-Landau equation:
$$\frac{2i\omega _q}\lambda \left( \frac \partial {\partial T}+\frac \partial
{\partial \theta }+ \gamma \right) \widetilde{E}=-\frac{3\left( 1+\sigma
_v^2\right) }{\omega _1^2}\frac{\partial ^2\widetilde{E}}{\partial \theta ^2}%
+\frac{2i}{\omega _1}\frac{\partial \widetilde{E}}{\partial \theta }-\left(
1-\frac \lambda {4\omega _q^2}W\right) W\widetilde{E}-$$
$$-\frac{2i}{\omega _1}\frac \lambda {4\omega _q^2}W\frac{\partial \widetilde{E%
}}{\partial \theta }+\frac i{2\omega _q}\left( 1-\frac \lambda {\omega
_o^2}\right) \widetilde{E}\frac{\partial W}{\partial \theta }-\frac{2i}{%
\omega _1}\left( 1-\frac \lambda {\omega _o^2}\right) \left( 1-\frac \lambda
{4\omega _q^2}W\right) \frac \partial {\partial \theta }\left( W\widetilde{E}%
\right) .$$
It can be further cast to a simpler form by introducing the rescaled independent and dependent variables according to
$$\tau =\frac{\lambda T}{2\omega _q}\quad ;\quad x=\frac{\omega _o}{\sqrt{%
3\left( 1+\sigma _v^2\right) }}\left( \theta -T+\frac{\lambda T}{\omega
_o\omega _q}\right) \quad ;\quad \Psi =\frac{\left| \lambda \right|
\widetilde{E}}{\omega _o\sigma _v}.$$
Then the generalized Ginzburg-Landau equation for the amplitude of the resonator voltage takes its final form:
$$i\frac{\partial \Psi }{\partial \tau } + ib \gamma \Psi =-\left( 1-\frac{%
2i\gamma }{\omega _o}\right) \frac{\partial ^2\Psi }{\partial x^2}+a\gamma
\frac{\partial \Psi }{\partial x}-\left( 1-b_1W\right) W\Psi -$$
$$-ab_1\left( \gamma +i\omega _o\right) W\frac{\partial \Psi }{\partial x}%
-a_1\left( \gamma +i\omega _o\right) \left( 1-b_1W\right) \frac \partial
{\partial x}\left( W\Psi \right) +\frac{ia_1b_1b\omega _o^2}2\Psi \frac{%
\partial W}{\partial x}, \label{gglequat}$$
where
$$a=\frac 2{\omega _o\sqrt{3\left( 1+\sigma _v^2\right) }}\qquad ;\qquad b=%
\frac{2\omega _q}\lambda ,$$
$$b_1=\frac 1{\lambda b^2}\qquad ;\qquad a_1=a\left( 1-4b_1\right) ,$$
and the function $W\left( \left| \Psi \right| \right) $ is given now by the simple expression:
$$W\left( \left| \Psi \right| \right) =\left( 1+i\gamma _L\right) \left(
1-\frac 12\left| \Psi \right| ^2\right) \left| \Psi \right| ^2.
\label{nlpotent}$$
The generalized Ginzburg-Landau equation (\[gglequat\]) is known [@cross] to provide the basic framework for the study of many properties of non equilibrium systems, such as existence and interaction of coherent structures, generic onset of travelling wave disturbance in continuous media, appearance of chaos. Recent experimental and numerical evidence (see e.g. [@tzenov], [@colestock] and the references therein) shows that similar behavior is consistent with the propagation of charged particle beams, and the generalized Ginzburg-Landau equation we have derived could represent the appropriate analytical model to study the above mentioned phenomena.
Concluding Remarks.
===================
As a result of the investigation performed we have shown that a coasting beam under the influence of a resonator impedance exhibits spatial-temporal patterns modulated by an envelope (amplitude) function, which varies slowly compared to the fast time and short wavelength scales of the pattern itself. Extracting long wavelength and slow time scale behavior of the system we have derived a set of coupled nonlinear evolution equations for the beam envelope distribution function and voltage amplitude. We have further shown that the amplitude of the nonlinear wave satisfies a generalized Ginzburg-Landau equation.
It is worthwhile to mention that the analytical framework presented here bears rather general features. It provides complete kinetic description of slow, fully nonlinear particle-wave interaction process, and allows higher order corrections to the generalized Ginzburg-Landau equation to be taken into account.
Acknowledgements.
=================
The author is indebted to Pat Colestock for many helpful discussions concerning the subject of the present paper, and to David Finley and Steve Holmes for their enlightened support of this work.
[9]{} *S.I. Tzenov and P.L. Colestock*, ”Solitary Waves on a Coasting High-Energy Stored Beam”, **Fermilab-Pub-98-258**, Fermilab, Batavia, 1998.\
**physics/9808032**, 24 August 1998.
*P.L. Colestock, L.K. Spentzouris and S.I. Tzenov*, ”Coherent Nonlinear Phenomena in High Energy Synchrotrons: Observations and Theoretical Models”,** International Symposium on Near Beam Physics**, Fermilab, September 22-24, 1997, R.A. Carrigan and N.V. Mokhov eds.,Fermilab, Batavia, 1998, pp 94-104.\
**physics/9808035**, 25 August 1998.
*L.-Y. Chen, N. Goldenfeld and Y. Oono*, ”Renormalization Group and Singular Perturbations: Multiple Scales, Boundary Layers and Reductive Perturbation Theory”, **Phys. Rev. E,** Vol.** 54** (1996) p. 376.
*T. Kunihiro*, ”The Renormalization Group Method Applied to Asymptotic Analysis of Vector Fields”, **Prog. of Theoretical Physics,** Vol.** 97** (1997) p. 179.
*J. Horvath*, **”Topological Vector Spaces and Distributions”**, Addison-Wesley, Reading, Massachusetts, 1966.
*Yu.L. Klimontovich*, **”Statistical Theory of Open Systems”**, Kluwer Academic Publishers, Dordrecht, 1995.
*M.C. Cross and P.C. Hohenberg*, ”Pattern Formation Outside of Equilibrium”, **Reviews of Modern Physics**, Vol. **65** (1993) p. 851.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We extend a previous bispectrum analysis of the Cosmic Microwave Background temperature anisotropy, allowing for the presence of correlations between different angular scales. We find a strong non-Gaussian signal in the “inter-scale” components of the bispectrum: their observed values concentrate close to zero instead of displaying the scatter expected from Gaussian maps. This signal is present over the range of multipoles $\ell=6 -18$, in contrast with previous detections. We attempt to attribute this effect to galactic foreground contamination, pixelization effects, possible anomalies in the noise, documented systematic errors studied by the COBE team, and the effect of assumptions used in our Monte Carlo simulations. Within this class of systematic errors the confidence level for rejecting Gaussianity varies between $97\%$ and $99.8\%$.'
author:
- João Magueijo
title: 'New non-Gaussian feature in COBE-DMR Four Year Maps'
---
Introduction
============
In a recent [*Letter*]{} [@fmg] found strong evidence for non-Gaussianity in the anisotropy of the Cosmic Microwave Background (CMB) temperature. This detection was followed by similar claims from [@nfs98] and [@pvl98], and caused considerable consternation among theorists (see [@nvs] for a discussion). These three groups employed very different statistical tools, but used the same dataset – the COBE-DMR 4 year maps. Recent work by [@bt99] confirmed these measurements, but [@banday] cast doubts upon the cosmological origin of the observed signals. We clearly do not fully understand some of the less conspicuous systematic errors associated with DMR maps. We feel that the origin of the observed non-Gaussian features will probably not be conclusively identified before an independent all-sky dataset becomes available.
In this Letter we revisit and complete the analysis of [@fmg]. In that work the possibility of departures from Gaussianity was examined in terms of the bispectrum. Given a full-sky map, $\frac{\Delta T}{T}({\bf n})$, this may be expanded into Spherical Harmonic functions: $$\begin{aligned}
\frac{\Delta T}{T}({\bf n})=\sum_{\ell m}a_{\ell m}Y_{\ell m}({\bf n})
\label{almdef} \end{aligned}$$ The coefficients $a_{\ell m}$ may then be combined into rotationally invariant multilinear forms (see [@santa] for a possible algorithm). The most general cubic invariant is the bispectrum, and is given by $$\begin{aligned}
{\hat B}_{\ell_1\ell_2\ell_3}&=&\alpha_{\ell_1\ell_2\ell_3}
\sum_{m_1m_2m_3}\left
( \begin{array}{ccc} \ell_1 & \ell_2 & \ell_3 \\ m_1 & m_2 & m_3
\end{array} \right ) a_{\ell_1 m_1}a_{\ell_2 m_2} a_{\ell_3 m_3}
\nonumber \\
\alpha_{\ell_1\ell_2\ell_3}&=&\frac{1}{(2\ell_1+1)^{\frac{1}{2}}
(2\ell_2+1)^{\frac{1}{2}}(2\ell_3+1)^{\frac{1}{2}}}\left (
\begin{array}{ccc} \ell_1
& \ell_2 & \ell_3 \\ 0 & 0 & 0
\end{array} \right )^{-1}
\label{bispec} \end{aligned}$$ where the $(\ldots)$ is the Wigner $3J$ symbol. In [@fmg] correlations between different multipoles were ignored, and so $\ell_1=\ell_2=\ell_3$. Here we consider bispectrum components sensitive to correlations between different scales. Selection rules require that $\ell_1+\ell_2+\ell_3$ be even. The simplest chain of correlators is therefore ${\hat A_\ell}=B_{\ell-1\, \ell \,\ell+1}$, with $\ell$ even. Other components, involving more distant multipoles, may be considered but they are very likely to be dominated by noise; it is natural to assume that possible non-Gaussian inter-scale correlations decay with $\ell$ separation. We shall consider a ratio $$\begin{aligned}
J^3_\ell &=& { {\hat A}_{\ell}
\over ({\hat C}_{\ell-1})^{1/2}({\hat C}_{\ell})^{1/2}
({\hat C}_{\ell+1})^{1/2}}
\label{defI} \end{aligned}$$ where ${\hat C}_\ell=\frac{1}{2\ell+1}\sum_m|a_{\ell m}|^2$. This quantity is dimensionless, and is invariant under rotations and parity. It extends the $I^3_\ell$ statistic used in [@fmg] [^1].
The theoretical importance of the bispectrum as a non-Gaussian qualifier has been recognized in a number of publications ([@luo94], [@peebles1], [@sg98], [@gs98], [@wang]). [@kog96a] measured the pseudocollapsed and equilateral three point function of the DMR four year data. The bispectrum may be regarded as the Fourier space counterpart of the three point function. The work presented in this [*Letter*]{}, combined with [@fmg] and [@heav], provides a complete set of signal dominated bispectrum components inferred from DMR maps.
The publicly released 4 year maps
=================================
We first consider the inverse-noise-variance-weighted, average maps of the 53A, 53B, 90A and 90B [*COBE*]{}-DMR channels, with monopole and dipole removed, at resolution 6, in galactic and ecliptic pixelization. We use the extended galactic cut of [@banday97], and [@benn96] to remove most of the emission from the plane of the Galaxy. To estimate the $J^3_\ell$s we set the value of the pixels within the galactic cut to 0 and the monopole and dipole of the cut map to zero. We then integrate the map multiplied with spherical harmonics to obtain the estimates of the $a_{\ell m}$s and apply equations \[bispec\] and \[defI\]. The observed $J^3_\ell$ are to be compared with their distribution $P(J^3_\ell)$ as inferred from Monte Carlo simulations in which Gaussian maps are subject to DMR noise and galactic cut. In simulating DMR noise we take into account the full noise covariance matrix, as described in [@corn]. This includes correlations between pixels $60^\circ$ degrees apart.
The results are displayed in Fig. \[fig1\]. Leaving aside the deviant $J^3_4$, it is blatantly obvious that the observed $J^3_\ell$ do not exhibit the scatter around zero implied by $P(J^3_\ell)$ in the range of scales $\ell=6-18$. This may be mathematically formalized by means of a goodness of fit statistic, such as the “chi squared”: $$\label{presc}
X^2={1\over N}{\sum_\ell X_\ell^2}=
{1\over N}{\sum_\ell (-2\log P_\ell(J^3_\ell)
+ \beta_\ell),}$$ where the constants $\beta_\ell$ are defined so that for each term of the sum $\langle X_\ell^2\rangle=1$. As explained in [@fmg] this quantity reduces to the usual chi squared when the distributions $P$ are Gaussian. When $P$ is not Gaussian, $X^2$ goes to infinite where $P$ goes to zero, reaches a minimum (usually around zero) at the peak of $P$, and has average 1. Hence $X^2$ does for a non-Gaussian $P$ what the usual chi squared does for a Gaussian $P$. A good fit is represented by $X^2\approx 1$. If $X^2\gg 1$ the data is plagued by deviants, that is observations far in the tail of the theoretical distribution. If $X^2\ll 1$ the observations fail to exhibit the scatter predicted by the theory, concentrating uncannily on the peak of the distribution. Both cases present grounds for rejecting the hypothesis embodied in $P(J^3_\ell)$, in our case Gaussianity.
We find $X^2=0.14$ and $X^2=0.22$ for data in galactic and ecliptic pixelization, respectively. To quantify the confidence level for rejecting Gaussianity we determine the distribution of $X^2$, $F(X^2)$, making use of further Monte Carlo simulations. The detailed procedure shadows that described in [@fmg]. We stress that in each realization a new sky is produced, from which a full set of $J_\ell$ is derived, for which the $X^2$ is computed. The result is plotted in Fig. \[fig2\], where we superimpose the observed $X^2$ and its distribution. We then compute the percentage of the population with a larger $X^2$ than the observed one. We find that $P(X^2>0.14)= 0.998$ (and $P(X^2>0.22)=0.985$) for maps in galactic (ecliptic) pixelization. The lack of scatter in the observed $J^3_\ell$ implies that Gaussianity may be rejected at the $99.8\%$ ($98.5\%$) confidence level.
A closer analysis reveals that this signal is mainly in the 53 GHz channel (see Table 1), which is also the least noisy channel. However the confidence level for rejecting Gaussianity increases (from $93.2\%$ to $99.8\%$ in galactic pixelization) when the 53Ghz channel is combined with the 90Ghz channel. Hence the overall signal is due to both channels. The reduced confidence levels in the separate channels merely reflect a lower signal to noise ratio, and the Gaussian nature of noise.
Combining different Gaussianity tests
=====================================
Woven into the above argument is a perspective on how to combine different Gaussianity tests which differs from that presented by [@bt99]. In that work the authors argue that if $n$ tests are made for a given hypothesis, and they return confidence levels for rejection $\{p_i\}$, then, if $p_{max}=\max\{p_i\}$, the actual confidence level for rejection is $p_{max}^n$.
While the above recipe is formally correct it cannot be applied when the hypothesis is Gaussianity. Let the various tests be a set of cumulants $\{\kappa_i\}$ ([@kos]). Suppose that all cumulants are consistent with Gaussianity except for a single cumulant, which prompts us to reject Gaussianity with confidence level $p_{max}$. Clearly the confidence level for rejecting Gaussianity is $p_{max}$, since it is enough for the distribution to have a single non-Gaussian cumulant for it to be non-Gaussian.
The point is that Gaussianity cannot be regarded by itself as an hypothesis, since the corresponding alternative hypothesis includes an infinity of independent degrees of freedom involving different moments and scales. The argument of [@bt99] is correct when applied to independent tests concerning the same non-Gaussian degree of freedom, for instance independent tests related to the skewness. However it cannot be true for different tests probing independent non-Gaussian degrees of freedom, say skewness and kurtosis.
In the context of our result (which returns $X^2\ll 1$), we notice that if we were to include into the analysis the results of [@fmg] (for which $X^2\gg 1$) we would get an average $X^2\approx 1$. Such procedure is obviously nonsensical: two wrongs don’t make a right. One should examine independent non-Gaussian features separately, in particular the $I^3_\ell$ and the $J^3_\ell$, or two ranges of $\ell$, one with $X^2\gg 1$ the other with $X^2\ll 1$. The only practical constraint is sample variance, forcing any analysis to include more than one degree of freedom so that $F(X^2)$ is sufficiently peaked.
The possibility of a non-cosmological origin
============================================
Could this signal have a non-cosmological origin? The possibility of foreground contamination was considered in two ways. Firstly we subject foreground templates to the same analysis. At the observing frequencies the obvious contaminant should be foreground dust emission. The DIRBE maps ([@boggess92]) supply us with a useful template on which we can measure the $J^3_\ell$s. We have done this for two of the lowest frequency maps, the $100$ $\mu$m and the $240$ $\mu$m maps. The estimate is performed in exactly the same way as for the DMR data (i.e. using the extended Galaxy cut). We performed a similar exercise with the Haslam 408Mhz ([@haslam]) map. The results are presented in Table 1. We find that the $J^3_\ell$ (in contrast to the $I^3_\ell$ used by [@fmg]) are capable of exposing the non-Gaussianity in these templates, even when smoothed by a $7^\circ$ beam. However none of the signatures found correlates with the DMR signal. DIRBE maps produce highly deviant $J^3_\ell$, whereas all $J^3_\ell>0$ for the Haslam map.
We also considered foreground corrected maps (see Table 1). In these one corrects the coadded 53 and 90 Ghz maps for the DIRBE correlated emission. We studied maps made in ecliptic and galactic frames, and also another map made in the ecliptic frame but with the DIRBE correction forced to have the same coupling as determined in the galactic frame. The confidence levels for rejecting Gaussianity are $97.9\%$, $98.5\%$, and $97.1\%$, respectively. The signal is therefore reduced, but not erased.
Could the observed signal be due to detector noise? The DMR noise is subtly non-Gaussian, due to its anisotropy and pixel-pixel correlations ([@corn]). These features were incorporated into the simulations leading to $P(J^3_\ell)$. However it could just happen that the noise in the particular realization we have observed turned out to be a fluke, concentrating the observed $J^3_\ell$ around zero. We examined this possibility by considering difference maps $(A-B)/2$. If the observed effect is the result of a noise fluke, it should be exacerbated in $(A-B)/2$ maps, rather than in $(A+B)/2$ maps. As can be seen in Table 1, one of the noise maps ($(A-B)/2$, 53Ghz, in ecliptic pixelization) is indeed unusually non-Gaussian — but with $X^2\gg 1$ rather than $X^2\ll 1$. This feature disappears in $(A-B)/2$, 53Ghz, maps made in the galactic pixelization. The $(A-B)/2$, 90GHz map, in galactic pixelization, has a low $X^2$ but far from significant.
[@bt99] have shown that removing a selected beam-size region from DMR maps deteriorates the $I^3_\ell$ non-Gaussian signal. This fact is of great interest as it highlights the possible spatial localization of what is a priori a “Fourier space” statistic. More recently [@banday] pointed out that removing single beam-size regions may also reinforce the $I^3_\ell$ non-Gaussian signal. We have subjected our $J^3_\ell$ analysis to this exercise. We found that $X^2$ from maps without a single beam sized region is [*very* ]{} sharply peaked around the uncut value 0.14. There is a region without which $X^2=0.22$ but it is also possible to remove a region so that $X^2=0.08$. Hence the $J^3_\ell$ signal can never be significantly deteriorated by means of this prescription, and for this reason we believe it to be essentially a Fourier space feature.
A number of systematic error templates were also examined ([@kog96c]). These provide estimates at the $95\%$ confidence level of errors due to the following: the effect of instrument susceptibility to the Earth magnetic field; any unknown effects at the spacecraft spin period; errors in the calibration associated with long-term drifts, and calibration errors at the orbit and spin frequency; errors due to incorrect removal of the COBE Doppler and Earth Doppler signals; errors in correcting for emissions from the Earth, and eclipse effects; artifacts due to uncertainty in the correction for the correlation created by the low-pass filter on the lock-in amplifiers (LIA) on each radiometer; errors due to emissions from the moon, and the planets. The systematic templates display strongly non-Gaussian structures, tracing the DMR scanning patterns. We added or subtracted these templates enhanced by a factor of up to 4 to DMR maps (see [@santa] for a better description of the procedure). The effect on the $J^3_\ell$ spectrum was always found to be small, leading to very small variations in the $X^2$.
[@banday] have recently claimed that the systematic errors due to eclipse effects may be larger than previously thought. They showed how the $I^3_\ell$ change dramatically when estimated from maps in which data collected in the two month eclipse season has been discarded. These maps are more noisy, and the $I^3_\ell$ are very sensitive to noise. Indeed the $I^3_\ell$ are cubic statistics, with a signal to noise proportional to (Number of Observations$)^{3/2}$, and so they are much more sensitive to noise than the power spectrum. Perhaps the variations in $I^3_\ell$ merely reflect a larger noise, and not a systematic effect. This possibility could be disproved if no striking variations in the $I^3_\ell$ were found in maps for which [*other*]{} two month data samples are excised. We have applied the $J^3_\ell$ analysis to maps without eclipse data, and found that the confidence level for rejecting Gaussianity does not drop below $99.2\%$ (see Table 1). Hence the result described in this Letter appears to be robust in this respect. A more detailed description of the impact of systematics upon the $I^3_\ell$ and $J^3_\ell$ will be the subject of a comprehensive publication ([@banday00]).
We finally subject our algorithm to a number of tests. Arbitrary rotations of the coordinate system (as opposed to the pixelization scheme) affects $J^3_\ell$ to less than a part in $10^5$. Possible residual offsets (resulting from the removal of the monopole and dipole on the cut map) do not destroy the signal found. Finally changing the various assumptions going into Monte Carlo simulations do not affect the estimated distributions $P(J^3_\ell)$. We found these distributions to be independent of the assumed shape of the power spectrum, of the exact shape of the DMR beam, or the inclusion of the pixel window function.
In summary the $J^3_\ell$ analysis appears to be more sensitive to shortcomings in DMR maps than the $I^3_\ell$. This fact is already obvious in the differences between ecliptic and galactic pixelizations in the publicly released maps. When all possible renditions of DMR data are considered the significance level of our detection may vary between $97\%$ and $99.8\%$. Therefore the various tests for systematics we have described do not leave the result unscathed, but neither do they rule out a cosmological origin. We should stress that, in line with all previous work in the field, we have employed a frequentist approach. One may therefore question the Bayesian meaning of the confidence levels quoted. A Bayesian treatment of the bispectrum remains unfeasible (see however [@contaldi]).
In comparison with [@fmg] the result we have described is more believable from a theoretical point of view. It spreads over a range of scales. Previous detections concentrate on a single mode. The result obtained is puzzling in that, rather than revealing the presence of deviants, it shows a perfect alignment of the observed $J^3_\ell$ on the top of their distribution for a Gaussian process. This is perhaps not as strange as it might seem at first: in [@fms] it was shown how non-Gaussianity may reveal itself not by non-zero average cumulants, but by abnormal errorbars around zero.
ACKNOWLEDGEMENTS {#acknowledgements .unnumbered}
================
I would like to thank K. Baskerville, P. Ferreira and K. Gorski for help with this project. Special thanks to T. Banday for supplying DMR maps without the eclipse season data. This work was performed on COSMOS, the Origin 2000 supercomputer owned by the UK-CCC and supported by HEFCE and PPARC. I thank the Royal Society for financial support.
[99]{} Banday, A.J. [*et al*]{} [*Ap.J.*]{}, [**475**]{}, 393 (1997). Banday, A.J., Zaroubi, S., and Górski, K.M., astro-ph/9908070. Banday, A.J., Ferreira, P.G., Górski, K.M., Magueijo, J., and Zaroubi, S., in preparation. Bennet, C.L. [*et al*]{} [*Ap. J.*]{} [**464**]{}, 1 (1996). Bogess et al. (1992) [*Ap. J*]{} [**397**]{} 420 Bromley, B. and Tegmark, M. [astro-ph/9904254]{}. Contaldi, C.R., Ferreira, P.G., Magueijo, J., Górski, K.M., “A Bayesian estimate of the skewness of the Cosmic Microwave Background”, astro-ph/9910138. Ferreira, P.G., Magueijo, J. and Silk, J. [*Phys.Rev.*]{} [**D56** ]{}, 4592-4603 (1997). Ferreira, P.G., Magueijo, J. and Górski, K.M.G. [*Ap.J.*]{} [**503**]{} L1(1998) Goldberg, D.M., and Spergel, D.N., astro-ph/9811251. Haslam, C.G.T et al (1982) [*Astron. Astrophys. Suppl. Ser.*]{}, [**63**]{}, 205. Heavens, A., astro-ph/9804222. M. Kamionkowski and A. Jaffe, [*Nature*]{} [**395**]{} 639 (1998). Stuart, A. and Ord, K., Kendall’s advanced theory of statistics, Wiley, London, 1994. Kogut, A., Banday, A. J., Bennett, C. L., Górski, K. M., Hinshaw, G., Smoot, G. F., Wright, E. L. (1996) [*Ap. J.*]{} [**464**]{} L29 Kogut, A., Banday, A. J., Bennett, C. L., Górski, K. M., Hinshaw, G., Smoot, G. F., Wright, E. L. (1996) [*Ap. J.*]{} [**470**]{} 653-673. Luo X. (1994) [*Ap. J.*]{} [**427**]{} L71. C. H. Lineweaver, G. F. Smoot, C. L. Bennett, E. L. Wright, L. Tenorio, A. Kogut, P. B. Keegstra, G. Hinshaw, A. J. Banday, astro-ph/9403021. Magueijo, J., Ferreira, P., and Górski, K.M., 1998, in the Proceedings of the Planck Meeting, Santender 98, astro-ph/9810414. Novikov, D., Feldman, H. and Shandarin, S. [astro-ph/9809238]{} Pando, J., Valls-Gabaud and Fang, L.Z., [*Phys.Rev.Lett*]{} [**81**]{} 4568 (1998). Peebles, P.J. [*Ap.J.*]{} [**483**]{}, L1 (1997). Peebles, P.J.; astro-ph/9805194 and astro-ph/9805212. Spergel, D.N. and Goldberg, D.M., astro-ph/9811252. Wang, L. and Kamionkowski, M,. astro-ph/9907431.
\[table1\]
[cccccccccccc]{} & $J^3_4$ & $J^3_6$ & $J^3_8$ & $J^3_{10}$ & $J^3_{12}$ & $J^3_{14}$ & $J^3_{16}$ & $J^3_{18}$ & $X^2$ & Reject %\
Gauss-rms&.263&.225&.194&.178&.162&.153&.144&.135& —&—\
DMR - ecl& -.521& .009 & -.022 &-.007 &.112 & .088& .054 & .045 & .22 & 98.5\
DMR - gal&-.502&.012&.029&-.088&.098&-.002&.013&-.030&.14& 99.8\
DMR - cor/ecl & -.554& .022 & -.091 &-.099 &.104 & .061& .071 & .057 & .25 & 97.9\
DMR - cor/ecg & -.555& .026 & -.105 &-.104 &.102 & .055& .076 & .059 & .27 & 97.1\
DMR - cor/gal&-.542&.042&-.068&-.140&.108&-.026&.031&-.019&.23& 98.5\
DMR - gal 90&-.513&.003&-.316&-.167&.140&.063&.042&-.049&.66& 71.2\
DMR - gal 53&-.497&-.053&-.009&.022&.172&-.076&.107&-.031&.36& 93.2\
DMR - gal/ne &-.448&.034&.042&-.022&.127&.009&.008&.042&.18& 99.2\
A-B 53 ecl &-.304&-.002&.126&-.008&-.426&.437&.170&-.047&2.41 & 98.4 $^*$\
A-B 53 gal &-.259&.034&.118&-.003&-.193&.248&.246&.019&1.07& 38.1\
A-B 90 ecl &-.134&-.151&-.195&.099&-.181&.091&-.273&.021&1.00& 43.3\
A-B 90 gal &-.085&-.158&-.173&.001&-.140&.047&-.180&.079&.57& 79.0\
DIRBE08 ecl &.076&.203&-.109&.404&.240&.027&-.293&-.107&1.93 & 94.1$^*$\
DIRBE10 ecl &.252&.124&-.267&.706&.134&.184&-.177&-.145&3.15 &99.7$^*$\
Haslam ecl&.279&.459&.062&.158&.196&.146&.049&.100&1.21&29.4\
Haslam gal &.258&.462&.044&.182&.163&.105&.037&.045&1.04&39.7\
[^1]: Taking the modulus is not necessary to ensure parity invariance, contrary to the statement made in [@fmg]. This does not affect any of the conclusions, since $P(I^3_\ell)=P(-I^3_\ell)$ for a Gaussian process.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have performed a comprehensive analysis of the type D group $D^{(1)}_{9n, 3n}$ as flavor symmetry and the generalized CP symmetry. All possible residual symmetries and their consequences for the prediction of the mixing parameters are studied. We find that only one type of mixing pattern is able to accommodate the measured values of the mixing angles in both “direct” and “variant of semidirect” approaches, and four types of mixing patterns are phenomenologically viable in the “semidirect” approach. The admissible values of the mixing angles as well as CP violating phases are studied in detail for each case. It is remarkable that the first two smallest $D^{(1)}_{9n,3n}$ groups with $n=1, 2$ can fits the experimental data very well. The phenomenological predictions for neutrinoless double beta decay are discussed.'
author:
- |
\
Cai-Chang Li[^1] , Chang-Yuan Yao[^2] , Gui-Jun Ding[^3]\
\
*[Department of Modern Physics, University of Science and Technology of China,]{}\
*[Hefei, Anhui 230026, China]{}\
**
title: |
\
**Lepton Mixing Predictions from Infinite Group Series $D^{(1)}_{9n, 3n}$ with Generalized CP**
---
\[sec:Int\]Introduction
=======================
The precise measurement of the reactor mixing angle $\theta_{13}$ [@Abe:2011sj; @Adamson:2011qu; @Abe:2011fz; @An:2012eh; @Ahn:2012nd] encourages the pursuit of the still missing results on leptonic CP violation and neutrino mass ordering as well as the characteristic neutrino nature. Some low-significance hints for a maximally CP-violating value of the Dirac phase $\delta_{CP}\simeq3\pi/2$ have been observed [@Abe:2015awa]. The global fits to lepton mixing parameters [@Capozzi:2013csa; @Forero:2014bxa; @Gonzalez-Garcia:2014bfa] also provide weak evidence for the existence of Dirac type CP violation in neutrino oscillation. In the case that neutrinos are Majorana particles, two more Majorana CP phases $\alpha_{21}$ and $\alpha_{31}$ would be present, and they are crucial to the neutrinoless double beta decay process. However, the present experimental data don’t impose any constraint on the values of the Majorana phases.
Finite discrete non-abelian flavor symmetries have been widely used to make predictions for lepton flavor mixing. Assuming the original flavor symmetry group is spontaneously broken to distinct abelian residual symmetries in the neutrino and charged lepton sectors at a low energy scale, one can then determine mixing patterns from the residual symmetries and the structure of discrete flavor symmetry groups. Please see Refs. [@Altarelli:2010gt; @Ishimori:2010au; @King:2013eh] for review on discrete flavor symmetries and the application in model building. For Majorana neutrinos, if the residual symmetries of the charged lepton and neutrino mass matrices originate from a finite flavor group, the lepton mixing matrix would be fully determined by residual symmetries up to independent row and column permutations. It turns out that the possible forms of the PMNS matrix are strongly constrained in this scenario such that the mixing patterns compatible with the data are of trimaximal form, and the Dirac CP phase is predicted to be $0$ or $\pi$ [@Fonseca:2014koa]. The same conclusion is reached for neutrinos being Dirac particles [@Yao:2015dwa]. We note that the neutrino masses are not constrained in this approach and consequently the both Majorana phases $\alpha_{21}$ and $\alpha_{31}$ are undetermined. Their values can be fixed by considering a specific model. If the residual flavor symmetries of the neutrino and charged lepton mass matrix are partially contained in the underlying flavor group, the PMNS matrix would contains at least two free continuous parameters. As a result, the predictivity of the model would be lessened to a certain extent.
Besides the extensively discussed residual flavor symmetries, the neutrino and charged lepton mass matrices also admit residual CP transformations, and the residual CP symmetries can be generated by performing two residual CP transformations [@Chen:2014wxa; @Chen:2015nha; @Everett:2015oka]. Analogous to residual flavor symmetries, the residual CP transformations can also constraint the lepton flavor mixing in particular the CP violating phases [@Chen:2014wxa]. The simplest nontrivial CP transformation is known as $\mu-\tau$ reflection which gives rise to maximal atmospheric mixing and maximal Dirac phase [@Harrison:2002kp; @Grimus:2003yn; @Farzan:2006vj]. The deviation from maximal atmospheric mixing and non-maximal Dirac CP violation can be naturally obtained from the so-called generalized $\mu-\tau$ reflection [@Chen:2015siy].
Recently the flavor symmetry has been extended to combine with the generalized CP symmetry [@Feruglio:2012cw; @Holthausen:2012dk]. This can lead to rather predictive scenario where both mixing angles and CP phases determined by a small number of (frequently only one) input parameters [@Feruglio:2012cw]. In this case, the CP transformation matrix is generally non-diagonal and it is also called generalized CP. The generalized CP symmetry and the corresponding constraints on quark mass matrices have been exploited about thirty year ago [@Ecker:1981wv; @Grimus:1995zi]. In this case the interplay between CP and flavor symmetries has to be carefully treated in order to make the theory consistent [@Feruglio:2012cw; @Holthausen:2012dk; @Chen:2014tpa]. There have been some models and model independent analysis of CP and flavor symmetries, such as $A_{4}$ [@Ding:2013bpa], $S_{4}$ [@Feruglio:2012cw; @Ding:2013hpa; @Li:2014eia; @Feruglio:2013hia; @Luhn:2013vna; @Li:2013jya], $\Delta(27)$ [@Branco:2015gna], $\Delta(48)$ [@Ding:2013nsa], $A_{5}$ [@Li:2015jxa; @DiIura:2015kfa; @Ballett:2015wia], $\Delta(96)$ [@Ding:2014ssa], and the group series $\Delta(3n^{2})$ [@Hagedorn:2014wha; @Ding:2015rwa] and $\Delta(6n^{2})$ [@King:2014rwa; @Hagedorn:2014wha; @Ding:2014ora] for general integer $n$. It is notable that smaller group for instance $A_4$ [@Ding:2013bpa], $S_4$ [@Feruglio:2012cw; @Ding:2013hpa; @Li:2014eia; @Feruglio:2013hia; @Luhn:2013vna; @Li:2013jya] and $A_5$ [@Li:2015jxa; @DiIura:2015kfa; @Ballett:2015wia] can already describe the experimentally measured values of the mixing angles, and the Dirac CP phase is predicted to be conserved or maximal while the Majorana phases are trivial. On the other hand, all the three CP violating phases generally depend on the free real parameter $\theta$ for $\Delta(3n^{2})$ [@Hagedorn:2014wha; @Ding:2015rwa] and $\Delta(6n^{2})$ [@King:2014rwa; @Hagedorn:2014wha; @Ding:2014ora] flavor symmetries.
In the present work, we shall thoroughly analyze the lepton mixing patterns which can be obtained from the breaking of $D^{(1)}_{9n, 3n}$ flavor symmetry and generalized CP. All possible residual symmetries in the “direct”, “semidirect” and “variant of semidirect” approaches and their consequences for the prediction of the mixing parameters are studied. We shall perform a detailed numerical analysis for all the possible mixing patterns. The admissible values of the mixing parameters for each $n$ and the possible values of the effective mass $|m_{ee}|$ will be explored.
The outline of this paper is as follows. In section \[sec:GCP\] we find the class-inverting automorphism of the $D^{(1)}_{9n,3n}$ group and the corresponding physically well-defined generalized CP transformations are determined by solving the consistency condition. In section \[sec:framework\] we review the approach to determining the lepton flavor mixing from residual flavor and CP symmetries of the neutrino and the charged lepton sectors. All possible residual symmetries and the consequences for the prediction of the flavor mixing are studied in the method of the direct approach in section \[sec:direct\_approach\]. The PMNS matrix is determined to be of the trimaximal form, both Dirac phase $\delta_{CP}$ and the Majorana phase $\alpha_{31}$ are conserved, and the values of $\alpha_{21}$ are integer multiple of $2\pi/(3n)$. We investigate the possible mixing patterns which can be derived from the semidirect approach and variant of semidirect approach in section \[sec:Z2xCP\_neutrino\] and section \[sec:Z2xCP\_charged\_lepton\]. The analytical expressions of the PMNS matrices, mixing angles and CP invariants are presented, the admissible values of the mixing angles and CP violation phases are analyzed numerically in detail, and phenomenological predictions for neutrinoless double beta decay are studied. For the lowest order $D^{(1)}_{9n ,3n}$ group with $n=1, 2$, we find all the mixing patterns that can describe the experimentally measured values of the mixing angles, and a $\chi^2$ analysis is performed. Finally we summarize and present our conclusions in section \[sec:Conclusion\]. The group theory of $D^{(1)}_{9n,3n}$ is presented in Appendix \[app:group\_theory\] including the conjugacy classes, the irreducible representations, the character table, the Kronecker products and the Clebsch-Gordan coefficients.
\[sec:GCP\]Generalized CP consistent with $D^{(1)}_{9n,3n}$ family symmetry
===========================================================================
The finite subgroups of $SU(3)$ have been systematically classified by mathematicians [@su(3)_subgroups_book] (see Refs. [@Grimus:2010ak; @Grimus:2011fk; @Grimus:2013apa] for recent work). It is well-established that all discrete subgroups of $SU(3)$ can be divided into five categories: type A, type B, type C, type D, and type E [@Grimus:2011fk; @Grimus:2013apa]. The type D group turns out to be particularly significant in flavor symmetry theory [@Fonseca:2014koa; @King:2013vna]. Type D group is isomorphic to $(Z_{m}\times Z_{n})\rtimes S_3$, and it can be generated by four generators $a$, $b$, $c$ and $d$ subject to the following rules [@Grimus:2013apa]: $$\begin{aligned}
\nonumber&&a^3=b^2=(ab)^2=c^m=d^n=1,\quad cd=dc,\\
\nonumber&&\quad~ aca^{-1}=c^kd,\quad ada^{-1}=c^{-m/n}d^{-(k+1)},\\ \quad \label{eq:group_relation}&&\qquad~~ bcb^{-1}=cd,\quad bdb^{-1}=d^{-1}\,.\end{aligned}$$ It is found that the type D group exists only for [@Grimus:2013apa] $$k=0, m=n ~~~~~ \text{or}~~~~~ k=1, m=3n\,.$$ In the case of $k=0$, $m=n$, the corresponding group denoted as $D^{(0)}_{n, n}$ is exactly the well-known $\Delta(6n^2)$ group [@Escobar:2008vc]. For another case of $k=1$, $m=3n$, the corresponding type D group denoted as $D^{1)}_{3n, n}$ is isomorphic to $Z_3\times\Delta(6n^2)$ if $n$ is not divisible by 3 [@Grimus:2013apa]. Therefore the representation of $D^{(1)}_{3n, n}$ can be obtained by multiplying the representation matrices of $\Delta(6n^2)$ with 1, $e^{2\pi i/3}$ and $e^{4\pi i/4}$ for $3 \nmid n$. As a consequence, the $D^{(1)}_{3n, n}$ group for $3 \nmid n$ would give rise to the same set of lepton flavor mixing as $\Delta(6n^2)$ group no matter whether the generalized CP symmetry is considered or not. The $\Delta(6n^2)$ as flavor symmetry group has been comprehensively explored in the literature [@King:2013vna; @Ding:2014ora; @Hagedorn:2014wha], we shall focus on the second independent type D infinite series of groups $D^{(1)}_{9n,3n}$ where $n$ is any positive integer. It is remarkable that $D^{(1)}_{9n,3n}$ can generate experimentally viable lepton and quark mixing simultaneously [@Yao:2015dwa]. In the present work, we shall include the generalized CP symmetry compatible with $D^{(1)}_{9n,3n}$ and investigate its predictions for lepton mixing angles and CP violating phases. The group theory of $D^{(1)}_{9n,3n}$, its irreducible representations and the Clebsch-Gordan coefficients are presented in Appendix \[app:group\_theory\].
$n$ [$G_f$]{} `GAP-Id` `Inn(G_f)` `Out(G_f)`
----- ------------------- -------------- ----------------------------------------------------------------- ------------ --
1 $D^{(1)}_{9,3}$ \[162,14\] $\left(\left(Z_3\times Z_3\right)\rtimes Z_3\right)\rtimes Z_2$ $Z_6$
2 $D^{(1)}_{18, 6}$ \[648,259\] $\left(\left(Z_6\times Z_6\right)\rtimes Z_3\right)\rtimes Z_2$ $Z_6$
3 $D^{(1)}_{27, 9}$ \[1458,659\] $\left(\left(Z_9\times Z_9\right)\rtimes Z_3\right)\rtimes Z_2$ $Z_{18}$
: \[tab:automorphism\]The automorphism groups of the $D^{(1)}_{9n, 3n}$ group with $n=1, 2, 3$, where `Inn(G_f)` and `Out(G_f)` denote inner automorphism group and outer automorphism group of `G_f` respectively. Note that each of these three groups has a unique class-inverting outer automorphism.
It is highly nontrivial to introduce the generalized CP symmetry in the presence of a discrete flavor symmetry $G_{f}$. In order to consistently combine the generalized CP symmetry with flavor symmetry, the following consistency condition has to be fulfilled [@Feruglio:2012cw; @Holthausen:2012dk; @Chen:2014tpa], $$\label{eq:consistency_condition}X_{\mathbf{r}}\rho^{*}_{\mathbf{r}}(g)X^{\dagger}_{\mathbf{r}}=\rho_{\mathbf{r}}(g^{\prime}),\quad
g,g^{\prime}\in G_{f}\,,$$ where $\rho_{\mathbf{r}}(g)$ is the representation matrix of the element $g$ in the irreducible representation $\mathbf{r}$ of $G_{f}$, and $X_{\mathbf{r}}$ is the generalized CP transformation. Obviously the CP transformation $X_{\mathbf{r}}$ maps $g$ into another group element $g^{\prime}$. Therefore the generalized CP symmetry corresponds the automorphism group of $G_{f}$. Moreover, it was shown that the physically well-defined CP transformations should be given by class-inverting automorphism of $G_f$ [@Chen:2014tpa]. We have exploited the computer algebra system `GAP` [@GAP4.4] to calculate the automorphism group of the first three $D^{(1)}_{9n,3n}$ groups with $n=1, 2, 3$, the results are listed in table \[tab:automorphism\]. Notice that larger $D^{(1)}_{9n,3n}$ group for $n\geq4$ is not stored in `GAP` at present. We see that the automorphism group of $D^{(1)}_{9n,3n}$ is quite complex but each one of $D^{(1)}_{9, 3}$, $D^{(1)}_{18, 6}$ and $D^{(1)}_{27, 9}$ has a unique class-inverting outer automorphism. Furthermore, we find a generic class-inverting automorphism $u$ of the $D^{(1)}_{9n,3n}$ group, and its actions on the generators $a$, $b$, $c$, $d$ are as follows $$\label{eq:physical_aut}a\stackrel{u}{\longmapsto}a,\quad
b\stackrel{u}{\longmapsto}b,\quad c\stackrel{u}{\longmapsto}c^{-1},\quad
d\stackrel{u}{\longmapsto}d^{-1}\,.$$ It is easy to check that $u$ indeed maps each element into the class of its inverse element for any value of the parameter $n$. We denote the physical CP transformation corresponding to the automorphism $u$ as $X_{\mathbf{r}}(u)$, and its explicit form is determined by the following consistency equations: $$\begin{aligned}
\nonumber&&X_{\mathbf{r}}\left(u\right)\rho^{*}_{\mathbf{r}}(a)X^{\dagger}_{\mathbf{r}}\left(u\right)=\rho_{\mathbf{r}}\left(u\left(a\right)\right)=\rho_{\mathbf{r}}(a)\,,\\
\nonumber&&X_{\mathbf{r}}\left(u\right)\rho^{*}_{\mathbf{r}}(b)X^{\dagger}_{\mathbf{r}}\left(u\right)=\rho_{\mathbf{r}}\left(u\left(b\right)\right)=\rho_{\mathbf{r}}\left(b\right)\,,\\
\nonumber&&X_{\mathbf{r}}\left(u\right)\rho^{*}_{\mathbf{r}}(c)X^{\dagger}_{\mathbf{r}}\left(u\right)=\rho_{\mathbf{r}}\left(u\left(c\right)\right)=\rho_{\mathbf{r}}\left(c^{-1}\right)\,,\\
\label{eq:conmsistency_equations}&&X_{\mathbf{r}}\left(u\right)\rho^{*}_{\mathbf{r}}(d)X^{\dagger}_{\mathbf{r}}\left(u\right)=\rho_{\mathbf{r}}\left(u\left(d\right)\right)=\rho_{\mathbf{r}}\left(d^{-1}\right)\,.\end{aligned}$$ In our working basis shown in Appendix \[app:group\_theory\], the representation matrices of $a$ and $b$ are real while the representation matrices of $c$ and $d$ are complex and diagonal for any irreducible representations of $D^{(1)}_{9n, 3n}$. Therefore the CP transformation $X_{\mathbf{r}}(u)$ is a unit matrix, i.e. $$\label{eq:gcp_trans}
X_{\mathbf{r}}\left(u\right)=\mathbb{1}_{\bf{r}}\,.$$ Given this CP transformation $X_{\mathbf{r}}(u)$, the matrix $\rho_{\mathbf{r}}(g)X_{\mathbf{r}}(u)=\rho_{\mathbf{r}}(g)$ is also an admissible CP transformation for any $g\in D^{(1)}_{9n, 3n}$. It corresponds to performing a conventional CP transformation followed by a group transformation $\rho_{\mathbf{r}}(g)$. As a consequence, we conclude that the generalized CP transformation compatible with the $D^{(1)}_{9n, 3n}$ family symmetry is of the same form as the flavor symmetry transformation in our basis, i.e. $$\label{eq:GCP_full}
X_{\mathbf{r}}=\rho_{\mathbf{r}}(g),\quad g\in D^{(1)}_{9n,3n}\,.$$ Note that other possible CP transformations can also be defined if a model contains only a subset of irreducible representations. Lepton mixing can be derived from the remnant symmetries in the charged lepton and neutrino mass matrices, while the mechanism of symmetry breaking is irrelevant. The basic procedure and the resulting master formulae are given in Refs. [@Ding:2013bpa; @Ding:2013hpa; @Ding:2014ora; @Chen:2014wxa; @Chen:2015nha]. In the following, we shall consider all possible remnant symmetries of the neutrino and charged lepton sectors and discuss the predictions for the PMNS matrix and the lepton mixing parameters.
\[sec:framework\]Framework
==========================
In the present work, the family symmetry is taken to be $D^{(1)}_{9n, 3n}$, and the generalized CP symmetry is considered in order to predict the lepton mixing parameters including the CP violating phases. Without loss of generality, we assume that the three left-handed leptons transform as a triplet $\mathbf{3}_{1, 0}$ under $D^{(1)}_{9n, 3n}$. For brevity we shall denote the faithful irreducible representation $\mathbf{3}_{1, 0}$ as $\mathbf{3}$. The representation matrices of the generators $a$, $b$, $c$ and $d$ in $\mathbf{3}_{1, 0}$ are given in Eq. . The light neutrinos are assumed to be Majorana particles. From the bottom-up perspective, the most general symmetry of a generic charged lepton mass matrices is $U(1)\times U(1)\times U(1)$, which has finite subgroups isomorphic to a cyclic group $Z_m$ for any integer $m$ or a direct product of several cyclic groups [@Chen:2014wxa; @Chen:2015nha; @Yao:2015dwa]. On the other hand, the largest possible symmetry of the neutrino mass matrix is $Z_2\times Z_2$ [@Chen:2014wxa; @Chen:2015nha; @Yao:2015dwa; @Lam:2007qc]. Moreover the neutrino and charged lepton mass matrices are invariant under a set of CP transformations, and both the $U(1)\times U(1)\times U(1)$ symmetry group of the charged-lepton mass term and the $Z_2\times Z_2$ symmetry of the neutrino mass terms can be generated by performing two CP symmetry transformations [@Chen:2014wxa; @Chen:2015nha]. Conversely, the lepton mass matrices are strongly constrained by the postulated remnant symmetry such that the lepton mixing matrix can be derived from the remnant symmetries in the charged lepton and neutrino sectors, while the mechanism of dynamically realizing the assumed remnant symmetries is irrelevant [@Chen:2014wxa; @Chen:2015nha]. From the view of the top-down method, the remnant flavor and CP symmetries of the neutrino and charged lepton mass matrices may originate from certain symmetry group implemented at high energy scales. In the present work, both flavor symmetry $D^{(1)}_{9n, 3n}$ and the generalized CP are imposed, i.e., the parent symmetry is $D^{(1)}_{9n, 3n}\rtimes H_{CP}$, where $H_{CP}$ denotes the generalized CP transformations consistent with $D^{(1)}_{9n, 3n}$ and it is given by Eq. . $D^{(1)}_{9n, 3n}\rtimes H_{CP}$ is assumed to be broken down into $G_{l}\rtimes H^{l}_{CP}$ and $G_{\nu}\times H^{\nu}_{CP}$ in the charged lepton and neutrino sectors respectively. The allowed forms of the neutrino and charged lepton mass matrices are constrained by the remnant symmetries, and subsequently we can diagonalize them to get the PMNS matrix.
The requirement that a subgroup $G_{l}\rtimes H^{l}_{CP}$ is preserved at low energies entails that the combination $m^{\dagger}_{l}m_{l}$ has to fulfill $$\begin{aligned}
\nonumber&&\rho^{\dagger}_{\mathbf{3}}(g_{l})m^{\dagger}_{l}m_{l}\rho_{\mathbf{3}}(g_{l})=m^{\dagger}_{l}m_{l},\quad g_{l}\in G_{l},\\
\label{eq:constr_clep}&&X^{\dagger}_{l\mathbf{3}}m^{\dagger}_{l}m_{l}X_{l\mathbf{3}}=\left(m^{\dagger}_{l}m_{l}\right)^{\ast},\quad X_{l\mathbf{3}}\in H^{l}_{CP}\,,\end{aligned}$$ where the charged lepton mass matrix $m_l$ is given in the convention $l^{c}m_{l}l$. The hermitian combination $m^{\dagger}_{l}m_{l}$ is diagonalized by the unitary transformation $U_{l}$ with $U^{\dagger}_{l}m^{\dagger}_{l}m_{l}U_{l}=\text{diag}(m^2_{e}, m^2_{\mu}, m^2_{\tau})$. The three charged leptons have distinct masses $m_{e}\neq m_{\mu}\neq m_{\tau}$. From Eq. , it is straightforward to derive that the remnant symmetry $G_{l}\rtimes H^{l}_{CP}$ leads to the following constraints on $U_{l}$ $$\begin{aligned}
\nonumber&&U^{\dagger}_{l}\rho_{\mathbf{3}}(g_{l})U_{l}=\rho^{diag}_{\mathbf{3}}(g_{l}),\quad g_{l}\in G_{l},\\
\label{eq:constr_Ul}&&U^{\dagger}_{l}X_{l\mathbf{3}}U^{\ast}_{l}=X^{diag}_{l\mathbf{3}},\quad X_{l\mathbf{3}}\in H^{l}_{CP}\,,\end{aligned}$$ where both $\rho^{diag}_{\mathbf{3}}(g_{l})$ and $X^{diag}_{l\mathbf{3}}$ are diagonal phase matrices. As a consequence, we see that $U_{l}$ also diagonalizes the residual flavor symmetry transformation matrix $\rho_{\mathbf{3}}(g_{l})$, the residual CP transformation $X_{l\mathbf{3}}$ is a symmetric matrix, and the following restricted consistency condition should be satisfied [@Li:2014eia], $$\label{eq:res_cons_cond_clep}X_{l\mathbf{r}}\rho^{\ast}_{\mathbf{r}}(g_{l})X^{-1}_{l\mathbf{r}}=\rho_{\mathbf{r}}(g^{-1}_{l}),\qquad
g_{l}\in G_{l},~X_{l\mathbf{r}}\in H^{l}_{CP} \,.$$ In the same fashion, the neutrino mass matrix is invariant under the action of the elements of the residual subgroup $G_{\nu}\times H^{\nu}_{CP}$: $$\begin{aligned}
\nonumber&&\rho^{T}_{\mathbf{3}}(g_{\nu})m_{\nu}\rho_{\mathbf{3}}(g_{\nu})=m_{\nu},\qquad g_{\nu}\in G_{\nu},\\
\label{eq:constr_nu}&&X^{T}_{\nu\mathbf{3}}m_{\nu}X_{\nu\mathbf{3}}=m^{*}_{\nu},\qquad
\quad X_{\nu\mathbf{3}}\in H^{\nu}_{CP}\,.\end{aligned}$$ We denote the unitary diagonalization matrix of $m_{\nu}$ as $U_{\nu}$ fulfilling $U^{T}_{\nu}m_{\nu}U_{\nu}=\text{diag}\left(m_1, m_2, m_3\right)$. Then $U_{\nu}$ would be subject to the following constraints from the postulated residual symmetry [@Chen:2014wxa; @Chen:2015nha; @Everett:2015oka]: $$\begin{aligned}
\nonumber&&U^{\dagger}_{\nu}\rho_{\mathbf{3}}(g_{\nu})U_{\nu}=\textrm{diag}(\pm1, \pm1, \pm1),\\
\label{eq:constr_Unu}&&U^{\dagger}_{\nu}X_{\nu\mathbf{3}}U^{*}_{\nu}=\text{diag}(\pm1, \pm1, \pm1)\,,\end{aligned}$$ where the “$\pm$” signs can be chosen independently. Therefore the residual CP transformation $X_{\nu\mathbf{3}}$ is a symmetric unitary matrix as well, and the restricted consistency condition on the neutrino sector takes the form [@Chen:2014wxa; @Chen:2015nha; @Everett:2015oka; @Feruglio:2012cw]: $$\label{eq:res_cons_cond_nu}X_{\nu\mathbf{r}}\rho^{*}_{\mathbf{r}}(g_{\nu})X^{-1}_{\nu\mathbf{r}}=\rho_{\mathbf{r}}(g_{\nu}),\qquad
g_{\nu}\in G_{\nu},~X_{\nu\mathbf{r}}\in H^{\nu}_{CP}\,.$$ Obviously $X_{\nu\mathbf{r}}$ maps any element $g_{\nu}$ of the neutrino residual flavor symmetry $G_{\nu}$ into itself. Hence the mathematical structure of the remnant subgroup comprising $G_{\nu}$ and $H^{\nu}_{CP}$ is generally a direct product instead of a semidirect product. Given a pair of well-defined remnant symmetries $G_{l}\rtimes H^{l}_{CP}$ and $G_{\nu}\times H^{\nu}_{CP}$ for which the consistency equations in Eqs. (\[eq:res\_cons\_cond\_clep\], \[eq:res\_cons\_cond\_nu\]) are fulfilled, the allowed forms of the mass matrices $m^{\dagger}_{l}m_{l}$ and $m_{\nu}$ can be determined from Eqs. (\[eq:constr\_clep\], \[eq:constr\_nu\]), and subsequently the prediction for the PMNS matrix $U_{\text{PMNS}}=U^{\dagger}_{l}U_{\nu}$ can be obtained by diagonalizing $m^{\dagger}_{l}m_{l}$ and $m_{\nu}$.
For two pair of remnant symmetry subgroups $\left\{G_{l}\rtimes H^{l}_{CP}, G_{\nu}\times H^{\nu}_{CP}\right\}$ and $\big\{G^{\prime}_{l}\rtimes H^{l^{\prime}}_{CP}$, $G^{\prime}_{\nu}\times H^{\nu^{\prime}}_{CP}\big\}$, if $G_{l}$, $G_{\nu}$ and $G^{\prime}_{l}$, $G^{\prime}_{\nu}$ are related by a similarity transformation, for example if they are conjugate, $$G^{\prime}_{l}=hG_{l}h^{-1},\qquad G^{\prime}_{\nu}=hG_{\nu}h^{-1},\qquad h\in D^{(1)}_{9n, 3n}\,.$$ The remnant CP would also be related by $$\label{eq:CP_conju}H^{l^{\prime}}_{CP}=\rho_{\mathbf{r}}(h)H^{l}_{CP}\rho^{T}_{\mathbf{r}}(h)
,\qquad H^{\nu^{\prime}}_{CP}=\rho_{\mathbf{r}}(h)H^{\nu}_{CP}\rho^{T}_{\mathbf{r}}(h)$$ in order to fulfill the consistency conditions in Eqs. (\[eq:res\_cons\_cond\_clep\], \[eq:res\_cons\_cond\_nu\]). That is to say the elements of $H^{l^{\prime}}_{CP}$ and $H^{\nu^{\prime}}_{CP}$ are given by $\rho_{\mathbf{r}}(h)X_{l\mathbf{r}}\rho^{T}_{\mathbf{r}}(h)$ and $\rho_{\mathbf{r}}(h)X_{\nu\mathbf{r}}\rho^{T}_{\mathbf{r}}(h)$ respectively, where $X_{\nu\mathbf{r}}\in H^{\nu}_{CP}$ and $X_{l\mathbf{r}}\in H^{l}_{CP}$. Notice that all the possible remnant CP transformations compatible with the remnant flavor symmetry have been considered in this work. Hence if $G_{l}\rtimes H^{l}_{CP}$ and $G_{\nu}\times H^{\nu}_{CP}$ fix the charged lepton and neutrino mass matrices to be $m^{\dagger}_{l}m_{l}$ and $m_{\nu}$, then $m^{\prime\dagger}_{l}m^{\prime}_{l}\equiv\rho_{\mathbf{3}}(h)m^{\dagger}_{l}m_{l}\rho^\dagger_{\mathbf{3}}(h)$ and $m^{\prime}_{\nu}\equiv\rho^{\ast}_{\mathbf{3}}(h)m_{\nu}\rho^{\dagger}_{\mathbf{3}}(h)$ would be invariant under the remnant symmetries $G^{\prime}_{l}\rtimes H^{l^{\prime}}_{CP}$ and $G^{\prime}_{\nu}\times H^{\nu^{\prime}}_{CP}$ respectively. As a result, two pair of remnant symmetries $\left\{G_{l}\rtimes H^{l}_{CP}, G_{\nu}\times H^{\nu}_{CP}\right\}$ and $\big\{G^{\prime}_{l}\rtimes H^{l^{\prime}}_{CP}$, $G^{\prime}_{\nu}\times H^{\nu^{\prime}}_{CP}\big\}$ would yield the same results for the PMNS matrix $U_{\text{PMNS}}$. In this work, we shall perform a comprehensive analysis of the mixing patterns which can be derived from the group $D^{(1)}_{9n, 3n}\rtimes H_{CP}$. It is sufficient to only analyze a few representative remnant symmetries which give rise to different results for $U_{\text{PMNS}}$ and lepton mixing parameters, as other possible choices for the remnant symmetry groups are related to the representative ones by similarity transformation and consequently no new results are obtained.
\[sec:direct\_approach\]Lepton mixing from direct approach
==========================================================
In the direct approach, the residual flavor symmetry $G_{\nu}$ is a Klein four subgroup, and the residual flavor symmetry $G_{l}$ is a cyclic group $Z_{m}$ with index $m\geq3$ or a product of cyclic groups. We assume that the residual flavor symmetry group $G_{l}$ can distinguish the three generations of charged lepton. In other words, the restricted representation of the triplet representation $\mathbf{3}$ on $G_{l}$ should decompose into three inequivalent 1-dimensional representations of $G_{l}$. From Eq. and Eq. , we see that $U_{l}$ not only diagonalizes the mass matrix $m^{\dagger}_{l}m_{l}$ but also the residual flavor symmetry transformation matrix $\rho_{\mathbf{3}}(g_{l})$ with $g_{l}\in G_{l}$. As a result, the requirement that $U^{\dagger}_{l}\rho_{\mathbf{3}}(g_{l})U_{l}=\rho^{diag}_{\mathbf{3}}(g_{l})$ is diagonal allows us to determine $U_{l}$ and no knowledge of $m^{\dagger}_{l}m_{l}$ is necessary. Notice that the remnant CP invariant condition in Eq. is automatically satisfied, the reason is that the residual CP transformation $X_{l\mathbf{3}}$ has to be compatible with residual flavor symmetry and its allowed form is strongly constrained by the restricted consistency condition of Eq. .
As shown in the Appendix \[app:group\_theory\], the group structure of the $D^{(1)}_{9n, 3n}$ has been studied in detail. The residual subgroup $G_{l}$ is an abelian subgroup, and it can be generated by the generator $c^{s}d^{t}$, $bc^{s}d^{t}$, $ac^{s}d^{t}$, $a^2c^{s}d^{t}$, $abc^{s}d^{t}$ or $a^2bc^{s}d^{t}$ with $s=0,1,\dots,9n-1,t=0,1,\dots,3n-1$. The diagonalization of $\rho_{\mathbf{3}}(g_{l})$ determines the unitary transformation $U_{l}$ up to permutations and phases of the column vectors if $\rho_{\mathbf{3}}(g_{l})$ has non-degenerate eigenvalues, where $g_{l}$ can be taken to be the generator of $G_{l}$. The explicit form of $U_{l}$ for different $G_{l}$ and the corresponding remnant CP transformations compatible with $G_{l}$ are summarized in table \[tab:cle\_diagonal\_matrix\]. If the eigenvalues of $\rho_{\mathbf{3}}(g_{l})$ are degenerate so that its diagonalization matrix $U_{l}$ can not be determined uniquely, we would extend $G_{l}$ from a single cyclic subgroup to a product of cyclic groups, for example $G_{l}=G_{1}\times G_{2}$ where the generators of $G_1$ and $G_2$ should be commutable with each other. If $G_1$ (or $G_2$) is sufficient to distinguish among the generations such that its eigenvalues are not degenerate, then another subgroup $G_2$ (or $G_1$) would not impose any new constraint on the lepton mixing. On the other hand, if the three eigenvalues of the generator of either $G_1$ or $G_2$ are completely degenerate, e.g. $G_1 (~\text{or}~G_2)=\langle c^{3n}\rangle$, its three-dimensional representation matrix would be proportional to a unit matrix. As a result, we shall concentrate on the case that the representation matrices of both $G_1$ and $G_2$ have two degenerate eigenvalues, therefore either $G_1$ or $G_2$ alone fixes only a column of $U_{l}$ and the third column can be determined by unitary condition. The possible extension of remnant flavor symmetry group $G_{l}$, the corresponding remnant CP transformations and the unitary transformations $U_{l}$ are collected in table \[tab:extension\_Gch\]. We see that the diagonalization matrix $U_{l}$ can only take five distinct forms $U_l^{(1)}$, $U_l^{(2)}$, $U_l^{(3)}$, $U_l^{(4)}$ or $U_l^{(5)}$ such that the constraints on $s$ and $t$ shown in table \[tab:cle\_diagonal\_matrix\] are relaxed.
[|c|c|c|c|]{} & & &\
\[-0.16in\] $G_{l}$ & $U_{l}$ & `Constraints` & $H^{l}_{CP}$\
& & &\
\[-0.16in\] & & &\
\[-0.16in\]
& & $t\neq0$ &\
$\langle c^{s}d^{t}\rangle$ & $U^{(1)}_l=\begin{pmatrix}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{pmatrix}$ & $s-t\neq0$ $\text{mod}(3n)$ & $\{c^{\gamma}d^{\delta}\}$\
& & $s-2t\neq0$ $\text{mod}(3n)$ &\
& & &\
\[-0.16in\] & & &\
\[-0.10in\]
& & & $\{c^{2t-s+2\delta+3n\tau}d^{\delta},$\
\[-0.25in\]
$\langle bc^{s}d^{t}\rangle$ & $U^{(2)}_l=\frac{1}{\sqrt{2}}
\begin{pmatrix}
\sqrt{2} &~ 0 ~& 0 \\
0 &~ -e^{\frac{i\pi(2t-s)}{3 n}} ~& e^{\frac{i\pi(2t-s)}{3 n}} \\
0 &~ 1 ~& 1 \\
\end{pmatrix}$ & $s\neq0,3n,6n$ &\
& & & $bc^{2\delta+3n\tau}d^{\delta}\}$\
& & &\
\[-0.10in\] & & &\
\[-0.16in\]
& & & $\{bc^{-2t+3n\tau}d^{-t},$\
$\langle ac^{s}d^{t}\rangle$ & $U^{(3)}_l=\frac{1}{\sqrt{3}}
\begin{pmatrix}
e^{-\frac{2 i \pi s}{9 n}} &~ \omega^2 e^{-\frac{2 i \pi s}{9 n}} ~& \omega e^{-\frac{2 i \pi s}{9 n}} \\
e^{\frac{2 i \pi (3t-2 s)}{9 n}} &~ \omega e^{\frac{2 i \pi (3t-2 s)}{9 n}} ~& \omega^2 e^{\frac{2 i \pi (3t-2 s)}{9 n}} \\
1 &~ 1 ~& 1 \\
\end{pmatrix}$ & — & $abc^{s-2t+3n\tau}d^{s-2t},$\
& & & $a^{2}bc^{t-s+3n\tau}\}$\
& & &\
\[-0.16in\] & & &\
\[-0.16in\]
& & & $\{bc^{2(t-s)+3n\tau}d^{t-s},$\
$\langle a^2c^{s}d^{t}\rangle$ & $U^{(3')}_l=\frac{1}{\sqrt{3}}\left(
\begin{array}{ccc}
e^{\frac{2 i \pi (3t-2s)}{9 n}} &~ \omega e^{\frac{2 i \pi (3t-2s)}{9 n}} ~& \omega^2 e^{\frac{2 i \pi (3t-2s)}{9 n}} \\
e^{\frac{2 i \pi (3t-s)}{9 n}} &~ \omega^2 e^{\frac{2 i \pi (3t-s)}{9 n}} ~& \omega e^{\frac{2 i \pi (3t-s)}{9 n}} \\
1 &~ 1 ~& 1 \\
\end{array}
\right)$ & — & $abc^{-t+3n\tau}d^{-t},$\
& & & $a^{2}bc^{2t-s+3n\tau}\}$\
& & &\
\[-0.16in\] & & &\
\[-0.16in\]
$\langle abc^{s}d^{t}\rangle$ & $U^{(4)}_l=\frac{1}{\sqrt{2}}
\begin{pmatrix}
e^{\frac{i \pi (t-s)}{3 n}} & 0 & -e^{\frac{i \pi (t-s)}{3 n}} \\
0 & \sqrt{2} & 0 \\
1 & 0 & 1 \\
\end{pmatrix}$ & $s-3t\neq0,3n,6n$ & $\{c^{\gamma}d^{\gamma+s-t},abc^{\gamma}d^{\gamma}\}$\
& & &\
\[-0.16in\] & & &\
\[-0.16in\]
$\langle a^2bc^{s}d^{t}\rangle$ & $U^{(5)}_l=\frac{1}{\sqrt{2}}
\begin{pmatrix}
-e^{-\frac{i \pi t}{3 n}} &~ e^{-\frac{i \pi t}{3 n}} ~& 0 \\
1 &~ 1 ~& 0 \\
0 &~ 0 ~& \sqrt{2} \\
\end{pmatrix}$ & $2s-3t\neq0,3n,6n$ & $\{c^{\gamma}d^{-t},a^{2}bc^{\gamma}\}$\
& & &\
\[-0.16in\]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\mathcal{G}_1$ $\mathcal{G}_2$ `Constraints on group parameters` `Form of` $U_{l}$ $H^{l}_{CP}$
------------------------------------------ --------------------------------------------------- ------------------------------------------------------------------------- ----------------------------- --------------------------------------------------------------------------------------------------------------
$\langle c^sd^t\rangle$ $\langle c^{s^{\prime}}d^{t^{\prime}}\rangle$ $\begin{array}{l} $U^{(1)}_{l}$ $\{c^\gamma d^\delta \}$
\left\{\begin{array}{c}
s-2t=0\pmod{3n}\\
s^{\prime}-t^{\prime}=0\pmod{3n}
\end{array}\right.\\[0.15in]
\hskip-0.16in\text{or} \left\{\begin{array}{c}
s-2t=0\pmod{3n}\\
t^{\prime}=0\pmod{3n}
\end{array}\right.\\[0.15in]
\hskip-0.16in\text{or}
\left\{\begin{array}{c}
s-t=0\pmod{3n}\\
t^{\prime}=0\pmod{3n}
\end{array}\right.\\[0.1in]
\qquad\quad(s\leftrightarrow s^{\prime},~~ t\leftrightarrow t^{\prime})
\end{array}$
\[-12pt\][$\langle bc^sd^t\rangle$]{} $\langle c^{s^{\prime}}d^{t^{\prime}}\rangle$ $\begin{array}{c} \[-12pt\][ $U^{(2)}_{l}$]{} \[-12pt\][$\begin{array}{l} \{c^{2t+2\delta+3n\tau}d^\delta,\\ bc^{2\delta+3n\tau}d^\delta\} \end{array}$]{}
s^{\prime}-2t^{\prime}=0\pmod{3n}\\
s=0\pmod{3n}
\end{array}$
$\langle bc^{s^{\prime}}d^{t^{\prime}}\rangle$ $\begin{array}{c}
(s-s^{\prime})-2(t-t^{\prime})=3l_1n\pmod{6n}\\
s=3l_2n\pmod{6n},\quad s^{\prime}=3l_3n\pmod{6n}
\end{array}$
\[-22pt\][$\langle abc^sd^t\rangle$]{} $\langle c^{s^{\prime}}d^{t^{\prime}}\rangle$ $\begin{array}{c} \[-22pt\][ $U^{(4)}_{l}$]{} \[-12pt\][$\begin{array}{l} \{c^{\gamma}d^{\gamma+2t},\\ abc^{\gamma}d^\gamma\} \end{array}$]{}
s^{\prime}-t^{\prime}=0\pmod{3n}\\
3t-s=0\pmod{3n}
\end{array}$
$\langle abc^{s^{\prime}}d^{t^{\prime}}\rangle$ $\begin{array}{c}
(s-s^{\prime})-(t-t^{\prime})=3l_1n\pmod{6n}\\
3t-s=3l_2n\pmod{6n}\\
3t^{\prime}-s^{\prime}=3l_3n\pmod{6n}
\end{array}$
\[-22pt\][$\langle a^2bc^sd^t\rangle$]{} $\langle c^{s^{\prime}}d^{t^{\prime}}\rangle$ $\begin{array}{c} \[-22pt\][ $U^{(5)}_{l}$]{} \[-12pt\][$\begin{array}{l} \{c^{\gamma}d^{-t},\\ a^2bc^{\gamma}\} \end{array}$]{}
t^{\prime}=0\pmod{3n}\\
2s-3t=0\pmod{3n}
\end{array}$
$\langle a^2bc^{s^{\prime}}d^{t^{\prime}}\rangle$ $\begin{array}{c}
t-t^{\prime}=3l_1n\pmod{6n}\\
2s-3t=3l_2n\pmod{6n}\\
2s^{\prime}-3t^{\prime}=3l_3n\pmod{6n}
\end{array}$
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[tab:extension\_Gch\]The product extension of the remnant flavor symmetry $G_{l}=G_1\times G_2$, the remnant CP transformation compatible with $G_{l}$, and the corresponding unitary transformation $U_{l}$. We require the column vectors fixed by $\mathcal{G}_1$ and $\mathcal{G}_2$ be different. Consequently we have the parameters $l_{1,2,3}=0,1$ and $l_1+l_2+l_3=1,3$. The values of parameters $s$, $t$, $s^\prime$, $t^\prime$, $\gamma$, $\delta$ and $\tau$ are $s, s^{\prime},\gamma=0,1,\cdots,9n-1$, $t,t^\prime,\delta=0,1,\cdots,3n-1$ and $\tau=0, 1, 2$.
[|c|c|]{} &\
\[-0.16in\] $G_{\nu}$ & $X_{\nu}$\
&\
\[-0.16in\] &\
\[-0.16in\]
$K^{(c^{9n/2},d^{3n/2})}_4$ & $\rho_{\mathbf{r}}(c^{\gamma}d^{\delta})$\
&\
\[-0.16in\] &\
\[-0.16in\]
$K^{(d^{3n/2},bd^{x})}_4$ & $\rho_{\mathbf{r}}(c^{2\delta+2x+3n\tau}d^{\delta}),
\rho_{\mathbf{r}}(bc^{2\delta+3n\tau}d^{\delta})$\
&\
\[-0.16in\] &\
\[-0.16in\]
$K^{(c^{9n/2}d^{3n/2},abc^{3y}d^{y})}_4$ & $\rho_{\mathbf{r}}(c^{\delta-2y-3n\tau}d^{\delta}),\rho_{\mathbf{r}}(abc^{\delta-3n\tau}d^{\delta})$\
&\
\[-0.16in\] &\
\[-0.16in\]
$K^{(c^{9n/2},a^2bc^{3z}d^{2z})}_4$ & $\rho_{\mathbf{r}}(c^{\gamma}d^{-2z}),\rho_{\mathbf{r}}(a^2bc^{\gamma})$\
In the direct approach, the flavor symmetry group $D^{(1)}_{9n, 3n}$ group is broken down to a Klein four subgroup in the neutrino sector. From Appendix \[app:group\_theory\], we see that $D^{(1)}_{9n, 3n}$ for even $n$ has only four Klein four subgroups: $$\begin{aligned}
\nonumber&&K^{(c^{9n/2},d^{3n/2})}_4\equiv\left\{1, c^{9n/2}, d^{3n/2}, c^{9n/2}d^{3n/2}\right\},\quad K^{(d^{3n/2},bd^{x})}_4\equiv\left\{1, d^{3n/2}, bd^{x}, bd^{x+3n/2} \right\},\\
\nonumber&&K^{(c^{9n/2}d^{3n/2}, abc^{3y}d^{y})}_4\equiv\left\{1, c^{9n/2}d^{3n/2}, abc^{3y}d^{y}, abc^{3y+9n/2}d^{y+3n/2}\right\},\\
\label{eq:K4_conjugate}&&K^{(c^{9n/2}, a^2bc^{3z}d^{2z})}_4\equiv\left\{1, c^{9n/2}, a^2bc^{3z}d^{2z}, a^2bc^{3z+9n/2}d^{2z}\right\}\,,\end{aligned}$$ where $x, y, z=0, 1, \ldots, 3n-1$. We note that $K^{(c^{9n/2},d^{3n/2})}_4$ is a normal subgroup of $D^{(1)}_{9n, 3n}$, and the remaining three $K_4$ subgroups are conjugate: $$\begin{aligned}
\nonumber&&(a^{2}c^{y-x+2\delta}d^{\delta})K^{(d^{3n/2},bd^{x})}_4(a^{2}c^{y-x+2\delta}d^{\delta})^{-1}=K^{(c^{9n/2}d^{3n/2}, abc^{3y}d^{y})}_4,\\
&&(ac^{-z-x+2\delta}d^{\delta})K^{(d^{3n/2},bd^{x})}_4(ac^{-z-x+2\delta}d^{\delta})^{-1}=K^{(c^{9n/2}, a^2bc^{3z}d^{2z})}_4\,,\end{aligned}$$ with $\delta=0, 1, \ldots, 3n-1$. Furthermore, the residual CP symmetry $H^{\nu}_{CP}$ in the neutrino sector has to be compatible with the remnant $K_4$ symmetry, and the following restricted consistency condition must be fulfilled, $$X_{\nu\mathbf{r}}\rho^{\ast}_{\mathbf{r}}(g)X^{-1}_{\nu\mathbf{r}}=\rho_{\mathbf{r}}(g),\qquad
g\in K_4\,.$$ Solving this equation, we can straightforwardly find the eligible remnant CP transformations for different $K_4$ subgroups. The results are collected in table \[tab:K4\_RCP\]. Then we proceed to determine the neutrino mass matrix $m_{\nu}$ invariant under the action of both remnant CP and remnant flavor symmetry for each case, i.e., $m_{\nu}$ is subject to the constraints in Eq. .
- [$G_{\nu}=K^{(c^{9n/2}, d^{3n/2})}_4$, $X_{\nu\mathbf{r}}=\rho_{\mathbf{r}}(c^{\gamma}d^{\delta})$]{}
In our working basis, the representation matrices for both $a$ and $c$ are diagonal with $$\rho_{\mathbf{3}}(c^{9n/2})=\left(\begin{array}{ccc}
-1 & 0 & 0 \\
0 & -1 & 0 \\
0 & 0 & 1
\end{array}\right),\quad \rho_{\mathbf{3}}(d^{3n/2})=\left(\begin{array}{ccc}
1 & 0 & 0 \\
0 & -1 & 0 \\
0 & 0 & -1
\end{array}\right)\,.$$ Consequently the residual flavor symmetry enforces the neutrino mass matrix to be diagonal as well. Taking into account the remnant CP symmetry further, we find $$m_{\nu}=\begin{pmatrix}
m_{11} e^{-2i\pi\frac{\gamma}{9n}} & 0 & 0 \\
0 & m_{22}e^{-2i\pi\frac{\gamma-3\delta}{9n}} & 0 \\
0 & 0 & m_{33}e^{2i\pi\frac{2\gamma-3\delta}{9n}}
\end{pmatrix}\,,$$ where $m_{11}$, $m_{22}$ and $m_{33}$ are real parameters. We can read out the neutrino diagonalization matrix $U_{\nu}$ as $$\label{eq:unu_K41}
U_{\nu}=\text{diag}\left(e^{i\pi\frac{\gamma}{9n}},e^{i\pi\frac{\gamma-3\delta}{9n}}, e^{-i\pi\frac{2\gamma-3\delta}{9 n}}\right)Q_{\nu}\,,$$ where $Q_{\nu}$ is a diagonal phase matrix with entry being $\pm1$ or $\pm i$, and it encodes the CP parity of the neutrino states. The light neutrino mass eigenvalues are $$m_1=\left|m_{11}\right|,\qquad m_2=\left|m_{22}\right|,\qquad
m_3=\left|m_{33}\right|\,.$$ Obviously the light neutrino masses depend on only three real parameters, and the order of the light neutrino masses can not be fixed by remnant symmetries. Therefore the unitary transformation $U_{\nu}$ is determined up to independent row and column permutations in the present framework, and the neutrino mass spectrum can be wither normal ordering (NO) or inverted ordering (IO).
- [$G_{\nu}=K^{(d^{3n/2},bd^{x})}_4$, $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{2\delta+2x+3n\tau}d^{\delta}),\rho_{\mathbf{r}}(bc^{2\delta+3n\tau}d^{\delta})\right\}$]{}
In the same fashion as previous case, we find that the light neutrino mass matrix takes the following form: $$m_{\nu}=\begin{pmatrix}
m_{11}e^{-2i\pi\frac{2x+2\delta+3n\tau}{9n}} & 0 &~ 0 \\
0 & m_{22}e^{-2i\pi\frac{2x-\delta+3n\tau}{9 n}} &~ m_{23}e^{i\pi\frac{2x+2\delta+3n\tau}{9n}} \\
0 & m_{23}e^{i\pi\frac{2x+2\delta+3n\tau}{9n}} &~ m_{22}e^{2i\pi\frac{4x+\delta-3n\tau}{9 n}} \\
\end{pmatrix}\,,$$ where $m_{11}$, $m_{22}$ and $m_{23}$ are real. It is diagonalized by the unitary transformation $U_{\nu}$ with $$\label{eq:unu_abcy_2}
U_{\nu}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
\sqrt{2}e^{i\pi\frac{2x+2\delta+3n\tau}{9n}} & 0 &~ 0 \\
0 & e^{i\pi\frac{2x-\delta+3n\tau}{9n}} &~ -e^{i\pi\frac{2x-\delta+3n\tau}{9n}} \\
0 & e^{-i\pi\frac{4x+\delta-3n\tau}{9n}} &~ e^{-i\pi\frac{4x+\delta-3n\tau}{9n}}
\end{pmatrix}\,,$$ where the matrix $Q_{\nu}$ is omitted for simplicity and we will also not explicitly write out this factor hereafter. The light neutrino masses are $$m_1=\left|m_{11}\right|,\qquad
m_2=\left|m_{22}+(-1)^{\tau}m_{23}\right|,\qquad m_3=\left|m_{22}-(-1)^{\tau}m_{23}\right|\,.$$
- [$G_{\nu}=K^{(c^{9n/2}d^{3n/2},abc^{3y}d^{y})}_4$, $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{\delta-2y-3n\tau}d^{\delta}),\rho_{\mathbf{r}}(abc^{\delta-3n\tau}d^{\delta})\right\}$]{}
In this case, we find that the light neutrino mass matrix takes the form $$m_{\nu}=\begin{pmatrix}
m_{11}e^{2i\pi\frac{2y-\delta+3n\tau}{9n}} & 0 & m_{13}e^{-i\pi\frac{2y+2\delta+3n\tau}{9n}} \\
0 & m_{22}e^{2i\pi\frac{2y+2\delta+3n\tau}{9n}} & 0 \\
m_{13}e^{-i\pi\frac{2y+2\delta+3n\tau}{9n}} & 0 & m_{11}e^{-2i\pi\frac{4y+\delta+6n\tau}{9n}}
\end{pmatrix}\,,$$ where $m_{11}$, $m_{13}$ and $m_{22}$ are real. Consequently the unitary transformation $U_{\nu}$ is $$U_{\nu}=\frac{1}{\sqrt{2}}\left(
\begin{array}{ccc}
-e^{i\pi\frac{-2y+\delta+6 n \tau }{9 n}} & 0 & e^{i \pi \frac{-2 y+\delta +6 n \tau }{9 n}} \\
0 & \sqrt{2} e^{-i\pi\frac{2y+2\delta+3n\tau}{9 n}} & 0 \\
e^{i\pi\frac{4y+\delta+6n\tau}{9n}} & 0 & e^{i \pi \frac{4y+\delta+6n\tau}{9n}}
\end{array}
\right)\,.$$ The light neutrino masses are $$m_1=\left|m_{11}-(-1)^{\tau}m_{13}\right|,\qquad m_2=\left|m_{22}\right|,\qquad
m_3=\left|m_{11}+(-1)^{\tau}m_{13}\right|\,.$$
- [$G_{\nu}=K^{(c^{9n/2},a^2 b c^{3z}d^{2z})}_4$, $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{\gamma}d^{-2z}),\rho_{\mathbf{r}}(a^2bc^{\gamma})\right\}$]{}
The light neutrino mass matrix $m_{\nu}$ is constrained by the remnant symmetry to be of the form $$\label{eq:mnu_k4_4}m_{\nu}=\left(
\begin{array}{ccc}
m_{11} e^{-2i \pi \frac{ \gamma }{9 n}} &~ m_{12} e^{-2i \pi\frac{3z+\gamma}{9n}} & 0 \\
m_{12} e^{-2i\pi\frac{3z+\gamma}{9n}} &~ m_{11} e^{-2i\pi\frac{6z+\gamma}{9n}} & 0 \\
0 &~ 0 & m_{33}e^{4i\pi\frac{3z+\gamma}{9n}}
\end{array}
\right)\,,$$ where $m_{11}$, $m_{12}$ and $m_{33}$ are real. The unitary matrix $U_{\nu}$ diagonalizing the above neutrino mass matrix is determined to be $$\label{eq:unu_k4_4}
U_{\nu}=\frac{1}{\sqrt{2}}
\left(\begin{array}{ccc}
-e^{i\pi\frac{\gamma}{9n}} &~ e^{i\pi\frac{\gamma}{9 n}} & 0 \\
e^{i\pi\frac{6z+\gamma}{9n}} &~ e^{i\pi\frac{6z+\gamma}{9n}} & 0 \\
0 & 0 & \sqrt{2} e^{-2 i \pi \frac{ 3 z+\gamma }{9 n}}
\end{array}
\right)\,.$$ The neutrino masses are given by $$m_1=\left|m_{11}-m_{12}\right|,\qquad m_2=\left|m_{11}+m_{22}\right|,\qquad
m_3=\left|m_{33}\right|\,.$$
Then we proceed to discuss the possible mixing patterns achievable in direct approach by combining the different remnant symmetries of the charged lepton sector with those of the neutrino sector. As shown in section \[sec:framework\], two pairs of subgroups $\left\{G_{l}, G_{\nu}\right\}$ and $\left\{G^{\prime}_{l}, G^{\prime}_{\nu}\right\}$ would yield the same results for the PMNS matrix after considering all the eligible residual CP transformations, if these two pairs of groups are conjugate. Notice the conjugate relations between distinct $K_{4}$ subgroups in Eq. and the identities $(bc^{2\epsilon}d^{2\epsilon})(abc^sd^t)(bc^{2\epsilon}d^{2\epsilon})^{-1}=a^2bc^sd^{s-t}$, $(bc^{2\epsilon}d^{2\epsilon})K^{(c^{9n/2}, d^{3n/2})}_{4}(bc^{2\epsilon}d^{2\epsilon})^{-1}=K^{(c^{9n/2}, d^{3n/2})}_{4}$ and $(bc^{2\epsilon}d^{2\epsilon})K^{(d^{3n/2},bd^x)}_{4}(bc^{2\epsilon}d^{2\epsilon})^{-1}=K^{(d^{3n/2},bd^x)}_{4}$ for any integer $\epsilon$, we find it is sufficient to only consider eight kinds of remnant symmetries with $G_{l}=\langle c^{s}d^{t}\rangle$, $\langle bc^{s}d^{t}\rangle$, $\langle ac^{s}d^{t}\rangle$, $\langle abc^{s}d^{t}\rangle$ and $G_{\nu}=K^{(c^{9n/2},d^{3n/2})}_4$, $K^{(d^{3n/2},bd^x)}_4$. In this scenarios, all mixing parameters including Majorana phases are completely fixed by remnant symmetries.
(1)
: [$G_{l}=\langle c^{s}d^{t}\rangle$, $G_{\nu}=K^{(c^{9n/2},d^{3n/2})}_4$, $X_{\nu\bf{r}}=\{\rho_{\bf{r}}(c^{\gamma}d^{\delta})\}$]{}\
In this case, the unitary transformation $U_{l}$ is a unit matrix, as shown in table \[tab:cle\_diagonal\_matrix\]. $U_{\nu}$ is a diagonal phase matrix and it is given by Eq. . As a result, the PMNS matrix is also a diagonal matrix up to row and column permutations, and obviously it doesn’t agree with the present neutrino oscillation data [@Capozzi:2013csa; @Forero:2014bxa; @Gonzalez-Garcia:2014bfa].
(2)
: [$G_{l}=\langle bc^{s}d^{t}\rangle$, $G_{\nu}=K^{(c^{9n/2},d^{3n/2})}_4$, $X_{\nu\bf{r}}=\{\rho_{\bf{r}}(c^{\gamma}d^{\delta})\}$]{}\
In this case, the postulated residual subgroups lead to the mixing pattern $$U_{\text{PMNS}}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
\sqrt{2} &~ 0 ~& 0 \\
0 &~ 1 ~& -e^{i \varphi_{1}} \\
0 &~ 1 ~& e^{i \varphi_{1}}
\end{pmatrix}\,,$$ with $$\varphi_{1}=-\frac{\pi(\gamma-2\delta+s-2t)}{3n}\,.$$ The lepton mixing angles are $\theta_{13}=\theta_{12}=0$, $\theta_{23}=45^\circ$, and therefore large corrections to both $\theta_{12}$ and $\theta_{13}$ are necessary in order to be compatible with experimental data.
(3)
: [$G_{l}=\langle ac^{s}d^{t}\rangle$, $G_{\nu}=K^{(c^{9n/2},d^{3n/2})}_4$, $X_{\nu\bf{r}}=\{\rho_{\bf{r}}(c^{\gamma}d^{\delta})\}$]{}\
This residual symmetry allows us to pin down the lepton mixing matrix as: $$U_{\text{PMNS}}=\frac{1}{\sqrt{3}}
\begin{pmatrix}
e^{i\varphi_1} &~ 1 ~& e^{i\varphi_2} \\
\omega e^{i\varphi_1} &~ 1 ~& \omega^2e^{i\varphi_2} \\
\omega^2e^{i\varphi_1} &~ 1 ~& \omega e^{i\varphi_2}
\end{pmatrix}\,,$$ where $$\varphi_{1}=\frac{\pi(3\gamma-3\delta+2 s)}{9n}, \qquad \varphi_{2}=\frac{\pi (3 \gamma -6 \delta +4 s-6 t)}{9 n}\,.$$ This pattern leads to $\sin^2\theta_{12}=\sin^2\theta_{23}=1/2$, $\sin^2\theta_{13}=1/3$ and a maximal Dirac CP phase $|\delta_{CP}|=\pi/2$. The solar as well as the reactor mixing angle have to acquire appropriate corrections in order to be in accordance with experimental data.
(4)
: [$G_{l}=\langle abc^{s}d^{t}\rangle$, $G_{\nu}=K^{(c^{9n/2},d^{3n/2})}_4$, $X_{\nu\bf{r}}=\{\rho_{\bf{r}}(c^{\gamma}d^{\delta})\}$]{}\
In this case we find the lepton mixing matrix is $$U_{\text{PMNS}}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
\sqrt{2} &~ 0 ~& 0 \\
0 &~ 1 ~& e^{i\varphi_1} \\
0 &~ -1 ~& e^{i\varphi_1}
\end{pmatrix}\,, \quad \mathrm{with} \quad \varphi_{1}=-\frac{\pi(\gamma-\delta+s-t)}{3n}\,,$$ which leads to $\theta_{12}=\theta_{13}=0$, $\theta_{23}=45^\circ$. Large corrections to $\theta_{12}$ and $\theta_{13}$ are needed to be compatible with experimental data.
(5)
: [$G_{l}=\langle c^{s}d^{t}\rangle$, $G_{\nu}=K^{(d^{3n/2},bd^x)}_4$, $X_{\nu\bf{r}}=\{\rho_{\bf{r}}(c^{2\delta+2x+3n\tau}d^{\delta}),\rho_{\bf{r}}(bc^{2\delta+3n\tau}d^{\delta})\}$]{}\
The unitary transformation $U_{\nu}$ is fixed by residual subgroup to be Eq. , and the PMNS matrix takes the form $$U_{\text{PMNS}}=
\begin{pmatrix}
1 &~ 0 ~& 0 \\
0 &~ \frac{1}{\sqrt{2}} ~& -\frac{1}{\sqrt{2}} \\
0 &~ \frac{1}{\sqrt{2}} ~& \frac{1}{\sqrt{2}}
\end{pmatrix}\,,$$ which lead to $\theta_{12}=\theta_{13}=0$, $\theta_{23}=45^\circ$. Again $\theta_{12}$ and $\theta_{13}$ require large corrections in order to be in the experimentally preferred range.
(6)
: [$G_{l}=\langle bc^{s}d^{t}\rangle$, $G_{\nu}=K^{(d^{3n/2},bd^x)}_4$, $X_{\nu\bf{r}}=\{\rho_{\bf{r}}(c^{2\delta+2x+3n\tau}d^{\delta}),\rho_{\bf{r}}(bc^{2\delta+3n\tau}d^{\delta})\}$]{}\
Using these residual symmetries, we can derive the lepton mixing matrix $$U_{\text{PMNS}}=
\begin{pmatrix}
1 &~ 0 ~& 0 \\
0 &~ -\sin \varphi_1 ~& \cos \varphi_1 \\
0 &~ \cos \varphi_1 ~& \sin\varphi_1
\end{pmatrix}\,, \quad \mathrm{with} \quad \varphi_1=\frac{\pi(s-2t+2x)}{6n}\,.$$ The mixing angles are $\theta_{12}=\theta_{13}=0$, $\sin^2\theta_{23}=\cos^2\varphi_1$ which is strongly disfavored by the experimental data [@Capozzi:2013csa; @Forero:2014bxa; @Gonzalez-Garcia:2014bfa].
(7)
: [$G_{l}=\langle ac^{s}d^{t}\rangle$, $G_{\nu}=K^{(d^{3n/2},bd^x)}_4$, $X_{\nu\bf{r}}=\{\rho_{\bf{r}}(c^{2\delta+2x+3n\tau}d^{\delta}),\rho_{\bf{r}}(bc^{2\delta+3n\tau}d^{\delta})\}$]{}\
In this case, the lepton mixing matrix is determined to be of the trimaximal form, i.e., the second column of the PMNS matrix is $(1, 1, 1)^{T}/\sqrt{3}$ with $$\label{eq:acsdt_K4}
U_{\text{PMNS}}=\frac{1}{\sqrt{3}}
\begin{pmatrix}
-\sqrt{2} e^{i\varphi_2}\cos\varphi_1 &~ 1 ~& \sqrt{2}e^{i\varphi_2}\sin\varphi_1 \\
\sqrt{2} e^{i\varphi_2} \sin \left(\frac{\pi }{6}+\varphi_1\right) &~ 1 ~& \sqrt{2} e^{i\varphi_2} \cos \left(\frac{\pi }{6}+\varphi_1\right) \\
\sqrt{2} e^{i\varphi_2} \sin \left(\frac{\pi }{6}-\varphi_1\right) &~ 1 ~& -\sqrt{2} e^{i\varphi_2} \cos \left(\frac{\pi }{6}-\varphi_1\right) \\
\end{pmatrix}\,,$$ where $$\label{eq:varphi_1_2_values}\varphi_1=\frac{2s-3t+3x}{9n}\pi,\qquad
\varphi_2=-\frac{\delta+t+x}{3n}\pi\,.$$ These two parameters $\varphi_1$ and $\varphi_2$ are independent from each other, and they can take the following discrete values $$\begin{aligned}
\nonumber&&\varphi_1~(\textrm{mod}~2\pi)=0, \frac{1}{9n}\pi, \frac{2}{9n}\pi, \ldots, \frac{18n-1}{9n}\pi\,,\\
&&\varphi_2~(\textrm{mod}~2\pi)=0, \frac{1}{3n}\pi, \frac{2}{3n}\pi, \ldots, \frac{6n-1}{3n}\pi\,.\end{aligned}$$ We can read out the mixing angles as $$\begin{aligned}
\nonumber && \sin^2\theta_{13}=\frac{2}{3}\sin^2\varphi_1, \quad
\sin^2\theta_{12}=\frac{1}{2+\cos2\varphi_{1}},\quad
\sin^2\theta_{23}=\frac{1+\cos(\frac{\pi}{6}+2\varphi_1)}{2+\cos2\varphi_1}\,.\end{aligned}$$ All possible predictions of $\sin\theta_{13}$ for each $D^{(1)}_{9n, 3n}$ of even $n$ are displayed in figure \[fig:theta13\_alpha21\]. It is remarkable that viable reactor mixing angle $\theta_{13}$ can always be achieved for each $n$. Moreover, the three mixing angles are closely related as follows $$3\sin^2\theta_{12}\cos^2\theta_{13}=1,\quad \sin^2\theta_{23}=\frac{1}{2}\pm\frac{1}{2}\tan\theta_{13}\sqrt{2-\tan^2\theta_{13}}\,.$$ Inputting the experimentally preferred $3\sigma$ range $0.0176\leq\sin^2\theta_{13}\leq0.0295$ [@Capozzi:2013csa], we obtain predictions for solar as well as atmospheric mixing angles: $$0.339\leq\sin^2\theta_{12}\leq0.343,\qquad 0.378\leq\sin^2\theta_{23}\leq0.406,~~\mathrm{or}~~0.594\leq\sin^2\theta_{23}\leq0.622\,.$$ From the PMNS matrix of Eq. , we can also extract the CP violating phases $$\begin{aligned}
\sin\delta_{CP}=\sin\alpha_{31}=0, \quad \tan\alpha_{21}=-\tan2\varphi_2\,,\end{aligned}$$ where the contribution of the CP parity matrix $Q_{\nu}$ is considered. We see that both Dirac phase $\delta_{CP}$ and the Majorana phase $\alpha_{31}$ are trivial, and another Majorana phase $\alpha_{21}$ is $$\alpha_{21}=-2\varphi_{2}~~\mathrm{or}~~ \alpha_{21}=\pi-2\varphi_{2}\,.$$ The admissible values of $\alpha_{21}$ are $$\alpha_{21}=0, \frac{2}{3n}\pi, \frac{4}{3n}\pi, \ldots, \frac{6n-2}{3n}\pi\,,$$ which are plotted in figure \[fig:theta13\_alpha21\]. Note that here the predictions for the CP phases are consistent with the general results of Ref. [@Chen:2015nha].
(8)
: [$G_{l}=\langle abc^{s}d^{t}\rangle$, $G_{\nu}=K^{(d^{3n/2},bd^x)}_4$, $X_{\nu\bf{r}}=\{\rho_{\bf{r}}(c^{2\delta+2x+3n\tau}d^{\delta}),\rho_{\bf{r}}(bc^{2\delta+3n\tau}d^{\delta})\}$]{}\
In this case, we find the lepton mixing matrix is the well-known bimaximal pattern $$U_{\text{PMNS}}=\frac{1}{2}
\begin{pmatrix}
-\sqrt{2} &~ \sqrt{2} ~& 0 \\
1 &~ 1 ~& \sqrt{2} e^{i\varphi_1} \\
1 &~ 1 ~& -\sqrt{2} e^{i\varphi_1}
\end{pmatrix}\,,
\quad \mathrm{with} \quad \varphi_1=\frac{\pi (s-2 t+2 x)}{6 n}\,.$$ The bimaximal mixing can be a valid first approximation in a model where corrections of order of the Cabibbo angle can naturally arise [@Li:2014eia; @Altarelli:2009gn].
\[sec:Z2xCP\_neutrino\]Lepton mixing from semidirect approach
=============================================================
In the semidirect approach, the original symmetry $D^{(1)}_{9n, 3n}\rtimes H_{CP}$ is broken at low energies into $G_{l}\rtimes H^{l}_{CP}$ in the charged lepton sector and to $Z_2\times H^{\nu}_{CP}$ in the neutrino sector. The PMNS matrix turns out to depend on only a single real parameter in this scenario. It is generally assumed that the residual flavor symmetry $G_{l}$ is able to distinguish the three generations of charged leptons such that the unitary matrix $U_{l}$ can be determined from the requirement that all the generators of $G_{l}$ should be simultaneously diagonalized by $U_{l}$. The possible candidates for the subgroup $G_{l}$, the remnant CP transformations compatible with $G_{l}$ and the corresponding unitary transformation $U_{l}$ are summarized in table \[tab:cle\_diagonal\_matrix\] and table \[tab:extension\_Gch\]. Then we turn to the neutrino sector. From the multiplication rules given in Eq. , we see that the order 2 elements of the $D^{(1)}_{9n, 3n}$ group are $$\label{eq:z2_1} bd^{x},\quad abc^{3y}d^{y}, \quad a^{2}bc^{3z}d^{2z},\quad x, y, z=0,1\ldots 3n-1\,,$$ and additionally $$\label{eq:z2_2}c^{9n/2},\quad d^{3n/2},\quad c^{9n/2}d^{3n/2}\,,$$ for even $n$. The residual CP transformation $X_{\nu\mathbf{r}}$ is a symmetric unitary matrix, and it should map the element of the neutrino residual flavor symmetry to itself, $$\label{eq:consistency_neutrino}
X_{\nu\mathbf{r}}\rho^{*}_{\mathbf{r}}(g_{\nu})X^{-1}_{\nu\mathbf{r}}=\rho_{\mathbf{r}}(g_{\nu}),\quad
g_{\nu}\in G_{\nu}\,.$$ The eligible residual CP transformations for different $Z_2$ subgroups are collected in table \[tab:Z2\_RCP\]. Furthermore, we notice that all the $Z_2$ elements in Eq. are conjugate: $$\label{eq:neutrino_conjugate_1}
\begin{array}{ll}
\left(c^{\gamma}d^{\delta}\right)bd^{x}\left(c^{\gamma}d^{\delta}\right)^{-1}=bd^{x^{\prime}},
~&~ \left(bc^{\gamma}d^{\delta}\right)bd^{x}\left(bc^{\gamma}d^{\delta}\right)^{-1}=bd^{-x^{\prime}}\,,\\
\left(ac^{\gamma}d^{\delta}\right)bd^{x}\left(ac^{\gamma}d^{\delta}\right)^{-1}=a^2bc^{-3 x^{\prime}}d^{-2 x^{\prime}},
~&~ \left(a^2c^{\gamma}d^{\delta}\right)bd^{x}\left(a^2c^{\gamma}d^{\delta}\right)^{-1}=abc^{3 x^{\prime}}d^{x^{\prime}}\,, \\
\left(abc^{\gamma}d^{\delta}\right)bd^{x}\left(abc^{\gamma}d^{\delta}\right)^{-1}=a^2bc^{3 x^{\prime}}d^{2 x^{\prime}},
~&~ \left(a^2bc^{\gamma}d^{\delta}\right)bd^{x}\left(a^2bc^{\gamma}d^{\delta}\right)^{-1}=abc^{-3 x^{\prime}}d^{-x^{\prime}}\,,
\end{array}$$ where $x^{\prime}=x+\gamma-2\delta$. Similarly the three elements in Eq. are also conjugate to each other: $$\label{eq:neutrino_conjugate_4}
\begin{array}{ll}
\left(c^{\gamma}d^{\delta}\right)c^{9n/2}\left(c^{\gamma}d^{\delta}\right)^{-1}=c^{9n/2},
~&~\left(bc^{\gamma}d^{\delta}\right)c^{9n/2}\left(bc^{\gamma}d^{\delta}\right)^{-1}=c^{9n/2}d^{3n/2}\,,\\
\left(ac^{\gamma}d^{\delta}\right)c^{9n/2}\left(ac^{\gamma}d^{\delta}\right)^{-1}=c^{9n/2}d^{3n/2},
~&~\left(a^2c^{\gamma}d^{\delta}\right)c^{9n/2}\left(a^2c^{\gamma}d^{\delta}\right)^{-1}=d^{3n/2}\,,\\
\left(abc^{\gamma}d^{\delta}\right)c^{n/2}\left(abc^{\gamma}d^{\delta}\right)^{-1}=d^{3n/2},
~&~\left(a^2bc^{\gamma}d^{\delta}\right)c^{9n/2}\left(a^2bc^{\gamma}d^{\delta}\right)^{-1}=c^{9n/2}\,.
\end{array}$$
-------------------------------------- ----------------------------------------------------------------------------------------------------------
\[-0.16in\] $G_{\nu}$ $X_{\nu}$
\[-0.16in\]
\[-0.16in\] $Z^{bd^x}_2$ $\rho_{\mathbf{r}}(c^{2\delta+2x+3n\tau}d^{\delta})$, $\rho_{\mathbf{r}}(bc^{2\delta+3n\tau}d^{\delta})$
\[-0.16in\]
\[-0.16in\] $Z^{abc^{3y}d^{y}}_2$ $\rho_{\mathbf{r}}(c^{\delta-2y-3n\tau}d^{\delta})$, $\rho_{\mathbf{r}}(abc^{\delta-3n\tau}d^{\delta})$
\[-0.16in\]
\[-0.16in\] $Z^{a^2bc^{3z}d^{2z}}_2$ $\rho_{\mathbf{r}}(c^{\gamma}d^{-2z})$, $\rho_{\mathbf{r}}(a^2bc^{\gamma})$
\[-0.16in\]
\[-0.16in\] $Z^{c^{9n/2}}_2$ $\rho_{\mathbf{r}}(c^{\gamma}d^{\delta})$, $\rho_{\mathbf{r}}(a^2bc^{\gamma})$
\[-0.16in\]
\[-0.16in\] $Z^{d^{3n/2}}_2$ $\rho_{\mathbf{r}}(c^{\gamma}d^{\delta})$, $\rho_{\mathbf{r}}(bc^{2\delta+3n\tau}d^{\delta})$
\[-0.16in\]
\[-0.16in\] $Z^{c^{9n/2}d^{3n/2}}_2$ $\rho_{\mathbf{r}}(c^{\gamma}d^{\delta})$, $\rho_{\mathbf{r}}(abc^{\delta+3n\tau}d^{\delta})$
-------------------------------------- ----------------------------------------------------------------------------------------------------------
: \[tab:Z2\_RCP\] Different types of remnant $Z_2$ subgroup $G_{\nu}$ and viable remnant CP transformations, where the superscript of the $Z_2$ subgroup denotes its generators. The allowed values of the parameters are $\gamma=0, 1, \ldots, 9n-1$, $x, y, z, \delta=0, 1, \ldots, 3n-1$, and $\tau=0, 1, 2$.
As a result, it is sufficient to consider the representative residual symmetry $G_{\nu}=Z^{bd^x}_2$, $Z^{c^{9n/2}}_2$ and $G_{l}=\left\langle c^{s}d^{t}\right\rangle$, $\left\langle
bc^{s}d^{t}\right\rangle$, $\left\langle ac^{s}d^{t}\right\rangle$, $\left\langle abc^{s}d^{t}\right\rangle$ and $\left\langle
a^2bc^{s}d^{t}\right\rangle$. Since only a $Z_2$ subgroup instead of a full Klein subgroup is preserved by the neutrino mass matrix, the postulated remnant flavor symmetries can only fix one column of the PMNS matrix. We list the explicit form of the determined columns for different remnant flavor symmetry in table \[tab:PMNS\_column\_Z2\]. Global analysis of the neutrino oscillation data gives the $3\sigma$ ranges on the absolute values of the elements of the PMNS matrix [@Capozzi:2013csa]: $$\label{eq:3sigma_ranges}||U_{\text{PMNS}}||=\left(
\begin{array}{ccc}
0.789\sim0.853 ~&~ 0.501\sim0.594 ~&~
0.133\sim0.172\\
0.194\sim0.558 ~&~ 0.408\sim0.735 ~&~
0.602\sim0.784 \\
0.194\sim0.558 ~&~ 0.408\sim0.735 ~&~
0.602\sim0.784 \\
\end{array}
\right)\,,$$ It is obvious that none entry of the PMNS matrix is vanishing [@Capozzi:2013csa; @Forero:2014bxa; @Gonzalez-Garcia:2014bfa]. Therefore if one element of the fixed column is predicted to be zero, it would be excluded by the experimental data. From table \[tab:PMNS\_column\_Z2\] we see that only three independent cases are viable with the residual flavor symmetries $(G_{\nu}, G_{l})=(Z^{bd^{x}}_{2}, \langle ac^{s}d^{t}\rangle)$, $(Z^{c^{9n/2}}_{2}, \langle ac^{s}d^{t}\rangle)$ and $(Z^{bd^{x}}_{2}, \langle abc^{s}d^{t}\rangle)$. In the following, the contribution of all admissible remnant CP transformations will be included further. We shall find the neutrino mass matrix invariant under the residual flavor and CP symmetries, and then the unitary transformation $U_{\nu}$ as well as the PMNS matrix $U_{\text{PMNS}}$ will be presented for each case.
[|c||c|c|]{} & &\
\[-0.15in\] & $G_{\nu}=Z^{bd^x}_2$ & $G_{\nu}=Z^{c^{9n/2}}_2$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_l=\langle c^{s}d^{t}\rangle$ & $\frac{1}{\sqrt{2}}\left(\begin{array}{c}
0 \\
-1\\
1
\end{array}\right)$& $\left(\begin{array}{c}
0\\
0\\
1
\end{array}\right)$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_l=\langle bc^{s}d^{t}\rangle$ & $\left(\begin{array}{c}
0\\
\cos\left(\frac{s-2t+2x}{6n}\pi\right)\\
\sin\left(\frac{s-2t+2x}{6n}\pi\right)
\end{array}
\right)$ & $\frac{1}{\sqrt{2}}\left(\begin{array}{c}
0\\
-1\\
1
\end{array}
\right)$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_l=\langle ac^{s}d^{t}\rangle$ & $\sqrt{\frac{2}{3}}\left(
\begin{array}{c}
\sin\left(\frac{2s-3t+3x}{9n}\pi\right) \\
\cos\left(\frac{\pi}{6}+\frac{2s-3t+3x}{9n}\pi\right) \\
\cos\left(\frac{\pi}{6}-\frac{2s-3t+3x}{9n}\pi\right)
\end{array}
\right)$ & $\frac{1}{\sqrt{3}}\left(\begin{array}{c}
1\\
1\\
1
\end{array}
\right)$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_l=\langle abc^{s}d^{t}\rangle$ & $\frac{1}{2}\left(\begin{array}{c}
1\\
1\\
\sqrt{2}
\end{array}
\right)$ & $\frac{1}{\sqrt{2}}\left(\begin{array}{c}
0\\
-1\\
1
\end{array}
\right)$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_l=\langle a^2bc^{s}d^{t}\rangle$ & $\frac{1}{2}\left(\begin{array}{c}
1\\
1\\
\sqrt{2}
\end{array}
\right)$ & $\left(\begin{array}{c}
0\\
0\\
1
\end{array}
\right)$\
& &\
\[-0.15in\]
()
: [ $G_{l}=\left\langle ac^{s}d^{t}\right\rangle$, $G_{\nu}=Z^{bd^x}_2$, $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{2\delta+2x+3n\tau}d^{\delta}),
\rho_{\mathbf{r}}(bc^{2\delta+2x+3n\tau}d^{\delta+x})\right\}$]{}\
The residual symmetry transformation $G_{\nu}\times H^{\nu}_{CP}$ of the neutrino fields leaves the neutrino mass term invariant. Therefore the neutrino mass matrix $m_{\nu}$ must satisfy $$\begin{aligned}
\nonumber&&\rho^{T}_{\mathbf{3}}(g_{\nu})m_{\nu}\rho_{\mathbf{3}}(g_{\nu})=m_{\nu},\qquad g_{\nu}\in G_{\nu},\\
\label{eq:constr_nu_again}&&X^{T}_{\nu\mathbf{3}}m_{\nu}X_{\nu\mathbf{3}}=m^{*}_{\nu},\qquad
\quad X_{\nu\mathbf{3}}\in H^{\nu}_{CP}\,.\end{aligned}$$ In our working basis, it is straightforward to find that the neutrino mass matrix is constrained to take the form $$\label{eq:mnu_I}m_{\nu}=
\begin{pmatrix}
e^{-\frac{4i\pi(x+\delta)}{9 n}}m_{11} &~ e^{-\frac{i\pi(4x+\delta)}{9n}}m_{12} &~ e^{\frac{i\pi(2 x-\delta)}{9n}}m_{12} \\
e^{-\frac{i\pi(4 x+\delta )}{9n}}m_{12} &~ e^{-\frac{2i\pi(2 x-\delta)}{9 n}}m_{22} &~ e^{\frac{2i\pi(x+\delta)}{9n}}m_{23} \\
e^{\frac{i\pi(2 x-\delta)}{9n}}m_{12} &~ e^{\frac{2i\pi(x+\delta)}{9 n}} m_{23} &~ e^{\frac{2i\pi(4x+\delta)}{9n}}m_{22}
\end{pmatrix}\,,$$ where $m_{11}$, $m_{12}$, $m_{22}$ and $m_{23}$ are real. It follows that the neutrino mass matrix $m_{\nu}$ can be diagonalized by $$U^{T}_{\nu}m_{\nu}U_{\nu}=\text{diag}\left(m_1,m_2,m_3\right)\,,$$ with the unitary transformation $$\label{eq:unu_bdx}
U_{\nu}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
\sqrt{2} e^{\frac{2 i \pi (x+\delta )}{9 n}} \cos \theta &~ 0 ~& \sqrt{2} e^{\frac{2 i \pi (x+\delta )}{9 n}} \sin\theta \\
-e^{\frac{i \pi (2 x-\delta )}{9 n}} \sin\theta &~ -e^{\frac{i \pi (2 x-\delta )}{9 n}} ~& e^{\frac{i \pi (2 x-\delta )}{9 n}} \cos\theta \\
-e^{-\frac{i \pi (4 x+\delta )}{9 n}} \sin\theta &~ e^{-\frac{i \pi (4 x+\delta )}{9 n}} ~& e^{-\frac{i \pi (4 x+\delta )}{9 n}} \cos\theta \\
\end{pmatrix}Q_{\nu}\,,$$ where the angle $\theta$ is $$\tan2\theta=\frac{2\sqrt{2}m_{12}}{m_{22}+m_{23}-m_{11}}\,.$$ The factor $Q_{\nu}$ is a diagonal phase matrix with elements equal to $\pm1$ and $\pm i$ and it is necessary to make the light neutrino masses positive definite. The neutrino mass eigenvalues are given by $$\begin{aligned}
\nonumber&&m_1=\frac{1}{2}\left|m_{11}+m_{22}+m_{23}-\frac{m_{22}+m_{23}-m_{11}}{\cos2\theta}\right|,\\
\nonumber&&m_2=\left|m_{22}-m_{23}\right|,\\
\nonumber&&m_3=\frac{1}{2}\left|m_{11}+m_{22}+m_{23}+\frac{m_{22}+m_{23}-m_{11}}{\cos2\theta}\right|\,.\end{aligned}$$ We see that the neutrino masses depend on four parameters $m_{11}$, $m_{12}$, $m_{22}$ and $m_{23}$, the experimentally measured mass squared differences could be easily accommodated. The order of the three neutrino masses $m_1$, $m_2$ and $m_3$ can not be pinned down in the present framework, hence the unitary matrix $U_{\nu}$ is determined up to permutations of the columns (the same holds true in the following cases), and the neutrino mass spectrum can be either normal ordering (NO) or inverted ordering (IO). Taking into account the corresponding charged lepton diagonalization matrix $U_{l}$ listed in table \[tab:cle\_diagonal\_matrix\] and table \[tab:extension\_Gch\], we find the PMNS matrix $U_{\text{PMNS}}\equiv U^{\dagger}_{l}U_{\nu}$ up to row and column permutations is [$$\begin{aligned}
U^{I}_{\text{PMNS}}=\frac{1}{\sqrt{3}}
\begin{pmatrix}
e^{i \varphi_{2}} \cos\theta-\sqrt{2} \cos\varphi_{1}\sin\theta &~ \sqrt{2} \sin\varphi_{1} ~& e^{i \varphi_{2}} \sin\theta+\sqrt{2} \cos\theta \cos\varphi_{1} \\
-e^{i \varphi_{2}}\cos\theta-\sqrt{2}\sin\theta\sin\left(\frac{\pi }{6}-\varphi_{1}\right) &~ \sqrt{2} \cos \left(\frac{\pi }{6}-\varphi_{1}\right) ~& -e^{i\varphi_{2}}\sin\theta+\sqrt{2}\cos\theta\sin\left(\frac{\pi }{6}-\varphi_{1}\right) \\
e^{i\varphi_{2}} \cos\theta+\sqrt{2} \sin\theta \sin \left(\frac{\pi }{6}+\varphi_{1}\right) &~ \sqrt{2} \cos \left(\frac{\pi }{6}+\varphi_{1}\right) ~& e^{i \varphi_{2}} \sin\theta-\sqrt{2} \cos\theta\sin \left(\frac{\pi }{6}+\varphi_{1}\right)
\end{pmatrix}Q_{\nu}\,,\end{aligned}$$]{} with $$\varphi_1=\frac{2s-3t+3x}{9n}\pi,\qquad
\varphi_2=\frac{\delta+t+x}{3n}\pi\,.$$ Both $\varphi_1$ and $\varphi_2$ are determined by the postulated remnant symmetries, they are independent of each other, and their values can be multiple of $\frac{\pi}{9n}$ and $\frac{\pi}{3n}$ respectively $$\begin{aligned}
\nonumber&&\varphi_1~(\mathrm{mod}~2\pi)=0, \frac{1}{9n}\pi, \frac{2}{9n}\pi, \ldots, \frac{18n-1}{9n}\pi\,,\\
\label{eq:para_values_caseI}&&\varphi_2~(\mathrm{mod}~2\pi)=0, \frac{1}{3n}\pi, \frac{2}{3n}\pi, \ldots, \frac{6n-1}{3n}\pi\,.\end{aligned}$$ We see that one column of the PMNS matrix is determined to be $$\sqrt{\frac{2}{3}}\left(\begin{array}{c}
\sin\varphi_1\\
\cos\left(\pi/6-\varphi_1\right)\\
\cos\left(\pi/6+\varphi_1\right)
\end{array}
\right)$$ in this case. As the neutrino mass ordering isn’t constrained in the present framework, this column vector can be any of the three column of the PMNS matrix. As a consequence, the PMNS matrix can take the following three possible forms: [$$\begin{aligned}
\nonumber U^{I,1}_{\text{PMNS}}&=&\frac{1}{\sqrt{3}}
\begin{pmatrix}
\sqrt{2} \sin\varphi_{1} &~ e^{i \varphi_{2}} \cos\theta-\sqrt{2} \cos\varphi_{1}\sin\theta &~ e^{i \varphi_{2}} \sin\theta+\sqrt{2} \cos\theta \cos\varphi_{1} \\
\sqrt{2} \cos \left(\frac{\pi }{6}-\varphi_{1}\right) & -e^{i \varphi_{2}}\cos\theta-\sqrt{2}\sin\theta\sin\left(\frac{\pi }{6}-\varphi_{1}\right) &~ -e^{i\varphi_{2}}\sin\theta+\sqrt{2}\cos\theta\sin\left(\frac{\pi }{6}-\varphi_{1}\right) \\
\sqrt{2} \cos \left(\frac{\pi }{6}+\varphi_{1}\right) & ~e^{i\varphi_{2}} \cos\theta+\sqrt{2} \sin\theta \sin \left(\frac{\pi }{6}+\varphi_{1}\right) &~ e^{i \varphi_{2}} \sin\theta-\sqrt{2} \cos\theta\sin \left(\frac{\pi }{6}+\varphi_{1}\right)
\end{pmatrix}\,,\\
\nonumber U^{I,2}_{\text{PMNS}}&=&\frac{1}{\sqrt{3}}
\begin{pmatrix}
e^{i \varphi_{2}} \cos\theta-\sqrt{2} \cos\varphi_{1}\sin\theta &~ \sqrt{2} \sin\varphi_{1} ~& e^{i \varphi_{2}} \sin\theta+\sqrt{2} \cos\theta \cos\varphi_{1} \\
-e^{i \varphi_{2}}\cos\theta-\sqrt{2}\sin\theta\sin\left(\frac{\pi }{6}-\varphi_{1}\right) &~ \sqrt{2} \cos \left(\frac{\pi }{6}-\varphi_{1}\right) ~& -e^{i\varphi_{2}}\sin\theta+\sqrt{2}\cos\theta\sin\left(\frac{\pi }{6}-\varphi_{1}\right) \\
e^{i\varphi_{2}} \cos\theta+\sqrt{2} \sin\theta \sin \left(\frac{\pi }{6}+\varphi_{1}\right) &~ \sqrt{2} \cos \left(\frac{\pi }{6}+\varphi_{1}\right) ~& e^{i \varphi_{2}} \sin\theta-\sqrt{2} \cos\theta\sin \left(\frac{\pi }{6}+\varphi_{1}\right)
\end{pmatrix}\,,\\
\nonumber U^{I,3}_{\text{PMNS}}&=&\frac{1}{\sqrt{3}}
\begin{pmatrix}
e^{i \varphi_{2}} \sin\theta+\sqrt{2} \cos\theta \cos\varphi_{1} &~ e^{i \varphi_{2}} \cos\theta-\sqrt{2} \cos\varphi_{1}\sin\theta &~ \sqrt{2} \sin\varphi_{1} \\
-e^{i\varphi_{2}}\sin\theta+\sqrt{2}\cos\theta\sin\left(\frac{\pi }{6}-\varphi_{1}\right) &~ -e^{i \varphi_{2}}\cos\theta-\sqrt{2}\sin\theta\sin\left(\frac{\pi }{6}-\varphi_{1}\right) &~ \sqrt{2} \cos \left(\frac{\pi }{6}-\varphi_{1}\right) \\
e^{i \varphi_{2}} \sin\theta-\sqrt{2} \cos\theta\sin \left(\frac{\pi }{6}+\varphi_{1}\right) &~ e^{i\varphi_{2}} \cos\theta+\sqrt{2} \sin\theta \sin \left(\frac{\pi }{6}+\varphi_{1}\right) &~ \sqrt{2} \cos \left(\frac{\pi }{6}+\varphi_{1}\right)
\end{pmatrix}\,.\label{eq:PMNS_caseI}\end{aligned}$$]{} The effect of row permutation is equivalent to redefinitions of the parameters $\theta$, $\varphi_1$ and $\varphi_2$, and no new possible values of $\varphi_1$ and $\varphi_2$ beyond those in Eq. are obtained. Then we proceed to discuss the phenomenological predictions of each mixing pattern. For $U^{I,1}_{\text{PMNS}}$ the three lepton mixing angles read as $$\begin{aligned}
\nonumber && \sin^{2}\theta_{13}=\frac{1}{3}\left(1+\cos^2\theta\cos2\varphi_1+\sqrt{2}\sin2\theta\cos\varphi_1\cos\varphi_2\right),\\
\nonumber&&\sin^2\theta_{12}=\frac{1+\sin^2\theta\cos2\varphi_1-\sqrt{2}\sin2\theta\cos\varphi_1\cos\varphi_2}
{2-\cos^2\theta\cos2\varphi_1-\sqrt{2}\sin2\theta\cos\varphi_1\cos\varphi_2},\\
\label{eq:mixing_para_caseI_1st}&&\sin^2\theta_{23}=\frac{1-\cos^2\theta\sin\left(\pi/6+2\varphi_1\right)-\sqrt{2}\sin2\theta\cos\varphi_2\sin\left(\pi/6-\varphi_1\right)}
{2-\cos^2\theta\cos2\varphi_1-\sqrt{2}\sin2\theta\cos\varphi_1\cos\varphi_2}\,,\end{aligned}$$ which yield the correlation $$3\cos^2\theta_{12}\cos^2\theta_{13}=2\sin^2\varphi_{1}\,,$$ In order to accommodate the experimentally favored $3\sigma$ ranges $0.259\leq\sin^2\theta_{12}\leq0.359$ and $0.0176\leq\sin^2\theta_{13}\leq0.0295$ from the global fit [@Capozzi:2013csa], we find the allowed region of the parameter $\varphi_1$ is $$\varphi_1\in\left[0.417\pi, 0.583\pi\right]\cup\left[1.417\pi, 1.583\pi\right]\,.$$ Obviously $\varphi_1$ should be around $\pi/2$ or $3\pi/2$. Furthermore, the three CP rephasing invariants $J_{CP}$, $I_1$ and $I_2$ are predicted to be $$\begin{aligned}
\nonumber && \left|J_{CP}\right|=\frac{1}{6\sqrt{6}}\left|\sin2\theta\sin3\varphi_1\sin\varphi_2\right| \,, \\
\nonumber && \left|I_1\right|=\frac{2\sqrt{2}}{9}\left|\sin ^2\varphi_1 \sin \varphi_2 \left(\sqrt{2}\cos^2 \theta \cos \varphi_2- \sin 2\theta \cos \varphi _1\right)\right| \,, \\
\label{eq:CP_para_caseI_1st} && \left|I_2\right|= \frac{2\sqrt{2}}{9}\left|\sin ^2\varphi_1 \sin \varphi_2 \left(\sqrt{2}\sin^2 \theta \cos \varphi_2+ \sin 2\theta \cos \varphi _1\right)\right| \,,\end{aligned}$$ where $J_{CP}$ is well-known Jarlskog invariant, and $I_1$ and $I_2$ are defined for the Majorana phases with $$\begin{aligned}
\nonumber J_{CP}&=&\mathrm{Im}\left[\left(U_{\text{PMNS}}\right)_{11}\left(U_{\text{PMNS}}\right)_{33}\left(U^{\ast}_{\text{PMNS}}\right)_{13}\left(U^{\ast}_{\text{PMNS}}\right)_{31}\right]\\
\nonumber&=&\frac{1}{8}\sin2\theta_{12}\sin2\theta_{13}\sin2\theta_{23}\cos\theta_{13}\sin\delta_{CP}\,,\\
\nonumber I_1&=&\text{Im}\left[\left(U_{\text{PMNS}}\right)^{\ast2}_{11}\left(U_{\text{PMNS}}\right)^2_{12}\right]=\frac{1}{4}\sin^22\theta_{12}\cos^4\theta_{13}\sin\alpha_{21}\,,\\
I_2&=&\text{Im}\left[\left(U_{\text{PMNS}}\right)^{\ast2}_{11}\left(U_{\text{PMNS}}\right)^2_{13}\right]=\frac{1}{4}\sin^22\theta_{13}\cos^2\theta_{12}\sin\alpha^{\prime}_{31}\,,\end{aligned}$$ where $\alpha^{\prime}_{31}\equiv\alpha_{31}-2\delta_{CP}$, $\delta_{CP}$ is the Dirac CP violating phase, $\alpha_{21}$ and $\alpha_{31}$ are the Majorana CP phases in the standard parameterization of the PMNS matrix [@Agashe:2014kda]. We show the absolute values of $J_{CP}$, $I_1$ and $I_2$ in Eq. , the reason is because the sign of the $J_{CP}$ depends on the ordering of rows and columns and the sign of $I_1$ and $I_2$ could be changed by the CP parity matrix $Q_{\nu}$. Moreover, if the lepton doublet fields are assigned to the triplet $\mathbf{3}_{9n-1, 0}$ instead of $\mathbf{3}_{1, 0}$, the prediction for $U_{\text{PMNS}}$ would be complex conjugated such that the signs of $J_{CP}$, $I_1$ and $I_2$ are all inversed. We show the possible predictions for the mixing parameters $\sin^2\theta_{12}$, $\sin\theta_{13}$, $\sin^2\theta_{23}$ as well as $\left|\sin\delta_{CP}\right|$, $\left|\sin\alpha_{21}\right|$ and $\left|\sin\alpha_{31}\right|$ for each $D^{(1)}_{9n, 3n}$ group in figure \[fig:caseI\_1cl\_mixing\_para\], where all the admissible values of $\varphi_1$ and $\varphi_2$ shown in Eq. are considered and all the three mixing angles are required to lie in the $3\sigma$ allowed regions adapted from [@Capozzi:2013csa]. It is notable that the solar mixing angle is predicted to be within the narrow interval of $0.313\leq\sin^2\theta_{12}\leq0.344$. The near future medium-baseline reactor neutrino oscillation experiments, such as JUNO [@An:2015jdp] and RENO-50 [@Kim:2014rfa] are expected to make very precise, sub-percent measurements of the solar mixing angle $\theta_{12}$. They provide the one of the most significant test of this mixing pattern. The allowed values of the CP violation phases increase with group index $n$ and they are strongly constrained for smaller $n$. From figure \[fig:caseI\_1cl\_mixing\_para\] we can read $0 \leq\left|\sin\delta_{CP}\right|\leq 0.226$, $0.847 \leq\left|\sin\alpha_{21}\right|\leq 0.873$ and $0 \leq\left|\sin\alpha'_{31}\right|\leq 0.488$ in the case of $n=1$. However, almost any values of the CP phases can be achieved for sufficient large value of $n$.
![\[fig:caseI\_1cl\_mixing\_para\]The possible values of $\sin^2\theta_{12}$, $\sin\theta_{13}$, $\sin^2\theta_{23}$, $\left|\sin\delta_{CP}\right|$, $\left|\sin\alpha_{21}\right|$ and $\left|\sin\alpha^{\prime}_{31}\right|$ with respect to $n$ for the mixing pattern $U^{I,1}_{\text{PMNS}}$ in the case I, where the three lepton mixing angles are required to be within the experimentally preferred $3\sigma$ ranges. The $1\sigma$ and $3\sigma$ regions of the three neutrino mixing angles are adapted from global fit [@Capozzi:2013csa].](ParameterCaseI123.pdf){width="99.00000%"}
Then we turn to the second mixing pattern $U^{I,2}_{\text{PMNS}}$ in which $\sqrt{\frac{2}{3}}\left(\sin\varphi_1, \cos\left(\frac{\pi}{6}-\varphi_1\right), \cos\left(\frac{\pi}{6}+\varphi_1\right)\right)^{T}$is the second column vector. Its predictions for the mixing angles are $$\begin{aligned}
\nonumber&&\sin^2\theta_{13}=\frac{1}{3}\left(1+\cos^2\theta\cos2\varphi_1+\sqrt{2}\sin2\theta\cos\varphi_1\cos\varphi_2\right),\\
\nonumber&&\sin^2\theta_{12}=\frac{2\sin^2\varphi_1}{2-\cos^2\theta\cos2\varphi_1-\sqrt{2}\sin2\theta\cos\varphi_1\cos\varphi_2}\,,\\
\label{eq:mixing_para_caseI_2nd}&&\sin^2\theta_{23}=\frac{1-\cos^2\theta\sin\left(\pi/6+2\varphi_1\right)-\sqrt{2}\sin2\theta\cos\varphi_2\sin\left(\pi/6-\varphi_1\right)}
{2-\cos^2\theta\cos2\varphi_1-\sqrt{2}\sin2\theta\cos\varphi_1\cos\varphi_2}\,.\end{aligned}$$ We see that the solar and reactor mixing angles are correlated as $$3\sin^2\theta_{12}\cos^2\theta_{13}=2\sin^2\varphi_1\,.$$ In order to accommodate the experimental results on $\theta_{12}$ and $\theta_{13}$, $\varphi_1$ should vary in the interval: $$\varphi_{1}\in\left[0.210\pi,0.259\pi\right]\cup\left[0.741\pi,0.790\pi\right]\cup \left[1.210\pi,1.259\pi\right]\cup \left[1.741\pi,1.790\pi\right]\,.$$ Consequently we have $$|\cos(\varphi_1-\frac{\pi}{6})|,|\cos(\varphi_1+\frac{\pi}{6})|\in[0.230,0.377]\cup[0.958,0.991]\,.$$ We see that both $(22)$ and $(32)$ entries of $U^{I,2}_{\text{PMNS}}$ are not in agreement with the experimental data given by Eq. . Hence this mixing pattern is phenomenologically disfavored.
![\[fig:caseI\_2cl\_mixing\_para\] The allowed values of $\sin^2\theta_{23}$ and $\sin\theta_{13}$ for the mixing pattern $U^{I,3}_{\text{PMNS}}$ in case I, where the first four smallest $D^{(1)}_{9n, 3n}$ group with $n=1, 2, 3, 4$ are considered. The $1\sigma$ and $3\sigma$ regions of the three neutrino mixing angles are adapted from global fit [@Capozzi:2013csa]. ](CorrelationCaseI789n1234.pdf){width="90.00000%"}
For the third possible arrangement of the rows and columns, the PMNS matrix is $U^{I, 3}_{\text{PMNS}}$. In this case, the third column of the PMNS matrix doesn’t depend on the continuous parameter $\theta$ and it is completely fixed the remnant CP symmetry. It is straightforward to extract the mixing angles. $$\begin{aligned}
\nonumber&&\sin^2\theta_{13}=\frac{2}{3}\sin^2\varphi_1,\qquad\quad \sin^2\theta_{23}=\frac{1+\sin\left(\pi/6+2\varphi_1\right)}{2+\cos2\varphi_1}\,,\\
\label{eq:mixing_para_caseI_3rd}&&\sin^2\theta_{12}=\frac{1+\sin^2\theta\cos2\varphi_1-\sqrt{2}\sin2\theta\cos\varphi_1\cos\varphi_2}{2+\cos2\varphi_1}\,.\end{aligned}$$ The experimental data $1.76\times10^{-2}\leq\sin^2\theta_{13}\leq2.95\times10^{-2}$ at $3\sigma$ level [@Capozzi:2013csa] can be accommodated for the following values of the parameter $\varphi_1$: $$\begin{split}
\varphi_1\in\left[0.0519\pi,0.0675\pi\right]\cup\left[0.933\pi,0.948\pi\right]\cup \left[1.0519\pi,1.0675\pi\right]\cup \left[1.933\pi,1.948\pi\right]\,.
\end{split}$$ As both $\theta_{13}$ and $\theta_{23}$ depend on a single parameter $\varphi_1$, we can derive a sum rule between them, $$2\sin^2\theta_{23}=1\pm\tan\theta_{13}\sqrt{2-\tan^2\theta_{13}}\;.$$ Given the experimental best fitting value of the reactor mixing angle $\sin^2\theta_{13}=2.34\times 10^{-2}$ [@Capozzi:2013csa], we have $$\sin^2\theta_{23}\simeq0.391,\quad\mathrm{or}\quad \sin^2\theta_{23}\simeq0.609\,,$$ which is within the $3\sigma$ range although it is non-maximal. For a given $D^{(1)}_{9n, 3n}$ group, the atmospheric and reactor mixing angles can only take a set of discrete values. The possible values of $\sin^2\theta_{23}$ and $\sin\theta_{13}$ for the first four smallest $n=1, 2, 3, 4$ are displayed in figure \[fig:caseI\_2cl\_mixing\_para\]. We see that the values $\varphi_1=\pm\pi/18, \pm17\pi/18$ in the case of $n=2, 4$ lead to $(\theta_{13}, \theta_{23})=(8.151^{\circ}, 50.813^{\circ})$ or $(8.151^{\circ}, 39.187^{\circ})$ which are compatible with the present experimental data [@Capozzi:2013csa]. The next generation of superbeam neutrino oscillation experiments would provide a high-precision determination of $\theta_{23}$. If no significant deviations from maximal mixing of $\theta_{23}$ will be detected, our present scheme will be excluded. Furthermore, we find that the CP invariants are $$\begin{aligned}
\nonumber && \left|J_{CP}\right|=\frac{1}{6\sqrt{6}}\left|\sin2\theta\sin3\varphi_1\sin\varphi_2\right| \,, \\
\nonumber && \left|I_1\right|=\frac{1}{9} \left|\cos\varphi_1\sin\varphi_2\left(4\cos2\theta\cos\varphi_1\cos \varphi_2-\sqrt{2}\sin2\theta\cos2\varphi_1\right)\right|\,, \\
\label{eq:CP_para_caseI_3rd}&&\left|I_2\right|=\frac{2\sqrt{2}}{9} \left|\sin^2\varphi_1 \sin \varphi_2 \left(\sqrt{2}\sin^2 \theta \cos \varphi_2+ \sin 2\theta \cos \varphi _1\right)\right| \,.\end{aligned}$$ Furthermore, we study the admissible values of mixing angles and CP phases for each $D^{(1)}_{9n, 3n}$ group. The numerical results are displayed in figure \[fig:caseI\_3cl\_mixing\_para\]. We easily see that the atmospheric mixing angle $\theta_{23}$ is not maximal and it is around the $3\sigma$ upper or lower bounds. Similar to the $\Delta(6n^2)$ group [@Ding:2014ora], maximal value of the Majorana phase $\alpha^{\prime}_{31}$ can not be achieved in this case and it is found to be in the range of $\left|\sin\alpha^{\prime}_{31}\right|\leq 0.910$ while almost any values of $\delta_{CP}$ and $\alpha_{21}$ can be possible for large $n$.
![\[fig:caseI\_3cl\_mixing\_para\]The possible values of $\sin^2\theta_{12}$, $\sin\theta_{13}$, $\sin^2\theta_{23}$, $\left|\sin\delta_{CP}\right|$, $\left|\sin\alpha_{21}\right|$ and $\left|\sin\alpha^{\prime}_{31}\right|$ with respect to $n$ for the mixing pattern $U^{I,3}_{\text{PMNS}}$ in the case I, where the three lepton mixing angles are required to be within the experimentally preferred $3\sigma$ ranges. The $1\sigma$ and $3\sigma$ regions of the three neutrino mixing angles are adapted from global fit [@Capozzi:2013csa].](ParameterCaseI789.pdf){width="99.00000%"}
As a concrete example, we shall study the first two smallest $D^{(1)}_{9n, 3n}$ group with $n=1$ and $n=2$. From the expression of the PMNS matrix, we know that $U^{I, 1}_{\text{PMNS}}$ has the following symmetry properties: $$\begin{aligned}
\nonumber&&U^{I, 1}_{\text{PMNS}}(\theta, \varphi_1, \pi+\varphi_2)=U^{I, 1}_{\text{PMNS}}(-\theta, \varphi_1, \varphi_2)\text{diag}(1, -1, 1),\\
\nonumber&&U^{I, 1}_{\text{PMNS}}(\theta, \varphi_1, \pi-\varphi_2)=[U^{I, 1}_{\text{PMNS}}(-\theta, \varphi_1, \varphi_2)]^{*}\text{diag}(1, -1, 1),\\
\nonumber&&U^{I, 1}_{\text{PMNS}}(\theta, \varphi_1, -\varphi_2)=[U^{I, 1}_{\text{PMNS}}(\theta, \varphi_1, \varphi_2)]^{*},\\
\label{eq:caseI_symmetry}&&U^{I, 1}_{\text{PMNS}}(\theta, \pi+\varphi_1, \varphi_2)=U^{I, 1}_{\text{PMNS}}(-\theta, \varphi_1, \varphi_2)\text{diag}(-1, 1, -1)\,,\end{aligned}$$ where the diagonal matrix can be absorbed into the matrix $Q_{\nu}$. Similar relations are satisfied for the PMNS matrix $U^{I, 3}_{\text{PMNS}}$. Note that the PMNS matrix would become its complex conjugation if the three generations of leptons are assigned to the triplet $\mathbf{3}_{9n-1, 0}\cong\mathbf{3}^{*}_{1, 0}$. As a result, without loss of generality, we shall focus on the case of $0\leq\varphi_1\leq\pi$ and $0\leq\varphi_2\leq\pi/2$. A conventional $\chi^2$ analysis is performed. Notice that we don’t include the information of the Dirac CP phase $\delta_{CP}$ into the $\chi^2$ function, since the evidence for a preferred value of $\delta_{CP}$ coming from both present experiments and the global fitting is rather weak. The numerical results are reported in table \[tab:caseI\_n12\], where we exclude all patterns that can not accommodate the experimental data at the best fitting point $\theta=\theta_{bf}$ for which the $\chi^2$ function is minimized. Since the global fit results of the mixing angles are slightly distinct for NO and IO neutrino mass spectrums [@Capozzi:2013csa], the $\chi^2$ function has been defined for NO and IO respectively. The values in the parentheses are the results for the IO case. Applying the symmetry transformations in Eq. , we can obtain other values of $\varphi_1$ and $\varphi_2$ which yield the same best fit values for the mixing angles such that the same $\chi^2_{min}$ is obtained. For both mixing patterns $U^{I, 1}_{\text{PMNS}}$ and $U^{I, 3}_{\text{PMNS}}$, we can check that the formulae in Eqs. (\[eq:mixing\_para\_caseI\_1st\],\[eq:mixing\_para\_caseI\_3rd\]) for the mixing angles $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$ are invariant while $\sin^2\theta_{23}$ turns into $\cos^2\theta_{23}$ under the transformation $\varphi_1\rightarrow\pi-\varphi_1$, $\theta\rightarrow\pi-\theta$. As a result, the sum of the best fitting value $\theta_{bf}$ for $\varphi_1$ and $\pi-\varphi_1$ is approximately equal to $\pi$. It is remarkable that even the smallest $D^{(1)}_{9n, 3n}$ group with $n=1$ allows a reasonable fit to the experimental data, for instance, the mixing patterns with $(\varphi_1, \varphi_2)=(4\pi/9, 0)$, $(4\pi/9, \pi/3)$, $(5\pi/9, 0)$ and $(5\pi/9, \pi/3)$ can describe the experimentally measured values of the mixing angles, as can be seen from table \[tab:caseI\_n12\]. In particular, the CP violating phases are neither conserved nor maximal in the case of $(\varphi_1, \varphi_2)=(4\pi/9, \pi/3)$ and $(5\pi/9, \pi/3)$. The PMNS matrix $U^{I, 1}_{\text{PMNS}}$ for $n=2$ as well as $(\varphi_1, \varphi_2)=(\pi/2, \pi/2)$ give rise to maximal atmospheric mixing and maximal Dirac phase. On the other hand, the group index $n$ should be equal or greater than 2 in order to obtain phenomenologically viable mixing pattern of the form $U^{I, 3}_{\text{PMNS}}$. Scrutinizing all the admissible cases listed in table \[tab:caseI\_n12\], we find that the predictions for $\theta_{13}$ are almost the same, nevertheless $\theta_{12}$, $\theta_{23}$ and $\delta_{CP}$ are predicted to be considerably different. The JUNO experiment will be capable of reducing the error of $\sin^2\theta_{12}$ to about $0.1^{\circ}$ or around $0.3\%$ [@An:2015jdp]. Future long baseline experiments such as DUNE [@Adams:2013qkq], LBNO [@::2013kaa], T2HK [@Kearns:2013lea] and possibly ESS$\nu$SB [@Baussan:2012cw] at the European Spallation Source can make very precise measurements of the oscillation parameters $\theta_{12}$, $\theta_{23}$ and $\delta_{CP}$. Therefore future neutrino facilities have the potential to discriminate between the above possible cases, or to rule them out entirely. Furthermore, we expect that a more ambitious facility such as the neutrino factory [@Geer:1997iz] could provide a more stringent tests of our approach.
[|c|c|c|c|c|c|c|c|c|c|c|]{}\
\
& $\varphi_1$ & $\varphi_2$ & $\theta_{bf}$ & $\chi^2_{min}$ & $\sin^2\theta_{13}$ & $\sin^2\theta_{12}$ & $\sin^2\theta_{23}$ & $|\sin\delta_{CP}|$ & $|\sin\alpha_{21}|$ & $|\sin\alpha^{\prime}_{31}|$\
& & & $0.0245$ & $3.789$ & $0.0243$ & $0.337$ & $0.419$ & & &\
&&&($0.0274$) & ($4.267$) & ($0.0248$) & ($0.337$) & ($0.421$) &&&\
&& & $0.0435$ & $3.928$ & $0.0242$ & $0.337$ & $0.417$ &$0.125$ & $0.857$ & $0.276$\
&&& ($0.0480$) & ($4.438$) & ($0.0247$) & ($0.337$) & ($0.419$) &($0.137$) & ($0.856$) & ($0.302$)\
& & & $3.108$ & $21.499$ & $0.0259$ & $0.336$ & $0.574$ & & &\
&&& ($3.117$) & ($3.822$) & ($0.0244$) & ($0.337$) & ($0.581$) &&&\
&& & $3.087$ & $22.307$ & $0.0255$ & $0.337$ & $0.578$ &$0.154$ & $0.854$ & $0.338$\
&& & ($3.097$) & ($3.849$) & ($0.0243$) & ($0.337$) & ($0.583$) & ($0.127$) & ($0.856$) & ($0.281$)\
\
& & & $0.0278$ & $3.807$ & $0.0243$ & $0.337$ & $0.419$ &$0.0462$ & $0.869$ & $0.103$\
&&& ($0.0311$) & ($4.289$) & ($0.0248$) & ($0.337$) & ($0.421$) &($0.0510$) & ($0.870$) & ($0.114$)\
&& & $0.108$ & $5.666$ & $0.0237$ & $0.338$ & $0.400$ &$0.362$ & $0.0532$ & $0.739$\
&&& ($0.116$) & ($6.131$) & ($0.0243$) & ($0.337$) & ($0.400$) &($0.384$) & ($0.0572$) & ($0.774$)\
& & &$3.104$ & $21.616$ & $0.0258$ & $0.336$ & $0.575$ &$0.0602$ & $0.871$ & $0.134$\
&&& ($3.113$) & ($3.826$) & ($0.0244$) & ($0.337$) & ($0.581$) &($0.0468$) & ($0.869$) & ($0.104$)\
&& & $3.033$ & $27.468$ & $0.0238$ & $0.337$ & $0.600$ &$0.365$ & $0.0537$ & $0.744$\
&&& ($3.026$) & ($4.087$) & ($0.0243$) & ($0.337$) & ($0.600$) &($0.383$) & ($0.0569$) & ($0.772$)\
& & &$0.261$ & $26.399$ & $0.0222$ & $0.318$ & $0.604$ &$0.885$ & $0.866$ & $0.866$\
&& & ($0.272$) & ($1.490$) & ($0.0240$) & ($0.317$) & ($0.608$) &($0.887$) & ($0.866$) & ($0.866$)\
&& &$2.877$ & $3.838$ & $0.0228$ & $0.318$ & $0.394$ &$0.886$ & $0.866$ & $0.866$\
&& & ($2.873$) & ($4.352$) & ($0.0234$) & ($0.317$) & ($0.393$) &($0.887$) & ($0.866$) & ($0.866$)\
&& &$0.269$ &$3.946$ &$0.0235$ & $0.317$ & $0.5$ & & &\
&&&($0.272$) &($0.380$) &($0.0241$) & ($0.317$) & ($0.5$) & & &\
& & & $0.0344$ & $27.637$ & & & & & &\
&&& ($0.0344$) & ($4.238$) & & & &&&\
&& &$0.0399$ & $27.637$ & & & &$0.0431$ & $0.881$ & $0.0279$\
&& & ($0.0399$) & ($4.238$) & & & & ($0.0431$) & ($0.881$) & ($0.0279$)\
&& &$0.0716$ & $27.637$ & & & &$0.134$ & $0.815$ & $0.0868$\
&& & ($0.0716$) & ($4.238$) & & & &($0.134$) & ($0.815$) & ($0.0868$ )\
&& & & $31.219$ & & $0.340$ & & & &\
&&&& ($7.820$) & & ($0.340$) & &&&\
& & & $3.107$ & $5.707$ & & & & & &\
&&& ($3.107$) & ($6.374$) & & & &&&\
&& &$3.102$ & $5.707$ & & & &$0.0431$ & $0.881$ & $0.0279$\
&&& ($3.102$) & ($6.374$) & & & &($0.0431$) & ($0.881$) & ($0.0279$)\
&& &$3.070$ & $5.707$ & & & &$0.134$ & $0.815$ & $0.0868$\
&& & ($3.070$) & ($6.374$) & & & &($0.134$) & ($0.815$) & ($0.0868$)\
&& & &$9.289$ & & $0.340$ & & & &\
&&&& ($9.955$) & & ($0.340$) & &&&\
Since the Majorana CP violating phases can be predicted in the present framework, we now discuss its phenomenological implications in the neutrinoless double beta ($0\nu\beta\beta$) decay. It is well-known that the $0\nu\beta\beta$ decay process is the most sensitive probe for Majorana neutrinos. Its observation would establish the Majorana nature of neutrinos irrespective of the underlying mass generation mechanism. The $0\nu\beta\beta$ decay rate is proportional to the square of the effective Majorana mass $|m_{ee}|$ which is given by [@Agashe:2014kda] $$\label{eq:mee}\left|m_{ee}\right|=\left|m_1\cos^2\theta_{12}\cos^2\theta_{13}+m_2\sin^2\theta_{12}\cos^2\theta_{13}e^{i \alpha_{21}}+m_3\sin^2\theta_{13}e^{i\alpha^{\prime}_{31}}\right|\,.$$ The values of $\left|m_{ee}\right|$ are dependent on both CP phases $\alpha_{21}$ and $\alpha^{\prime}_{31}\equiv\alpha_{31}-2\delta_{CP}$. For the mixing pattern $U^{I, 1}_{\text{PMNS}}$, $|m_{ee}|$ is of the form $$\begin{aligned}
\nonumber&&|m_{ee}|=\frac{1}{3}\Big|2m_1\sin^2\varphi_1+q_1m_2(e^{i \varphi_{2}} \cos\theta-\sqrt{2} \cos\varphi_{1}\sin\theta)^2\\
&&\qquad\qquad+q_2m_3(e^{i \varphi_{2}} \sin\theta+\sqrt{2} \cos\theta \cos\varphi_{1})^2\Big|\,,\end{aligned}$$ where $q_1, q_2=\pm1$ appears due to the undetermined CP parity of the neutrino states encoded in the matrix $Q_{\nu}$. For another admissible mixing pattern $U^{I, 3}_{\text{PMNS}}$, $|m_{ee}|$ is given by $$\begin{aligned}
\nonumber&&|m_{ee}|=\frac{1}{3}\Big|2m_3\sin^2\varphi_{1}+q_1m_2( e^{i \varphi_{2}} \cos\theta-\sqrt{2} \cos\varphi_{1}\sin\theta)^2\\
&&\qquad\qquad+q_2m_1(e^{i \varphi_{2}} \sin\theta+\sqrt{2} \cos\theta \cos\varphi_{1})^2\Big|\,.\end{aligned}$$ The achievable values of the effective mass $|m_{ee}|$ for both $n\rightarrow\infty$ and $n=2$ are plotted in figure \[fig:mee\_CaseI\]. Here we require the three mixing angle be within their $3\sigma$ allowed values while the neutrino mass-squared splittings are fixed at their best-fit values from Ref. [@Capozzi:2013csa]. We see that the majority of the experimentally allowed $3\sigma$ region of $|m_{ee}|$ can be reproduced in the limit $n\rightarrow\infty$. In the case of $n=2$, it is remarkable that the effective mass $|m_{ee}|$ obtained from $U^{I, 1}_{\text{PMNS}}$ is found to be around 0.0155eV, 0.0175eV, 0.0279eV, 0.0423eV, or 0.0484eV for IO neutrino mass spectrum. These predictions are beyond the reach of the present $0\nu\beta\beta$ experiments such as GERDA [@Agostini:2013mzu], EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm]. However, the proposed facilities nEXO and KamLAND2-Zen [@Piquemal:2013uaa] etc aim to increase the sensitivity to cover the full IO region, such that all of our patterns with this mass spectrum could be tested. For NO the effective mass $|m_{ee}|$ is much smaller than the IO case and it can even vanish for certain values of the lightest neutrino mass because of a cancellation between different terms in Eq. . Obviously exploring the NH region experimentally is beyond the reach of any planned $0\nu\beta\beta$ experiment. Even if the signals of $0\nu\beta\beta$ decays are not observed and the neutrino masses spectrum are measured to be NO by upcoming neutrino oscillation experiments [@An:2015jdp; @Kim:2014rfa], one can still extract useful information on the Majorana phases $\alpha_{21}$ and $\alpha^{\prime}_{31}$ by combining the cosmological data on the absolute neutrino mass scale and the improved measurement of $\theta_{12}$, $\theta_{23}$ and $\delta_{CP}$ from a number of complementary neutrino oscillation experiments.
![\[fig:mee\_CaseI\]The possible values of the effective Majorana mass $|m_{ee}|$ as a function of the lightest neutrino mass in the case I. The left and right panels are for the mixing patterns $U^{I, 1}_{\text{PMNS}}$ and $U^{I, 3}_{\text{PMNS}}$ respectively. The red (blue) dashed lines indicate the most general allowed regions for IO (NO) neutrino mass spectrum obtained by varying the mixing parameters over the $3\sigma$ ranges [@Capozzi:2013csa]. The orange (cyan) areas denote the achievable values of $|m_{ee}|$ in the limit of $n\rightarrow\infty$ assuming IO (NO) spectrum. The purple and green regions are the theoretical predictions for the $D^{(1)}_{9n ,3n}$ group with $n=2$. Notice that the purple (green) region overlaps the orange (cyan) one. The present most stringent upper limits $|m_{ee}|<0.120$ eV from EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] is shown by horizontal grey band. The vertical grey exclusion band represents the current bound coming from the cosmological data of $\sum m_i<0.230$ eV at $95\%$ confidence level obtained by the Planck collaboration [@Ade:2013zuv]. ](meeI123_v1.pdf "fig:"){width="0.46\linewidth"} ![\[fig:mee\_CaseI\]The possible values of the effective Majorana mass $|m_{ee}|$ as a function of the lightest neutrino mass in the case I. The left and right panels are for the mixing patterns $U^{I, 1}_{\text{PMNS}}$ and $U^{I, 3}_{\text{PMNS}}$ respectively. The red (blue) dashed lines indicate the most general allowed regions for IO (NO) neutrino mass spectrum obtained by varying the mixing parameters over the $3\sigma$ ranges [@Capozzi:2013csa]. The orange (cyan) areas denote the achievable values of $|m_{ee}|$ in the limit of $n\rightarrow\infty$ assuming IO (NO) spectrum. The purple and green regions are the theoretical predictions for the $D^{(1)}_{9n ,3n}$ group with $n=2$. Notice that the purple (green) region overlaps the orange (cyan) one. The present most stringent upper limits $|m_{ee}|<0.120$ eV from EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] is shown by horizontal grey band. The vertical grey exclusion band represents the current bound coming from the cosmological data of $\sum m_i<0.230$ eV at $95\%$ confidence level obtained by the Planck collaboration [@Ade:2013zuv]. ](meeI789_v1.pdf "fig:"){width="0.46\linewidth"}
()
: $G_{l}=\left\langle abc^{s}d^{t}\right\rangle$, $G_{\nu}=Z^{bd^x}_2$, $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{2\delta+2x+3n\tau}d^{\delta}),
\rho_{\mathbf{r}}(bc^{2\delta+2x+3n\tau}d^{\delta+x})\right\}$\
This case differs from the previous one in the residual flavor symmetry $G_{l}$. From table \[tab:cle\_diagonal\_matrix\] and table \[tab:extension\_Gch\], we know that the charged lepton diagonalization matrix is exactly $U^{(4)}_{l}$. Since the neutrino mass matrix preserves the same remnant symmetry as case , the neutrino mass matrix should take the form of Eq. , and it is diagonalized by the unitary transformation $U_{\nu}$ in Eq. . Using the freedom in exchanging rows and columns, we find the phenomenologically viable lepton mixing matrix is $$\label{eq:PMNS_case_II_1}
U^{II,1}_{\text{PMNS}}=\frac{1}{2}
\begin{pmatrix}
-\sin\theta-\sqrt{2} e^{i \varphi_{3}} \cos\theta &~ 1 ~& \cos\theta-\sqrt{2} e^{i \varphi_{3}} \sin\theta \\
-\sin\theta+\sqrt{2} e^{i\varphi_{3}} \cos\theta &~ 1 ~& \cos\theta+\sqrt{2} e^{i\varphi_{3}} \sin\theta \\
-\sqrt{2} \sin \theta &~ -\sqrt{2} ~& \sqrt{2} \cos\theta
\end{pmatrix}Q_{\nu}\,,$$ or $$\label{eq:PMNS_case_II_2}
U^{II,2}_{\text{PMNS}}=\frac{1}{2}
\begin{pmatrix}
-\sin\theta-\sqrt{2} e^{i \varphi_{3}} \cos\theta &~ 1 ~& \cos\theta-\sqrt{2} e^{i \varphi_{3}} \sin\theta \\
-\sqrt{2} \sin \theta &~ -\sqrt{2} ~& \sqrt{2} \cos\theta \\
-\sin\theta+\sqrt{2} e^{i\varphi_{3}} \cos\theta &~ 1 ~& \cos\theta+\sqrt{2} e^{i\varphi_{3}} \sin\theta
\end{pmatrix}Q_{\nu}\,,$$ where $$\varphi_{3}=\frac{s-t+2x+\delta}{3n}\pi\,,$$ and its possible values are $$\varphi_{3}~(\mathrm{mod}~2\pi)=0, \frac{1}{3n}\pi, \frac{2}{3n}\pi, \ldots, \frac{6n-1}{3n}\pi\,.$$ It is easy to check that $U^{II, 1}_{\text{PMNS}}$ as well as $U^{II, 2}_{\text{PMNS}}$ have the symmetry property $$\begin{aligned}
\nonumber&&U^{II, 1}_{\text{PMNS}}(\theta, \varphi_3+\pi)=U^{II, 1}_{\text{PMNS}}(-\theta, \varphi_3)\text{diag}(-1, 1, 1),\\
\label{eq:UPMNS_sym_caseII}&&U^{II, 2}_{\text{PMNS}}(\theta, \varphi_3+\pi)=U^{II, 2}_{\text{PMNS}}(-\theta, \varphi_3)\text{diag}(-1, 1, 1)\,.\end{aligned}$$ We see that the second column of the PMNS matrix is $(1, 1, -\sqrt{2})/2$ or $(1, -\sqrt{2}, 1)/2$ in this case. For the mixing pattern $U^{II,1}_{\text{PMNS}}$, the three lepton mixing angles are found to be $$\begin{aligned}
\nonumber&&\sin^2\theta_{13}=\frac{1}{8}
\left(3-\cos2\theta-2\sqrt{2}\sin2\theta\cos\varphi_3\right),\\
\nonumber&&\sin^2\theta_{12}=\frac{2}{5+\cos2\theta+2\sqrt{2}\sin2\theta\cos\varphi_3},\\
\label{eq:mixing_para_caseII_1st}&&\sin^2\theta_{23}=\frac{3-\cos2\theta+2\sqrt{2}\sin2\theta\cos\varphi_3}{5+\cos2\theta+2\sqrt{2}\sin2\theta\cos\varphi_3}\,,\end{aligned}$$ which fulfill the following sum rules $$\begin{aligned}
\nonumber&&\hskip4cm 4\sin^2\theta_{12}\cos^2\theta_{13}=1, \\ &&\hskip-1cm\cos^2\theta_{13}\cos^2\theta_{23}=\frac{\cos2\theta_{13}+2\cos^2\varphi_3\pm2\cos\varphi_3\sqrt{6\sin^2\theta_{13}-8\sin^4\theta_{13}-\sin^2\varphi_3}}{1+8\cos^2\varphi_3}\,.\end{aligned}$$ Given the $3\sigma$ range of $\theta_{13}$, the solar mixing angle $\theta_{12}$ is determined to lie in the region of $0.254\leq\sin^2\theta_{12}\leq0.258$ which is rather close to its $3\sigma$ lower limit 0.259 [@Capozzi:2013csa]. However, this mixing pattern is a good leading order approximation because accordance with the experimental data could be easily achieved in a concrete model after higher order corrections contributions are included. We plot the $1\sigma$, $2\sigma$ and $3\sigma$ contour regions for $\sin^2\theta_{ij}$ with $ij=12, 13, 23$ in the $\varphi_3-\theta$ plane in figure \[fig:caseII\_contour\_mixing\_para\]. Obviously the most stringent constraint comes from the precisely measured reactor angle $\theta_{13}$. Moreover, the three CP rephasing invariants are given by $$\begin{aligned}
\nonumber && \left|J_{CP}\right|=\frac{1}{8 \sqrt{2}}\left|\sin 2 \theta\sin\varphi_3\right|\,, \\
\nonumber && \left|I_1\right|=\frac{1}{8\sqrt{2}}\left|(\sin2\theta+2\sqrt{2}\cos^2\theta\cos\varphi_3)\sin\varphi_3\right|\,,\\
\label{eq:CP_para_caseII_1st}&&\left|I_2\right|=\frac{1}{8\sqrt{2}}\left|(\sin2\theta-2\sqrt{2}\cos2\theta\cos\varphi_3)\sin\varphi_3\right|\,.\end{aligned}$$
![\[fig:caseII\_contour\_mixing\_para\]The contour regions of the three mixing angles in the case II. The red, blue and green areas denote the predictions for $\sin^2\theta_{13}$, $\sin^2\theta_{12}$ and $\sin^2\theta_{23}$ respectively. The allowed $1\sigma$, $2\sigma$ and $3\sigma$ regions of each mixing angle are represented by different shadings. Here we take the $3\sigma$ lower limit of $\sin^2\theta_{12}$ to be 0.254 instead of 0.259 given by Ref. [@Capozzi:2013csa]. The best fit values of the mixing angles are indicated by dotted lines.](C2regionAllNHFir.pdf){width="50.00000%"}
The three CP violation phases extracted from these invariants depend on $\theta$ and $\varphi_3$. The predictions for $|\sin\delta_{CP}|$, $|\sin\alpha_{21}|$ and $|\sin\alpha^{\prime}_{31}|$ are plotted in figure \[fig:caseII\_contour\_CP\_para\], where the black areas represent the regions in which all three lepton mixing angles are in the experimentally preferred $3\sigma$ ranges. To accommodate the experimental data of mixing angles [@Capozzi:2013csa], both $\delta_{CP}$ and $\alpha_{21}$ can not be maximal. The values of $|\sin\delta_{CP}|$ and $|\sin\alpha_{21}|$ are bounded from above with $|\sin\delta_{CP}|\leq 0.895$ and $|\sin\alpha_{21}|\leq 0.545$.
![\[fig:caseII\_contour\_CP\_para\]The contour plots of $|\sin\delta_{CP}|$, $|\sin\alpha_{21}|$ and $|\sin\alpha^{\prime}_{31}|$ in the $\varphi_3-\theta$ plane in the case II. The black areas represent the regions in which the lepton mixing angles are compatible with experimental data at $3\sigma$ level, and it can be read out from figure \[fig:caseII\_contour\_mixing\_para\].](C2DeltaCPNHAll.pdf "fig:"){width="49.50000%"} ![\[fig:caseII\_contour\_CP\_para\]The contour plots of $|\sin\delta_{CP}|$, $|\sin\alpha_{21}|$ and $|\sin\alpha^{\prime}_{31}|$ in the $\varphi_3-\theta$ plane in the case II. The black areas represent the regions in which the lepton mixing angles are compatible with experimental data at $3\sigma$ level, and it can be read out from figure \[fig:caseII\_contour\_mixing\_para\].](C2AlphaNHAll.pdf "fig:"){width="49.50000%"} ![\[fig:caseII\_contour\_CP\_para\]The contour plots of $|\sin\delta_{CP}|$, $|\sin\alpha_{21}|$ and $|\sin\alpha^{\prime}_{31}|$ in the $\varphi_3-\theta$ plane in the case II. The black areas represent the regions in which the lepton mixing angles are compatible with experimental data at $3\sigma$ level, and it can be read out from figure \[fig:caseII\_contour\_mixing\_para\].](C2Alpha31NHAll.pdf "fig:"){width="49.50000%"}
The second PMNS matrix $U^{II, 2}_{\text{PMNS}}$ can be obtained from $U^{II, 1}_{\text{PMNS}}$ by exchanging the second and third rows. Therefore $U^{II, 2}_{\text{PMNS}}$ and $U^{II, 1}_{\text{PMNS}}$ give rise to the same reactor and solar mixing angles and the Majorana phases, while the atmospheric angle changes from $\theta_{23}$ to $\pi/2-\theta_{23}$ and the Dirac phase changes from $\delta_{CP}$ to $\pi+\delta_{CP}$. The achievable values of the mixing parameters for each $D^{(1)}_{9n, 3n}$ group are displayed in figure \[fig:caseII\_mixing\_para\].
For the first two smallest $D^{(1)}_{9n, 3n}$ group with $n=1, 2$. The possible values of $\varphi_3$ are $0,\frac{\pi}{3}, \ldots,\frac{5\pi}{3}$ for $n=1$ and $0,\frac{\pi}{6}, \ldots,\frac{11\pi}{6}$ for $n=2$. We find that agreement with experimental data can be achieved for $\varphi_3=0$ or $\pi$. Due to symmetry relation in Eq. , $\varphi_3=0$ and $\varphi_3=\pi$ should give rise to the same predictions for the mixing parameters. Therefore it is sufficient to focus on $\varphi_3=0$, and the best fitting results are listed in table \[tab:caseII\_n12\]. Notice that all the three CP phases are predicted to take CP conserving values $\{\delta_{CP}, \alpha_{21}, \alpha_{31}\}\subseteq\{0, \pi\}$. The same conclusion can be drawn from figure \[fig:caseII\_mixing\_para\].
[|c|c|c|c|c|c|c|c|c|c|]{}\
\
& $\varphi_3$ & $\theta_{bf}$ & $\chi^2_{min}$ & $\sin^2\theta_{13}$ & $\sin^2\theta_{12}$ & $\sin^2\theta_{23}$ & $|\sin\delta_{CP}|$ & $|\sin\alpha_{21}|$ & $|\sin\alpha^{\prime}_{31}|$\
& & $0.433$ & $27.807$ & $0.0246$& $0.256$ & $0.578$ & & &\
& & $(0.435)$ & $(10.086)$ & $(0.0242)$ & $(0.256)$ & $(0.579)$ &&&\
& & $0.436$ & $9.865$ & $0.0238$ & $0.256$ & $0.421$ & & &\
& & $(0.434)$ & $(10.455)$ & $(0.0244)$ & $(0.256)$& $(0.422)$ & & &\
![\[fig:caseII\_mixing\_para\]The possible values of $\sin^2\theta_{12}$, $\sin\theta_{13}$, $\sin^2\theta_{23}$, $\left|\sin\delta_{CP}\right|$, $\left|\sin\alpha_{21}\right|$ and $\left|\sin\alpha^{\prime}_{31}\right|$ with respect to $n$ for the mixing pattern $U^{II, 1}_{\text{PMNS}}$ and $U^{II, 2}_{\text{PMNS}}$ in the case II, where the three lepton mixing angles are required to be within the experimentally preferred $3\sigma$ ranges. The $1\sigma$ and $3\sigma$ regions of the three neutrino mixing angles are adapted from global fit [@Capozzi:2013csa]. Here we take the $3\sigma$ lower limit of $\sin^2\theta_{12}$ to be 0.254 instead of 0.259 given by Ref. [@Capozzi:2013csa].](ParameterCaseII_New.pdf){width="99.00000%"}
As regards the neutrinoless double beta decay, both $U^{II, 1}_{\text{PMNS}}$ and $U^{II, 2}_{\text{PMNS}}$ yield the same effective Majorana mass: $$|m_{ee}|=\frac{1}{4}\Big|m_1(\sin\theta+\sqrt{2} e^{i\varphi_{3}} \cos\theta)^2+q_1m_2+q_2m_3(\cos\theta-\sqrt{2} e^{i \varphi_{3}} \sin\theta)^2\Big|$$ with $q_1, q_2=\pm1$. We show the predicted values of $|m_{ee}|$ in figure \[fig:mee\_CaseII\]. Notice that for IO spectrum $|m_{ee}|$ can be either 0.0233eV or 0.0483eV which are accessible to the next generation $0\nu\beta\beta$ experiments. In the case of NO spectrum, $|m_{ee}|$ strongly depends on the lightest neutrino mass and CP parity, and it can be vanishing for certain values of the lightest neutrino mass.
![\[fig:mee\_CaseII\]The possible values of the effective Majorana mass $|m_{ee}|$ as a function of the lightest neutrino mass in the case II. The red (blue) dashed lines indicate the most general allowed regions for IO (NO) neutrino mass spectrum obtained by varying the mixing parameters over the $3\sigma$ ranges [@Capozzi:2013csa]. The orange (cyan) areas denote the achievable values of $|m_{ee}|$ in the limit of $n\rightarrow\infty$ assuming IO (NO) spectrum. The purple and green regions are the theoretical predictions for the $D^{(1)}_{9n ,3n}$ group with $n=2$. Notice that the purple (green) region overlaps the orange (cyan) one. The present most stringent upper limits $|m_{ee}|<0.120$ eV from EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] is shown by horizontal grey band. The vertical grey exclusion band represents the current bound coming from the cosmological data of $\sum m_i<0.230$ eV at $95\%$ confidence level obtained by the Planck collaboration [@Ade:2013zuv]. ](meeII_v1.pdf){width="0.60\linewidth"}
()
: $G_{l}=\left\langle ac^{s}d^{t}\right\rangle$, $G_{\nu}=Z^{c^{9n/2}}_2$, $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{\gamma}d^{\delta})\right\}$\
In this case, $n$ should be even in order to have a $Z_2$ subgroup generated by $c^{9n/2}$. The neutrino mass matrix invariant under the assumed residual symmetry is found to take the form $$m_{\nu}=
\begin{pmatrix}
m_{11} e^{-\frac{i2 \pi \gamma }{9 n}} & ~m_{12} e^{-\frac{i \pi (2 \gamma -3 \delta )}{9n}} ~& 0 \\
m_{12} e^{-\frac{i \pi (2 \gamma -3 \delta )}{9n}} &~m_{22} e^{-\frac{i 2\pi (\gamma -3 \delta )}{9 n}} ~& 0 \\
0 &~ 0 ~& m_{33} e^{\frac{i2 \pi (2 \gamma -3 \delta )}{9 n}}
\end{pmatrix}\,,$$ where $m_{11}$, $m_{12}$, $m_{13}$ and $m_{22}$ are real. It is diagonalized by the unitary matrix $$\label{eq:unu_c9n/2}
U_{\nu}=
\begin{pmatrix}
e^{\frac{i\pi\gamma }{9n}}\cos\theta &~ e^{\frac{i\pi\gamma }{9n}}\sin\theta ~& 0 \\
-e^{\frac{i \pi (\gamma -3 \delta )}{9n}}\sin\theta &~ e^{\frac{i \pi (\gamma -3 \delta )}{9n}}\cos\theta ~& 0 \\
0 &~ 0 ~& e^{-\frac{i \pi (2 \gamma -3 \delta )}{9n}}
\end{pmatrix}Q_{\nu}\,,$$ with the rotation angle $\theta$ satisfying $$\tan2\theta=\frac{2m_{12}}{m_{22}-m_{11}}\,.$$ The light neutrino masses are $$\begin{aligned}
\nonumber&&m_1=\frac{1}{2}\left|m_{11}+m_{22}-\frac{m_{22}-m_{11}}{\cos2\theta}\right|,\\
\nonumber&&m_2=\frac{1}{2}\left|m_{11}+m_{22}+\frac{m_{22}-m_{11}}{\cos2\theta}\right|,\\
\nonumber&&m_3=\left|m_{33}\right|\,.\end{aligned}$$ As the residual flavor symmetry in the charged lepton sector is $G_{l}=\left\langle ac^{s}d^{t}\right\rangle$, the charged lepton diagonalization matrix is $U^{(3)}_{l}$ shown in table \[tab:cle\_diagonal\_matrix\]. Thus the lepton mixing matrix is determined to be $$\label{eq:PMNS_case_III}
U^{III}_{\text{PMNS}}=\frac{1}{\sqrt{3}}
\begin{pmatrix}
\cos\theta-e^{i\varphi_4} \sin\theta &~ 1 ~&\sin\theta+ e^{i\varphi_4} \cos \theta \\
\omega \cos\theta-\omega^2e^{i\varphi_4}\sin\theta &~ 1 ~& \omega\sin\theta+\omega^2 e^{i\varphi_4}\cos\theta \\
\omega^2\cos\theta-\omega e^{i\varphi_4} \sin\theta &~ 1 ~& \omega^2\sin\theta+\omega e^{i\varphi_4} \cos\theta
\end{pmatrix}\text{diag}(e^{i\varphi_5}, 1, e^{i\varphi_5})Q_{\nu}\,,$$ where $$\varphi_4=\frac{2s-6t-3\delta}{9n}\pi,\qquad \varphi_{5}=\frac{2s+3\gamma-3\delta}{9n}\pi\,.$$ Notice that $\varphi_4$ and $\varphi_{5}$ are not completely independent, and they can take the following discrete values: $$\begin{aligned}
\nonumber&&\varphi_4~(\mathrm{mod}~2\pi)=0, \frac{1}{9n}\pi, \frac{2}{9n}\pi, \ldots, \frac{18n-1}{9n}\pi\,,\\
&&\varphi_4-\varphi_{5}~(\mathrm{mod}~2\pi)=0, \frac{1}{3n}\pi, \frac{2}{3n}\pi, \ldots, \frac{6n-1}{3n}\pi\,.\end{aligned}$$ We easily see that one column of the PMNS matrix is $(1, 1, 1)^{T}/\sqrt{3}$ which can only be the second column vector in order to accommodate the experimental data of lepton mixing angles. The permutations of the PMNS matrix which leave the second column unchanged don’t lead to physically different results. From Eq. we can extract the mixing angles $$\begin{aligned}
\nonumber&&\sin^2\theta_{13}=\frac{1}{3}\left[1+\sin2\theta\cos\varphi_4\right]\,,\\
\nonumber&&\sin^2\theta_{12}=\frac{1}{2-\sin2\theta\cos\varphi_4}\,,\\
\label{eq:mixing_para_caseIII}&&\sin^2\theta_{23}=\frac{1-\sin2\theta\sin\left(\varphi_4+\pi/6\right)}{2-\sin2\theta\cos\varphi_4}\,.\end{aligned}$$ Then we can derive the following sum rules among the mixing angles
$$\begin{aligned}
\label{eq:corre_caseIII_1}&&\hskip3cm 3\sin^2\theta_{12}\cos^2\theta_{13}=1,\\
\label{eq:corre_caseIII_2}&&
(1-3\sin^2\theta_{13})\tan\varphi_4+\sqrt{3}\cos^2\theta_{13}\cos2\theta_{23}=0\,.\end{aligned}$$
The correlation of Eq. yields $\sin^2\theta_{12}\simeq0.341$ for the best fit value $\sin^2\theta_{13}=0.0234$ [@Capozzi:2013csa]. Inputting the $3\sigma$ ranges of the atmospheric as well reactor mixing angles in Eq. , we find the phase difference $\varphi_4$ should vary in the interval $$\label{eq:varphi4_range}\varphi_4\in\left[0,0.138\pi\right]\cup\left[0.862\pi,1.138\pi\right]\cup\left[1.862\pi,2\pi\right]\,.$$ As shown in Eq. , all the three mixing angles are expressed in terms of $\varphi_4$ and $\theta$. The contour regions for $\sin^2\theta_{ij}$ in the plane of $\varphi_4$ and $\theta$ are displayed in figure \[fig:caseIII\_contour\_mixing\_para\]. One can see that agreement with experimental data can be achieved for appropriate values of $\varphi_4$ and $\theta$.
![\[fig:caseIII\_contour\_mixing\_para\] The contour regions of the three mixing angles in the case III. The red, blue and green areas denote the predictions for $\sin^2\theta_{13}$, $\sin^2\theta_{12}$ and $\sin^2\theta_{23}$ respectively. The allowed $1\sigma$, $2\sigma$ and $3\sigma$ regions of each mixing angle are represented by different shadings. The best fit values of the mixing angles are indicated by dotted lines. Note that both the $1\sigma$ range and the best fit value of $\theta_{12}$ can not be achieved in this case because of the sum rule in Eq. .](C3regionAllNH_v2.pdf){width="50.00000%"}
Furthermore, the CP invariants are given by $$\begin{aligned}
\nonumber && \left|J_{CP}\right|=\frac{1}{6 \sqrt{3}}\left|\cos 2 \theta\right| \,, \\
\nonumber && \left|I_1\right|=\frac{1}{9}\left|\cos^2\theta \sin2\varphi_5-\sin2\theta \sin(\varphi_4+2\varphi_5)+\sin^2\theta\sin(2\varphi_4+2\varphi_5)\right| \,, \\
&& \left|I_2\right|= \frac{1}{9}\left|\cos2\theta\sin2\varphi_4\right| \,.\end{aligned}$$ Thus both $\delta_{CP}$ and $\alpha^{\prime}_{31}$ only depend on $\varphi_4$ and $\theta$ while the Majorana phase $\alpha_{21}$ is dependent on all the three parameters $\varphi_4$, $\varphi_5$ and $\theta$. The predictions for $|\sin\delta_{CP}|$ and $|\sin\alpha^{\prime}_{31}|$ are shown in the plane $\theta$ versus $\varphi_4$ in figure \[fig:caseIII\_contour\_CP\_para\]. One can see that all values of the CP phases are possible in the regions where the lepton mixing angles are compatible with the experimental data at $3\sigma$ level. Moreover, the possible values of the mixing angles and CP phases for each $D^{(1)}_{9n, 3n}$ group until $n=50$ are plotted in figure \[fig:caseIII\_mixing\_para\].
![\[fig:caseIII\_contour\_CP\_para\]The contour plots of $|\sin\delta_{CP}|$ and $|\sin\alpha^{\prime}_{31}|$ in the case III. The black areas represent the regions in which the lepton mixing angles are compatible with experimental data at $3\sigma$ level, and it can be read out from figure \[fig:caseIII\_contour\_mixing\_para\].](C3DeltaCPNH_v2.pdf "fig:"){width="49.50000%"} ![\[fig:caseIII\_contour\_CP\_para\]The contour plots of $|\sin\delta_{CP}|$ and $|\sin\alpha^{\prime}_{31}|$ in the case III. The black areas represent the regions in which the lepton mixing angles are compatible with experimental data at $3\sigma$ level, and it can be read out from figure \[fig:caseIII\_contour\_mixing\_para\].](C3Alpha31NH_v2.pdf "fig:"){width="49.50000%"}
![\[fig:caseIII\_mixing\_para\]The possible values of $\sin^2\theta_{12}$, $\sin\theta_{13}$, $\sin^2\theta_{23}$, $\left|\sin\delta_{CP}\right|$, $\left|\sin\alpha_{21}\right|$ and $\left|\sin\alpha^{\prime}_{31}\right|$ with respect to $n$ for the mixing pattern $U^{III}_{\text{PMNS}}$ in the case III, where the three lepton mixing angles are required to be within the experimentally preferred $3\sigma$ ranges. The $1\sigma$ and $3\sigma$ regions of the three neutrino mixing angles are adapted from global fit [@Capozzi:2013csa].](ParameterCaseIII_New.pdf){width="99.00000%"}
Then we proceed to study the phenomenologically viable mixing patterns which can be derived from the $D^{(1)}_{9n, 3n}$ group with $n=2$. Note that the index $n$ has to be even in this case. We can check that the PMNS matrix given by Eq. has the following symmetry properties $$\begin{aligned}
\nonumber&&U^{III}_{\text{PMNS}}(\theta, \pi+\varphi_4, \varphi_5)=U^{III}_{\text{PMNS}}(-\theta, \varphi_4, \varphi_5)\text{diag}(1, 1, -1),\\
&&U^{III}_{\text{PMNS}}(\theta, \varphi_4, \pi/2+\varphi_5)=U^{III}_{\text{PMNS}}(\theta, \varphi_4, \varphi_5)\text{diag}(i, 1, i)\,,\end{aligned}$$ where the diagonal matrix on the right-handed side can be absorbed into $Q_{\nu}$. That is to say, both $U^{III}_{\text{PMNS}}(\theta, \pi+\varphi_4, \varphi_5)$ and $U^{III}_{\text{PMNS}}(\theta, \varphi_4, \pi/2+\varphi_5)$ give rise to the same predictions for the lepton mixing parameters as $U^{III}_{\text{PMNS}}(\theta, \varphi_4, \varphi_5)$ up to redefinition of the free parameter $\theta$. Hence we can take the fundamental intervals of $\varphi_4$ and $\varphi_5$ to be $[0, \pi)$ and $[0, \pi/2)$ respectively. The allowed values of $\varphi_4$ are $0$, $\pi/18$, $\pi/9$, $\ldots$, $17\pi/9$, $35\pi/18$. However, only $\varphi_4~(\mathrm{mod}~\pi)=0$, $\pi/18$, $\pi/9$, $8\pi/9$ and $17\pi/18$ are within the range of Eq. such that they can give a good fit to the experimental data. The results of the $\chi^2$ analysis are summarized in table \[tab:caseIII\_n12\]. Notice that the best fitting values of the mixing angles and $|\sin\delta_{CP}|$, $|\sin\alpha^{\prime}_{31}|$ are dependent on $\varphi_4$ while the best fitting value of $|\sin\alpha_{21}|$ depends on $\varphi_4$ as well as $\varphi_5$. The mixing patterns with the same $\varphi_4$ but different $\varphi_5$ are expected to be distinguished by some rare processes which are sensitive to the Majorana phases such as the neutrinoless double decay and the radiative emission of neutrino pair in atoms [@Yoshimura:2006nd]. In this case, the effective Majorana mass $|m_{ee}|$ is predicted to be $$|m_{ee}|=\frac{1}{3}\Big|m_1(\cos\theta-e^{i\varphi_4} \sin\theta)^2+q_1m_2e^{-2i\varphi_5}+q_2m_3(\sin\theta+ e^{i\varphi_4}\cos \theta)^2\Big|\,,$$ where $q_1, q_2=\pm1$. The numerical results are shown in figure \[fig:mee\_CaseIII\].
[|c|c|c|c|c|c|c|c|c|c|]{}\
\
$\varphi_4$ & $\varphi_5$ & $\theta_{bf}$ & $\chi^2_{min}$ & $\sin^2\theta_{13}$ & $\sin^2\theta_{12}$ & $\sin^2\theta_{23}$ & $|\sin\delta_{CP}|$ & $|\sin\alpha_{21}|$ & $|\sin\alpha^{\prime}_{31}|$\
& &$2.168$ &$7.480$ & $0.0233$& $0.341$ & $0.5$ & &$|\sin2\varphi_5|$ &\
&& ($2.166$) & ($3.987$) & ($0.0238$) & ($0.341$) & ($0.5$) && ($|\sin2\varphi_5|$)&\
& & & & & & & & $0.525$ &\
&& & & & & & & ($0.526$) &\
& &&&&& & & $0.9996$ &\
& &&&&& & & ($0.9996$) &\
& &&&&& & & $0.474$ &\
& &&&&& & & ($0.474$) &\
& & & & & & & & $0.474$ &\
& & & & & & & & ($0.474$) &\
& &&&&& & & $0.9996$ &\
& &&&&& & & ($0.9996$) &\
& &&&&& & & $0.525$ &\
& &&&&& & & ($0.526$) &\
& & & & & & & & $0.853$ &\
& & & & & & & & ($0.852$) &\
& &&&&& & & $0.879$ &\
& &&&&& & & ($0.879$) &\
& &&&&& & & $0.0255$ &\
& &&&&& & & ($0.0272$) &\
& & & & & & & & $0.0253$ &\
& & & & & & & & $(0.0274$) &\
& &&&&& & & $0.878$ &\
& &&&&& & & ($0.879$) &\
& &&&&& & & $0.853$ &\
& &&&&& & & ($0.852$) &\
![\[fig:mee\_CaseIII\]The possible values of the effective Majorana mass $|m_{ee}|$ as a function of the lightest neutrino mass in the case III. The red (blue) dashed lines indicate the most general allowed regions for IO (NO) neutrino mass spectrum obtained by varying the mixing parameters over the $3\sigma$ ranges [@Capozzi:2013csa]. The orange (cyan) areas denote the achievable values of $|m_{ee}|$ in the limit of $n\rightarrow\infty$ assuming IO (NO) spectrum. The purple and green regions are the theoretical predictions for the $D^{(1)}_{9n ,3n}$ group with $n=2$. Notice that the purple (green) region overlaps the orange (cyan) one. The present most stringent upper limits $|m_{ee}|<0.120$ eV from EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] is shown by horizontal grey band. The vertical grey exclusion band represents the current bound coming from the cosmological data of $\sum m_i<0.230$ eV at $95\%$ confidence level obtained by the Planck collaboration [@Ade:2013zuv]. ](meeIII_v1.pdf){width="0.60\linewidth"}
()
: $G_{l}=\left\langle ac^{s}d^{t}\right\rangle$, $G_{\nu}=Z^{c^{9n/2}}_2$, $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(a^{2}bc^{\gamma})\right\}$\
This case differs from the case III in the residual CP transformation of the neutrino sector. The group index $n$ has to be an even integer as well. In the same way, the neutrino mass matrix invariant under the assumed remnant symmetry is determined to be $$m_{\nu}=\begin{pmatrix}
m_{11} e^{- i \left(\theta+\frac{4 \pi\gamma}{9n}\right)}
&~ m_{12}e^{-\frac{2 i \pi \gamma }{9 n}} ~& 0 \\
m_{12}e^{-\frac{2 i\pi\gamma}{9 n}} &~ m_{11}e^{i\theta} ~& 0 \\
0 &~ 0 ~& m_{33}e^{\frac{4i\pi\gamma}{9n}}
\end{pmatrix}\,,$$ where $m_{11}$, $m_{12}$, $m_{33}$ and $\theta$ are real parameters. The unitary transformation $U_{\nu}$ which diagonalizes $m_{\nu}$ is of the form $$U_{\nu}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
e^{ i \left(\frac{\theta}{2}+\frac{2\pi\gamma}{9n}\right)} &~ e^{ i \left(\frac{\theta}{2}+\frac{2\pi\gamma }{9n}\right)} ~& 0 \\
-e^{-i\frac{\theta}{2}} &~ e^{-i\frac{\theta}{2}} ~& 0 \\
0 & 0 & \sqrt{2}e^{-\frac{2 i \pi \gamma }{9 n}}
\end{pmatrix}Q_{\nu}\,.$$ The light neutrino masses are given by $$m_1=\left|m_{11}-m_{12}\right|,\quad m_2=\left|m_{11}+m_{12}\right|,\quad m_{3}=\left|m_{33}\right|\,.$$ Then we find that the PMNS matrix takes the form $$\label{eq:PMNS_case_IV}
U^{IV}_{\text{PMNS}}=\frac{1}{\sqrt{3}}\left(
\begin{array}{ccc}
\sqrt{2}e^{i\varphi_7}\sin\left(\varphi_6+\frac{\theta}{2}\right) ~&~ 1
~&~ \sqrt{2}e^{i\varphi_7}\cos\left(\varphi_6+\frac{\theta}{2}\right) \\
\sqrt{2}e^{i\varphi_7}\cos\left(\varphi_6+\frac{\theta}{2}+\frac{\pi}{6}\right)
~&~1~&~-\sqrt{2}e^{i\varphi_7}\sin\left(\varphi_6+\frac{\theta}{2}+\frac{\pi}{6}\right)\\
-\sqrt{2}e^{i\varphi_7}\cos\left(\varphi_6+\frac{\theta}{2}-\frac{\pi}{6}\right)
~&~1~&~\sqrt{2}e^{i\varphi_7}\sin\left(\varphi_6+\frac{\theta}{2}-\frac{\pi}{6}\right)
\end{array}
\right)Q_{\nu}\,,$$ with $$\varphi_6=\frac{\gamma-s+3t}{9n}\pi ,\qquad
\varphi_7=\frac{\gamma+s-t}{3n}\pi\,.$$ Obviously the second column of the PMNS matrix is $\left(1/\sqrt{3}, 1/\sqrt{3}, 1/\sqrt{3}\right)^{T}$ as well. The mixing parameters extracted from Eq. are: $$\begin{aligned}
\nonumber&&\sin^2\theta_{13}=\frac{1}{3}\left[1+\cos(\theta+2\varphi_6)\right],\quad
\sin^2\theta_{12}=\frac{1}{2-\cos(\theta+2\varphi_6)}\,,\\
\nonumber&&\sin^2\theta_{23}=\frac{1+\sin\left(\theta+2\varphi_6-\pi/6\right)}{2-\cos(\theta+2\varphi_6)}\,,\\
\label{eq:mixing_parameters_case_XIII}&&\sin\delta_{CP}=\sin\alpha_{31}=0,\qquad
\left|\sin\alpha_{21}\right|=|\sin(2\varphi_7)|\,.\end{aligned}$$ Notice that the mixing angles depend on the combination $\theta+2\varphi_6$ so that the values of the parameter $\varphi_6$ is irrelevant. Moreover, we can see that the mixing angles fulfill the following sum rules $$3\cos^2\theta_{13}\sin^2\theta_{12}=1,\qquad
2\sin^2\theta_{23}=1\pm\tan\theta_{13}\sqrt{2-\tan^2\theta_{13}}\,.$$ The $3\sigma$ range of the reactor angle $0.0176\leq\sin^2\theta_{13}\leq0.0295$ [@Capozzi:2013csa] can be reproduced for $$\theta+2\varphi_6\in[0.865\pi, 0.896\pi]\cup[1.104\pi, 1.135\pi]\,.$$ Thus the solar and atmospheric mixing angles are determined to be within the intervals $$\begin{aligned}
\nonumber&&\hskip2.3cm 0.339\leq\sin^2\theta_{12}\leq0.343\,,\\
&&0.378\leq\sin^2\theta_{23}\leq0.406,~~\mathrm{or}~~0.594\leq\sin^2\theta_{23}\leq0.622\,,\end{aligned}$$ which are in accordance with the experimentally measured values. Note that the atmospheric angle $\theta_{23}$ deviates from maximal mixing. These predictions can be test by JUNO [@An:2015jdp] and forthcoming long baseline neutrino oscillation experiments. As regards the CP violating phases, both Dirac phase $\delta_{CP}$ and the Majorana phase $\alpha_{31}$ are conserved while another Majorana phase $\alpha_{21}$ can take the discrete values of $0$, $\frac{2}{3n}\pi$, $\frac{4}{3n}\pi$, $\ldots$, $\frac{6n-2}{3n}\pi$. In this case, the effective Majorana mass $|m_{ee}|$ takes a simple form, $$|m_{ee}|=\frac{1}{3}\Big|2m_1\sin^2\left(\varphi_6+\theta/2\right)+q_1m_2e^{2i\varphi_7}+2q_2m_3\cos^2\left(\varphi_6+\theta/2\right)\Big|\,.$$ The predictions on $|m_{ee}|$ are plotted in figure \[fig:mee\_CaseIV\]. For the IO spectrum and $n=2$, we find $|m_{ee}|$ can take a few discrete values and these results can be tested in forthcoming $0\nu\beta\beta$ experiments.
![\[fig:mee\_CaseIV\]The possible values of the effective Majorana mass $|m_{ee}|$ as a function of the lightest neutrino mass in the case IV. The red (blue) dashed lines indicate the most general allowed regions for IO (NO) neutrino mass spectrum obtained by varying the mixing parameters over the $3\sigma$ ranges [@Capozzi:2013csa]. The orange (cyan) areas denote the achievable values of $|m_{ee}|$ in the limit of $n\rightarrow\infty$ assuming IO (NO) spectrum. The purple and green regions are the theoretical predictions for the $D^{(1)}_{9n ,3n}$ group with $n=2$. Notice that the purple (green) region overlaps the orange (cyan) one. The present most stringent upper limits $|m_{ee}|<0.120$ eV from EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] is shown by horizontal grey band. The vertical grey exclusion band represents the current bound coming from the cosmological data of $\sum m_i<0.230$ eV at $95\%$ confidence level obtained by the Planck collaboration [@Ade:2013zuv]. ](meeIV_v1.pdf){width="0.60\linewidth"}
\[sec:Z2xCP\_charged\_lepton\]Lepton mixing from a variant of semidirect approach
=================================================================================
In contrast with semidirect approach discussed in section \[sec:Z2xCP\_neutrino\], we shall assume that the original symmetry $D^{(1)}_{9n,3n}\rtimes H_{CP}$ is broken down to $Z_2\times CP$ in the charged lepton sector, and the residual symmetry of the neutrino mass matrix is $K_{4}\rtimes H^{\nu}_{CP}$, where $K_4$ is a Klein subgroup of $D^{(1)}_{9n,3n}$. Since each order 2 element of the $D^{(1)}_{9n, 3n}$ group is conjugate to either $bd^{x}$ or $c^{9n/2}$, as shown in Eq. and Eq. , it is sufficient to discuss the representative remnant symmetry $G_{l}=Z^{bd^{x}}_2, Z^{c^{9n/2}}_2$ and $G_{\nu}=K^{(c^{9n/2},d^{3n/2})}_4$, $K^{(d^{3n/2},bd^{x})}_4$, $K^{(c^{9n/2}d^{3n/2},abc^{3y}d^{y})}_4$ and $K^{(c^{9n/2},a^2bc^{3z}d^{2z})}_4$. In this variant of the semidirect approach, the PMNS matrix turns out to depend on only one real continuous parameter besides the discrete parameters specifying the remnant symmetries, and one row of the PMNS matrix would be completely fixed by the assumed remnant symmetries. The fixed row vectors for different representative residual flavor symmetries are listed in table \[tab:PMNS\_row\]. We find that essentially only one type of residual symmetry with $(G_{\nu}, G_{l})=(K^{(c^{9n/2},a^2bc^{3z}d^{2z})}_4, Z^{bd^{x^{\prime}}}_2)$ is phenomenologically viable in this scenario.
[|c||c|c|]{} & &\
\[-0.15in\] & $G_{l}=Z^{bd^{x^{\prime}}}_2$ & $G_{l}=Z^{c^{9n/2}}_2$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_{\nu}=K^{(c^{9n/2},d^{3n/2})}_4$ & $\frac{1}{\sqrt{2}}\left(\begin{array}{c}
0\\
-1\\
1
\end{array}
\right)^{T}$& $\left(\begin{array}{c}
0\\
0\\
1
\end{array}\right)^{T}$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_{\nu}=K^{(d^{3n/2},bd^{x})}_4$ & $\left(\begin{array}{c}
0\\
\cos\left(\frac{x+x'}{3n}\pi\right)\\
-i\sin\left(\frac{x+x'}{3n}\pi\right)
\end{array}
\right)^{T}$& $\frac{1}{\sqrt{2}}\left(\begin{array}{c}
0\\
1\\
1
\end{array}
\right)^{T}$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_{\nu}=K^{(c^{9n/2}d^{3n/2},abc^{3y}d^{y})}_4$ & $\frac{1}{2}\left(\begin{array}{c}
-\sqrt{2}\\
1\\
1
\end{array}
\right)^{T}$& $\frac{1}{\sqrt{2}}\left(\begin{array}{c}
0\\
1\\
1
\end{array}
\right)^{T}$\
& &\
\[-0.15in\] & &\
\[-0.15in\]
$G_{\nu}=K^{(c^{9n/2},a^2bc^{3z}d^{2z})}_4$ & $\frac{1}{2}\left(\begin{array}{c}
-\sqrt{2}\\
1 \\
1
\end{array}\right)^{T}$ & $\left(\begin{array}{c}
1\\
0\\
0
\end{array}
\right)^{T}$\
& &\
\[-0.15in\]
()
: $G_{l}=\left\{1,bd^{x}\right\}$, $X_{l\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{2x+2\delta+3n\tau}d^{\delta}), \rho_{\mathbf{r}}(bc^{2x+2\delta+3n\tau}d^{x+\delta})\right\}$, $G_{\nu}=K^{(c^{9n/2},a^2bc^{3z}d^{2z})}_4$ and $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{\gamma}d^{-2z}),
\rho_{\mathbf{r}}(a^2bc^{\gamma})\right\}$\
Here we would like to recall that the residual CP transformations are determined by the restricted consistency conditions in Eqs. (\[eq:res\_cons\_cond\_clep\], \[eq:res\_cons\_cond\_nu\]). The parameter $n$ should be even in order to have a residual Klein subgroup. The phenomenological constraints of the residual flavor symmetry $G_{\nu}=K^{(c^{9n/2},a^2bc^{3z}d^{2z})}_4$ as well as the residual CP transformation $X_{\nu\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{\gamma}d^{-2z}),
\rho_{\mathbf{r}}(a^2bc^{\gamma})\right\}$ has been studied in section \[sec:direct\_approach\]. The light neutrino mass matrix $m_{\nu}$ and its diagonalization matrix $U_{\nu}$ are found to be given by Eq. and Eq. respectively. Then we proceed to the charged lepton sector. The invariance of the charged lepton mass matrix under the residual symmetry $G_{l}=\left\{1,bd^{x}\right\}$ and $X_{l\mathbf{r}}=\left\{\rho_{\mathbf{r}}(c^{2x+2\delta+3n\tau}d^{\delta}), \rho_{\mathbf{r}}(bc^{2x+2\delta+3n\tau}d^{x+\delta})\right\}$ implies that the hermitian matrix $m^{\dagger}_{l}m_{l}$ has to fulfill the invariant condition of Eq. , i.e. $$\begin{aligned}
\nonumber&&\hskip1.8cm\rho^{\dagger}_{\mathbf{3}}(bd^{x})m^{\dagger}_{l}m_{l}\rho_{\mathbf{3}}(bd^{x})=m^{\dagger}_{l}m_{l},\\
&&\rho^{\dagger}_{\mathbf{3}}(c^{2x+2\delta+3n\tau}d^{\delta})m^{\dagger}_{l}m_{l}\rho_{\mathbf{3}}(c^{2x+2\delta+3n\tau}d^{\delta})=\big(m^{\dagger}_{l}m_{l}\big)^{*}\,,\end{aligned}$$ which lead to $$m^{\dagger}_{l}m_{l}=\left(
\begin{array}{ccc}
\widetilde{\mathfrak{m}}_{11} &~ \widetilde{\mathfrak{m}}_{12}e^{i\pi\frac{\delta}{3n}} &~ \widetilde{\mathfrak{m}}_{12}e^{i\pi\frac{2x+\delta}{3n}} \\
\widetilde{\mathfrak{m}}_{12}e^{-i\pi\frac{\delta}{3n}} &~ \widetilde{\mathfrak{m}}_{22} &~ \widetilde{\mathfrak{m}}_{23}e^{i\pi\frac{2x}{3n}}\\
\widetilde{\mathfrak{m}}_{12}e^{-i\pi\frac{2x+\delta}{3n}} &~ \widetilde{\mathfrak{m}}_{23}e^{-i\pi\frac{2x}{3n}} &~ \widetilde{\mathfrak{m}}_{22}
\end{array}
\right)\,,$$ where $\widetilde{\mathfrak{m}}_{11}$, $\widetilde{\mathfrak{m}}_{12}$, $\widetilde{\mathfrak{m}}_{22}$ and $\widetilde{\mathfrak{m}}_{23}$ are real, and they have dimension of squared mass. It can be diagonalized by the unitary matrix $$\label{eq:ul_bdx_2}U_{l}=\frac{1}{\sqrt{2}}
\left(
\begin{array}{ccc}
0 &~ -\sqrt{2} \sin\theta &~ \sqrt{2} \cos \theta \\
-e^{i \pi \frac{2 x}{3 n}} &~ e^{-i \pi \frac{\delta }{3n}} \cos \theta &~ e^{-i\pi\frac{\delta}{3n}}\sin\theta \\
1 &~ e^{-i \pi \frac{2 x+\delta }{3 n}} \cos \theta &~ e^{-i \pi \frac{ 2 x+\delta }{3 n}} \sin \theta
\end{array}
\right)\,,$$ with the angle $\theta$ satisfying $$\tan2\theta=\frac{2\sqrt{2}\,\widetilde{\mathfrak{m}}_{12}}{\widetilde{\mathfrak{m}}_{11}-\widetilde{\mathfrak{m}}_{22}-\widetilde{\mathfrak{m}}_{23}}\,.$$ The squared charged lepton masses are determined to be of the form $$\begin{aligned}
\nonumber&&m^2_{l_1}=\widetilde{\mathfrak{m}}_{22}-\widetilde{\mathfrak{m}}_{23}\,,\\
\nonumber&&m^2_{l_2}=\frac{1}{2}\left[\widetilde{\mathfrak{m}}_{11}+\widetilde{\mathfrak{m}}_{22}+\widetilde{\mathfrak{m}}_{23}-\frac{\widetilde{\mathfrak{m}}_{11}-\widetilde{\mathfrak{m}}_{22}-\widetilde{\mathfrak{m}}_{23}}{\cos 2\theta}\right]\,,\\
&&m^2_{l_3}=\frac{1}{2}\left[\widetilde{\mathfrak{m}}_{11}+\widetilde{\mathfrak{m}}_{22}+\widetilde{\mathfrak{m}}_{23}+\frac{\widetilde{\mathfrak{m}}_{11}-\widetilde{\mathfrak{m}}_{22}-\widetilde{\mathfrak{m}}_{23}}{\cos 2\theta}\right]\,.\end{aligned}$$ Notice that the order of the masses $m^2_{l_1}$, $m^2_{l_2}$ and $m^2_{l_3}$ can not be pinned down by remnant symmetry, therefore the matrix $U_{l}$ in Eq. is determined up to permutations and phases of its column vectors. The lepton flavor mixing originates from the mismatch between the unitary transformations $U_{l}$ in Eq. and $U_{\nu}$ in Eq. , and the PMNS matrix can take the form $$\label{eq:PMNS_case_V_1}U^{V, 1}_{\text{PMNS}}=\frac{1}{2}\left(
\begin{array}{ccc}
\sin\theta+\sqrt{2}e^{i\varphi_8}\cos\theta ~&~
\sin\theta-\sqrt{2}e^{i\varphi_8}\cos\theta ~&~
\sqrt{2}e^{i\varphi_9}\sin\theta \\
1 ~&~ 1 ~&~ -\sqrt{2}e^{i\varphi_9} \\
\cos\theta-\sqrt{2}e^{i\varphi_8}\sin\theta ~&~
\cos\theta+\sqrt{2}e^{i\varphi_8}\sin\theta ~&~
\sqrt{2}e^{i\varphi_9}\cos\theta
\end{array}
\right)\,,$$ or $$\label{eq:PMNS_case_V_2}U^{V, 2}_{\text{PMNS}}=\frac{1}{2}\left(
\begin{array}{ccc}
\sin\theta+\sqrt{2}e^{i\varphi_8}\cos\theta ~&~
\sin\theta-\sqrt{2}e^{i\varphi_8}\cos\theta ~&~
\sqrt{2}e^{i\varphi_9}\sin\theta \\
\cos\theta-\sqrt{2}e^{i\varphi_8}\sin\theta ~&~
\cos\theta+\sqrt{2}e^{i\varphi_8}\sin\theta ~&~
\sqrt{2}e^{i\varphi_9}\cos\theta \\
1 ~&~ 1 ~&~ -\sqrt{2}e^{i\varphi_9}
\end{array}
\right)\,,$$ where $$\varphi_8=-\frac{2z+\delta}{3n}\pi,\qquad
\varphi_9=\frac{2x-4z-\gamma}{3n}\pi\,.$$ Obviously the values of both $\varphi_8$ and $\varphi_9$ are integer multiple of $\frac{\pi}{3n}$, i.e. $$\varphi_8, \varphi_9~(\mathrm{mod}~2\pi)=0, \frac{1}{3n}\pi, \frac{2}{3n}\pi,\ldots, \frac{6n-1}{3n}\pi\,.$$ The mixing patterns $U^{V, 1}_{\text{PMNS}}$ and $U^{V, 2}_{\text{PMNS}}$ are related through the exchange of the second and third rows in the PMNS mixing matrix. Other permutations of rows and columns don’t lead to new patterns consistent with experimental data. We can extract the following results for the mixing angles, $$\begin{aligned}
\nonumber&&\sin^2\theta_{13}=\frac{1}{2}\sin^2\theta,\qquad
\sin^2\theta_{12}=\frac{1}{2}-\frac{\sqrt{2}\sin2\theta\cos\varphi_8}{3+\cos2\theta},\\
&&\sin^2\theta_{23}=\frac{2}{3+\cos2\theta}~~\mathrm{for}~~U^{V, 1}_{\text{PMNS}},\qquad \sin^2\theta_{23}=\frac{1+\cos2\theta}{3+\cos2\theta}~~\mathrm{for}~~U^{V, 2}_{\text{PMNS}}\,.\end{aligned}$$ The $3\sigma$ range of $\sin^2\theta_{13}$ can be reproduced for $$\theta\in\left[0.060\pi,0.078\pi\right]\cup\left[0.922\pi,0.940\pi\right]\,.$$ We can check that the mixing angles fulfill the following sum rules,
$$\begin{aligned}
\label{eq:correlation_case_V_1}&&\qquad\cos2\theta_{12}=\pm2\tan\theta_{13}\sqrt{1-\tan^2\theta_{13}}\cos\varphi_{8}\,,\\
\label{eq:correlation_case_V_2}&&2\cos^2\theta_{13}\sin^2\theta_{23}=1,\quad \text{or}\quad 2\cos^2\theta_{13}\sin^2\theta_{23}=\cos2\theta_{13}\,,\end{aligned}$$
where the “+” and “$-$” signs are valid for $0<\theta<\pi/2$ and $\pi/2<\theta<\pi$ respectively. In order to accommodate the experimental data on solar and reactor mixing angles, the first sum rule of Eq. implies that the parameter $\varphi_8$ should vary in the interval $$\label{eq:varphi8_cons}\varphi_8\in[0, 0.193\pi]\cup[0.807\pi, 1.193\pi]\cup[1.807\pi, 2\pi]\,.$$ From the correlation of Eq. , we can derive $$\label{eq:theta_23_V}0.509\leq\sin^2\theta_{23}\leq0.515,~~\text{or}~~0.485\leq\sin^2\theta_{23}\leq0.491\,.$$
![\[fig:caseV\_contour\_mixing\_para\] The contour regions of the three mixing angles in the case V. The red, blue and green areas denote the predictions for $\sin^2\theta_{13}$, $\sin^2\theta_{12}$ and $\sin^2\theta_{23}$ respectively. The allowed $1\sigma$, $2\sigma$ and $3\sigma$ regions of each mixing angle are represented by different shadings. The best fit values of the mixing angles are indicated by dotted lines. In this case the atmospheric angle $\theta_{23}$ is predicted to be in the interval of Eq. such that neither the $1\sigma$ range nor its best fit value can be achieved.](C5regionAllNH.pdf){width="60.00000%"}
The contour plots for $\sin^2\theta_{ij}$ is shown in figure \[fig:caseV\_contour\_mixing\_para\]. Since both reactor angle $\theta_{13}$ and the atmospheric angle $\theta_{23}$ only depend on the parameter $\theta$, the corresponding contour regions are horizontal bands. There exist three small regions in which all the three mixing angles are within the experimentally preferred $3\sigma$ ranges. Furthermore, we find the following expressions for the CP invariants, $$\begin{aligned}
\nonumber && \left|J_{CP}\right|=\frac{1}{8 \sqrt{2}}\left|\sin 2 \theta\sin\varphi_8\right|\,,\qquad \left|I_1\right|=\frac{1}{8 \sqrt{2}}\left|(1+3 \cos2 \theta )\sin2\theta \sin \varphi_8\right| \,, \\
&& \left|I_2\right|= \frac{\sin^2\theta}{8}\left|\sqrt{2} \sin2\theta\sin (2\varphi_9-\varphi_8)+2\cos ^2\theta \sin2(\varphi_9-\varphi_8)+\sin^2\theta\sin2 \varphi _9\right| \,,\end{aligned}$$ from which we know that both $\delta_{CP}$ and $\alpha_{21}$ are only dependent on $\theta$ and $\varphi_8$ while the value of $\alpha^{\prime}_{31}$ depends on three parameters $\theta$, $\varphi_8$ and $\varphi_9$. We display the predictions for $|\sin\delta_{CP}|$ and $|\sin\alpha_{21}|$ in the $\varphi_8-\theta$ plane in figure \[fig:caseV\_contour\_CP\_para\]. One can see that both $\delta_{CP}$ and $\alpha_{21}$ can not be maximal if the three mixing angles are required to be consistent with the experimental data.
![\[fig:caseV\_contour\_CP\_para\]The contour plots of $|\sin\delta_{CP}|$ and $|\sin\alpha_{21}|$ in the case V. The black areas represent the regions in which the lepton mixing angles are compatible with experimental data at $3\sigma$ level, and it can be read out from figure \[fig:caseV\_contour\_mixing\_para\].](C5DeltaCPNH.pdf "fig:"){width="49.50000%"} ![\[fig:caseV\_contour\_CP\_para\]The contour plots of $|\sin\delta_{CP}|$ and $|\sin\alpha_{21}|$ in the case V. The black areas represent the regions in which the lepton mixing angles are compatible with experimental data at $3\sigma$ level, and it can be read out from figure \[fig:caseV\_contour\_mixing\_para\].](C5Alpha21NH.pdf "fig:"){width="49.50000%"}
In analogy to previous cases, we numerically study the possible values of the mixing parameters for each $D^{(1)}_{9n, 3n}$ group. We can read from figure \[fig:caseV\_mixing\_para\] that a bit larger $\theta_{12}$ (still in the $3\sigma$ range) is favored with $0.328\leq\sin^2\theta_{12}\leq0.359$, and the atmospheric angle $\sin^2\theta_{23}$ is predicted to be around 0.487 and 0.513. These results can be testable at forthcoming neutrino oscillation facilities. The same conclusions on CP phases are reached as those from figure \[fig:caseV\_contour\_CP\_para\]. We find the upper bounds of $|\sin\delta_{CP}|$ and $|\sin\alpha_{21}|$ are $\left|\sin\delta_{CP}\right|\leq 0.594$ and $\left|\sin\alpha_{21}\right|\leq0.399$ respectively. On the other hand, any value of the Majorana phase $\alpha_{31}$ is possible for large value of $n$.
![\[fig:caseV\_mixing\_para\]The possible values of $\sin^2\theta_{12}$, $\sin\theta_{13}$, $\sin^2\theta_{23}$, $\left|\sin\delta_{CP}\right|$, $\left|\sin\alpha_{21}\right|$ and $\left|\sin\alpha^{\prime}_{31}\right|$ with respect to $n$ for the mixing pattern $U^{V, 1}_{\text{PMNS}}$ and $U^{V, 2}_{\text{PMNS}}$ in the case V, where the three lepton mixing angles are required to be within the experimentally preferred $3\sigma$ ranges. The $1\sigma$ and $3\sigma$ regions of the three neutrino mixing angles are adapted from global fit [@Capozzi:2013csa]. Note that the group index $n$ should be even in this case.](ParameterCaseV_New.pdf){width="99.00000%"}
Now we discuss the lepton mixing patterns which can be obtained from the $D^{(1)}_{9n, 3n}$ group with $n=2$. Note that the smallest $D^{(1)}_{9n, 3n}$ group for $n=1$ doesn’t comprise the required Klein subgroup. The PMNS matrices $U^{V, 1}_{\text{PMNS}}$ and $U^{V, 2}_{\text{PMNS}}$ fulfill the following relations
\[eq:PMNS\_sym\_caseV\] $$\begin{aligned}
\label{eq:PMNS_sym_caseV1}&&U^{V}_{\text{PMNS}}(\theta, \varphi_8+\pi, \varphi_9)=\text{diag}(-1, 1, 1)U^{V}_{\text{PMNS}}(-\theta, \varphi_8, \varphi_9),\\
\label{eq:PMNS_sym_caseV2}&&U^{V}_{\text{PMNS}}(\theta, \varphi_8, \varphi_9+\pi/2)=U^{V}_{\text{PMNS}}(\theta, \varphi_8, \varphi_9)\text{diag}(1, 1, i),\\
\label{eq:PMNS_sym_caseV3}&&U^{V}_{\text{PMNS}}(\theta, \pi-\varphi_8, \varphi_9)=\text{diag}(-1, 1, 1)[U^{V}_{\text{PMNS}}(-\theta, \varphi_8, -\varphi_9)]^{\ast}\,,\end{aligned}$$
where $U^{V}_{\text{PMNS}}$ refers to $U^{V, 1}_{\text{PMNS}}$ and $U^{V, 2}_{\text{PMNS}}$. The diagonal matrices on the left-handed and right-handed sides can be absorbed by the charged lepton fields and $Q_{\nu}$ respectively. Therefore the shifts of $\varphi_8$ into $\varphi_8+\pi$ and $\varphi_9$ into $\varphi_9+\pi/2$ don’t lead to physically new results. For $n=2$ the values of $\varphi_8$ and $\varphi_9$ can be $0$, $\pi/6$, $\pi/3$, $\ldots$, $11\pi/6$. Considering the constraint on the parameter $\varphi_8$ given by Eq. , we find only $\varphi_8~(\mathrm{mod}~\pi)=0$, $\pi/6$ and $5\pi/6$ can describe the data on lepton mixing. The results of our $\chi^2$ analysis are displayed in table \[tab:caseV\_n2\]. Since the mixing angles $\sin^2\theta_{ij}$ and the CP invariants $J_{CP}$ and $I_{1}$ are expressed in terms of $\theta$ and $\varphi_8$, and the parameter $\varphi_9$ only enters into the expression of $I_2$, the relation in Eq. implies that $\varphi_8$ and $\pi-\varphi_8$ give rise to the same best fitting values of mixing parameters except $|\sin\alpha^{\prime}_{31}|$. This is exactly the reason why the numerical results for $\varphi_8=\pi/6$ and $\varphi_8=5\pi/6$ are only different in the values of $|\sin\alpha^{\prime}_{31}|$. Finally we plot the predictions for the effective mass $|m_{ee}|$ with respect to the lightest neutrino mass in figure \[fig:mee\_CaseV\]. One sees that the values of $|m_{ee}|$ are rather close to the lower or upper boundary of the $3\sigma$ region for IO.
[|c|c|c|c|c|c|c|c|c|c|c|]{}\
\
s & $\varphi_8$ & $\varphi_9$ & $\theta_{bf}$ & $\chi^2_{min}$ & $\sin^2\theta_{13}$ & $\sin^2\theta_{12}$ & $\sin^2\theta_{23}$ & $|\sin\delta_{CP}|$ & $|\sin\alpha_{21}|$ & $|\sin\alpha^{\prime}_{31}|$\
& & $0$ & & & & & & & & $0$ ($0$)\
& & $\frac{\pi}{6}$ & & & & & & & & $0.866$ ($0.866$)\
& & $\frac{\pi}{3}$ & & & & & & & & $0.866$ ($0.866$)\
& & $0$ & & & & & & & & $0.786$ ($0.785$)\
& & $\frac{\pi}{6}$ & & & & & & & & $0.142$ ($0.144$)\
& & $\frac{\pi}{3}$ & & & & & & & & $0.928$ ($0.929$)\
& & $0$ & & & & & & & &$0.786$ ($0.785$)\
& & $\frac{\pi}{6}$ & & & & & & & & $0.928$ ($0.929$)\
& & $\frac{\pi}{3}$ & & & & & & & & $0.142$ ($0.144$)\
& & $0$ & & & & & & & & $0$ ($0$)\
& &$\frac{\pi}{6}$& & & & & & & & $0.866$ ($0.866$)\
& & $\frac{\pi}{3}$ & & & & & & & & $0.866$ ($0.866$)\
& & $0$ & & & & & & & & $0.786$($0.785$)\
& &$\frac{\pi}{6}$ & & & & & & & & $0.143$ ($0.144$)\
& & $\frac{\pi}{3}$ & & & & & & & & $0.929$ ($0.929$)\
& & $0$ & & & & & & & & $0.786$ ($0.785$)\
& &$\frac{\pi}{6}$ & & & & & & & & $0.929$ ($0.929$)\
& & $\frac{\pi}{3}$ & & & & & & & & $0.143$ ($0.144$)\
![\[fig:mee\_CaseV\]The possible values of the effective Majorana mass $|m_{ee}|$ as a function of the lightest neutrino mass in the case V. The red (blue) dashed lines indicate the most general allowed regions for IO (NO) neutrino mass spectrum obtained by varying the mixing parameters over the $3\sigma$ ranges [@Capozzi:2013csa]. The orange (cyan) areas denote the achievable values of $|m_{ee}|$ in the limit of $n\rightarrow\infty$ assuming IO (NO) spectrum. The purple and green regions are the theoretical predictions for the $D^{(1)}_{9n ,3n}$ group with $n=2$. Notice that the purple (green) region overlaps the orange (cyan) one. The present most stringent upper limits $|m_{ee}|<0.120$ eV from EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] is shown by horizontal grey band. The vertical grey exclusion band represents the current bound coming from the cosmological data of $\sum m_i<0.230$ eV at $95\%$ confidence level obtained by the Planck collaboration [@Ade:2013zuv]. ](meeV_v1.pdf){width="0.60\linewidth"}
\[sec:Conclusion\]Conclusions
=============================
The type D finite subgroup of $SU(3)$ has two independent series: $D^{(0)}_{n, n}\cong\Delta(6n^2)$ and $D^{(1)}_{9n, 3n}\cong(Z_{9n}\times Z_{3n})\rtimes S_3$. The $\Delta(6n^2)$ flavor symmetry with or without CP symmetry and its predictions for the lepton flavor mixing has been discussed in the literature. In the present work, we have performed a comprehensive analysis of the mixing patterns which can be derived from another type D group series $D^{(1)}_{9n ,3n}$ and the generalized CP. The phenomenological consequence of the “direct” approach, “semidirect” approach and “variant of semidirect” approach are studied in a model independent way. The three approaches differ in the residual symmetries preserved by the neutrino and charged lepton sectors.
The mathematical structure of $D^{(1)}_{9n, 3n}$ has been investigated. Using the method of induced representations, we find all the irreducible representations of $D^{(1)}_{9n, 3n}$ group and its character table for arbitrary $n$. We have derived the Kronecker products and constructed the Clebsch-Gordan coefficients. These details would be necessary and particularly useful for model builders aiming at construction of flavor models based on the group $D^{(1)}_{9n, 3n}$. Furthermore, we have identified the class-inverting automorphisms of the $D^{(1)}_{9n, 3n}$ group, and show that the corresponding CP transformations are of the same form as the flavor symmetry transformations in our working basis.
In the “direct” approach, the original symmetry $D^{(1)}_{9n,3n}\rtimes H_{CP}$ is broken down to $K_{4}\times H^{\nu}_{CP}$ in the neutrino sector and to $G_{l}\rtimes H^{l}_{CP}$ in the charged lepton sector, where $G_{l}$ is an abelian subgroup which allows to distinguish the three generations of leptons. In this scenario, all the lepton mixing parameters including the Majorana CP phases are completely fixed by the residual symmetries. We have considered all the possible residual subgroups $K_4$, $G_{l}$ and the residual CP transformations that can be consistently combined. We find that the lepton mixing matrices compatible with the data are of the trimaximal form. Both Dirac phase $\delta_{CP}$ and the Majorana phase $\alpha_{31}$ are predicted to be conserved, and the values of the Majorana phase $\alpha_{21}$ are $0$, $\frac{2}{3n}\pi$, $\frac{4}{3n}\pi$, $\ldots$, $\frac{6n-2}{3n}\pi$.
In contrast with the “direct” approach, the residual symmetry preserved by the neutrino mass matrix is $Z_2\times H^{\nu}_{CP}$ in the “semidirect” approach. Since the remnant flavor symmetry of the neutrino sector is $Z_2$ instead of $K_4$, it would fix only one column of the PMNS matrix. Taking into account the remnant CP transformations further, all the lepton mixing angles as well as the CP violating phases would be predicted in terms of a continuous free parameter $\theta$ besides the parameters characterizing the residual symmetries. We find that only four types of mixing patterns named as cases I, II, III and IV can accommodate the experimental data on lepton mixing angles for certain values of the continuous parameter $\theta$ and the discrete parameter $\varphi_i$ determined by the postulated residual symmetries. For cases III and IV, the residual $Z_2$ subgroup is chosen to be generated by the element $c^{9n/2}$ such that the group index $n$ has to be even. We have performed a detailed analytical and numerical analysis. It is remarkable that either the solar mixing angle $\theta_{12}$ or the atmospheric mixing angle $\theta_{23}$ is bounded within certain intervals for arbitrary $n$. As a consequence, these predictions can be testable by the next generation of reactor neutrino experiments and long baseline experiments. The admissible values of the mixing angles and CP phases for each $D^{(1)}_{9n,3n}$ group until $n=50$ have been studied. Interestingly enough, the first two smallest $D^{(1)}_{9n,3n}$ groups with $n=1, 2$ already allow a good fit to the data on lepton mixing angles, and the CP violating phases can be conserved, maximal or some other irregular values. Moreover, the phenomenological predictions for the neutrinoless double beta decay are exploited.
In the so-called “variant of semidirect” approach, the remnant symmetries of the neutrino and the charged lepton mass matrices are assumed to be $K_{4}\times H^{\nu}_{CP}$ and $Z_{2}\times H^{l}_{CP}$ respectively. We find only one type of mixing pattern named as case V is phenomenologically viable in this scenario. One row of the PMNS matrix is determined to be $(1/2, 1/2, -e^{i\varphi_9}/\sqrt{2})$. The solar mixing angle is predicted to lie in the interval $0.328\leq\sin^2\theta_{12}\leq0.359$, and the atmospheric angle is in the range of $0.510\leq\sin^2\theta_{23}\leq0.515$ or $0.485\leq\sin^2\theta_{23}\leq0.490$. Moreover, both Dirac phase and the Majorana phase $\alpha_{21}$ are bounded from above $\left|\sin\delta_{CP}\right|\leq0.594$ and $\left|\sin\alpha_{21}\right|\leq0.399$ respectively.
In our framework, the obtained results for lepton flavor mixing only depend on the structure of flavor symmetry group and the postulated residual symmetries, and they are independent of the breaking mechanism that how the required vacuum alignment needed to achieve the remnant symmetries is dynamically realized. It would be interesting to construct concrete models in which the breaking of the symmetry group to the residual symmetries are spontaneous due to the non-vanishing vacuum expectation values of some flavon fields.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is supported by the National Natural Science Foundation of China under Grant Nos. 11275188, 11179007 and 11522546.
Appendix {#appendix .unnumbered}
========
\[app:group\_theory\]Group theory of $D^{(1)}_{9n,3n}\cong(Z_{9n}\times Z_{3n})\rtimes S_{3}$ {#sclass}
=============================================================================================
The group $D^{(1)}_{9n,3n}$ for a generic integer $n$ is a non-Abelian finite subgroup of $SU(3)$ of type D [@Grimus:2013apa]. Its order is $162n^2$. It is isomorphic to the semidirect product of the $S_3$, the smallest non-Abelian finite group, with $(Z_{9n} \times Z_{3n})$, i.e. $D^{(1)}_{9n,3n}\cong(Z_{9n}\times Z_{3n})\rtimes S_{3}$. The $D^{(1)}_{9n,3n}$ group can be defined in terms of four generators $a$, $b$, $c$ and $d$ fulfilling the following relations [@Grimus:2013apa; @Yao:2015dwa]: $$\begin{aligned}
\nonumber&&\qquad\quad a^3=b^2=(ab)^2=c^{9n}=d^{3n}=1,\quad cd=dc,\\
\label{eq:multiplication_rules}&&aca^{-1}=cd,\quad ada^{-1}=c^{-3}d^{-2},\quad bcb^{-1}=cd,\quad bdb^{-1}=d^{-1}\,.\end{aligned}$$ One can see that $a$ and $b$ generate $S_3$, and $c$ and $d$ generate the $Z_{9n}$ and $Z_{3n}$ subgroups respectively. Any group element $g\in D^{(1)}_{9n, 3n}$ can be written as a product of powers of the generators $a$, $b$, $c$ and $d$, $$\label{eq:genform1}
g=a^{\alpha}b^{\beta}c^{\gamma}d^{\delta}\,,$$ with $\alpha=0, 1, 2$, $\beta=0, 1$, $\gamma=0, 1, \ldots, 9n-1$ and $\delta=0, 1,\ldots, 3n-1$. From the multiplication rules in Eq. , the following useful relations can be obtained, $$\begin{aligned}
\nonumber&&ca=ac^{-2}d^{-1},\qquad da=ac^3d,\qquad ba=a^2b,\quad ca^2=a^2cd,\\
\label{eq:useful_relations}&& da^2=a^2c^{-3}d^{-2},\qquad ba^2=ab,\qquad cb=bcd,\qquad db=bd^{-1}\,.\end{aligned}$$ Utilizing Eqs. (\[eq:genform1\], \[eq:useful\_relations\]), we find that the elements of $D^{(1)}_{9n, 3n}$ group belong to the following conjugacy classes: $$\begin{aligned}
\label{eq:classes}
\nonumber 1 &:&1\mathcal{C}_{1}=\left\{1\right\},\\
\nonumber 2 &:&1\mathcal{C}_{1}^{(\nu)}=\left\{c^{\nu}\right\},\quad \nu=\mbox{$3n,6n$}, \\
\nonumber 9n-3 &:& 3\mathcal{C}_{1}^{(\rho)}=\left\{c^{\rho}, c^{\rho}d^{\rho}, c^{-2\rho}d^{-\rho}\right \},\quad\rho\neq \mbox{$0,3n,6n$}, \\
\nonumber\frac{27n(n-1)+6}{6}&:&6\mathcal{C}_{1}^{(\rho,\sigma)}=\left\{c^{\rho}d^{\sigma}, c^{\rho-3\sigma}d^{\rho-2\sigma}, c^{3\sigma-2\rho}d^{\sigma-\rho}, c^{\rho}d^{\rho-\sigma}, c^{3\sigma-2\rho}d^{2\sigma-\rho}, c^{\rho-3\sigma}d^{-\sigma} \right \},\\
\nonumber 3 &:& 18n^{2}\mathcal{C}_{2}^{(\tau)}=\{ac^{\tau+3x}d^{y}, a^2c^{\tau+3x}d^{\tau+3x-y}|\;x,y=0,1,\ldots,3n-1\},~\tau=0,1,2 \\
\nonumber 9n &:& 9n\mathcal{C}_{3}^{(\rho)}=\left\{bc^{\rho}d^{x}, a^2bc^{\rho-3x}d^{\rho-2x}, abc^{\rho-3x}d^{-x}| x=0,1,\ldots,3n-1 \right\}, \rho=0,1,\ldots,9n-1\,,\end{aligned}$$ where the quantity on the left of the colon denotes the number of classes and the quantity on the right of the colon refers to the number of elements contained in the classes. The parameters $\rho$ and $\sigma$ in the conjugacy class $6\mathcal{C}_{1}^{(\rho,\sigma)}$ can take the values $\rho=0,1,\ldots,9n-1$, $\sigma=0,1,\ldots, 3n-1$, and the following possibilities are excluded, $$\rho-2\sigma=0~\text{mod}(3n),\qquad \rho-\sigma=0~\text{mod}(3n),\qquad \sigma=0\,.$$ As a result, the $D^{(1)}_{9n, 3n}$ group totally has $1+2+(9n-3)+\frac{27n(n-1)+6}{6}+3+9n=(3n+1)(3n+8)/2$ different conjugacy classes. Furthermore, we can check that the center of the $D^{(1)}_{9n,3n}$ group is $Z(D^{(1)}_{9n,3n})=\left\{1, c^{3n}, c^{6n}\right\}$.
\[sec:representation\]Irreducible representations of $D^{(1)}_{9n, 3n}$ group
-----------------------------------------------------------------------------
Now we proceed to construct all the irreducible representations of the $D^{(1)}_{9n,3n}$ group. Firstly we concentrate on the one-dimensional representations in which all generators are represented by pure numbers and they are commutable with each other. From Eq. , we see that $$a=d=1,\quad b^2=1,\quad c^3=1\,.$$ Hence $D^{(1)}_{9n,3n}$ group has six singlet representations given by $$\label{eq:one_rep}
\begin{array}{lll}
\mathbf{1_{0,0}} &:& a=b=c=d=1 \,,\\
\mathbf{1_{0,1}} &:& a=b=d=1,\quad c=\omega\,,\\
\mathbf{1_{0,2}} &:& a=b=d=1,\quad c=\omega^{2} \,,\\
\mathbf{1_{1,0}} &:& a=c=d=1, \quad b=-1 \,, \\
\mathbf{1_{1,1}} &:& a=d=1, \quad b=-1, \quad c=\omega \,, \\
\mathbf{1_{1,2}} &:& a=d=1, \quad b=-1, \quad c=\omega^{2} \,,
\end{array}$$ for any integer $n$, where $\omega\equiv e^{2\pi i/3}$. These one-dimensional representations differ in the values of the generators $b$ and $c$, and they can be neatly written as $$\mathbf{1}_{i,j}~:~ a=d=1,\quad b=(-1)^i,\quad c=\omega^j,~~ \text{with}~~i=0,1,~j=0,1,2\,.$$ As far as we know, the representations of the $D^{(1)}_{9n,3n}$ group has not been worked out in the literature. It is a nontrivial task. In the following, we shall use the method of induced representations to build the remaining irreducible representations. The induced representation can be commonly found in the literature. In the following, we first briefly review the basic idea of the induced representation. Let $\mathcal{G}$ be a finite group and $\mathcal{H}$ any subgroup of $\mathcal{G}$ with index $n$. The index of $\mathcal{H}$ in $\mathcal{G}$ is the number of cosets of $\mathcal{H}$ in $\mathcal{G}$, i.e. $n=|\mathcal{G}|/|\mathcal{H}|$ where $|\mathcal{G}|$ and $|\mathcal{H}|$ denote the order of $\mathcal{G}$ and $\mathcal{H}$ respectively. We denote $x_1$, $x_{2}$, $\ldots$, $x_n$ as a full set of representatives in $\mathcal{G}$ of the cosets in $\mathcal{G}/\mathcal{H}$, i.e. $$\mathcal{G}/\mathcal{H}=x_{1}(x_{1}\equiv 1)\mathcal{H}~\oplus~x_2\mathcal{H}~\oplus~\cdots~\oplus ~x_{n}\mathcal{H}\,.$$ Furthermore, let $\varrho$ be a $d$-dimensional irreducible representation of $\mathcal{H}$ with $\varrho: H\rightarrow GL(V)$, where $V$ is the representation space of dimension $d$ and $GL(V)$ is the group of non-singular linear maps on $V$. Supposing $\left\{e_{1},\ldots, e_{d}\right\}$ is a basis of the vector space $V$, the action of any element $h\in\mathcal{H}$ on the basis vector $e_{i}$ is $$h:~e_{i}~\mapsto~ \varrho(h)_{ji}e_{j}\,.$$ The induced representation can be thought of as acting on the following space: $$W=\bigoplus^{n}_{i=1}x_{i}V\,,$$ where each $x_{i}V$ is an isomorphic copy of the vector space $V$. The basis vector of the space $W$ can be taken to be $$x_{k}e_{i}\equiv e_{k,i}\,,\quad\text{with}\quad k=1, 2,\ldots,n,~~i=1, 2,\ldots,d\,.$$ According to the definition of coset, any $g\in\mathcal{G}$ will then send each $x_{k}$ to a unique $x_{m}h$ with $h\in\mathcal{H}$ such that $gx_{k}=x_{m}h$ where $k, m=1, 2,\ldots,n$. In the induced representation, an element $g\in\mathcal{G}$ acts on the vector space $W$ as follows $$\label{eq:def_induced_rep}g:~e_{k, i}~\mapsto~ge_{k, i}=gx_{k}e_{i}=x_{m}he_{i}
=\varrho(h)_{ji}x_me_{j}
=\varrho(h)_{ji}e_{m,j}\,.$$ Thus we see that $\mathcal{G}$ acts linearly on $W$, and its action is thus represented by a $(dn\times dn)$ matrix. Notice that the induced representation is not necessarily irreducible.
We now apply this method to the group $D^{(1)}_{9n,3n}=\mathcal{G}$, and take the subgroup to be $\mathcal{H}=Z_{9n}\times Z_{3n}$. The index of $\mathcal{H}$ in $\mathcal{G}$ is $n=6$. Since $\mathcal{H}$ is an abelian subgroup, its irreducible representations $\varrho$ can only be one-dimensional. $e_1$ is the basis for the representation space of $\mathcal{H}$, the generators $c$ and $d$ act on $e_1$ as follows $$\label{eq:eigenvalue_cd}
c e_1=\eta^le_{1},\qquad d e_{1}=\eta^{-3k}e_{1}\,,$$ where $\eta=e^{\frac{2\pi i}{9n}}$, the values of the parameters $l$ and $k$ are $l=0, 1, \dots, 9n-1$ and $k=0, 1, \dots, 3n-1$. The six representative elements of the coset $\mathcal{G}/\mathcal{H}$ can be chosen to be $$x_1=1,\quad x_2=a^2,\quad x_3=a,\quad x_4=b,\quad x_5=ab,\quad x_6=a^2b\,.$$ As a consequence, we can obtain the basis of the vector space $W$ on which the induced representation is defined, $$\begin{aligned}
\nonumber&& e_{1}\equiv x_1e_1=e_1,~\quad e_{2}\equiv x_2e_1=a^{2}e_1,
\quad e_{3}\equiv x_3e_1=ae_1,\\
\label{eq:six_basis}&&e_{4}\equiv x_4e_1=be_1, \quad e_{5}\equiv x_5e_1=abe_1,\quad e_{6}\equiv x_6e_1=a^{2}be_{1}\,.\end{aligned}$$ According to Eq. , the actions of the generators $a$, $b$, $c$ and $d$ on the above six basis vectors can be straightforwardly derived by utilizing the useful identities in Eq. : $$\label{eq:linear1}
\begin{array}{lll}
ae_{1}=e_{3}, ~~&~~ ae_{2}=e_{1}, ~~&~~ ae_{3}=e_{2}, \\
ae_{4}=e_{5}, ~~&~~ ae_{5}=e_{6}, ~~&~~ ae_{6}=e_{4}, \\
be_{1}=e_{4}, ~~&~~ be_{2}=e_{5}, ~~&~~ be_{3}=e_{6}, \\
be_{4}=e_{1}, ~~&~~ be_{5}=e_{2}, ~~&~~ be_{6}=e_{3}, \\
ce_{1}=\eta^{l}e_{1}, ~~&~~ ce_{2}=\eta^{l-3k}e_{2}, ~~&~~ ce_{3}=\eta^{-2l+3k}e_{3}, \\
ce_{4}=\eta^{l-3k}e_{4}, ~~&~~ ce_{5}=\eta^{-2l+3k}e_{5}, ~~&~~ ce_{6}=\eta^{l}e_{6}, \\
de_{1}=\eta^{-3k}e_{1}, ~~&~~ de_{2}=\eta^{-3l+6k}e_{2}, ~~&~~ de_{3}=\eta^{3l-3k}e_{3}, \\
de_{4}=\eta^{3k}e_{4}, ~~&~~ de_{5}=\eta^{3l-6k}e_{5}, ~~&~~ de_{6}=\eta^{-3l+3k}e_{6}\,.
\end{array}$$ Then we can read out the representation matrices as follows $$\label{eq:6drep_induced}
\mathbf{6}_{(l, k)}:\quad a=\left(\begin{array}{cc} a_{1}& 0 \\ 0 & a_{2} \end{array}\right),\qquad
b=\left(\begin{array}{cc} 0 & \mathbb{1}_{3} \\ \mathbb{1}_{3} & 0 \end{array}\right),\qquad
c=\left(\begin{array}{cc} c_{1} & 0 \\ 0 & c_{2}
\end{array}\right),\qquad
d=\left(\begin{array}{cc} d_{1} & 0 \\ 0 & d_{2}
\end{array}\right)\,,$$ where $\mathbb{1}_{3}$ refers to a $3\times3$ unit matrix, and the different submatrices are given by $$\begin{aligned}
\nonumber&&a_{1}=\left(\begin{array}{ccc}
0 & 1 & 0 \\
0 & 0 & 1 \\
1 & 0 & 0
\end{array}\right),\qquad\quad
a_{2}=\left(\begin{array}{ccc}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array}\right)\,, \\
\nonumber&&\hskip-0.4in
c_{1}=\left(\begin{array}{ccc}
\eta^{l}& 0 & 0 \\
0 & \eta^{l-3k} & 0 \\
0 & 0 & \eta^{-2l+3k}
\end{array}\right),\qquad\quad
c_{2}=\left(\begin{array}{ccc}
\eta^{l-3k} & 0 & 0 \\
0 & \eta^{-2l+3k} & 0 \\
0 & 0 & \eta^{l}
\end{array}\right)\,, \\
\label{eq:6drep_subMatrix}&&\hskip-0.4in d_{1}=\left(\begin{array}{ccc}
\eta^{-3k}& 0 & 0 \\
0 & \eta^{-3l+6k} & 0 \\
0 & 0 & \eta^{3l-3k}
\end{array}\right),\qquad
d_{2}=d^{-1}_{1}=\left(\begin{array}{ccc}
\eta^{3k}& 0 & 0 \\
0 & \eta^{3l-6k} & 0 \\
0 & 0 & \eta^{-3l+3k}
\end{array}\right)\,.\end{aligned}$$ The above different representations labelled by $(l, k)$ may be equivalent. If we perform the similarity transformations generated by $$\label{eq:simtra1}
S=\left(\begin{array}{cc}
a_{1} & 0 \\
0 & a_{1} \\
\end{array}\right), \qquad
T=\left(\begin{array}{cc}
0 & t \\
t & 0 \\
\end{array}\right),\qquad S^{3}=T^{2}=(ST)^{2}=1\,,$$ where $$t=\left(\begin{array}{ccc}
0 & 0 & 1\\
0 & 1 & 0\\
1 & 0 & 0\\
\end{array}\right)\,,$$ the representations matrices for $a$ and $b$ are kept intact while the diagonal elements of both $c$ and $d$ are interchanged. As a result, the same representation is labeled in six different ways $$\label{eq:6pairs}
\left(\begin{array}{c} l \\ k
\end{array}\right),~
\left(\begin{array}{c} l-3k \\ l-2k
\end{array}\right),~
\left(\begin{array}{c}-2l+3k \\ -l+k
\end{array}\right),~
\left(\begin{array}{c} l \\ l-k
\end{array}\right),~
\left(\begin{array}{c} -2l+3k \\ -l+2k
\end{array}\right),~
\left(\begin{array}{c} l-3k \\ -k
\end{array}\right)\,.$$ The six pairs above can be compactly written into the form $$M^{p}_{s}\left(\begin{array}{c} l \\ k
\end{array}\right)\,,$$ where $$\label{eq:defineM}
M^{p}_{s}=\left(\begin{array}{cc}
1 ~&~ -3 \\
1 ~&~ -2
\end{array}\right)^{p}
\left(\begin{array}{cc}
1 ~&~ 0 \\
1 ~&~ -1
\end{array}\right)^{s},\quad\mathrm{with}\quad p=0, 1, 2, ~s=0, 1\,.$$ Now we proceed to study whether the six-dimensional representations constructed by the induced representation method are irreducible or not by the famous **Mackey theorem** in math [@Mackey:1951; @Simon:1996; @Serre:1977]. If any one of them is reducible, we further decompose it into the direct sum of the irreducible representations of $D^{(1)}_{9n, 3n}$ group.
**Theorem (Mackey’s Irreducibility Criterion):** Let $\mathcal{H}\subset\mathcal{G}$ and $\varrho$ be a representation of $\mathcal{H}$. For $s\in\mathcal{G}$, we define $$\mathcal{H}_{s}\equiv\mathcal{H}\cap s\mathcal{H}s^{-1}, \qquad \varrho^{s}(h)\equiv\varrho(s^{-1}hs)$$ for $h\in \mathcal{H}_{s}$ such that $\varrho^{s}$ is a representation of $\mathcal{H}_{s}$. Then the induced representation $\text{Ind}^{\mathcal{G}}_{\mathcal{H}}(\varrho)$ is irreducible if and only if
(1)
: [$\varrho$ is irreducible.]{}
(2)
: [For all $s\in \mathcal{G}\setminus\mathcal{H}$, $\varrho^{s}$ and $\text{Res}^{\mathcal{H}_{s}}_{\mathcal{H}}(\varrho)$ are disjoint.]{}
Here $\text{Ind}^{\mathcal{G}}_{\mathcal{H}}(\varrho)$ denotes a representation of $G$ and it is induced from a representation $\varrho$ on a subgroup $\mathcal{H}$. $\text{Res}^{\mathcal{H}_{s}}_{\mathcal{H}}(\varrho)$ is the restriction of the representation $\varrho$ on $\mathcal{H}$ to $\mathcal{H}_{s}$. The notation $\mathcal{G}\setminus\mathcal{H}$ denotes the group elements in $\mathcal{G}$ but not in $\mathcal{H}$. Two representations $\varrho$ and $\varrho^{\prime}$ of a group are said to be disjoint if and only if they contain no equivalent subrepresentations, equivalently if and only if their characters are orthogonal. From this theorem, it is easy to further obtain a useful corollary.\
**Corollary:** Suppose $\mathcal{H}$ is a normal subgroup of $\mathcal{G}$, then we have $\mathcal{H}_{s}=\mathcal{H}$ and $\text{Res}^{\mathcal{H}_{s}}_{\mathcal{H}}(\varrho)=\varrho$. In order that $\text{Ind}^{\mathcal{G}}_{\mathcal{H}}(\varrho)$ be irreducible, it is necessary and sufficient that $\varrho$ is irreducible and not isomorphic to any of its conjugate $\rho^{s}$ for $s\notin H$.\
This implies that the representation $\text{Ind}^{\mathcal{G}}_{\mathcal{H}}(\varrho)$ would be reducible if there is a $s\in\mathcal{G}\setminus\mathcal{H}$ leading to $\varrho^s(\mathcal{H})\cong \varrho(\mathcal{H})$ for normal subgroup $\mathcal{H}$. The corollary can be exploited to determine whether the six-dimensional representations $\mathbf{6}_{(l,k)}$ of the $D^{(1)}_{9n,3n}$ group in Eq. are reducible or not. The subgroup $\mathcal{H}=Z_{9n}\times Z_{3n}$ is a normal subgroup of $D^{(1)}_{9n,3n}$, and it is abelian such that its irreducible representation $\varrho$ is one-dimensional and specified by Eq. . From the above corollary of the Mackey theorem, we know that the six-dimensional representation $\mathbf{6}_{(l,k)}$ is reducible if and only if $\varrho^{s}(\mathcal{H})$ and $\varrho(\mathcal{H})$ are equivalent representations for an element $s\in D^{(1)}_{9n,3n} \setminus \mathcal{H}$. In order to obtain the conditions in which the six-dimensional representations $\mathbf{6}_{(l,k)}$ is reducible, we only need to consider the value of $s$ is $b$, $ab$, $a^2b$, $a$ and $a^2$ respectively. The results are collected in table \[tab:re\_six\_irrep\].
$s$ $\varrho^s(\mathcal{H})\cong \varrho(\mathcal{H})$ `Reducible conditions`
------------ ----------------------------------------------------------------------------------------------- ---------------------------------
$b$ $\varrho(b^{-1}cb)=\varrho(c)$, $\varrho(b^{-1}db)=\varrho(d)$ $3k=0$ $(\text{mod}~9n)$
$ab$ $\varrho((ab)^{-1}c(ab))=\varrho(c)$, $\varrho((ab)^{-1}d(ab))=\varrho(d)$ $3l-3k=0$ $\text{mod}(9n)$
$a^2b$ $\varrho\left((a^2b)^{-1}c(a^2b)\right)=\varrho(c)$, $\varrho((a^2b)^{-1}d(a^2b))=\varrho(d)$ $3l-6k=0$ $\text{mod}(9n)$
$a$($a^2$) $\varrho(a^{-1}ca)=\varrho(c)$, $\varrho(a^{-1}da)=\varrho(d)$ $3k=0$, $3l=0$ $\text{mod}(9n)$
: \[tab:re\_six\_irrep\] The reducible conditions for the six-dimensional representations $\mathbf{6}_{(l, k)}$. The one-dimensional representation $\varrho$ of $\mathcal{H}$ is given by Eq. , i.e. $\varrho(c)=\eta^{l}$ and $\varrho(d)=\eta^{-3k}$. The values of parameters $l$ and $k$ are $l=0, 1, \cdots, 9n-1$ and $k=0, 1, \cdots, 3n-1$.
- [**Six-dimensional representations**]{}
Six-dimensional representations of the $D^{(1)}_{9n, 3n}$ group have been constructed by the method of induced representation, as shown in Eq. and Eq. . From table \[tab:re\_six\_irrep\], we find that the induced representation $\mathbf{6}_{(l, k)}$ in Eq. would be reducible when any of the following conditions are met $$\begin{cases}
9n ~~:~~k=0,~~l=0, 1, \ldots, 9n-1, \\
9n-3 ~~:~~3l-3k=0~\text{mod}(9n),~~ k\neq 0\\
9n-3 ~~:~~3l-6k=0~\text{mod}(9n),~~ k\neq 0\\
\end{cases}\,,$$ where the quantity on the left of the colon is the number of $(l, k)$ values of the properties on the right of the colon. Excluding these values for $l$ and $k$, there should be $9n\times3n-9n-(9n-3)-(9n-3)=27n(n-1)+6$ different pairs of $(l, k)$. Furthermore, taking into account the over counting issue shown in Eq. , we essentially have $\frac{27n(n-1)+6}{6}$ six-dimensional irreducible representations, and the representation matrices of the generators are given in Eq. .
- [**Three-dimensional representations**]{}
Once the conditions $3k=0$, $3l-3k=0~\text{mod}(9n)$ or $3l-6k=0~\text{mod}(9n)$ are fulfilled, the six-dimensional induced representation $6_{(l, k)}$ could be decomposed into the direct sum of three and two-dimensional representations. Firstly we concentrate on the case of $3k=0$ or equivalently $k=0$ . From Eq. we see that the eigenvalues of $c$ on the three pairs of basis vectors $(e_1, e_4)$, $(e_2, e_6)$ and $(e_3, e_5)$ are $\eta^{l}$, $\eta^l$ and $\eta^{-2l}$ respectively, and the eigenvalues of $d$ on the three pairs vectors $(e_1, e_4)$, $(e_2, e_6)$ and $(e_3, e_5)$ are $1$, $\eta^{-3l}$ and $\eta^{3l}$ respectively. Hence we recombine the six vectors $e_1, \ldots, e_6$ into $$e^{\prime}_1=x_1e_1+y_1e_4, \quad e^{\prime}_2=x_2e_2+y_2e_6, \quad e^{\prime}_3=x_3e_3+y_3e_5\,.$$ In the case of $3k=0$ and $l\neq0,3n,6n$, the three vectors $e^{\prime}_1$, $e^{\prime}_2$ and $e^{\prime}_3$ can be distinguished from each other by the actions of $c$ and $d$, and they must be closed under the action of $a$ and $b$. Considering the effect of $a$, we find $$ae^{\prime}_1=e^{\prime}_3, \quad ae^{\prime}_2=e^{\prime}_1, \quad ae^{\prime}_3=e^{\prime}_2\,,$$ which yields $$x_1=x_2=x_3\equiv x, \qquad y_1=y_2=y_3\equiv y\,.$$ Furthermore, closeness under the action of $b$ implies $$be^{\prime}_1=\pm e^{\prime}_1, \quad be^{\prime}_2=\pm e^{\prime}_3, \quad be^{\prime}_3=\pm e^{\prime}_2\,.$$ In the case of $be^{\prime}_1=e^{\prime}_1$, i.e. $x=y$, then the normalized basis vectors of a three-dimensional subspace can be chosen to be $$\label{eq:three_basis1}
e^{\prime}_{1}=\frac{1}{\sqrt{2}}(e_{1}+e_{4}),\quad
e^{\prime}_{2}=\frac{1}{\sqrt{2}}(e_{2}+e_{6})=a^{2}e^{\prime}_{1},
\quad e^{\prime}_{3}=\frac{1}{\sqrt{2}}(e_{3}+e_{5})=ae^{\prime}_{1}\,.$$ In the case of $be^{\prime}_1=-e^{\prime}_1$, i.e. $x=-y$, we define the three normalized orthogonal vectors as $$\label{eq:three_basis2}
e^{\prime}_{4}=\frac{1}{\sqrt{2}}(e_{1}-e_{4}), \quad
e^{\prime}_{5}=\frac{1}{\sqrt{2}}(e_{2}-e_{6})=a^{2}e^{\prime}_{4}, \quad e^{\prime}_{6}=\frac{1}{\sqrt{2}}(e_{3}-e_{5})=ae^{\prime}_{4}\,.$$ It is easy to check that $e^{\prime}_{4}$, $e^{\prime}_{5}$ and $e^{\prime}_{6}$ span another three-dimensional subspace. The basis transformation matrix from $e_{i}$ to $e^{\prime}_{i}$ is denoted by $\Omega$, i.e. $$e^{\prime}_{i}=\sum^{6}_{j=1}{e_{j}\Omega_{ji}}\,,$$ where the similarity transformation $\Omega$ reads $$\Omega=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}
\mathbb{1}_{3} & \mathbb{1}_{3} \\
\varpi & -\varpi
\end{array}\right), \qquad
\varpi=
\left(\begin{array}{ccc}
1&0&0\\
0&0&1\\
0&1&0
\end{array}\right),$$ In the new basis, the representation matrices of the generators are of the following form $$\begin{array}{cc}
a^{\prime}\equiv\Omega^{-1}a\Omega=\left(\begin{array}{cc}
a_{\mathbf{3}_{l, 0}} & 0 \\ 0 & a_{\mathbf{3}_{l, 1}}
\end{array}\right), ~~&~~
b^{\prime}\equiv\Omega^{-1}b\Omega=\left(\begin{array}{cc}
b_{\mathbf{3}_{l, 0}} & 0 \\ 0 & b_{\mathbf{3}_{l, 1}}
\end{array}\right), \\[-7pt]
\\[4pt]
c^{\prime}\equiv\Omega^{-1}c\Omega=\left(\begin{array}{cc}
c_{\mathbf{3}_{l, 0}} & 0 \\ 0 & c_{\mathbf{3}_{l, 1}}
\end{array}\right), ~~&~~
d^{\prime}\equiv\Omega^{-1}d\Omega=\left(\begin{array}{cc}
d_{\mathbf{3}_{l, 0}} & 0 \\ 0 & d_{\mathbf{3}_{l, 1}}
\end{array}\right)\,,
\end{array}$$ where $$\begin{aligned}
\nonumber&& a_{\mathbf{3}_{l, 0}}=\left(\begin{array}{ccc}
0 &1 &0 \\
0&0&1 \\
1&0&0\end{array}\right),~
b_{\mathbf{3}_{l, 0}}=\left(\begin{array}{ccc}
1 &0 &0 \\
0&0&1 \\
0&1&0\end{array}\right),~
c_{\mathbf{3}_{l, 0}}=\left(\begin{array}{ccc}
\eta^{l}&0 &0 \\
0&\eta^{l}&0 \\
0&0&\eta^{-2l}
\end{array}\right),~
d_{\mathbf{3}_{l, 0}}=\left(\begin{array}{ccc}
1 &0 &0 \\
0&\eta^{-3l}&0 \\
0&0&\eta^{3l}
\end{array}\right),\\
\label{eq:3dimrep1}&&a_{\mathbf{3}_{l, 1}}~=~a_{\mathbf{3}_{l, 0}},\qquad
b_{\mathbf{3}_{l, 1}}~=~-b_{\mathbf{3}_{l, 0}},\qquad
c_{\mathbf{3}_{l, 1}}~=~c_{\mathbf{3}_{l, 0}},\qquad
d_{\mathbf{3}_{l, 1}}~=~d_{\mathbf{3}_{l, 0}}\,.\end{aligned}$$ This means that the six-dimensional representation $\mathbf{6}_{l,0}$ breaks up into two three-dimensional irreducible representation $\mathbf{3}_{l, 0}$ and $\mathbf{3}_{l, 1}$ which differ in the overall sign of $b$. Notice that the values of $l=0$, $3n$, $6n$ should be excluded, since both triplet representations $\mathbf{3}_{l, 0}$ and $\mathbf{3}_{l, 1}$ then could be decomposed into one-dimensional and two-dimensional representation.
Next we proceed to consider the case of $3l-3k=0~\text{mod}(9n)$ and $l\neq0, 3n, 6n$. We can construct the eigenstates of the generators $c$ and $d$ as follows $$e^{\prime}_1=x_1e_1+y_1e_5, \quad e^{\prime}_2=x_2e_2+y_2e_4, \quad e^{\prime}_3=x_3e_3+y_3e_6\,.$$ Note that $e^{\prime}_1$, $e^{\prime}_2$ and $e^{\prime}_3$ are mapped into each other under the action of $a$ and $b$, Taking into account the normalization condition further, we have $x_1=x_2=x_3=y_1=y_2=y_3=1/\sqrt{2}$ such that $$\label{eq:three_basis3}
e^{\prime}_{1}=\frac{1}{\sqrt{2}}(e_{1}+e_{5}), \quad
e^{\prime}_{2}=\frac{1}{\sqrt{2}}(e_{2}+e_{4})=a^{2}e^{\prime}_{1},
\quad e^{\prime}_{3}=\frac{1}{\sqrt{2}}(e_{3}+e_{6})=ae^{\prime}_{1}\,,$$ or $x_1=x_2=x_3=-y_1=-y_2=y_3=-1/\sqrt{2}$ which leads to $$\label{eq:three_basis4}
e^{\prime}_{4} =\frac{1}{\sqrt{2}}(e_{1}-e_{5}), \quad
e^{\prime}_{5}=\frac{1}{\sqrt{2}}(e_{2}-e_{4})=a^{2}e^{\prime}_{4},
\qquad e^{\prime}_{6}=\frac{1}{\sqrt{2}}(e_{3}-e_{6})=ae^{\prime}_{4}\,.$$ It is straightforward to check the following equations are fulfilled $$\begin{aligned}
\nonumber&&\hskip-0.25in be^{\prime}_1=e^{\prime}_2,\quad be^{\prime}_2=e^{\prime}_1, \quad be^{\prime}_3=e^{\prime}_3,\quad ce^{\prime}_1=\eta^{l}e^{\prime}_1,\quad ce^{\prime}_2=\eta^{-2l}e^{\prime}_2,\quad ce^{\prime}_3
=\eta^{l}e^{\prime}_3,\\
\nonumber&&\hskip-0.25in de^{\prime}_1=\eta^{-3l}e^{\prime}_1,\quad de^{\prime}_2=\eta^{3l}e^{\prime}_2,\quad de^{\prime}_3=e^{\prime}_3,\quad be^{\prime}_4=-e^{\prime}_5,\quad be^{\prime}_5=-e^{\prime}_4, \quad be^{\prime}_6=-e^{\prime}_6,\\
&&\hskip-0.25in ce^{\prime}_4=\eta^{l}e^{\prime}_4,~~ ce^{\prime}_5=\eta^{-2l}e^{\prime}_5,~~ ce^{\prime}_6
=\eta^{l}e^{\prime}_6,~~ de^{\prime}_4=\eta^{-3l}e^{\prime}_4,~~ de^{\prime}_5=\eta^{3l}e^{\prime}_5,~~ de^{\prime}_6=e^{\prime}_6\,.\end{aligned}$$ As a result, the induced representation $\mathbf{6}_{l,\,l\, \text{mod}(3n)}$ for $l\neq0$, $3n$, $6n$ can be split into two three-dimensional representations. The unitary transformation from the $e_{i}$ basis to the $e^{\prime}_{i}$ basis is $$e^{\prime}_{i}=\sum^{6}_{j=1}\Omega_{ji}e_{j}\,,$$ with $$\Omega=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}
\mathbb{1}_{3} & \mathbb{1}_{3} \\
\varpi & -\varpi
\end{array}\right), \qquad
\varpi=
\begin{pmatrix}
0&1&0\\
1&0&0\\
0&0&1
\end{pmatrix}\,.$$ Performing the similarity transformation $\Omega$, the representation matrices in the new basis are given by $$\begin{array}{cc}
a^{\prime}=\Omega^{-1}a\Omega=\left(\begin{array}{cc}
a_{\mathbf{3}_{l, 2}} & 0 \\ 0 & a_{\mathbf{3}_{l, 3}}
\end{array}\right), ~~&~~
b^{\prime}=\Omega^{-1}b\Omega=\left(\begin{array}{cc}
b_{\mathbf{3}_{l, 2}} & 0 \\ 0 & b_{\mathbf{3}_{l, 3}}
\end{array}\right), \\[-7pt]
\\[4pt]
c^{\prime}=\Omega^{-1}c\Omega=\left(\begin{array}{cc}
c_{\mathbf{3}_{l, 2}} & 0 \\ 0 & c_{\mathbf{3}_{l, 3}}
\end{array}\right), ~~&~~
d^{\prime}=\Omega^{-1}d\Omega=\left(\begin{array}{cc}
d_{\mathbf{3}_{l, 2}} & 0 \\ 0 & d_{\mathbf{3}_{l, 3}}
\end{array}\right)\,,
\end{array}$$ where $$\begin{aligned}
\nonumber&&
a_{\mathbf{3}_{l, 2}}=\left(\begin{array}{ccc}
0 &1 &0 \\
0&0&1 \\
1&0&0\end{array}\right),~
b_{\mathbf{3}_{l, 2}}=\left(\begin{array}{ccc}
0 &1 &0 \\
1&0&0 \\
0&0&1\end{array}\right),~
c_{\mathbf{3}_{l, 2}}=\left(\begin{array}{ccc}
\eta^{l}&0 &0 \\
0&\eta^{-2l}&0 \\
0&0&\eta^{l}\end{array}\right),~
d_{\mathbf{3}_{l, 2}}=\left(\begin{array}{ccc}
\eta^{-3l} &0 &0 \\
0&\eta^{3l}&0 \\
0&0&1\end{array}\right), \\
\label{eq:3dimrep2}&& a_{\mathbf{3}_{l, 3}}~=~a_{\mathbf{3}_{l, 2}},\qquad
b_{\mathbf{3}_{l, 3}}~=~-b_{\mathbf{3}_{l, 2}},\qquad
c_{\mathbf{3}_{l, 3}}~=~c_{\mathbf{3}_{l, 2}},\qquad
d_{\mathbf{3}_{l, 3}}~=~d_{\mathbf{3}_{l, 2}}\,.\end{aligned}$$ Therefore $\mathbf{6}_{l,\,l\, \text{mod}(3n)}$ for $l\neq0$, $3n$, $6n$ is the direct sum of three-dimensional irreducible representations $\mathbf{3}_{l, 3}$ and $\mathbf{3}_{l, 4}$.
Finally we consider the case of $3l-6k=0~\text{mod}(9n)$ with $k\neq0$ and $l\neq0, 3n, 6n$. In the same fashion as previous cases, we first recombine the basis vectors into $$e^{\prime}_1=x_1e_1+y_1e_6, \quad e^{\prime}_2=x_2e_2+y_2e_5, \quad e^{\prime}_3=x_3e_3+y_3e_4\,,$$ which are eigenstates of both $c$ and $d$ and fulfill $$ce^{\prime}_1=\eta^{-2l^{\prime}}e^{\prime}_1,~ ce^{\prime}_2=\eta^{l^{\prime}}e^{\prime}_2,~ ce^{\prime}_3=\eta^{l^{\prime}}e^{\prime}_3,~ de^{\prime}_1=\eta^{3l^{\prime}}e^{\prime}_1,~
de^{\prime}_2=e^{\prime}_2,~
de^{\prime}_3=\eta^{-3l^{\prime}}e^{\prime}_3\,,$$ for any values of $x_i$ and $y_i$ ($i=1, 2, 3$), where $l^{\prime}=l-3k=3n-k, 6n-k, 9n-k$ with $k=0, 1, \ldots, 3n-1$ such that the value of $l^{\prime}$ can be $0, 1, \ldots, 9n-1$. Taking into account the action of the remaining two generators $a$ and $b$, we find two three-dimensional subspaces would be generated. The basis of the first subspace can be chosen to be $$\label{eq:three_basis3}
e^{\prime}_{1}=\frac{1}{\sqrt{2}}(e_{1}+e_{6}), \quad
e^{\prime}_{2}=\frac{1}{\sqrt{2}}(e_{2}+e_{5})=a^{2}e^{\prime}_{1},
\quad e^{\prime}_{3}=\frac{1}{\sqrt{2}}(e_{3}+e_{4})=ae^{\prime}_{1}\,,$$ The basis vectors of the second three-dimensional subspace are $$\label{eq:three_basis4}
e^{\prime}_{4}=\frac{1}{\sqrt{2}}(e_{1}-e_{6}), \quad
e^{\prime}_{5}=\frac{1}{\sqrt{2}}(e_{2}-e_{5})=a^{2}e^{\prime}_{4},
\quad e^{\prime}_{6}=\frac{1}{\sqrt{2}}(e_{3}-e_{4})=ae^{\prime}_{4}\,,$$ We can read out the unitary basis transformation $$\Omega=\frac{1}{\sqrt{2}}\left(\begin{array}{cc}
\mathbb{1}_{3} & \mathbb{1}_{3} \\
\varpi & -\varpi
\end{array}\right), \qquad
\varpi=
\left(\begin{array}{ccc}
0 & 0 & 1\\
0 & 1 & 0\\
1 & 0 & 0
\end{array}\right)\,.$$ The representation matrices for the generators $a$, $b$, $c$ and $d$ transform as $$\begin{array}{cc}
a^{\prime}=\Omega^{-1}a\Omega=\left(\begin{array}{cc}
a_{\mathbf{3}_{l^{\prime}, 4}} & 0 \\
0 & a_{\mathbf{3}_{l^{\prime}, 5}}
\end{array}\right), ~~&~~
b^{\prime}=\Omega^{-1}b\Omega=\left(\begin{array}{cc}
b_{\mathbf{3}_{l^{\prime}, 4}} & 0 \\ 0 & b_{\mathbf{3}_{l^{\prime}, 5}}
\end{array}\right), \\[-7pt] \\[4pt]
c^{\prime}=\Omega^{-1}c\Omega=\left(\begin{array}{cc}
c_{\mathbf{3}_{l^{\prime}, 4}} & 0 \\ 0 & c_{\mathbf{3}_{l^{\prime}, 5}}
\end{array}\right), ~~&~~
d^{\prime}=\Omega^{-1}d\Omega=\left(\begin{array}{cc}
d_{\mathbf{3}_{l^{\prime}, 4}} & 0 \\ 0 & d_{\mathbf{3}_{l^{\prime}, 5}}
\end{array}\right)\,,
\end{array}$$ where $$\begin{aligned}
\nonumber&& a_{\mathbf{3}_{l^{\prime}, 4}}=\left(\begin{array}{ccc}
0 & 1 & 0 \\
0 & 0 & 1 \\
1 & 0 & 0 \end{array}\right),~
b_{\mathbf{3}_{l^{\prime}, 4}}=\left(\begin{array}{ccc}
0 & 0 & 1 \\
0 & 1 & 0 \\
1 & 0 & 0\end{array}\right),~
c_{\mathbf{3}_{l^{\prime}, 4}}=\left(\begin{array}{ccc} \eta^{-2l^{\prime}} & 0 & 0 \\
0 & \eta^{l^{\prime}} & 0 \\
0 & 0 & \eta^{l^{\prime}}\end{array}\right),~
d_{\mathbf{3}_{l^{\prime}, 4}}=\left(\begin{array}{ccc}
\eta^{3l^{\prime}} & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 &\eta^{-3l^{\prime}}
\end{array}\right), \\
\label{eq:3dimrep3}&&
a_{\mathbf{3}_{l^{\prime}, 5}}~=~a_{\mathbf{3}_{l^{\prime}, 4}},\qquad
b_{\mathbf{3}_{l^{\prime}, 5}}~=~-b_{\mathbf{3}_{l^{\prime}, 4}},\qquad
c_{\mathbf{3}_{l^{\prime}, 5}}~=~c_{\mathbf{3}_{l^{\prime}, 4}},\qquad
d_{\mathbf{3}_{l^{\prime}, 5}}~=~d_{\mathbf{3}_{l^{\prime}, 4}}\,.\end{aligned}$$ Note that both triplet representations $\mathbf{3}_{l^{\prime}, 4}$ and $\mathbf{3}_{l^{\prime}, 5}$ would be reducible for $l^{\prime}=0, 3n, 6n$.
So far we have obtained six three-dimensional irreducible representations $\mathbf{3}_{l, 0}$, $\mathbf{3}_{l, 1}$, $\mathbf{3}_{l, 2}$, $\mathbf{3}_{l, 3}$, $\mathbf{3}_{l, 4}$, $\mathbf{3}_{l, 5}$. However, only two of them are inequivalent because they are related with each other by similarity transformations as follows : $$\begin{array}{llll}
a_{\mathbf{3}_{l, 2}}=U^{\dagger} a_{\mathbf{3}_{l, 0}}U, ~&~ b_{\mathbf{3}_{l, 2}}=U^{\dagger} b_{\mathbf{3}_{l, 0}}U, ~&~
c_{\mathbf{3}_{l, 2}}=U^{\dagger} c_{\mathbf{3}_{l, 0}}U, ~&~ d_{\mathbf{3}_{l, 2}}=U^{\dagger} d_{\mathbf{3}_{l, 0}}U, \\
a_{\mathbf{3}_{l, 3}}=U^{\dagger} a_{\mathbf{3}_{l, 1}}U, ~&~ b_{\mathbf{3}_{l, 3}}=U^{\dagger} b_{\mathbf{3}_{l, 1}}U, ~&~
c_{\mathbf{3}_{l, 3}}=U^{\dagger} c_{\mathbf{3}_{l, 1}}U, ~&~ d_{\mathbf{3}_{l, 3}}=U^{\dagger} d_{\mathbf{3}_{l, 1}}U, \\
a_{\mathbf{3}_{l, 4}}=U a_{\mathbf{3}_{l, 0}}U^{\dagger}, ~&~ b_{\mathbf{3}_{l, 4}}=U b_{\mathbf{3}_{l, 0}}U^{\dagger}, ~&~
c_{\mathbf{3}_{l, 4}}=U c_{\mathbf{3}_{l, 0}}U^{\dagger}, ~&~ d_{\mathbf{3}_{l, 4}}=U d_{\mathbf{3}_{l, 0}}U^{\dagger}, \\
a_{\mathbf{3}_{l, 5}}=U a_{\mathbf{3}_{l, 1}}U^{\dagger}, ~&~ b_{\mathbf{3}_{l, 5}}=U b_{\mathbf{3}_{l, 1}}U^{\dagger}, ~&~
c_{\mathbf{3}_{l, 5}}=U c_{\mathbf{3}_{l, 1}}U^{\dagger}, ~&~
d_{\mathbf{3}_{l, 5}}=U d_{\mathbf{3}_{l, 1}}U^{\dagger}\,,
\end{array}$$ where the unitary transformation $U$ is $$U=\left(\begin{array}{ccc}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array}\right)\,.$$ Hence we conclude that the $D^{(1)}_{9n, 3n}$ group totally has $2(9n-3)$ inequivalent three-dimensional irreducible representations which can be chosen to be $\mathbf{3}_{l, 0}$ and $\mathbf{3}_{l, 1}$ with $l\neq0, 3n, 6n$.
- [**Two-dimensional representations**]{}
In the following, we shall show that both triplet representations $\mathbf{3}_{l, 0}$ and $\mathbf{3}_{l, 1}$ for $l=0, 3n, 6n$ would be reduced into the direct sum of one-dimensional representation and two-dimensional representation. Firstly we concentrate on $\mathbf{3}_{l, 0}$. In this case, the three basis vectors $e^{\prime}_1$, $e^{\prime}_2$ and $e^{\prime}_3$ in Eq. can not be distinguished by the actions of $c$ and $d$, the eigenstates of the generator $a$ are $$e^{\prime\prime}_1=\frac{1}{\sqrt{3}}(e^{\prime}_1+e^{\prime}_2+e^{\prime}_3), \quad
e^{\prime\prime}_2=\frac{1}{\sqrt{3}}(e^{\prime}_1+\omega e^{\prime}_2+\omega^2e^{\prime}_3), \quad
e^{\prime\prime}_3=\frac{1}{\sqrt{3}}(e^{\prime}_1+\omega^2e^{\prime}_2+\omega e^{\prime}_3)\,,$$ with $ae^{\prime\prime}_1=e^{\prime\prime}_1$, $ae^{\prime\prime}_2=\omega e^{\prime\prime}_2$ and $ae^{\prime\prime}_3=\omega^2e^{\prime\prime}_3$. Under the action of the generator $b$, $e^{\prime\prime}_1$ is mapped into itself and $e^{\prime\prime}_2$ and $e^{\prime\prime}_3$ are interchanged. Therefore the representation space of $\mathbf{3}_{l, 0}$ is split into one-dimensional subspace proportional to $e^{\prime\prime}_1$ and two-dimensional invariant subspaces spanned by $e^{\prime\prime}_2$ and $e^{\prime\prime}_3$. However, the representation matrix for $b$ is off-diagonal in the two-dimensional representation. In the present work, we would like to work in a basis where the representation matrix of $b$ is diagonal in the doublet representation such that all the relevant clebsch-gordan coefficients are real, as shown in Appendix \[sec:App\_CGC\]. Consequently we choose the basis vectors as follows $$\begin{aligned}
\nonumber&&e^{\prime\prime\prime}_1=e^{\prime\prime}_1=\frac{1}{\sqrt{3}}(e^{\prime}_1+e^{\prime}_2+e^{\prime}_3),\\
\nonumber&&e^{\prime\prime\prime}_2=\frac{1}{\sqrt{2}}(e^{\prime\prime}_2+e^{\prime\prime}_3)=\frac{1}{\sqrt{6}}(2e^\prime_1-e^\prime_2-e^\prime_3),\\
&&e^{\prime\prime\prime}_3=\frac{i}{\sqrt{2}}(e^{\prime\prime}_2-e^{\prime\prime}_3)=\frac{-1}{\sqrt{2}}(e^\prime_2-e^\prime_3)\,.\end{aligned}$$ Then we can read out the unitary basis transformation matrix as $$\mathcal{S}=\frac{1}{\sqrt{6}}\left(\begin{array}{ccc}
\sqrt{2} &~ 2 ~& 0 \\
\sqrt{2} &~ -1 ~& -\sqrt{3} \\
\sqrt{2} &~ -1 ~& \sqrt{3} \\
\end{array}\right)\,,$$ with $e^{\prime\prime\prime}_{i}=\sum^{3}_{j=1}e^{\prime}_{j}\mathcal{S}_{ji}$. In this set of new basis, the representation matrices for the generators $a$, $b$, $c$ and $d$ are $$\begin{array}{ll}
a^{\prime\prime}=\mathcal{S}^{-1}a_{\mathbf{3}_{l, 0}}\mathcal{S}=\left(\begin{array}{cc}
1 & 0 \\ 0 & a_{\mathbf{2}_{0}}
\end{array}\right), ~~&~~
b^{\prime\prime}=\mathcal{S}^{-1}b_{\mathbf{3}_{l, 0}}\mathcal{S}=\left(\begin{array}{cc}
1 & 0 \\ 0 & b_{\mathbf{2}_{0}}
\end{array}\right), \\[-7pt] \\[4pt]
c^{\prime\prime}=\mathcal{S}^{-1}c_{\mathbf{3}_{l, 0}}\mathcal{S}=\eta^{l}\mathbb{1}_{3}, ~~&~~
d^{\prime\prime}=\mathcal{S}^{-1}d_{\mathbf{3}_{l, 0}}\mathcal{S}=\mathbb{1}_{3},
\end{array}$$ with $$\label{eq:a2_b2}
a_{\mathbf{2}_{0}}=\frac{1}{2}\left(\begin{array}{cc}
-1 &~ -\sqrt{3} \\ \sqrt{3} &~-1
\end{array}\right), \quad
b_{\mathbf{2}_{0}}=\left(\begin{array}{cc}
1 &~ 0 \\ 0 &~-1
\end{array}\right)\,,$$ Note that $\eta^{l}=1, \omega,\omega^{2}$ for $l=0, 3n, 6n$ respectively.
Now we turn to another set of reducible triplet representations $\mathbf{3}_{l, 1}$ with $l=0, 3n, 6n$. In the same way as previous case, the new basis vectors are taken to be $$e^{\prime\prime}_4=\frac{1}{\sqrt{3}}(e^\prime_4+e^\prime_5+e^\prime_6)\,, \quad
e^{\prime\prime}_5=\frac{i}{\sqrt{2}}(e^\prime_5-e^\prime_6)\,, \quad
e^{\prime\prime}_6=\frac{i}{\sqrt{6}}(2e^\prime_4-e^\prime_5-e^\prime_6)\,,$$ where $e^{\prime}_4$, $e^{\prime}_5$ and $e^{\prime}_6$ are specified by Eq. . The unitary transformation for this basis change is $$\mathcal{S}=\frac{1}{\sqrt{6}}
\left(\begin{array}{ccc}
\sqrt{2} &~ 0 ~& 2i \\
\sqrt{2} &~ i\sqrt{3} ~& -i \\
\sqrt{2} &~ -i\sqrt{3} ~& -i \\
\end{array}\right)\,.$$ The corresponding representation matrices are given by $$\begin{array}{ll}
a^{\prime\prime}=\mathcal{S}^{-1}a_{\mathbf{3}_{l, 1}}\mathcal{S}=\begin{pmatrix}
1 & 0 \\ 0 & a_{\mathbf{2}_{0}}
\end{pmatrix}, ~~&~~
b^{\prime\prime}=\mathcal{S}^{-1}b_{\mathbf{3}_{l,1}}\mathcal{S}=\begin{pmatrix}
-1 & 0 \\ 0 & b_{\mathbf{2}_{0}}
\end{pmatrix}, \\[-7pt] \\[4pt]
c^{\prime\prime}=\mathcal{S}^{-1}c_{\mathbf{3}_{l, 1}}\mathcal{S}=\eta^{l}\mathbb{1}_{3}, ~~&~~
d^{\prime\prime}=\mathcal{S}^{-1}d_{\mathbf{3}_{l, 1}}\mathcal{S}=\mathbb{1}_{3}\,,
\end{array}$$ where $a_{\mathbf{2}_{0}}$ and $b_{\mathbf{2}_{0}}$ are shown in Eq. . Hence by performing similarity transformation on the reducible triplet representations $\mathbf{3}_{l, 0}$ and $\mathbf{3}_{l, 1}$ for $l=0, 3n, 6n$, we can obtain three inequivalent two-dimensional irreducible representations and six one-dimensional representations given in Eq. . The three two-dimensional representations differ in the representation matrix of $c$: $$\begin{aligned}
\nonumber&& a_{\mathbf{2}_{0}}=\frac{1}{2}\left(\begin{array}{cc}
-1 &~ -\sqrt{3} \\ \sqrt{3} &~-1
\end{array}\right),
\quad
b_{\mathbf{2}_{0}}=\left(\begin{array}{cc}
1 &~ 0 \\ 0 &~-1
\end{array}\right),
\quad
c_{\mathbf{2}_{0}}=d_{\mathbf{2}_{0}}=\mathbb{1}_{2}, \\
\nonumber&& a_{\mathbf{2}_{1}}= a_{\mathbf{2}_{0}},
\quad
b_{\mathbf{2}_{1}}= b_{\mathbf{2}_{0}},
\quad
c_{\mathbf{2}_{1}}=\omega\mathbb{1}_{2},\quad d_{\mathbf{2}_{1}}=\mathbb{1}_{2},\\
\label{eq:2_dimension}&& a_{\mathbf{2}_{2}}= a_{\mathbf{2}_{0}},
\quad
b_{\mathbf{2}_{2}}= b_{\mathbf{2}_{0}},
\quad
c_{\mathbf{2}_{2}}=\omega^{2}\mathbb{1}_{2},\quad d_{\mathbf{2}_{2}}=\mathbb{1}_{2}\,,\end{aligned}$$ which can also be sententiously written as $$a_{\mathbf{2}_{i}}=\frac{1}{2}\begin{pmatrix}
-1 &~ -\sqrt{3} \\
\sqrt{3} &~-1
\end{pmatrix},
~
b_{\mathbf{2}_{i}}= \begin{pmatrix}
1 &~ 0 \\
0 &~-1
\end{pmatrix},
~
c_{\mathbf{2}_{i}}=\omega^i\mathbb{1}_{2},~ d_{\mathbf{2}_{i}}=\mathbb{1}_{2}, ~i=0, 1, 2\,.$$ There are no more irreducible representations as we see that the number of irreducible representations is already equal to the number of conjugacy classes: $$\frac{27n(n-1)+6}{6}+2(9n-3)+3+6=1+2+(9n-3)+\frac{27n(n-1)+6}{6}+3+9n\,.$$ Furthermore, we find that the sum of the squares of the dimensions of the irreducible representations is really equal to the order of the group, i.e. $$\frac{27n(n-1)+6}{6}\times6^2+2(9n-3)\times3^2+3\times2^2+6\times1^2=162n^2\,.$$ We can derive the $D^{(1)}_{9n,3n}$ character table by taking traces over the relevant representation matrices. The results are displayed in table \[tab:characterAa\], where $\tilde{M}^{p}_{s}$ refers to $$\label{eq:Mtilde_nota}\tilde{M}^{p}_{s}\equiv\mbox{$
\left(\begin{array}{cc}
1 ~&~ 1 \\
-3 ~&~ -2
\end{array}\right)
^{p}
\left(\begin{array}{cc}
1 ~&~ 0 \\
-3 ~&~ -1
\end{array}\right)
^{s}$},\quad \mathrm{with} ~~ p=0,1,2, ~~s=0,1,$$
With the character table, it is easy to calculate the Kronecker products of two irreducible representations of the $D^{(1)}_{9n, 3n}$ group as follows: $$\begin{aligned}
\nonumber&&\mathbf{1}_{i,j}\otimes\mathbf{1}_{p,q}=\mathbf{1}_{f, g},\quad \mathbf{1}_{i,j}\otimes\mathbf{2}_{q}=\mathbf{2}_{g},\quad \mathbf{1}_{i,j}\otimes\mathbf{3}_{l, p}=\mathbf{3}_{l+3nj, f},\quad \mathbf{1}_{i,j}\otimes\mathbf{6}_{(l, k)}=\mathbf{6}_{(l+3nj, k)},\\
\nonumber&&\mathbf{2}_{j}\otimes\mathbf{2}_{q}=\mathbf{1}_{0,g}\oplus\mathbf{1}_{1,g}\oplus\mathbf{2}_{g},\quad
\mathbf{2}_{j}\otimes\mathbf{3}_{l, i}=\mathbf{3}_{l+3nj, 0}\oplus\mathbf{3}_{l+3nj, 1},\\
\nonumber&&\mathbf{2}_{j}\otimes\mathbf{6}_{(l, k)}~=\mathbf{6}_{(l+3nj, k)}\oplus\mathbf{6}_{(l+3nj,k)},\qquad \mathbf{3}_{l, i}\otimes\mathbf{3}_{ l^{\prime}, p}~=\mathbf{3}_{l+l^{\prime}, f}\oplus\mathbf{6}_{(l+l^{\prime}, l^{\prime})}\end{aligned}$$
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$1\mathcal{C}_1$ $1\mathcal{C}_1^{(\nu)}$ $3\mathcal{C}_1^{(\rho)}$ $6\mathcal{C}_1^{(\rho,\sigma)}$ $18n^2\mathcal{C}_2^{(\tau)}$ $\!\!9n\mathcal{C}_3^{(\rho)}\!\!$
----------------------- ------------------ -------------------------- --------------------------------------------------------- -------------------------------------------------------- ------------------------------- ------------------------------------
$\mathbf{1_{0,0}}$ $1$ $1$ $1$ $1$ $1$ $1$
$\mathbf{1_{0,1}}$ $1$ $1$ $\omega^{\rho}$ $\omega^{\rho}$ $\omega^{\tau}$ $\omega^{\rho}$
$\mathbf{1_{0,2}}$ $1$ $1$ $\omega^{2\rho}$ $\omega^{2\rho}$ $\omega^{2\tau}$ $\omega^{2\rho}$
$\mathbf{1_{1,0}}$ $1$ $1$ $1$ $1$ $1$ $-1$
$\mathbf{1_{1,1}}$ $1$ $1$ $\omega^{\rho}$ $\omega^{\rho}$ $\omega^{\tau}$ $-\omega^{\rho}$
$\mathbf{1_{1,2}}$ $1$ $1$ $\omega^{2\rho}$ $\omega^{2\rho}$ $\omega^{2\tau}$ $-\omega^{2\rho}$
$\mathbf{2_0}$ $2$ $2$ $2$ $2$ $-1$ $0$
$\mathbf{2_1}$ $2$ $2$ $2\omega^{\rho}$ $2\omega^{\rho}$ $-\omega^{\tau}$ $0$
$\mathbf{2_2}$ $2$ $2$ $2\omega^{2\rho}$ $2\omega^{2\rho}$ $-\omega^{2\tau}$ $0$
$\mathbf{3}_{l, 0}$ $3$ $ 3\eta^{l\nu}$ $\sum_{p}\eta^{(\rho,0)\tilde{M}^{p}_{0} $\sum_{p}\eta^{(\rho,\sigma)\tilde{M}^{p}_{0} $0$ $\!\!\eta^{l\rho}\!\!$
\mbox{\tiny $\begin{pmatrix} l \\0 \end{pmatrix}$}} $ \mbox{\tiny $\begin{pmatrix} l \\0 \end{pmatrix}$}}$
$\mathbf{3}_{l, 1}$ $3$ $3\eta^{l\nu}$ $\sum_{p}\eta^{(\rho,0)\tilde{M}^{p}_{0} $\sum_{p}\eta^{(\rho,\sigma)\tilde{M}^{p}_{0} $0$ $\!\!-\eta^{l\rho}\!\!$
\mbox{\tiny $\begin{pmatrix} l \\0 \end{pmatrix}$}} \mbox{\tiny $\begin{pmatrix} l \\0 \end{pmatrix}$}}$
$
$\mathbf{6}_{(l, k)}$ $6$ $6\eta^{l\nu}$ $\sum_{p,s}\eta^{(\rho,0)\tilde{M}^{p}_{s} $\sum_{p,s}\eta^{(\rho,\sigma)\tilde{M}^{p}_{s} 0 $0$
\mbox {\tiny $\begin{pmatrix} l\\-3k \end{pmatrix}$}} $ \mbox{ \tiny $\begin{pmatrix} l\\-3k \end{pmatrix}$}}$
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[tab:characterAa\]The character table of the $D^{(1)}_{9n,3n}$ group. The different conjugacy classes are presented in Eq. . The notation $\tilde{M}^{p}_{s}$ is explained in Eq. .
\[sec:App\_CGC\]Clebsch-Gordan Coefficients of the $D^{(1)}_{9n,3n}$ group
--------------------------------------------------------------------------
In the following, we shall decompose the product of two irreducible representations into a sum of irreducible representations of $D^{(1)}_{9n,3n}$. Under the action of the generators $a$, $b$ and $c$, different $D^{(1)}_{9n,3n}$ vector multiplets transform as follows: $$\begin{aligned}
\nonumber\mathbf{6}_{(l, k)} & :&~~
\left(\begin{array}{c} \alpha_1 \\ \alpha_2 \\ \alpha_3 \\ \alpha_4 \\ \alpha_5 \\ \alpha_6
\end{array}\right)\stackrel{a}{\mapsto}
\left(\begin{array}{c} \alpha_2 \\ \alpha_3 \\ \alpha_1 \\ \alpha_6 \\ \alpha_4 \\ \alpha_5
\end{array}\right),~~~\stackrel{b}{\mapsto}
\left(\begin{array}{c} \alpha_4 \\ \alpha_5 \\ \alpha_6 \\ \alpha_1 \\ \alpha_2 \\ \alpha_3
\end{array}\right),~~~\stackrel{c}{\mapsto}
\left(\begin{array}{c} \eta^{l}\alpha_1 \\ \eta^{l-3k}\alpha_2 \\ \eta^{3k-2l}\alpha_3 \\ \eta^{l-3k}\alpha_4 \\ \eta^{3k-2l}\alpha_5 \\ \eta^{l}\alpha_6
\end{array}\right), \\
\nonumber\mathbf{3}_{l, 0} &:&~~
\left(\begin{array}{c}\alpha_1 \\ \alpha_2 \\ \alpha_3
\end{array}\right)\stackrel{a}{\mapsto}
\left(\begin{array}{c} \alpha_2 \\ \alpha_3 \\ \alpha_1
\end{array}\right),~~~~\stackrel{b}{\mapsto}
\left(\begin{array}{c} \alpha_1 \\ \alpha_3 \\ \alpha_2
\end{array}\right),~~~\stackrel{c}{\mapsto}
\left(\begin{array}{c} \eta^{l}\alpha_1 \\ \eta^{l}\alpha_2 \\ \eta^{-2l}\alpha_3
\end{array}\right), \\
\nonumber\mathbf{3}_{l, 1} &:&~~
\left(\begin{array}{c} \alpha_1 \\ \alpha_2 \\ \alpha_3
\end{array}\right)\stackrel{a}{\mapsto}
\left(\begin{array}{c} \alpha_2 \\ \alpha_3 \\ \alpha_1
\end{array}\right),~~~~\stackrel{b}{\mapsto}
\left(\begin{array}{c} -\alpha_1 \\ -\alpha_3 \\ -\alpha_2
\end{array}\right),~~\stackrel{c}{\mapsto}
\left(\begin{array}{c} \eta^{l}\alpha_1 \\ \eta^{l}\alpha_2 \\ \eta^{-2l} \alpha_3 \end{array}\right), \\
\nonumber\mathbf{2_0} &:&~~
\left(\begin{array}{c} \alpha_1 \\ \alpha_2
\end{array}\right)\stackrel{a}{\mapsto}
\frac{1}{2}\left(\begin{array}{c}
-\alpha_1-\sqrt{3}\alpha_{2} \\ \sqrt{3}\alpha_{1}-\alpha_2
\end{array}\right),~~\stackrel{b}{\mapsto}
\left(\begin{array}{c} \alpha_1 \\ -\alpha_2
\end{array}\right),~~\stackrel{c}{\mapsto}
\left(\begin{array}{c} \alpha_1 \\ \alpha_2
\end{array}\right),\\
\nonumber\mathbf{2_1} &:&~~
\left(\begin{array}{c} \alpha_1 \\ \alpha_2
\end{array}\right)\stackrel{a}{\mapsto}
\frac{1}{2}\left(\begin{array}{c}
-\alpha_1-\sqrt{3}\alpha_{2} \\ \sqrt{3}\alpha_{1}-\alpha_2
\end{array}\right),~~\stackrel{b}{\mapsto}
\left(\begin{array}{c} \alpha_1 \\ -\alpha_2
\end{array}\right),~~\stackrel{c}{\mapsto}
\omega\left(\begin{array}{c} \alpha_1 \\ \alpha_2
\end{array}\right), \\
\label{eq:repsa}\mathbf{2_2} &:&~~
\left(\begin{array}{c} \alpha_1 \\ \alpha_2
\end{array}\right)\stackrel{a}{\mapsto}
\frac{1}{2}\left(\begin{array}{c}
-\alpha_1-\sqrt{3}\alpha_{2} \\ \sqrt{3}\alpha_{1}-\alpha_2
\end{array}\right),~~\stackrel{b}{\mapsto}
\left(\begin{array}{c} \alpha_1 \\ -\alpha_2
\end{array}\right),~~\stackrel{c}{\mapsto}
\omega^{2}\left(\begin{array}{c} \alpha_1 \\ \alpha_2
\end{array}\right)\,,\end{aligned}$$ where the action of the generator $d$ is not considered because it can be expressed in terms of $a$, $b$ and $c$, as shown in Eq. . Starting from these set of transformations rules, we can build a set of terms which define a space of an irreducible representation. Henceforth all Clebsch-Gordan (CG) coefficients would be reported in the form of $\alpha\otimes\beta$. We shall use $\alpha_{i}$ to denote the elements of first representation and $\beta_{i}$ stands for the elements of the second representation of the tensor product. Moreover, we shall denote $f\equiv i+p~(\text{mod}~2)$ and $g\equiv j+q~(\text{mod}~3)$ for simplicity of notation.
- [$\mathbf{1}_{i,j}\otimes\mathbf{1}_{p,q}~=\mathbf{1}_{f, g}$]{}\
$$\mathbf{1}_{f, g}\sim\alpha\beta\,.$$
- [$\mathbf{1}_{i,j}\otimes\mathbf{2}_{q}~=\mathbf{2}_{g}$]{}\
$$i=0~:~\mathbf{2}_{g}~\sim\left(\begin{array}{c}
\alpha\beta_1 \\
\alpha\beta_2
\end{array}\right), ~\qquad
i=1~:~ \mathbf{2}_{g}~\sim\left(\begin{array}{c}
\alpha\beta_2 \\
-\alpha\beta_1
\end{array}\right)\,.$$
- [$\mathbf{1}_{i,j}\otimes\mathbf{3}_{l, p}~=\mathbf{3}_{l+3nj, f}$]{}\
$$\mathbf{3}_{l+3nj, f}~\sim\left(\begin{array}{c}
\alpha\beta_1 \\
\alpha\beta_2 \\
\alpha\beta_3
\end{array}\right)\,.$$
- [$\mathbf{1}_{i,j}\otimes\mathbf{6}_{(l, k)}~=\mathbf{6}_{(l+3nj, k)}$]{}\
$$\mathbf{6}_{(l+3nj, k)}~\sim\left(\begin{array}{c}
\alpha\beta_1 \\
\alpha\beta_2 \\
\alpha\beta_3 \\
(-1)^{i}\alpha\beta_4 \\
(-1)^{i}\alpha\beta_5 \\
(-1)^{i}\alpha\beta_6 \\
\end{array}\right)\,.$$
- [$\mathbf{2}_{j}\otimes\mathbf{2}_{q}~=\mathbf{1}_{0,g}\oplus\mathbf{1}_{1,g}\oplus\mathbf{2}_{g}$]{}\
$$\mathbf{1}_{0,g}\sim\alpha_1\beta_1+\alpha_2\beta_2,\qquad
\mathbf{1}_{1,g}\sim\alpha_1\beta_2-\alpha_2\beta_1,\qquad
\mathbf{2}_{g}\sim\left(\begin{array}{c}
\alpha_1\beta_1-\alpha_2\beta_2 \\
-\alpha_1\beta_2-\alpha_2\beta_1\\
\end{array}\right)\,.$$
- [$\mathbf{2}_{j}\otimes\mathbf{3}_{l, i}~=\mathbf{3}_{l+3nj, i}\oplus\mathbf{3}_{l+3nj, m}$, where $m=i+1$ $(\text{mod}~2)$]{}\
$$\mathbf{3}_{l+3nj, i}~\sim\left(\begin{array}{c}
2\alpha_{1}\beta_{1}\\
-(\alpha_{1}+\sqrt{3}\alpha_{2})\beta_{2} \\
(-\alpha_{1}+\sqrt{3}\alpha_{2})\beta_{3} \\
\end{array}\right), \qquad
\mathbf{3}_{l+3nj, m}~\sim\left(\begin{array}{c}
2\alpha_{2}\beta_{1}\\
(\sqrt{3}\alpha_{1}-\alpha_{2})\beta_{2} \\
-(\sqrt{3}\alpha_{1}+\alpha_{2})\beta_{3} \\
\end{array}\right)\,.$$
- [$\mathbf{2}_{j}\otimes\mathbf{6}_{(l, k)}~=\mathbf{6}_{(l+3nj, k)}\oplus\mathbf{6}_{(l+3nj,k)}$]{}\
$$\mathbf{6}_{(l+3nj, k)}~\sim
\begin{pmatrix}
2\alpha_{1}\beta_{1}\\
-(\alpha_{1}+\sqrt{3}\alpha_{2})\beta_{2} \\
(-\alpha_{1}+\sqrt{3}\alpha_{2})\beta_{3} \\
2\alpha_{1}\beta_{4}\\
(-\alpha_1+\sqrt{3}\alpha_2)\beta_{5} \\
-(\alpha_1+\sqrt{3}\alpha_2)\beta_{6} \\
\end{pmatrix}, \qquad
\mathbf{6}_{(l+3nj, k)}~\sim
\begin{pmatrix}
2\alpha_{2}\beta_{1}\\
(\sqrt{3}\alpha_{1}-\alpha_{2})\beta_{2} \\
-(\sqrt{3}\alpha_{1}+\alpha_{2})\beta_{3} \\
-2\alpha_{2}\beta_{4}\\
(\sqrt{3}\alpha_{1}+\alpha_{2})\beta_{5} \\
-(\sqrt{3}\alpha_{1}-\alpha_{2})\beta_{6} \\
\end{pmatrix}\,.$$
- [$\mathbf{3}_{l, i}\otimes\mathbf{3}_{ l^{\prime}, p}~=\mathbf{3}_{l+l^{\prime}, f}\oplus\mathbf{6}_{(l+l^{\prime}, l^{\prime})}$]{}\
$$\begin{array}{ll}
\mathbf{3}_{l+l^{\prime}, f}~\sim\left(\begin{array}{c}
\alpha_{1}\beta_{1}\\
\alpha_{2}\beta_{2} \\
\alpha_{3}\beta_{3} \\
\end{array}\right),
~~&~~
\mathbf{6}_{(l+l^{\prime}, l^{\prime})}~\sim\left(\begin{array}{c}
\alpha_{1}\beta_{2}\\
\alpha_{2}\beta_{3} \\
\alpha_{3}\beta_{1} \\
(-1)^{i-p}\alpha_{1}\beta_{3}\\
(-1)^{i-p}\alpha_{3}\beta_{2} \\
(-1)^{i-p}\alpha_{2}\beta_{1} \\
\end{array}\right)\,.
\end{array}$$
- [$\mathbf{3}_{l, i}\otimes\mathbf{6}_{(l^{\prime}, k^{\prime})}~
=\mathbf{6}_{(l+l^{\prime}, k^{\prime})}\oplus
\mathbf{6}_{(l+l^{\prime}-3k^{\prime}, l^{\prime}-2k^{\prime})}
\oplus\mathbf{6}_{(l-2l^{\prime}+3k^{\prime}, k^{\prime}-l^{\prime})}$]{}\
$$\begin{array}{lll}
\mathbf{6}_{(l+l^{\prime}, k^{\prime})}\sim\begin{pmatrix}
\alpha_{1}\beta_{1}\\
\alpha_{2}\beta_{2} \\
\alpha_{3}\beta_{3} \\
(-1)^{i}\alpha_{1}\beta_{4}\\
(-1)^{i}\alpha_{3}\beta_{5} \\
(-1)^{i}\alpha_{2}\beta_{6}
\end{pmatrix},
&~ \mathbf{6}_{(l+l^{\prime}-3k^{\prime}, l^{\prime}-2k^{\prime})}\sim\begin{pmatrix}
\alpha_{1}\beta_{2}\\
\alpha_{2}\beta_{3} \\
\alpha_{3}\beta_{1} \\
(-1)^{i}\alpha_{1}\beta_{5}\\
(-1)^{i}\alpha_{3}\beta_{6} \\
(-1)^{i}\alpha_{2}\beta_{4}
\end{pmatrix},
~&\mathbf{6}_{(l-2l^{\prime}+3k^{\prime}, k^{\prime}-l^{\prime})}
\sim\begin{pmatrix}
\alpha_{1}\beta_{3}\\
\alpha_{2}\beta_{1} \\
\alpha_{3}\beta_{2} \\
(-1)^{i}\alpha_{1}\beta_{6}\\
(-1)^{i}\alpha_{3}\beta_{4} \\
(-1)^{i}\alpha_{2}\beta_{5}
\end{pmatrix}\,.
\end{array}$$
- [$\mathbf{6}_{(l, k)}\otimes\mathbf{6}_{(l^{\prime}, k^{\prime})}~
=\mathbf{6}_{(l+l^{\prime}, k+k^{\prime})}\oplus
\mathbf{6}_{(l+l^{\prime}-3k^{\prime}, k-2k^{\prime}+l^{\prime})}
\oplus\mathbf{6}_{(l-2l^{\prime}+3k^{\prime}, k+k^{\prime}-l^{\prime})}$]{}\
$\oplus\mathbf{6}_{(l+l^{\prime}-3k^{\prime}, k-k^{\prime})}
\oplus\mathbf{6}_{(l-2l^{\prime}+3k^{\prime}, k+2k^{\prime}-l^{\prime})}
\oplus\mathbf{6}_{(l+l^{\prime},
k-k^{\prime}+l^{\prime})}$\
$$\begin{array}{lll}
\mathbf{6}_{(l+l^{\prime}, k+k^{\prime})}\sim\begin{pmatrix}
\alpha_{1}\beta_{1}\\
\alpha_{2}\beta_{2} \\
\alpha_{3}\beta_{3} \\
\alpha_{4}\beta_{4}\\
\alpha_{5}\beta_{5} \\
\alpha_{6}\beta_{6}
\end{pmatrix},
& \mathbf{6}_{(l+l^{\prime}-3k^{\prime},
k-2k^{\prime}+l^{\prime})}\sim
\begin{pmatrix}
\alpha_{1}\beta_{2}\\
\alpha_{2}\beta_{3} \\
\alpha_{3}\beta_{1} \\
\alpha_{4}\beta_{5}\\
\alpha_{5}\beta_{6} \\
\alpha_{6}\beta_{4}
\end{pmatrix},
& \mathbf{6}_{(l-2l^{\prime}+3k^{\prime},
k+k^{\prime}-l^{\prime})}\sim\begin{pmatrix}
\alpha_{1}\beta_{3}\\
\alpha_{2}\beta_{1} \\
\alpha_{3}\beta_{2} \\
\alpha_{4}\beta_{6}\\
\alpha_{5}\beta_{4} \\
\alpha_{6}\beta_{5}
\end{pmatrix},\\[-7pt]\\[4pt]
\mathbf{6}_{(l+l^{\prime}-3k^{\prime},
k-k^{\prime})}\sim\begin{pmatrix}
\alpha_{1}\beta_{4}\\
\alpha_{2}\beta_{6} \\
\alpha_{3}\beta_{5} \\
\alpha_{4}\beta_{1}\\
\alpha_{5}\beta_{3} \\
\alpha_{6}\beta_{2}
\end{pmatrix}, &
\mathbf{6}_{(l-2l^{\prime}+3k^{\prime},
k+2k^{\prime}-l^{\prime})}\sim
\begin{pmatrix}
\alpha_{1}\beta_{5}\\
\alpha_{2}\beta_{4} \\
\alpha_{3}\beta_{6} \\
\alpha_{4}\beta_{2}\\
\alpha_{5}\beta_{1} \\
\alpha_{6}\beta_{3}
\end{pmatrix},
& \mathbf{6}_{(l+l^{\prime},
k-k^{\prime}+l^{\prime})}\sim
\begin{pmatrix}
\alpha_{1}\beta_{6}\\
\alpha_{2}\beta_{5} \\
\alpha_{3}\beta_{4} \\
\alpha_{4}\beta_{3}\\
\alpha_{5}\beta_{2} \\
\alpha_{6}\beta_{1}
\end{pmatrix}\,.
\end{array}$$ We would like to point out that certain three-dimensional and six-dimensional representations in the above tensor product decompositions may be reducible, and accordingly it should be reduced into smaller irreducible representations of the $D^{(1)}_{9n, 3n}$ group. The reducible conditions and corresponding reduction formulae are summarized in table \[tab:reducible\].
------------------------------------------------------ --------------------------------------------------------------------- --------------------------------------------------------------------- ---------------------------------------------------------------------------
$\mathbf{3}_{l, i}~\sim\begin{pmatrix}
\gamma_{1}\\
\gamma_{2} \\
\gamma_{3}\\
\end{pmatrix}$
$\left\{\begin{array}{c} $\left\{\begin{array}{c} $\left\{\begin{array}{c}
3k=0~(\text{mod}~9n)\\ 3l-3k=0~(\text{mod}~9n)\\ 3l-6k=0~(\text{mod}~9n)\\
l\neq0~(\text{mod}~3n) l\neq0~(\text{mod}~3n) l, k \neq0~(\text{mod}~3n)
\end{array}\right.$ \end{array}\right.$ \end{array}\right.$
\[-0.18in\]
\[-0.18in\] $\mathbf{6}_{(l , k)}\cong\mathbf{3}_{l, 0}\oplus\mathbf{3}_{l, 1}$ $\mathbf{6}_{(l , k)}\cong\mathbf{3}_{l, 0}\oplus\mathbf{3}_{l, 1}$ $\mathbf{6}_{(l , k)}\cong\mathbf{3}_{l-3k, 0}\oplus\mathbf{3}_{l-3k, 1}$
\[-0.18in\] $\mathbf{6}_{(l, k)}\sim \begin{pmatrix} $\begin{array}{c} $\begin{array}{c} $\begin{array}{c}
\gamma_{1}\\ \mathbf{3}_{l, 0}\sim \begin{pmatrix} \mathbf{3}_{l, 0}\sim \begin{pmatrix} \mathbf{3}_{l-3k, 0}\sim \begin{pmatrix}
\gamma_{2} \\ \gamma_{1}+\gamma_{4}\\ \gamma_{3}+\gamma_{6}\\ \gamma_{2}+\gamma_{5}\\
\gamma_{3} \\ \gamma_{2}+\gamma_{6} \\ \gamma_{1}+\gamma_{5} \\ \gamma_{3}+\gamma_{4} \\
\gamma_{4}\\ \gamma_{3}+\gamma_{5} \gamma_{2}+\gamma_{4} \gamma_{1}+\gamma_{6}
\gamma_{5} \\ \end{pmatrix}\\ \end{pmatrix}\\ \end{pmatrix}\\
\gamma_{6} \\[-0.1in] \\[-0.1in] \\[-0.1in]
\end{pmatrix}$ \mathbf{3}_{l, 1}\sim \begin{pmatrix} \mathbf{3}_{l, 1}\sim \begin{pmatrix} \mathbf{3}_{l-3k, 1}\sim \begin{pmatrix}
\gamma_{1}-\gamma_{4}\\ \gamma_{3}-\gamma_{6}\\ \gamma_{2}-\gamma_{5}\\
\gamma_{2}-\gamma_{6} \\ \gamma_{1}-\gamma_{5} \\ \gamma_{3}-\gamma_{4} \\
\gamma_{3}-\gamma_{5} \gamma_{2}-\gamma_{4} \gamma_{1}-\gamma_{6}
\end{pmatrix} \end{pmatrix} \end{pmatrix}
\end{array}$ \end{array}$ \end{array}$
\[-0.18in\]
$\mathbf{6}_{(l, k)}\sim \begin{pmatrix}
\gamma_{1}\\
\gamma_{2} \\
\gamma_{3} \\
\gamma_{4}\\
\gamma_{5} \\
\gamma_{6}
\end{pmatrix}$
\[-0.18in\]
------------------------------------------------------ --------------------------------------------------------------------- --------------------------------------------------------------------- ---------------------------------------------------------------------------
: \[tab:reducible\]The reducible conditions for $\mathbf{3}_{l, 0}$, $\mathbf{3}_{l, 1}$ and $\mathbf{6}_{(l, k)}$, and the decomposition of reducible three-dimensional and six-dimensional representations into smaller irreducible representations of $D^{(1)}_{9n, 3n}$. Notice that the expression for the doublet vector $\mathbf{2}_{\lambda}$ is not unique in the decomposition of $\mathbf{6}_{(3n\lambda, 0)}\cong\mathbf{1}_{0, \lambda}\oplus\mathbf{1}_{1, \lambda}\oplus\mathbf{2}_{\lambda}\oplus\mathbf{2}_{\lambda}$.
K. Abe [*et al.*]{} \[T2K Collaboration\], Phys. Rev. Lett. [**107**]{}, 041801 (2011) doi:10.1103/PhysRevLett.107.041801 \[arXiv:1106.2822 \[hep-ex\]\].
P. Adamson [*et al.*]{} \[MINOS Collaboration\], Phys. Rev. Lett. [**107**]{}, 181802 (2011) doi:10.1103/PhysRevLett.107.181802 \[arXiv:1108.0015 \[hep-ex\]\].
Y. Abe [*et al.*]{} \[Double Chooz Collaboration\], Phys. Rev. Lett. [**108**]{}, 131801 (2012) doi:10.1103/PhysRevLett.108.131801 \[arXiv:1112.6353 \[hep-ex\]\]; Phys. Rev. D [**86**]{}, 052008 (2012) doi:10.1103/PhysRevD.86.052008 \[arXiv:1207.6632 \[hep-ex\]\].
F. P. An [*et al.*]{} \[Daya Bay Collaboration\], Phys. Rev. Lett. [**108**]{}, 171803 (2012) doi:10.1103/PhysRevLett.108.171803 \[arXiv:1203.1669 \[hep-ex\]\]; Chin. Phys. C [**37**]{}, 011001 (2013) doi:10.1088/1674-1137/37/1/011001 \[arXiv:1210.6327 \[hep-ex\]\].
J. K. Ahn [*et al.*]{} \[RENO Collaboration\], Phys. Rev. Lett. [**108**]{}, 191802 (2012) doi:10.1103/PhysRevLett.108.191802 \[arXiv:1204.0626 \[hep-ex\]\]. K. Abe [*et al.*]{} \[T2K Collaboration\], Phys. Rev. D [**91**]{}, no. 7, 072010 (2015) doi:10.1103/PhysRevD.91.072010 \[arXiv:1502.01550 \[hep-ex\]\].
F. Capozzi, G. L. Fogli, E. Lisi, A. Marrone, D. Montanino and A. Palazzo, Phys. Rev. D [**89**]{}, 093018 (2014) doi:10.1103/PhysRevD.89.093018 \[arXiv:1312.2878 \[hep-ph\]\]. D. V. Forero, M. Tortola and J. W. F. Valle, Phys. Rev. D [**90**]{}, no. 9, 093006 (2014) doi:10.1103/PhysRevD.90.093006 \[arXiv:1405.7540 \[hep-ph\]\]. M. C. Gonzalez-Garcia, M. Maltoni and T. Schwetz, JHEP [**1411**]{}, 052 (2014) doi:10.1007/JHEP11(2014)052 \[arXiv:1409.5439 \[hep-ph\]\]. G. Altarelli and F. Feruglio, Rev. Mod. Phys. [**82**]{}, 2701 (2010) doi:10.1103/RevModPhys.82.2701 \[arXiv:1002.0211 \[hep-ph\]\].
H. Ishimori, T. Kobayashi, H. Ohki, Y. Shimizu, H. Okada and M. Tanimoto, Prog. Theor. Phys. Suppl. [**183**]{}, 1 (2010) doi:10.1143/PTPS.183.1 \[arXiv:1003.3552 \[hep-th\]\].
S. F. King and C. Luhn, Rept. Prog. Phys. [**76**]{}, 056201 (2013) doi:10.1088/0034-4885/76/5/056201 \[arXiv:1301.1340 \[hep-ph\]\]; S. F. King, A. Merle, S. Morisi, Y. Shimizu and M. Tanimoto, New J. Phys. [**16**]{}, 045018 (2014) doi:10.1088/1367-2630/16/4/045018 \[arXiv:1402.4271 \[hep-ph\]\]; S. F. King, J. Phys. G [**42**]{}, 123001 (2015) doi:10.1088/0954-3899/42/12/123001 \[arXiv:1510.02091 \[hep-ph\]\].
R. M. Fonseca and W. Grimus, JHEP [**1409**]{}, 033 (2014) doi:10.1007/JHEP09(2014)033 \[arXiv:1405.3678 \[hep-ph\]\].
C. Y. Yao and G. J. Ding, Phys. Rev. D [**92**]{}, no. 9, 096010 (2015) doi:10.1103/PhysRevD.92.096010 \[arXiv:1505.03798 \[hep-ph\]\].
P. Chen, C. C. Li and G. J. Ding, Phys. Rev. D [**91**]{}, 033003 (2015) doi:10.1103/PhysRevD.91.033003 \[arXiv:1412.8352 \[hep-ph\]\].
P. Chen, C. Y. Yao and G. J. Ding, Phys. Rev. D [**92**]{}, no. 7, 073002 (2015) doi:10.1103/PhysRevD.92.073002 \[arXiv:1507.03419 \[hep-ph\]\].
L. L. Everett, T. Garon and A. J. Stuart, JHEP [**1504**]{}, 069 (2015) doi:10.1007/JHEP04(2015)069 \[arXiv:1501.04336 \[hep-ph\]\].
P. F. Harrison and W. G. Scott, Phys. Lett. B [**535**]{}, 163 (2002) doi:10.1016/S0370-2693(02)01753-7 \[hep-ph/0203209\]; P. F. Harrison and W. G. Scott, Phys. Lett. B [**547**]{}, 219 (2002) doi:10.1016/S0370-2693(02)02772-7 \[hep-ph/0210197\]; P. F. Harrison and W. G. Scott, Phys. Lett. B [**594**]{}, 324 (2004) doi:10.1016/j.physletb.2004.05.039 \[hep-ph/0403278\].
W. Grimus and L. Lavoura, Phys. Lett. B [**579**]{}, 113 (2004) doi:10.1016/j.physletb.2003.10.075 \[hep-ph/0305309\]; W. Grimus and L. Lavoura, Fortsch. Phys. [**61**]{}, 535 (2013) doi:10.1002/prop.201200118 \[arXiv:1207.1678 \[hep-ph\]\].
Y. Farzan and A. Y. Smirnov, JHEP [**0701**]{}, 059 (2007) doi:10.1088/1126-6708/2007/01/059 \[hep-ph/0610337\].
P. Chen, G. J. Ding, F. Gonzalez-Canales and J. W. F. Valle, arXiv:1512.01551 \[hep-ph\].
F. Feruglio, C. Hagedorn and R. Ziegler, JHEP [**1307**]{}, 027 (2013) doi:10.1007/JHEP07(2013)027 \[arXiv:1211.5560 \[hep-ph\]\]. M. Holthausen, M. Lindner and M. A. Schmidt, JHEP [**1304**]{}, 122 (2013) doi:10.1007/JHEP04(2013)122 \[arXiv:1211.6953 \[hep-ph\]\].
G. Ecker, W. Grimus and W. Konetschny, Nucl. Phys. B [**191**]{}, 465 (1981) doi:10.1016/0550-3213(81)90309-6; G. Ecker, W. Grimus and H. Neufeld, Nucl. Phys. B [**247**]{}, 70 (1984) doi:10.1016/0550-3213(84)90373-0; J. Bernabeu, G. C. Branco and M. Gronau, Phys. Lett. B [**169**]{}, 243 (1986) doi:10.1016/0370-2693(86)90659-3; G. Ecker, W. Grimus and H. Neufeld, J. Phys. A [**20**]{}, L807 (1987) doi:10.1088/0305-4470/20/12/010; H. Neufeld, W. Grimus and G. Ecker, Int. J. Mod. Phys. A [**3**]{}, 603 (1988) doi:10.1142/S0217751X88000254.
W. Grimus and M. N. Rebelo, Phys. Rept. [**281**]{}, 239 (1997) doi:10.1016/S0370-1573(96)00030-0 \[hep-ph/9506272\]. M. C. Chen, M. Fallbacher, K. T. Mahanthappa, M. Ratz and A. Trautner, Nucl. Phys. B [**883**]{}, 267 (2014) doi:10.1016/j.nuclphysb.2014.03.023 \[arXiv:1402.0507 \[hep-ph\]\].
G. J. Ding, S. F. King and A. J. Stuart, JHEP [**1312**]{}, 006 (2013) doi:10.1007/JHEP12(2013)006 \[arXiv:1307.4212 \[hep-ph\]\]. G. J. Ding, S. F. King, C. Luhn and A. J. Stuart, JHEP [**1305**]{}, 084 (2013) doi:10.1007/JHEP05(2013)084 \[arXiv:1303.6180 \[hep-ph\]\].
F. Feruglio, C. Hagedorn and R. Ziegler, Eur. Phys. J. C [**74**]{}, 2753 (2014) doi:10.1140/epjc/s10052-014-2753-2 \[arXiv:1303.7178 \[hep-ph\]\].
C. Luhn, Nucl. Phys. B [**875**]{}, 80 (2013) doi:10.1016/j.nuclphysb.2013.07.003 \[arXiv:1306.2358 \[hep-ph\]\]. C. C. Li and G. J. Ding, Nucl. Phys. B [**881**]{}, 206 (2014) doi:10.1016/j.nuclphysb.2014.02.002 \[arXiv:1312.4401 \[hep-ph\]\].
C. C. Li and G. J. Ding, JHEP [**1508**]{}, 017 (2015) doi:10.1007/JHEP08(2015)017 \[arXiv:1408.0785 \[hep-ph\]\].
G. C. Branco, I. de Medeiros Varzielas and S. F. King, Nucl. Phys. B [**899**]{}, 14 (2015) doi:10.1016/j.nuclphysb.2015.07.024 \[arXiv:1505.06165 \[hep-ph\]\]; G. C. Branco, I. de Medeiros Varzielas and S. F. King, Phys. Rev. D [**92**]{}, no. 3, 036007 (2015) doi:10.1103/PhysRevD.92.036007 \[arXiv:1502.03105 \[hep-ph\]\].
G. J. Ding and Y. L. Zhou, Chin. Phys. C [**39**]{}, no. 2, 021001 (2015) doi:10.1088/1674-1137/39/2/021001 \[arXiv:1312.5222 \[hep-ph\]\]; G. J. Ding and Y. L. Zhou, JHEP [**1406**]{}, 023 (2014) doi:10.1007/JHEP06(2014)023 \[arXiv:1404.0592 \[hep-ph\]\].
C. C. Li and G. J. Ding, JHEP [**1505**]{}, 100 (2015) doi:10.1007/JHEP05(2015)100 \[arXiv:1503.03711 \[hep-ph\]\].
A. Di Iura, C. Hagedorn and D. Meloni, JHEP [**1508**]{}, 037 (2015) doi:10.1007/JHEP08(2015)037 \[arXiv:1503.04140 \[hep-ph\]\].
P. Ballett, S. Pascoli and J. Turner, Phys. Rev. D [**92**]{}, no. 9, 093008 (2015) doi:10.1103/PhysRevD.92.093008 \[arXiv:1503.07543 \[hep-ph\]\].
G. J. Ding and S. F. King, Phys. Rev. D [**89**]{}, no. 9, 093020 (2014) doi:10.1103/PhysRevD.89.093020 \[arXiv:1403.5846 \[hep-ph\]\].
C. Hagedorn, A. Meroni and E. Molinaro, Nucl. Phys. B [**891**]{}, 499 (2015) doi:10.1016/j.nuclphysb.2014.12.013 \[arXiv:1408.7118 \[hep-ph\]\].
G. J. Ding and S. F. King, arXiv:1510.03188 \[hep-ph\].
G. J. Ding, S. F. King and T. Neder, JHEP [**1412**]{}, 007 (2014) doi:10.1007/JHEP12(2014)007 \[arXiv:1409.8005 \[hep-ph\]\].
S. F. King and T. Neder, Phys. Lett. B [**736**]{}, 308 (2014) doi:10.1016/j.physletb.2014.07.043 \[arXiv:1403.1758 \[hep-ph\]\]. G.A. Miller, H.F. Blichfeldt and L.E. Dickson, *Theory and applications of finite groups* (John Wiley & Sons, New York, 1916).
W. Grimus and P. O. Ludl, J. Phys. A [**43**]{}, 445209 (2010) doi:10.1088/1751-8113/43/44/445209 \[arXiv:1006.0098 \[hep-ph\]\].
W. Grimus and P. O. Ludl, J. Phys. A [**47**]{}, no. 7, 075202 (2014) doi:10.1088/1751-8113/47/7/075202 \[arXiv:1310.3746 \[math-ph\]\].
W. Grimus and P. O. Ludl, J. Phys. A [**45**]{}, 233001 (2012) doi:10.1088/1751-8113/45/23/233001 \[arXiv:1110.6376 \[hep-ph\]\].
S. F. King, T. Neder and A. J. Stuart, Phys. Lett. B [**726**]{}, 312 (2013) doi:10.1016/j.physletb.2013.08.052 \[arXiv:1305.3200 \[hep-ph\]\].
J. A. Escobar and C. Luhn, J. Math. Phys. [**50**]{}, 013524 (2009) doi:10.1063/1.3046563 \[arXiv:0809.0639 \[hep-th\]\].
The GAP Group, [*[GAP – Groups, Algorithms, and Programming, Version 4.4.12]{}*]{}, 2008. <http://www.gap-system.org>.
C. S. Lam, Phys. Lett. B [**656**]{}, 193 (2007) doi:10.1016/j.physletb.2007.09.032 \[arXiv:0708.3665 \[hep-ph\]\]. G. Altarelli, F. Feruglio and L. Merlo, JHEP [**0905**]{}, 020 (2009) doi:10.1088/1126-6708/2009/05/020 \[arXiv:0903.1940 \[hep-ph\]\].
K. A. Olive [*et al.*]{} \[Particle Data Group Collaboration\], Chin. Phys. C [**38**]{}, 090001 (2014). doi:10.1088/1674-1137/38/9/090001
F. An [*et al.*]{} \[JUNO Collaboration\], arXiv:1507.05613 \[physics.ins-det\].
S. B. Kim, Nucl. Part. Phys. Proc. [**265-266**]{}, 93 (2015) doi:10.1016/j.nuclphysbps.2015.06.024 \[arXiv:1412.2199 \[hep-ex\]\].
C. Adams [*et al.*]{} \[LBNE Collaboration\], arXiv:1307.7335 \[hep-ex\]; M. Bass [*et al.*]{} \[LBNE Collaboration\], Phys. Rev. D [**91**]{}, no. 5, 052015 (2015) doi:10.1103/PhysRevD.91.052015 \[arXiv:1311.0212 \[hep-ex\]\]. S. K. Agarwalla [*et al.*]{} \[LAGUNA-LBNO Collaboration\], JHEP [**1405**]{}, 094 (2014) doi:10.1007/JHEP05(2014)094 \[arXiv:1312.6520 \[hep-ph\]\]; L. Agostino [*et al.*]{}, arXiv:1409.4405 \[physics.ins-det\]; S. K. Agarwalla [*et al.*]{} \[LAGUNA-LBNO Collaboration\], arXiv:1412.0593 \[hep-ph\]; S. K. Agarwalla [*et al.*]{} \[LAGUNA-LBNO Collaboration\], arXiv:1412.0804 \[hep-ph\]. E. Kearns [*et al.*]{} \[Hyper-Kamiokande Working Group Collaboration\], arXiv:1309.0184 \[hep-ex\].
E. Baussan, M. Dracos, T. Ekelof, E. F. Martinez, H. Ohman and N. Vassilopoulos, arXiv:1212.5048 \[hep-ex\]; E. Baussan [*et al.*]{} \[ESSnuSB Collaboration\], Nucl. Phys. B [**885**]{}, 127 (2014) doi:10.1016/j.nuclphysb.2014.05.016 \[arXiv:1309.7022 \[hep-ex\]\].
S. Geer, Phys. Rev. D [**57**]{}, 6989 (1998) \[Phys. Rev. D [**59**]{}, 039903 (1999)\] doi:10.1103/PhysRevD.57.6989, 10.1103/PhysRevD.59.039903 \[hep-ph/9712290\]; A. De Rujula, M. B. Gavela and P. Hernandez, Nucl. Phys. B [**547**]{}, 21 (1999) doi:10.1016/S0550-3213(99)00070-X \[hep-ph/9811390\]; A. Bandyopadhyay [*et al.*]{} \[ISS Physics Working Group Collaboration\], Rept. Prog. Phys. [**72**]{}, 106201 (2009) doi:10.1088/0034-4885/72/10/106201 \[arXiv:0710.4947 \[hep-ph\]\].
M. Agostini [*et al.*]{} \[GERDA Collaboration\], Phys. Rev. Lett. [**111**]{}, no. 12, 122503 (2013) doi:10.1103/PhysRevLett.111.122503 \[arXiv:1307.4720 \[nucl-ex\]\].
M. Auger [*et al.*]{} \[EXO-200 Collaboration\], Phys. Rev. Lett. [**109**]{}, 032505 (2012) doi:10.1103/PhysRevLett.109.032505 \[arXiv:1205.5608 \[hep-ex\]\]. J. B. Albert [*et al.*]{} \[EXO-200 Collaboration\], Nature [**510**]{}, 229 (2014) doi:10.1038/nature13432 \[arXiv:1402.6956 \[nucl-ex\]\]. A. Gando [*et al.*]{} \[KamLAND-Zen Collaboration\], Phys. Rev. Lett. [**110**]{}, no. 6, 062502 (2013) doi:10.1103/PhysRevLett.110.062502 \[arXiv:1211.3863 \[hep-ex\]\].
F. Piquemal, Nucl. Phys. Proc. Suppl. [**235-236**]{}, 273 (2013). doi:10.1016/j.nuclphysbps.2013.04.022
P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], Astron. Astrophys. [**571**]{}, A16 (2014) doi:10.1051/0004-6361/201321591 \[arXiv:1303.5076 \[astro-ph.CO\]\].
M. Yoshimura, Phys. Rev. D [**75**]{}, 113007 (2007) doi:10.1103/PhysRevD.75.113007 \[hep-ph/0611362\].
G. W. Mackey, Amer. J. Math. [**73**]{}, 576 (1951); G. W. Mackey, Acta Math. [**99**]{}, 265 (1958); L. Corwing, Proc. Amer. Math. Soc. [**47**]{}, 279 (1975);
B. Simon, *Representations of finite and compact groups*, (American Mathematical Soc., 1996).
J. -P. Serre, *Linear representations of finite groups*, (Springer Science & Business Media, 2012).
[^1]: E-mail: [[email protected]]{}
[^2]: E-mail: [[email protected]]{}
[^3]: E-mail: [[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present spatially and spectrally-resolved observations of CH$_3$OH emission from comet C/2012 K1 (PanSTARRS) using The Atacama Large Millimeter/submillimeter Array (ALMA) on 2014 June 28-29. Two-dimensional maps of the line-of-sight average rotational temperature ($T_{rot}$) were derived, covering spatial scales $0.3''''-1.8''''$ (corresponding to sky-projected distances $\rho\sim500$-2500 km). The CH$_3$OH column density distributions are consistent with isotropic, uniform outflow from the nucleus, with no evidence for extended sources of CH$_3$OH in the coma. The $T_{rot}(\rho)$ radial profiles show a significant drop within a few thousand kilometers of the nucleus, falling from about 60 K to 20 K between $\rho=0$ and 2500 km on June 28, whereas on June 29, $T_{rot}$ fell from about 120 K to 40 K between $\rho=$ 0 km and 1000 km. The observed $T_{rot}$ behavior is interpreted primarily as a result of variations in the coma kinetic temperature due to adiabatic cooling of the outflowing gas, as well as radiative cooling of the CH$_3$OH rotational levels. Our excitation model shows that radiative cooling is more important for the $J=7-6$ transitions (at 338 GHz) than for the $K=3-2$ transitions (at 252 GHz), resulting in a strongly sub-thermal distribution of levels in the $J=7-6$ band at $\rho\gtrsim1000$ km. For both bands, the observed temperature drop with distance is less steep than predicted by standard coma theoretical models, which suggests the presence of a significant source of heating in addition to the photolytic heat sources usually considered.'
author:
- 'M. A. Cordiner, N. Biver, J. Crovisier, D. Bockel[é]{}e-Morvan, M. J. Mumma, S. B. Charnley, G. Villanueva, L. Paganini, D. C. Lis, S. N. Milam, A. J. Remijan, I. M. Coulson, Y.-J. Kuan, J. Boissier'
bibliography:
- 'refs.bib'
title: 'Thermal physics of the inner coma: ALMA studies of the methanol distribution and excitation in comet C/2012 K1 (PanSTARRS)'
---
Introduction {#sec:intro}
============
Comets are considered fossils of the early Solar System — frozen relics containing ice, dust and debris from the protoplanetary accretion disk. Having existed in a relatively quiescent state since their formation [[*e.g.*]{} @dav16], cometary compositions can provide unique information on the thermal and chemical characteristics of the early Solar System. Most of our knowledge on cometary compositions comes from remote (ground-based) observations of their gaseous atmospheres/comae [@coc15], for which the (typically relatively low) angular resolution and incomplete spatial coverage limits the amount of information that can be obtained. A particular problem is the lack of understanding regarding the physical and chemical structure of the near-nucleus coma, at distances less than a few thousand kilometres from the comet’s surface.
With the advent of the Atacama Large Millimeter/submillimeter Array, high-sensitivity, high angular-resolution millimeter-wave interferometry of typical, moderately bright comets has become possible. ALMA’s unique capabilities allow us to probe the physical and chemical structure of the innermost regions of the coma in unprecedented detail, leading to new insights into the properties of the coma and the nucleus. The first cometary observations using ALMA were reported by @cor14, who measured the distributions of HCN, HNC and H$_2$CO in comets C/2012 F6 (Lemmon) and C/2012 S1 (ISON) and demonstrated unequivocally that HNC and H$_2$CO are released in the coma (as a result of photolytic and/or thermal processes), whereas HCN originates from (or very near to) the nucleus.
For simplicity, a constant coma kinetic temperature is commonly assumed during analysis of microwave and sub-mm cometary observations. However, theoretical and observational studies increasingly show that temperatures can vary substantially over short distance scales in the coma. Strong variations in temperature within a few hundred kilometers of the nucleus are predicted by coma hydrodynamic/Monte Carlo simulations [[*e.g.*]{} @kor87; @com99b; @com99; @rod02; @ten08], but these models are largely untested due to a lack of comparative observational data. Early observational reports of inner coma temperature variations were based on long-slit infrared spectroscopy of HCN and CO in the unusually active comet C/1995 O1 (Hale-Bopp) [@mag99; @dis01]. In a far-IR study of H$_2$O and HDO emission from comet C/2009 P1 (Garradd) by @boc12, a strongly variable temperature law was suggested — reaching a minimum at (4-20)$\times10^3$ km from the nucleus and rising to 150 K in the outer coma. Advances in ground-based infrared instrumentation and data analysis techniques permitted @bon07 [@bon08; @bon13; @bon14] to detect significant variations in the H$_2$O rotational excitation temperature over distance scales $\sim20$-1500 km in four comets, which led to the first detailed (quantitative) comparisons with theory for comet 73P/Schwassmann-Wachmann 3 [@fou12].
To-date, a general lack of information concerning spatial variations in the excitation of cometary molecules impedes the accuracy of important results in the cometary literature, hindering the derivation of accurate cometary mixing ratios, as well as introducing errors into the parent scale lengths of distributed sources. Thus, detailed measurements of coma temperatures are required in order to confirm and expand upon the previous observational findings, to stimulate revision and refinement of theoretical models.
By virtue of its large abundance in comets (typically on the order of a few percent with respect to H$_2$O), and its strong rotational bands throughout the mm/sub-mm, methanol (CH$_3$OH), is an ideal molecule for mapping the coma temperature distribution. Here we present results exploting the high resolution and sensitivity of ALMA to provide new information on the distribution and excitation of CH$_3$OH in the inner coma of the Oort-cloud comet C/2012 K1 (PanSTARRS), [and interpret our observations using non-LTE radiative transfer models]{}.
ALMA CH$_3$OH observations
==========================
ALMA observations of C/2012 K1 (PanSTARRS) were obtained pre-perihelion during 2014-06-28 19:07-20:05 and 2014-06-29 17:37-18:26, while the comet was at a heliocentric distance $r_H=1.42$-1.43 AU and geocentric distance $\Delta=1.96$-1.97 AU (the comet reached perihelion at $r_H=1.05$ AU on 2014-08-28). The CH$_3$OH $K=3-2$ rotational band near 251.9 GHz (in ALMA receiver band 6) was observed on June 28 and the $J=7-6$ band near 338.5 GHz (in band 7) was observed on June 29. Atmospheric conditions were excellent throughout (with zenith PWV $<0.4$ mm). Thirty 12-m antennae, with baseline lengths 20-650 m, resulted in an angular resolution of $0.80''\times0.43''$ at 252 GHz and $0.71''\times0.33''$ at 338 GHz. These beam dimensions correspond to $1100\times610$ km and $1000\times470$ km, respectively, at the distance of the comet and the measured RMS noise levels per channel were 1.8 mJybm$^{-1}$ and 3.0 mJybm$^{-1}$. The correlator was configured to simultaneously observe as many strong CH$_3$OH lines as possible (spanning different upper-state energy levels) in a single 976 kHz spectral window, using a spectral resolution of 488 kHz [(0.58 [kms$^{-1}$]{} in band 6 and 0.43 [kms$^{-1}$]{} in band 7)]{}.
The observing sequence on each date consisted of an interleaved series of scans of the science target and a continuum source (quasar) for phase calibration, alternating between 7 minutes integration on the comet and 30 s on the phase calibrator, for a total of 46 min on-source at 252 GHz and 38 min on-source at 338 GHz. The comet was tracked, and the position of the array phase center updated in real-time using the latest JPL Horizons orbital solution (sampled at 15 s intervals and interpolated using a 4th-order polynomial). The data were flagged, calibrated and cleaned using standard CASA routines [see for example @cor14], with Ganymede as the flux calibrator. Imaging was performed using a grid size of $768\times768$ pixels, with $0.05''$ pixel scale. The resulting data cubes were corrected for the response of the ALMA primary beam and then transformed from celestial coordinates to sky-projected distances with respect to the center of the comet. Spectral fluxes per beam were subsequently converted to the Rayleigh-Jeans brightness temperature scale ($T_B$) for further analysis.
Nançay OH observations {#sec:oh}
======================
Production rates for the dominant volatile H$_2$O are required for the measurement of CH$_3$OH mixing ratios and the derivation of the collisional excitation rates. The strengths of the 18 cm OH lines (at 1667 and 1665 MHz) in comet C/2012 K1 (PanSTARRS) were monitored using the Nançay radio telescope during the period 2014-04-03 to 2014-09-27. The observational procedure and data analysis were described previously by @cro02, and an expansion velocity of 0.70 [kms$^{-1}$]{} was retrieved following the analysis method of @tse07. The resulting OH production rates ($Q({\rm OH})$, for the period 2014-06-14 to 2014-07-17) are shown as a function of heliocentric distance (pre-perihelion) in Figure \[fig:oh\], including error bars due to thermal noise. The OH line inversion parameter was large for these observations, therefore errors on the OH production rate due to excitation uncertainties are expected to be negligible. There is no clear trend in $Q({\rm OH})$ *vs.* $r_H$ [around the time of our ALMA observations]{}, so we take the mean value of $9.0\times10^{28}$ s$^{-1}$ [(with a standard deviation of $1.7\times10^{28}$ s$^{-1}$). Assuming H$_2$O is the sole parent of OH, and adopting a branching ratio of 0.86 for the H$_2$O + $h\nu$ $\longrightarrow$ OH + H photolysis channel [@hue92]]{}, we obtain a water production rate of $(1.05\pm0.20)\times10^{29}$ s$^{-1}$.
![Production rates for OH in comet C/2012 K1 (PanSTARRS) as a function of heliocentric distance obtained using the Nançay radio telescope. [The values shown are pre-perihelion, and cover the period 2014-06-14 to 2014-07-17]{}. Dashed horizontal line shows the mean value of $9.0\times10^{28}$ s$^{-1}$. Vertical dotted line corresponds to the epoch of the ALMA observations. \[fig:oh\]](Nancay_Q_OH_rH.pdf){width="\columnwidth"}
Results
=======
Spectra and maps
----------------
Forteen individual emission lines of the $K=3-2$ band were detected on 28 June and sixteen lines of the $J=7-6$ band were detected on 29 June. The complete list of detected CH$_3$OH transitions and their upper-state energies is given in Table \[tab:k1lines\], and the observed spectra are shown in Figure \[fig:spec\]. These spectra were extracted by averaging the data over a circular aperture $1.5''$ in diameter, centered on the peak of the CH$_3$OH emission. Spectrally-integrated maps of the CH$_3$OH emission (shown in Figure \[fig:maps\]) were obtained by integrating over the spectral ranges covered by the lines in Table \[tab:k1lines\]. The CH$_3$OH emission peaks were approximately spatially coincident on both dates, offset by $1.2''$ south-west from the ALMA phase center, which is within the typical range of uncertainty for the position of the nucleus using optically-derived cometary ephemerides. No continuum emission was detected and a $3\sigma$ upper limit of 0.3 mJy was obtained for the average continuum flux (in band 7).
------ -------------------------------------- -------- --------- ----------------- --
ALMA Transition Freq. $E_{u}$ $W_p$
Band (MHz) (K) K[kms$^{-1}$]{}
6 $10 _{3} -10 _{2}\ A^{-+} $ 251164 177.5 0.14 (0.10)
$9 _{3} - 9 _{2}\ A^{-+} $ 251360 154.3 0.54 (0.10)
$8 _{3} - 8 _{2}\ A^{-+} $ 251517 133.4 0.71 (0.10)
$7 _{3} - 7 _{2}\ A^{-+} $ 251642 114.8 0.62 (0.10)
$6 _{3} - 6 _{2}\ A^{-+} $ 251738 98.5 0.47 (0.10)
$5 _{3} - 5 _{2}\ A^{-+} $ 251812 84.6 1.07 (0.10)
$4 _{3} - 4 _{2}\ A^{-+} $ 251867 73.0 0.60 (0.10)
$5 _{3} - 5 _{2}\ A^{+-} $ 251891 84.6 0.78 (0.10)
$6 _{3} - 6 _{2}\ A^{+-} $ 251896 98.5 1.04 (0.10)
$4 _{3} - 4 _{2}\ A^{-+} $ 251900 73.0 0.56 (0.10)
$3 _{3} - 3 _{2}\ A^{-+} $ 251906 63.7 0.44 (0.10)
$3 _{3} - 3 _{2}\ A^{+-} $ 251917 63.7 0.38 (0.10)
$7 _{3} - 7 _{2}\ A^{+-} $ 251924 114.8 0.84 (0.10)
$8 _{3} - 8 _{2}\ A^{+-} $ 251985 133.4 0.68 (0.10)
7 $7 _{0 } - 6 _{0 }\ E $ 338124 78.1 0.65 (0.11)
$7 _{-1} - 6_{-1 }\ E $ 338345 70.6 1.08 (0.11)
$7 _{0 } - 6 _{0 }\ A^+ $ 338409 65.0 1.22 (0.11)
$7 _{-4} - 6_{-4 }\ E $ 338504 152.9 0.50 (0.11)
$7 _{4} - 6_{4 }\ A^- $ 338513 145.3 1.46 (0.11)
$7 _{4} - 6_{4 }\ A^+ $ 338513 145.3 $B$
$7 _{2 } - 6 _{2 }\ A^- $ 338513 102.7 $B$
$7 _4 - 6 _4\ E $ 338530 161.0 0.32 (0.11)
$7 _{3 } - 6 _{3 }\ A^+ $ 338541 114.8 1.28 (0.14)
$7 _{3 } - 6 _{3 }\ A^- $ 338543 114.8 $B$
$7 _{-3} - 6_{-3 }\ E $ 338560 127.7 0.39 (0.12)
$7 _{3 } - 6 _{3 }\ E $ 338583 112.7 0.30 (0.12)
$7 _{1 } - 6 _{1 }\ E $ 338615 86.1 0.83 (0.12)
$7 _{2 } - 6 _{2 }\ A^+ $ 338640 102.7 0.64 (0.12)
$7 _{2 } - 6 _{2 }\ E $ 338722 87.3 1.63 (0.14)
$7 _{-2} - 6_{-2 }\ E $ 338723 90.9 $B$
------ -------------------------------------- -------- --------- ----------------- --
: CH$_3$OH lines detected in C/2012 K1 (PanSTARRS)\[tab:k1lines\]
{width="45.00000%"} {width="45.00000%"}
{width="45.00000%"} {width="45.00000%"}
The CH$_3$OH maps show a compact distribution with a strong central peak. The flux falls rapidly with distance from the center, which is consistent with the combined effects of (isotropic) outflow expansion and photodissociation due to Solar irradiation. Due to the lack of ALMA baselines shorter than 20 m, the largest angular scale detectable in our observations is $\approx7.4''$ (or $\approx10^4$ km), which results in filtering out of coma structures larger than this size. Based on a simple Haser parent model for K1/PanSTARRS (see Section \[sec:radial\]), spatial filtering is expected to have a negligible impact on the peak CH$_3$OH flux (less than a few percent), whereas 2000 km ($1.4''$) from the nucleus, only $\sim75$% of the predicted flux per beam may be recovered.
Rotational diagrams
-------------------
Detailed spectral modeling is required to interpret the observed multi-line CH$_3$OH data. We begin by using the method outlined by @boc94 to construct rotational excitation diagrams to obtain the internal (rotational) temperature of the CH$_3$OH molecules ($T_{rot}$), averaged along the line of sight. Figure \[fig:rotdiag\] shows rotational diagrams for June 28 (left) and June 29 (right), obtained from spectra extracted at the CH$_3$OH column density peak. The integrated line intensities (along with their statistical $1\sigma$ errors) for this position are given in Table \[tab:k1lines\]. The gradient of the rotational diagram is equal to $-1/T_{rot}$ and the intercept is equal to $ln(N/Q)$ where $Q$ is the partition function and $N$ is the column density. Errors on $T_{rot}$ and $N$ were obtained from the errors on the gradient and intercept, resulting in values of $T_{rot}=63.0\pm6.8$ K, $N=(2.8\pm0.7)\times10^{14}$ cm$^{-2}$ on June 28 and $T_{rot}=119\pm23$ K, $N=(2.7\pm1.0)\times10^{14}$ cm$^{-2}$ on June 29. No significant differences were detected between the abundances of the $A$ and $E$ nuclear spin states of CH$_3$OH, consistent with an equilibrium (high-temperature) spin distribution as observed in comet Hale-Bopp by @par07. Thus, we assumed equal abundances of $A$ and $E$ CH$_3$OH in our analysis from here on.
{width="45.00000%"} {width="45.00000%"} {width="45.00000%"} {width="45.00000%"}
Unfortunately, the rotational diagram method suffers from several shortcomings: blended lines with differing [upper-state energies]{} ($E_u$), and lines with signal-to-noise $<1$ cannot be included in the diagram, which limits the accuracy of the results. Furthermore, the full frequency extent of each line is not known a-priori, which hinders the accuracy of the individual line measurements — in determining the integration widths for weak lines, a balance has to be found between including the full line flux and excluding noisy data in the line wings, which inevitably results in some reduction in the overall signal-to-noise for each line (for our data, we used integration limits of $\pm1.5$ [kms$^{-1}$]{} for all lines). In order to obtain the most reliable temperatures and column densities, spectral line modeling is employed to maximise the available information in our ALMA data.
Line modeling {#sec:modeling}
-------------
Using a modified version of the spectral line modeling algorithm previously applied to interstellar rotational emission lines by @cor13, the spectra at each point in our CH$_3$OH maps were extracted and fitted (using a nonlinear least-squares algorithm) to determine $T_{rot}$ and $N$ as a function of the spatial coordinate. The technique works by fitting Gaussian optical depth curves to each emission line (as a function of frequency), parameterised by $T_{rot}$, $N$ and the line FWHM. The observed CH$_3$OH line intensity ratios encode the rotational temperature, and the absolute intensity ($T_B$) scaling encodes the column density, so the best fit to each spectrum can be found for a unique pair of \[$T_{rot}$, $N$\] values, as determined by the minimum of the sum-of-squares of the residuals between the observation and model (or $\chi^2$ value). Error estimates were obtained using a Monte Carlo noise-resampling technique, whereby for each spectrum, 300 synthetic, Gaussian, random noise spectra were generated (with standard deviation equal to the RMS noise of the observed spectrum), which were then added one at a time to the (noise-free) best-fitting model spectrum. The same least-squares fitting procedure was then repeated for each of the 300 synthetic datasets. The $1\sigma$ errors on $T_{rot}$, $N$ were obtained from the $\pm68$% percentiles of the resulting set of fit parameters.
This procedure assumes a Gaussian shape for the emission lines, which can be a poor approximation for cometary lines observed at high spectral resolution — indeed, many of our observed CH$_3$OH line profiles show slight asymmetries (consistent with asymmetric outgassing). To avoid problems with the fitting due to such non-Gaussianity, each extracted spectrum was first convolved with a Gaussian broadening kernel of FWHM 3 [kms$^{-1}$]{}. This is sufficiently broad with respect to the FWHM of the observed lines ($\approx0.9-1.2$ [kms$^{-1}$]{}) that it effectively results in the smoothing out of their profiles, producing a line shape practically indistinguishable from Gaussian, and independent of the specific outflow geometry of the comet. This smoothing also has the benefit that any small variations in the line FWHM and Doppler shift of the gas over the ALMA field of view ($\lesssim\pm0.25$ [kms$^{-1}$]{}) can be neglected, allowing these values to be held fixed during the fitting to further improve the accuracy of the results, which is particularly useful for the noisier data towards the edge of the field of view.
Examples of the line modeling results are shown in the lower panels of Figure \[fig:rotdiag\], obtained using spectra extracted at the CH$_3$OH column density peak position. The quality of these fits is [very good]{} (with reduced $\chi^2$ values in the range 1.0-1.3), and the $T_{rot}$ and $N$ values derived using this method ($T_{rot}=62.3\pm5.3$ K, $N=(2.8\pm0.1)\times10^{14}$ cm$^{-2}$ on June 28 and $T_{rot}=116\pm18$ K, $N=(2.7\pm0.4)\times10^{14}$ cm$^{-2}$ on June 29) are in good agreement with those obtained using the rotational diagram method. Due to improved utilization of the noisier data, this line modeling technique results in significantly improved accuracy of the derived parameters, and is therefore adopted as the preferred method of analysis from here on. The uncertainties derived from Monte Carlo noise replication are also expected to be more reliable because they implicitly account for correlations between the errors on $T_{rot}$ and $N$, whereas in the rotational diagram analysis, the (partially correlated) errors on the gradient and intercept cannot be easily disentangled.
Temperature and column density maps {#sec:maps}
-----------------------------------
Spectra were extracted pixel-by-pixel from the CH$_3$OH data cubes within an area of $2''\times2''$ [($40\times40$ pixels)]{} centered on the integrated emission peak. Using the line fitting method described in Section \[sec:modeling\], temperatures and column densities were obtained for each pixel to create the maps shown in Figure \[fig:ltemaps\]. These maps only show the derived $T_{rot}$ and $N$ values with uncertainties of less than 50% — values with larger errors have been masked (and are shown in white). The mean errors on these $T_{rot}$ and $N$ maps are $\pm14$ K and $\pm1.8\times10^{13}$ cm$^{-2}$ for June 28, $\pm15$ K and $\pm1.7\times10^{13}$ cm$^{-2}$ for June 29. Naturally, a higher CH$_3$OH column density leads to a stronger spectrum and hence lower uncertainties on $T_{rot}$ and $N$, so the regions closer to the centers of the maps are more reliable. The maps have been positioned [with their origins]{} at the respective column density peaks, which are marked with white crosses.
{width="45.00000%"} {width="45.00000%"} {width="45.00000%"} {width="45.00000%"}
The CH$_3$OH column density maps show a single, dominant peak as expected for spherically-symmetric outgassing from a compact nucleus. The shapes of the column density peaks are strongly influenced by the elliptical telescope beam, which hinders interpretation of the true structure of the innermost coma. [On both dates, the column density maps show a significant extension towards the top left, which is in addition to the dominant outflow structure. The direction of this extension coincides approximately with the direction of the comet’s orbital trail, perhaps implying some CH$_3$OH release from trailing material, although the presence of a directional CH$_3$OH jet/vent cannot be ruled out.]{}
Interestingly, whereas on June 29 the $T_{rot}$ distribution shows a strong central peak [about the nominal position of the nucleus]{} (falling from $116\pm18$ K to $\approx50\pm10$ K within about 500 km), on June 28 the main source of CH$_3$OH appears to be located within a temperature trough, which forms part of a much broader region of high temperatures that extends down the whole map in a roughly north-south direction. The central temperature on June 28 was $62\pm5$ K and the two flanking temperature maxima have $T_{rot}\approx85\pm14$ K on the left and $\approx73\pm12$ K on the right. However, the statistical significance of these peaks relative to the central temperature is low ($\sim1$-$2\sigma$), so a monotonically decreasing (or flat) temperature as a function distance from the nucleus in this region cannot be ruled out. Similarly, although some of the temperature structure towards the bottom left of the central peak on June 29 could be real, the uncertainties preclude a robust interpretation of the other features in these maps.
Radial profiles {#sec:radial}
---------------
{width="45.00000%"} {width="45.00000%"}
To help average out any clumpiness or small-scale coma structures and provide a clearer picture of the dominant physical and chemical processes as a function of distance from the nucleus, we perform azimuthal averaging of the ALMA data. Such azimuthally averaged data also benefit from a significantly improved signal-to-noise ratio, particularly at large radii, allowing us to probe the conditions further from the nucleus than in the 2-D maps.
Taking the CH$_3$OH column density peak as the center, our data were binned into a series of $0.1''$-wide annuli, and the average flux per spectral channel was taken in each annulus, resulting in azimuthally-averaged spectra ($\bar{S_{\nu}}$) as a function of sky-projected radius $\rho$. These $\bar{S_{\nu}}(\rho)$ were subject to the same fitting procedure outlined in Section \[sec:modeling\] to derive radial \[$T_{rot}$, $N$\] profiles, shown in Figure \[fig:profiles\].
The 1-D radial temperature profiles show a general trend for falling $T_{rot}$ as a function of distance. Similar to the 2-D maps, on June 28 $T_{rot}$ shows a possible ($1\sigma$ confidence) increase from $62\pm5$ K to $68\pm6$ K between $\rho=0$-500 km, followed by a relatively slow, steady decrease to $22\pm9$ K at $\rho=2500$ km. Conversely, on June 29 the temperature decreased sharply from $116\pm18$ K to $42\pm4$ K between $\rho=0$-1000 km. In addition to starting at least 30 K hotter near the nucleus, overall, the CH$_3$OH rotational temperatures remained at least 20 K higher throughout the observed coma on the 29th than the 28th, implying the presence of a substantial additional coma heat source on the 29th (22.5 hr later).
The possible temperature increase at distances $\rho>1500$ km on June 29 identified by @cor16 is less clear in our new results; the large error bars on our $T_{rot}$ profile at $\rho>2000$ km indicate that this result may have been at least partly due to an underestimate of the errors by @cor16, who used a rotational diagram method rather than spectral line fitting. Furthermore, the azimuthal profiles in the present study were taken about the CH$_3$OH column density peak, whereas those of @cor16 were taken about the integrated emission line peak; these positions were coincident on June 28, but on June 29 the column density peak is offset north-east by $0.16''$, leading to differences in the radial temperature profiles. [The elliptical shape of the telescope beam introduces some additional uncertainty into our radial temperature and column density profiles, which would be particularly problematic for any asymmetric coma features that happen to line up with the long axis of the beam. The (long-axis) spatial resolution HWHM of $550$ km on June 28 and $500$ km on June 29 should thus be considered as upper limits for the possible radial error margins on these azimuthally-averaged profiles]{}.
On both dates the azimuthally-averaged column densities decreased smoothly with distance from the peak value of about $2.8\times10^{14}$ cm$^{-2}$, with a shape closely consistent with uniform, spherically-symmetric outflow. This is confirmed by comparison of the observed $N(\rho)$ profiles in Figure \[fig:profiles\] with profiles obtained from a Haser-type spherical outflow model, in which CH$_3$OH is assumed to flow isotropically from the nucleus at a constant velocity of 0.5 [kms$^{-1}$]{} (consistent with the average CH$_3$OH line Doppler FWHM of $\approx1.0$ [kms$^{-1}$]{}), and is photodissociated at a rate of $10^{-5}$ s$^{-1}$ [@hue92]. To match the observational point spread functions, the model column density maps were run through the ALMA simulator in CASA with array configuration, imaging and cleaning parameters matching those of our observationions. Good fits to the observed $N(\rho)$ profiles were obtained for CH$_3$OH production rates of $1.9\times10^{27}$ s$^{-1}$ on June 28 and $1.4\times10^{27}$ s$^{-1}$ on June 29. Using the H$_2$O production rate of $(1.05\pm0.20)\times10^{29}$ s$^{-1}$ (section \[sec:oh\]), these results imply that CH$_3$OH was sublimated directly from the nucleus of comet K1/PanSTARRS with mixing ratios of 1.1-2.2%.
Our model column density profiles (red dots overlaid on the blue curves in Figure \[fig:profiles\]) provide an excellent fit to the observations on June 28, and a reasonably good fit on June 29. As with the study of @cor14, this confirms that the Haser model can be usefully applied to azimuthally-averaged ALMA observations of the inner coma. Slight discrepancies may be explained by the presence of clumpy structure, jets or variations in the outflow velocity. Some acceleration of the coma gas within our ALMA field of view is plausible. Indeed, the derived CH$_3$OH expansion velocity of 0.5 [kms$^{-1}$]{} in our $\sim2000$ km field of interest is smaller than the value of 0.7 [kms$^{-1}$]{} obtained for OH within a $\sim10^5$ km beam, as expected if the outflow speed increased with cometocentric distance. This possibility may be explored through detailed analysis of the CH$_3$OH outflow dynamics in a future study.
Discussion
==========
In the range $\rho=0$-2000 km, our ALMA observations indicate, as expected, a general trend for decreasing CH$_3$OH rotational temperatures as a function of cometocentric distance in the coma. To interpret this result requires an understanding of the meaning of $T_{rot}$. As explained by @boc94 [@boc04], the temperature $T_{rot}$ obtained from the analysis of CH$_3$OH rotational line ratios is representative of the distribution of internal rotational level populations. Unless the molecules are in local thermodynamic equilibrium (LTE, which occurs in the dense, inner coma through collisional equipartition), $T_{rot}$ can deviate significantly from the kinetic temperature of the gas ($T_{kin}$). The relationship between $T_{kin}$ and $T_{rot}$ depends on the competing influences of microscopic collisional and radiative processes. As a result, the prevalence of LTE depends primarily on the coma density, the spontaneous radiative decay rate of the gas and the flux of external (Solar) radiation. Within a few hundred kilometers of the nucleus where densities are high, the collisional rates between molecules are usually sufficient to maintain LTE, but as the molecules flow out into the less dense, outer coma regions (at $r_c\gtrsim1000$ km), collisions become less frequent, allowing the rotational levels to radiatively cool so that $T_{rot}$ falls below $T_{kin}$. Pumping of molecular ro-vibrational levels by Solar radiation can also have an important effect on $T_{rot}$ at large distances (beyond those considered in the present study).
Using the radiative transfer model of @biv99 [@biv15], which computes the CH$_3$OH level populations that result from collisions with H$_2$O and electrons, and pumping by solar radiation, we calculated the departure of $T_{rot}$ from $T_{kin}$ with distance from the nucleus. Using dashed line styles, Figure \[fig:trotModel\] shows $T_{rot}$ as a function of projected distance ($\rho$) for the $J=7-6$ and $K=3-2$ bands, under the assumption of constant kinetic temperatures of 60 K (left panel) and 120 K (right panel). The H$_2$O production rate was taken to be $10^{29}$ s$^{-1}$ and outflow velocity 0.5 [kms$^{-1}$]{}, with a collisional cross section between H$_2$O and CH$_3$OH of $5\times10^{-14}$ cm$^{-2}$. In this figure, $T_{rot}(K=3-2)$ and $T_{rot}(J=7-6)$ both show increasing departures from LTE with increasing distance from the nucleus. However, whereas the $K=3-2$ band remains relatively close to LTE (with $T_{rot}>0.75T_{kin}$ for $\rho<2500$ km), for the $J=7-6$ band, $T_{rot}$ falls much rapidly with distance, reaching $0.25T_{kin}$ by $\rho=2500$ km in the 120 K model. This is due to the larger Einstein $A$ coefficients of the $J=7-6$ transitions, which result in more rapid rotational cooling. Although this trend is qualitatively similar to the observed $T_{rot}$ pattern for the two CH$_3$OH bands in Figure \[fig:profiles\], the observed $T_{rot}$ curves both fall significantly more rapidly than expected with a coma kinetic temperature that remains constant as a function of distance. Sub-thermal excitation is therefore insufficient to fully explain the observed $T_{rot}$ behaviour in comet K1/PanSTARRS, [and a variable $T_{kin}(r)$ profile is required.]{}
[A variety of different kinetic temperature profiles were tested in our radiative transfer model, with the aim of reproducing the general behaviour of the observed $T_{rot}(r)$ profiles. For the $K=3-2$ profile on June 28, a good fit to the observed $T_{rot}(r)$ was obtained using a single-slope $T_{kin}(r)$ profile, starting at 90 K at the nucleus and falling to 35 K at $r_c=2500$ km (shown by the solid black line in the left panel of Figure \[fig:trotModel\]). For the $J=7-6$ data from June 29, however, it was more difficult to obtain a good fit to the measured $T_{rot}(r)$ data. Even with a rapidly falling $T_{kin}(r)$ profile in the inner coma, the very steep initial drop in $T_{rot}(r)$ could not be accurately reproduced. This is partly due to line-of-sight averaging, because cooler, more distant parts of the coma can dominate the temperature contribution from a compact, warm inner region (combined with the need to keep the inner coma temperatures physically reasonable). Our best fit to the $T_{rot}(J=7-6)$ profile was obtained using the $T_{kin}(r)$ shown in black in the right panel of Figure \[fig:trotModel\], beginning at a relatively high temperature of 150 K, falling to 40 K at $r_c=1000$ km, then rising back to 150 K by $r_c=2500$ km. Difficulties in modeling the $T_{rot}(r)$ profile on June 29 may be due to coma asymmetries or other physical factors not included in our model. Future attempts to more robustly derive the coma kinetic temperatures from these data may require modeling the coma structure in 3-D, as well as exploring the effects of variations in electron density profiles and temperatures, and gas collisional cross sections.]{}
{width="45.00000%"} {width="45.00000%"}
Compared with previous generations of chemical/hydrodynamic coma models (see @rod04 for a review), the temperature trends observed in K1/PanSTARRS differ from past theoretical expectations. In those models the neutral gas temperature falls rapidly as it travels outward from the nucleus, starting at $\sim100$ K and reaching a minimum of $\sim10$ K by 100 km as a result of adiabatic cooling. Around a cometocentric distance $r_c\sim100$ km, the onset of photoionisation due to Solar irradiation results in the production of a population of fast (hot) electrons, ions and neutrals. Collisions with these hot photoproducts heats the parent gases. However, the strong heating trends apparent in these conventional multi-fluid hydrodynamic models — which predict temperatures above a few hundred kelvins for cometocentric distances $>1000$ km — primarily reflects the fact that the neutral gas temperature is obtained by calculating the weighted average temperature over all coma neutrals, including the fast-moving photoproducts (primarily O, H, H$_2$ and OH from the photolysis of H$_2$O). Because the kinetic temperatures of these photoproducts can reach thousands of kelvins in the outer coma, their presence creates a strong bias in the average temperature of neutrals, which is then no longer properly representative of the temperature of parent gases. To account for this problem, the temperatures of the parents and photoproducts must be treated separately, as in the direct simulation Monte Carlo (DSMC) models of *e.g.* @ten08 and @fou12. These models show a clear separation in the kinetic temperatures of parents and photoproducts and confirm the expectation that parent species are dominated by adiabatic cooling for distances $\lesssim10^5$ km.
However, standard DSMC models still predict a very rapid drop in the temperature of parent molecules with distance that is inconsistent with our ALMA CH$_3$OH observations. For example, in the model for comet 67P/Churyumov-Gerasimenko by @ten08, the H$_2$O temperature falls by an order of magnitude within 100 km of the nucleus, [whereas on June 28, $T_{kin}(r)$ in K1/PanSTARRS fell by only a few percent over this range]{}. To resolve this discrepancy, an additional source of coma heating is required in the models. @fou12 considered the sublimation of (dirty) ice grains, which can significantly raise the H$_2$O rotational temperature in the region $r_c\lesssim1000$ km provided their mass represents a significant fraction of the total gas-phase H$_2$O production. Such sublimation heating could also raise the CH$_3$OH rotational temperature, either through injection of a source of heated CH$_3$OH into the coma, or by collisions with the heated H$_2$O molecules. Other possible heating sources that may be considered in future models include suprathermal electrons and ions that could be produced through interaction of the coma with the solar wind, UV and X-rays.
The temperature trend in K1/PanSTARRS is compatible with the results of other observational studies. The steadily decreasing $T_{kin}$ behaviour found on June 28 is comparable to the linear temperature slope deduced for the inner few thousand kilometers of comet P1/Garrad [@boc12]. [Further, on June 29 the falling kinetic temperature in the inner coma, rising back to $\sim150$ K at larger radii is qualitatively similar to that study]{}. Our observations on both dates are also broadly consistent with the general trend for decreasing H$_2$O rotational temperatures with distance over scales $\rho\sim10-1000$ km in comets C/2004 Q2 (Machholz) and 73P-B/Schwassmann-Wachmann 3 [@bon07; @bon08]. Comets 103P/Hartley 2 and C/2012 S1 (ISON) showed a more complex temperature behavior. At $r_H=0.53$ AU, comet ISON’s $T_{rot}$ fell from $\sim120$ K to $\sim85$ K within 1000 km of the nucelus, whereas at $r_H=0.35$ AU, evidence for a double-peaked temperature structure was observed, rising to maxima at cometocentric distances $\sim500-1000$ km [@bon14]. For 103P/Hartley 2, $T_{rot}$ was observed to decrease with distance on the sunward side of the nucleus, whereas on the anti-sunward side (as projected on the sky plane at a phase angle of $54^{\circ}$), evidence was found for a significant increase in $T_{rot}$ between $\rho=0$-75 km [@bon13]. This $T_{rot}$ behavior is analogous to the possible double-peaked temperature structure identified in K1/Panstarrs on 28 June (Figure \[fig:ltemaps\]). As discussed by @bon14, it may be that such unexpected temperature peaks in the coma can arise as a result of photolytic or sublimation heating, but additional theoretical studies will be required to confirm this hypothesis.
The behavior of the coma temperature as a function of time can provide more information on the nature of the coma heating processes, and is possible because of the 22.5 hr separation between our $K=3-2$ and $J=7-6$ CH$_3$OH observations. Although our excitation model shows that $T_{rot}(J=7-6)$ may still be somewhat sub-thermal in the inner coma, the $J=7-6$ and $K=3-2$ bands are both expected to be close to the coma kinetic temperature at around $r_c=0$. The implied dramatic increase in coma kinetic temperature between June 28th and 29th is therefore surprising (given the expected relative constancy of Solar radiation input), and is likely to have been caused by an increase in the heating rate of the inner coma, for instance, due to an increase in the supply of hot electrons or sublimating icy grains. [If the electron temperature dropped below the threshold for dissociative electron impacts with H$_2$O, this could also result in a sudden increase in the coma heating rate, as a larger fraction of the electron energy can then be converted into kinetic energy [as discussed by @bod16]. Interactions with the Solar wind could also lead to short-timescale variability in the coma energy balance]{}.
Large variations in the CH$_3$OH rotational temperature were observed over timescales of several hours in comet 103P/Hartley 2 by @dra12. [These variations were explained as due, in part, to the theorized correlation between water production rate and coma heating efficiency [see @boc87; @com04]]{}. However, [the relative constancy of the CH$_3$OH production rate during our observations appears to rule out this possibility in K1/PanSTARRS, and]{} the full explanation for such strong, transient variations in coma heating requires further investigation.
Several review studies have drawn comparison between interstellar, protostellar and cometary ice abundances (*e.g.* @ehr00; @mum11; @boo15). To-date, every coma species detected in the radio has also been found in interstellar clouds, and the dominant cometary ice constituents (CO, CO$_2$ and CH$_3$OH) show abundances with respect to H$_2$O that are generally within the range of values observed in protostellar environments. @boc94 identified the similarity between cometary and interstellar CH$_3$OH/H$_2$O ice ratios, and our values of 1.1-2.2% in C/2012 K1 (PanSTARRS), compared with 1-30% in low-mass protostars [@mum11], confirm this result. Recent detailed models for disk gas and ice chemistry confirm the plausibility of a close chemical relationship between cometary and protoplanetary material — *e.g.* @dro16 predict CH$_3$OH/H$_2$O $\sim1$-4% in the mid-plane ices for low-mass protoplanetary disks. The CH$_3$OH abundance in K1/PanSTARRS is also consistent with the detection of this molecule for the first time in a protoplanetary disk by @wal16, who obtained a gas-phase CH$_3$OH/H$_2$O ratio $\sim1$-5% in the disk surrounding the low-mass TW Hya system (at a distance of 54 pc). The fact that CH$_3$OH appears to be depleted [in comets and protoplanetary disks compared with]{} the median protostellar abundance of 6% in the nearby Galaxy [@boo15], [implies that significant processing of interstellar/protostellar envelope material occurs during or after its passage to the accretion disk, thus confirming the importance of cometary ices as a record for the physical and chemical processes occurring during the formation of the Solar System.]{}
Conclusion
==========
Using ALMA observations of C/2012 K1 (PanSTARRS), the first instantaneous spatial/spectral maps of CH$_3$OH rotational emission have been obtained in a cometary coma. Through rotational excitation analysis, 2-D spatial maps of the CH$_3$OH column density and rotational temperatures averaged along the line of sight have been derived, revealing new information on the physics and chemistry of the coma on scales 500-5000 km. We find that the $T_{rot}(J=7-6)$ and $T_{rot}(K=3-2)$ radial profiles both exhibit a relatively rapid drop with distance, which cannot be explained purely through sub-thermal excitation and must therefore be due to falling coma kinetic temperatures with distance from the nucleus. The observed temperature behavior is more consistent with the DSMC model of @fou12 (that includes coma heating from sublimating dirty ice grains), than the behavior seen in standard multi-fluid models, highlighting a need for continued research into coma heating (and cooling) mechanisms.
The CH$_3$OH radial column density profile is in good agreement with spherically-symmetric, uniform outflow from the nucleus. Accordingly, no evidence is found for significant production of CH$_3$OH in the coma, either from icy grain sublimation or photochemistry. The CH$_3$OH mixing ratios of 1.1-2.2% in K1/PanSTARRS are consistent with previous observations of comets at infrared and radio wavelengths. Combined with the observation of CH$_3$OH outgassing directly from the nucleus, our results confirm the utility of radio interferometric observations as a probe for the abundances of complex organic molecules in cometary ice. Our CH$_3$OH abundance adds to the evidence confirming a close chemical similarity between protostellar/protoplanetary and cometary ices.
The derivation of accurate coma [kinetic]{} temperatures (using CH$_3$OH or other molecules with a high density of rotational lines in the mm/sub-mm such as H$_2$CO or CH$_3$CN), combined with detailed theoretical modeling, is necessary to provide new constraints on the coma physics and further elucidate the gas heating and cooling mechanisms. Continued ALMA observations [(including observations at higher angular resolution and higher sensitivity),]{} will therefore play a crucial role in improving [our knowledge of coma energetics]{}, as well as leading to higher accuracy in cometary molecular production rates, parent scale lengths and gas outflow velocities. Improved accuracy and statistics in measurements of cometary ice abundances are key requirements for ongoing studies on the origin and evolution of icy materials in planetary systems.
This work was supported by NASA’s Planetary Atmospheres and Planetary Astronomy Programs and by the National Science Foundation under Grant No. AST-1616306. It makes use of the following ALMA data: ADS/JAO.ALMA\#2013.1.01061.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. The Nançay Radio Observatory is the Unit[é]{} scientifique de Nançay of the Observatoire de Paris, associated as USR No. B704 to the CNRS. The Nançay Observatory also gratefully acknowledges the financial support of the Conseil r[é]{}gional of the R[é]{}gion Centre in France. We acknowledge the advice of Dr. Boncho Bonev on the measurement of spatial variability in coma rotational temperatures.
[*Facilities:*]{}
| {
"pile_set_name": "ArXiv"
} |
---
address:
- 'Department of Biostatistics, University of North Carolina at Chapel Hill'
- North Carolina State University Bioinformatics Research Center
author:
- Pratyaydipta Rudra
- 'Fred A. Wright'
bibliography:
- 'rankcover.bib'
title: |
[Supplementary Article for]{}\
[ A procedure to detect general association based on concentration of ranks]{}
---
Details of the testing procedure {#sec:details}
================================
Let $(x_{k},y_{k})$ denote the ranks of the $k$th sample pair, $k=1,2,...,n$. We define
$d(i,j,x_k,y_k)$ = Distance between the point $(i,j)$ on the grid and $(x_k,y_k)$; $d_{ij}=\min_{k} d(i,j,x_k,y_k)$.
The RankCover method measures the concentration of ranks using the test statistic
$$\label{eq:1}
T(\delta) = \frac{1}{n^2}\sum_{i=1-\lceil \delta \rceil}^{n+\lceil \delta \rceil} \sum_{j=1-\lceil \delta \rceil}^{n+\lceil \delta \rceil} {I(d_{ij} \leq \delta)},$$
where $\delta$ is the disc radius (for Manhattan distance, $\delta$ is half the diagonal of each square). $\lceil t \rceil$ denotes the smallest integer greater than or equal to $t$.
![Figure showing the grid used to calculate RankCover statistic: $n=20$, $ \lceil \delta_{opt} \rceil=6$, all the points on the grid are used to calculate the statistic rather than using just the points in the inner square[]{data-label="fig:s1"}](gridplot.pdf){width="35.00000%"}
An $(n+2\lceil \delta \rceil) \times (n+2\lceil \delta \rceil)$ grid is considered which is an outward extension of the $n \times n$ grid $\{1,2,...,n \} \times \{1,2,...,n \}$ ().
In order to do the test, one can pre-compute the threshold based on a large number of simulations. The use of ranks enables such pre-computation as the distribution of our test statistic under null doesn’t depend on the distributions of $x$ and $y$. In , we have presented a table of such pre-computed thresholds for some sample sizes.
Choice of the disk size {#sec:discsize}
=======================
The choice of the disc size $\delta$ is an important consideration. We have proposed the use of a single optimum choice of $\delta$ as opposed to the whole $\delta$ versus $F(\delta)$ curve used by Diggle. The argument for choosing $\delta_{opt}= \sqrt{n}$ for Euclidean distance and $\delta=\frac{\pi}{2}\sqrt{n}$ is somewhat heuristic, but based on empirical observations for several sample sizes.
![Showing the mean and sd of $T(\delta)$ for sample sizes 50 and 100 (Euclidean distance is used)[]{data-label="fig:s2"}](euclidcurvesup.pdf){width="60.00000%"}
![Showing the mean and sd of $T(\delta)$ for sample sizes 50 and 100 (Manhattan distance is used)[]{data-label="fig:s3"}](manhattancurvesup.pdf){width="60.00000%"}
To understand the idea, we examine the expectation and standard deviation of $T(\delta)$ under null for varying $\delta$. These curves calculated based on 1000 simulations under null are shown in for Euclidean distance and for Manhattan distance. There is a clear change of curvature in the expectation in the vicinity of $\delta=\sqrt{n}$, and also we note that the standard deviation exhibits a local maximum and minimum in the vicinity. We reason that the local minimum of the standard deviation represents a good choice for $\delta$. We also note that the point where the expectation curve changes the curvature is approximately the same point as the local minimum of the standard deviation. However, there is no closed form expression for this point of local minimum. From simulations under different sample sizes, we have established that such local minima occur near $\delta= \sqrt{n} $ for Euclidian distance, and propose it as our choice of $\delta_{opt}$. Also, it is clear from these simulations that if the distance metric is symmetric (eg Euclidian, Manhattan etc), the shape of these curves depend on $\delta$ only through the area of the disk, and so we use $\delta_{opt}=\sqrt{\frac{\pi}{2}n}$ for the Manhattan distance.
A single $\delta_{opt}$ vs. the entire curve {#sec:singledelta}
============================================
shows an illustrative power comparison of our approach using a single optimum value of $\delta$ and the approach using the whole $\delta$ versus $F(\delta)$ curve. The second approach uses the area under curve as the test statistic. We have demonstrated the power comparisons for three different types of relationships: linear, quadratic and circular. It is clear from that the use of a single $\delta$ doesn’t reduce power substantially, but greatly reduces computation time.
![Showing the power comparison of RankCover using $\delta_{opt}$ and the Area Under Curve method for three different types of relationship[]{data-label="fig:s4"}](wholecurve.pdf){width="\textwidth"}
Details of the analysis of simulated data {#sec:simulation}
=========================================
This section explains the details of the analysis of simulated data in Section 3.1. We have used Manhattan distance throughout all the analyses due to the ease of tail area computation (). RankCover procedure with Manhattan distance appears to give similar results to that with Euclidean distance (See ).
The sample size is 50 (for other sample sizes see ) and we used 1000 simulations under the null for RankCover and MIC. For dCor and HHG, 1000 permutations are used. The power curves are obtained based on 500 simulations. The independent variable $x$ is simulated as $U(0,1)$. The dependent variable $y$ is calculated using the equation
$$y=f(x)+ \nu \times error,
\label{eq:2}$$
where $\nu$ is the noise scale parameter and increases from 0.1 to 1 as in Figure 4. The error distribution was chosen to be normal. However, as in @simon2014, the variance of the error distribution was considered differently for different forms of relationship. shows how the results are similar with other distributions also. The details of the forms of the function $f(.)$ and the error distributions are as below.
- Linear: $f(x)=x$ , error distribution is $N(0,1)$
- Quadratic: $f(x)=4(x-1/2)^2$ , error distribution is $N(0,1)$
- Cubic: $f(x)=128(x-1/3)^3-48(x-1/3)^2-12(x-1/3)$ , error distribution is $N(0,100)$
- Sine: $f(x)=sin(4\pi x)$ , error distribution is $N(0,4)$
- $X^{1/4}$: $f(x)=x^{1/4}$ , error distribution is $N(0,1)$
- Circle: $f(x)=(2r-1)\sqrt{1-(2x-1)^2}$ , error distribution is $N(0,1/16)$, where $r$ is a Bernoulli($1/2$) variable
- Two curves: $f(x)=2rx + (1-r)\sqrt{x}/2$ , error distribution is $N(0,1/4)$, where $r$ is a Bernoulli($1/2$) variable
- X-function: $f(x)=rx + (1-r)(1-x)$ , error distribution is $N(0,1/25)$, where $r$ is a Bernoulli($1/2$) variable
- Diamond: $f(x)= r_1 I(x<0.5) + r_2 I(x \geq 0.5)$ , error distribution is $N(0,1/100)$, where $r_1$ is a $U(0.5-x,0.5+x)$ variable and $r_2$ is a $U(x-0.5,1.5-x)$ variable
![Showing the power of different methods for Euclidean distance, $n=50$[]{data-label="fig:s5"}](simplotwhhgcircle.pdf){width="90.00000%"}
Choice of distance metric {#sec:distance}
=========================
We have explored two distance metrics: Euclidean distance and Manhattan distance. The performance of RankCover does not vary much based on the choice of the distance metric. shows the power analysis on simulated data using Euclidean distance. The results are not much different from those obtained using Manhattan distance (Figure 4). However, we recommend Manhattan distance since it has the advantage of more easily approximating the tail area ().
![Showing the power comparison of dCor and Spearman’s rank correlation[]{data-label="fig:s5prime"}](simplot_also_spearman.pdf){width="90.00000%"}
Comparison of dCor with Spearman’s rank correlation {#sec:alsospearman}
===================================================
The only cases where RankCover is dominated by some other method are all monotone relationships (linear, $X^{1/4}$, Two curves) and in all those cases dCor appears to be the best choice. However, we have shown () that even Spearman’s rank correlation is equally powerful in those cases. Therefore, if we have prior knowledge that the relationship is monotone, then we do not gain anything by using the fancier methods anyway, and could use Spearman’s rank correlation instead. We note that Spearman’s rank correlation does not have much “generality” in the sense that it is not powerful against non-monotone alternatives. However, dCor has also been shown to have similar limitations.
Simulation results for some other sample sizes {#sec:othern}
==============================================
and show simulation results based on sample sizes 25 and 100 respectively. As an augment to Figure 4, the usefulness of RankCover is reflected in these figures.
![Showing the power of different methods for $n=25$ (Manhattan distance)[]{data-label="fig:s6"}](simplotwhhg25.pdf){width="90.00000%"}
![Showing the power of different methods for $n=100$ (Manhattan distance)[]{data-label="fig:s7"}](simplotwhhg100.pdf){width="90.00000%"}
Simulation results for different marginal distributions of $x$ and $y$ {#sec:otherdist}
======================================================================
We have carried out the simulation analysis for different marginal distributions of $x$ and different error distributions. Three distributions of different shapes are used for the marginal distribution of $X$: uniform, truncated normal (a normal distribution with mean $1/2$ and variance $1/12$ truncated between 0 and 1)and a U-shaped beta (beta($1/2,1/2$)). The choices for the error distributions are normal, U(0,1) and beta($1/2,1/2$) with appropriate shift of origin and scale so that the mean and variance of the error distributions are 0 and 1 respectively.
The results of these nine cases show that RankCover has reasonable power in all these cases. It has very high power in some cases () and the power is not as high but still competitive in some other cases (). shows a summary of all the cases. The mean power over all the noise levels are shown for each case. Since the power curves rarely cross each other, the mean power (which is approximately proportional to area under the power curve) appears to be a good indicator of performance.
[cccccccccc]{}
\
& Linear & Quadratic & Cubic & Sine & $X^{1/4}$ & Circle & 2-Curves & X-function & Diamond\
& Linear & Quadratic & Cubic & Sine & $X^{1/4}$ & Circle & 2-Curves & X-function & Diamond\
\
\
dCor & 0.90 & 0.48 & 0.67 & 0.47 & 0.67 & 0.11 & 1.00 & 0.20 & 0.09\
RankCover & 0.97 & 0.94 & 0.91 & 0.63 & 0.85 & 1.00 & 1.00 & 0.95 & 0.84\
Hybrid & 0.95 & 0.90 & 0.87 & 0.56 & 0.79 & 1.00 & 1.00 & 0.93 & 0.76\
MIC & 0.88 & 0.50 & 0.55 & 0.43 & 0.69 & 0.71 & 0.96 & 0.50 & 0.14\
HHG & 0.94 & 0.72 & 0.74 & 0.47 & 0.77 & 0.97 & 1.00 & 0.89 & 0.76\
\
dCor & 0.90 & 0.52 & 0.69 & 0.51 & 0.69 & 0.10 & 1.00 & 0.19 & 0.09\
RankCover & 0.75 & 0.72 & 0.75 & 0.48 & 0.54 & 1.00 & 0.97 & 0.94 & 0.81\
Hybrid & 0.86 & 0.67 & 0.74 & 0.49 & 0.62 & 1.00 & 0.99 & 0.91 & 0.74\
MIC & 0.70 & 0.61 & 0.61 & 0.51 & 0.46 & 0.73 & 0.93 & 0.51 & 0.15\
HHG & 0.81 & 0.65 & 0.70 & 0.46 & 0.59 & 0.98 & 0.99 & 0.89 & 0.77\
\
dCor & 0.89 & 0.49 & 0.67 & 0.46 & 0.66 & 0.11 & 1.00 & 0.20 & 0.09\
RankCover & 0.85 & 0.80 & 0.81 & 0.50 & 0.63 & 1.00 & 0.99 & 0.94 & 0.83\
Hybrid & 0.86 & 0.74 & 0.77 & 0.47 & 0.62 & 1.00 & 0.99 & 0.91 & 0.75\
MIC & 0.74 & 0.51 & 0.56 & 0.44 & 0.47 & 0.71 & 0.93 & 0.50 & 0.15\
HHG & 0.82 & 0.62 & 0.69 & 0.41 & 0.60 & 0.97 & 0.99 & 0.88 & 0.76\
\
dCor & 0.71 & 0.42 & 0.38 & 0.34 & 0.40 & 0.05 & 0.94 & 0.46 & 0.05\
RankCover & 0.86 & 0.76 & 0.68 & 0.81 & 0.58 & 0.87 & 0.90 & 0.89 & 0.63\
Hybrid & 0.81 & 0.68 & 0.59 & 0.72 & 0.51 & 0.82 & 0.90 & 0.84 & 0.49\
MIC & 0.69 & 0.32 & 0.44 & 0.47 & 0.42 & 0.33 & 0.73 & 0.33 & 0.08\
HHG & 0.77 & 0.64 & 0.55 & 0.50 & 0.49 & 0.57 & 0.90 & 0.92 & 0.64\
\
dCor & 0.73 & 0.44 & 0.40 & 0.35 & 0.42 & 0.05 & 0.94 & 0.47 & 0.04\
RankCover & 0.56 & 0.50 & 0.41 & 0.61 & 0.30 & 0.85 & 0.85 & 0.88 & 0.63\
Hybrid & 0.65 & 0.45 & 0.40 & 0.55 & 0.36 & 0.79 & 0.88 & 0.84 & 0.50\
MIC & 0.48 & 0.38 & 0.37 & 0.58 & 0.25 & 0.33 & 0.69 & 0.35 & 0.08\
HHG & 0.58 & 0.53 & 0.43 & 0.48 & 0.32 & 0.57 & 0.89 & 0.93 & 0.63\
\
dCor & 0.70 & 0.41 & 0.37 & 0.34 & 0.40 & 0.06 & 0.93 & 0.47 & 0.05\
RankCover & 0.65 & 0.57 & 0.49 & 0.69 & 0.36 & 0.84 & 0.85 & 0.87 & 0.62\
Hybrid & 0.65 & 0.50 & 0.44 & 0.60 & 0.36 & 0.78 & 0.87 & 0.83 & 0.50\
MIC & 0.50 & 0.33 & 0.35 & 0.50 & 0.26 & 0.33 & 0.67 & 0.33 & 0.08\
HHG & 0.59 & 0.53 & 0.42 & 0.45 & 0.33 & 0.55 & 0.88 & 0.91 & 0.63\
\
dCor & 0.82 & 0.47 & 0.32 & 0.44 & 0.51 & 0.06 & 0.98 & 0.32 & 0.07\
RankCover & 0.93 & 0.88 & 0.74 & 0.76 & 0.71 & 0.98 & 0.98 & 0.94 & 0.77\
Hybrid & 0.90 & 0.81 & 0.65 & 0.67 & 0.64 & 0.97 & 0.97 & 0.91 & 0.68\
MIC & 0.79 & 0.42 & 0.38 & 0.46 & 0.55 & 0.50 & 0.87 & 0.42 & 0.10\
HHG & 0.87 & 0.68 & 0.52 & 0.48 & 0.61 & 0.78 & 0.97 & 0.93 & 0.73\
\
dCor & 0.81 & 0.46 & 0.27 & 0.43 & 0.51 & 0.12 & 0.97 & 0.26 & 0.04\
RankCover & 0.66 & 0.62 & 0.51 & 0.59 & 0.39 & 0.98 & 0.94 & 0.93 & 0.78\
Hybrid & 0.76 & 0.57 & 0.45 & 0.54 & 0.46 & 0.96 & 0.96 & 0.90 & 0.68\
MIC & 0.59 & 0.51 & 0.42 & 0.58 & 0.33 & 0.50 & 0.82 & 0.44 & 0.09\
HHG & 0.71 & 0.60 & 0.44 & 0.46 & 0.43 & 0.80 & 0.96 & 0.94 & 0.76\
\
dCor & 0.80 & 0.46 & 0.32 & 0.44 & 0.50 & 0.07 & 0.98 & 0.33 & 0.06\
RankCover & 0.75 & 0.70 & 0.57 & 0.64 & 0.47 & 0.97 & 0.95 & 0.92 & 0.78\
Hybrid & 0.76 & 0.63 & 0.50 & 0.58 & 0.46 & 0.95 & 0.96 & 0.89 & 0.68\
MIC & 0.61 & 0.43 & 0.36 & 0.48 & 0.35 & 0.49 & 0.82 & 0.42 & 0.10\
HHG & 0.71 & 0.59 & 0.43 & 0.44 & 0.43 & 0.76 & 0.96 & 0.92 & 0.74\
\[table:nineplot\]
![Showing the power of different methods when marginal of $x$ is beta and error distribution is beta ($n=50$)[]{data-label="fig:s8"}](simplotwhhg_betabeta.pdf){width="90.00000%"}
![Showing the power of different methods when marginal of $x$ is normal and error distribution is normal ($n=50$))[]{data-label="fig:s9"}](simplotwhhg_normnorm.pdf){width="90.00000%"}
Details of real data analyses {#sec:dataanalysis}
=============================
Example 1: Eckerle4 data {#sec:eckerle4}
------------------------
100,000 simulations were used for RankCover and MIC. 100,000 permutations were used for dCor and HHG. The estimates of $\beta_1, \beta_2, \beta_3$ obtained from NIST website are used for plotting the fitted curve in Figure 5. Source of data: [NIST StRD for non-linear regression](http://www.itl.nist.gov/div898/strd/nls/data/eckerle4.shtml).
Example 2: Aircraft data {#sec:aircraft}
------------------------
100,000 simulations were used for RankCover and MIC. 100,000 permutations were used for dCor and HHG. Source of data: `sm` Package in R [@bowman2013r].
Example 3: ENSO data {#sec:enso}
--------------------
100,000 simulations were used for RankCover and MIC. 100,000 permutations were used for dCor and HHG. The estimates of $\beta_1, \beta_2,..., \beta_9$ obtained from NIST website are used for plotting the fitted curve in Figure 7. Source of data: [NIST StRD for non-linear regression](http://www.itl.nist.gov/div898/strd/nls/data/enso.shtml).
Example 4: Yeast data {#sec:yeast}
---------------------
100,000 simulations were used for RankCover and MIC. 100,000 permutations were used for dCor and HHG. The data was pre-processed before analysis as follows. The data contained several missing observations. Since the sample size is small (24), we removed all the genes that had more than 3 missing observations. All other missing observations were imputed using KNN imputation [@troyanskaya2001]. Then quantile normalization was used to normalize the data. Unlike @reshef2011, we didn’t remove any of the time points and didn’t use any interpolation to find expression values for intermediate timepoints. Source of data: [Comprehensive Identification of Cell Cycle regulated Genes of the Yeast Saccharomyces cerevisiae by Microarray Hybridization](http://www.molbiolcell.org/content/9/12/3273/suppl/DC1).
Tables of pre-computed thresholds {#sec:tables}
=================================
The use of ranks in our procedure enables us to build tables of pre-computed thresholds for the test. Such pre-computed thresholds for RankCover method with Manhattan distance are given in and those for the hybrid method are given in . 100000 simulations were used to calculate the thresholds in each case. For the Manhattan metric, the rejection thresholds follow a sawtooth pattern (), with jump points occurring at the values of $n$ where $[\delta]$ changes. Simulations were performed for $n=20,...,100$. For large values of $n$, to reduce computation, tables were generated by (1) performing direct simulation for the values of $n$ at, and just prior to, the jump points, followed by (2) linear interpolation for remaining values of $n$.
![Showing the 5th percentiles of the Rankcover statistic for Manhattan distance for $n=50,...,500$. Similar pattern is observed for other percentiles also.[]{data-label="fig:sawtooth"}](sawtooth.pdf){width="85.00000%"}
[ccccccc]{}
\
Sample Sizes & p=0.1 & p=0.05 & p=0.025 & p=0.01 & p=0.001 & p=0.0001\
\
20 & 1.31000 & 1.28500 & 1.26000 & 1.23250 & 1.16500 & 1.10500\
21 & 1.27211 & 1.24717 & 1.22449 & 1.19501 & 1.12925 & 1.07936\
22 & 1.23554 & 1.21281 & 1.19215 & 1.16529 & 1.10331 & 1.06405\
23 & 1.39698 & 1.37240 & 1.35161 & 1.32325 & 1.26087 & 1.20227\
24 & 1.36111 & 1.33854 & 1.31771 & 1.29167 & 1.23090 & 1.16493\
25 & 1.32960 & 1.30720 & 1.28640 & 1.26240 & 1.20640 & 1.14720\
26 & 1.30030 & 1.27959 & 1.25888 & 1.23373 & 1.17899 & 1.12574\
27 & 1.27298 & 1.25240 & 1.23320 & 1.20850 & 1.15501 & 1.10151\
28 & 1.24745 & 1.22832 & 1.20918 & 1.18622 & 1.13520 & 1.08291\
29 & 1.22473 & 1.20452 & 1.18668 & 1.16290 & 1.11415 & 1.07134\
30 & 1.20222 & 1.18333 & 1.16444 & 1.14222 & 1.09222 & 1.04889\
31 & 1.18106 & 1.16233 & 1.14464 & 1.12279 & 1.07700 & 1.03018\
32 & 1.30469 & 1.28613 & 1.26855 & 1.24609 & 1.19922 & 1.15625\
33 & 1.28375 & 1.26538 & 1.24885 & 1.22865 & 1.18182 & 1.14509\
34 & 1.26384 & 1.24567 & 1.22924 & 1.20934 & 1.16263 & 1.11938\
35 & 1.24490 & 1.22776 & 1.21061 & 1.19102 & 1.14286 & 1.10286\
36 & 1.22685 & 1.20988 & 1.19367 & 1.17361 & 1.13040 & 1.09259\
37 & 1.21110 & 1.19430 & 1.17823 & 1.15997 & 1.11395 & 1.08400\
38 & 1.19453 & 1.17798 & 1.16274 & 1.14474 & 1.10319 & 1.06856\
39 & 1.17883 & 1.16239 & 1.14727 & 1.12821 & 1.08613 & 1.05523\
40 & 1.16375 & 1.14813 & 1.13375 & 1.11563 & 1.07437 & 1.04375\
41 & 1.26413 & 1.24866 & 1.23379 & 1.21594 & 1.17668 & 1.14456\
42 & 1.24943 & 1.23413 & 1.21995 & 1.20181 & 1.16213 & 1.12132\
43 & 1.23580 & 1.22012 & 1.20606 & 1.18875 & 1.15035 & 1.10871\
44 & 1.22159 & 1.20713 & 1.19318 & 1.17717 & 1.13998 & 1.10795\
45 & 1.20889 & 1.19407 & 1.18074 & 1.16395 & 1.12444 & 1.09284\
46 & 1.19660 & 1.18195 & 1.16824 & 1.15217 & 1.11531 & 1.08932\
47 & 1.18470 & 1.17021 & 1.15708 & 1.14079 & 1.10457 & 1.07062\
48 & 1.17231 & 1.15842 & 1.14497 & 1.12934 & 1.09115 & 1.06510\
49 & 1.16160 & 1.14744 & 1.13369 & 1.11828 & 1.08330 & 1.05373\
50 & 1.15080 & 1.13640 & 1.12400 & 1.10760 & 1.07360 & 1.04520\
51 & 1.13995 & 1.12611 & 1.11342 & 1.09765 & 1.06267 & 1.03114\
52 & 1.22337 & 1.20969 & 1.19749 & 1.18158 & 1.14756 & 1.11501\
53 & 1.21289 & 1.19972 & 1.18726 & 1.17230 & 1.13884 & 1.11463\
54 & 1.20302 & 1.18964 & 1.17764 & 1.16324 & 1.12929 & 1.10391\
55 & 1.19273 & 1.17983 & 1.16793 & 1.15372 & 1.12066 & 1.09421\
56 & 1.18367 & 1.17060 & 1.15912 & 1.14445 & 1.11129 & 1.07175\
57 & 1.17421 & 1.16159 & 1.15020 & 1.13573 & 1.10249 & 1.07572\
58 & 1.16498 & 1.15250 & 1.14090 & 1.12634 & 1.09304 & 1.06421\
59 & 1.15628 & 1.14392 & 1.13243 & 1.11807 & 1.08475 & 1.05918\
60 & 1.14778 & 1.13528 & 1.12361 & 1.11000 & 1.07805 & 1.05333\
61 & 1.13948 & 1.12739 & 1.11610 & 1.10266 & 1.06987 & 1.04139\
62 & 1.13137 & 1.11889 & 1.10744 & 1.09417 & 1.06322 & 1.03668\
63 & 1.12371 & 1.11111 & 1.10028 & 1.08743 & 1.05694 & 1.03452\
64 & 1.19434 & 1.18237 & 1.17188 & 1.15845 & 1.12524 & 1.10181\
65 & 1.18627 & 1.17467 & 1.16426 & 1.15101 & 1.12260 & 1.09870\
66 & 1.17906 & 1.16736 & 1.15657 & 1.14371 & 1.11524 & 1.08655\
67 & 1.17153 & 1.15995 & 1.14970 & 1.13678 & 1.11049 & 1.08599\
68 & 1.16436 & 1.15268 & 1.14208 & 1.12954 & 1.10208 & 1.07656\
69 & 1.15690 & 1.14556 & 1.13506 & 1.12329 & 1.09494 & 1.07498\
70 & 1.15041 & 1.13898 & 1.12878 & 1.11612 & 1.08898 & 1.06510\
71 & 1.14323 & 1.13212 & 1.12200 & 1.10930 & 1.08173 & 1.05237\
72 & 1.13657 & 1.12558 & 1.11555 & 1.10359 & 1.07485 & 1.05112\
73 & 1.12986 & 1.11878 & 1.10884 & 1.09758 & 1.06943 & 1.04447\
74 & 1.12381 & 1.11304 & 1.10299 & 1.09112 & 1.06410 & 1.03853\
75 & 1.11769 & 1.10702 & 1.09707 & 1.08533 & 1.05778 & 1.04000\
76 & 1.11165 & 1.10059 & 1.09107 & 1.07877 & 1.05315 & 1.03116\
77 & 1.10525 & 1.09462 & 1.08501 & 1.07320 & 1.04638 & 1.02749\
78 & 1.16650 & 1.15631 & 1.14678 & 1.13560 & 1.11012 & 1.08695\
79 & 1.16055 & 1.15014 & 1.14084 & 1.12931 & 1.10383 & 1.08348\
80 & 1.15453 & 1.14437 & 1.13484 & 1.12344 & 1.09922 & 1.07344\
81 & 1.14906 & 1.13900 & 1.12986 & 1.11888 & 1.09282 & 1.07194\
82 & 1.14307 & 1.13311 & 1.12433 & 1.11288 & 1.08864 & 1.06856\
83 & 1.13805 & 1.12818 & 1.11932 & 1.10814 & 1.08434 & 1.06474\
84 & 1.13265 & 1.12259 & 1.11338 & 1.10247 & 1.07851 & 1.05782\
85 & 1.12720 & 1.11696 & 1.10754 & 1.09689 & 1.07170 & 1.05190\
86 & 1.12196 & 1.11195 & 1.10289 & 1.09248 & 1.06963 & 1.04070\
87 & 1.11692 & 1.10715 & 1.09856 & 1.08733 & 1.06223 & 1.04082\
88 & 1.11170 & 1.10176 & 1.09310 & 1.08226 & 1.05850 & 1.03719\
89 & 1.10668 & 1.09670 & 1.08787 & 1.07650 & 1.05328 & 1.03055\
90 & 1.10198 & 1.09235 & 1.08346 & 1.07284 & 1.04901 & 1.02914\
91 & 1.09733 & 1.08767 & 1.07910 & 1.06883 & 1.04516 & 1.02210\
92 & 1.15064 & 1.14130 & 1.13268 & 1.12228 & 1.09983 & 1.07999\
93 & 1.14591 & 1.13666 & 1.12880 & 1.11840 & 1.09481 & 1.07619\
94 & 1.14113 & 1.13207 & 1.12381 & 1.11374 & 1.09110 & 1.07209\
95 & 1.13651 & 1.12720 & 1.11889 & 1.10903 & 1.08632 & 1.06825\
96 & 1.13184 & 1.12250 & 1.11404 & 1.10406 & 1.08203 & 1.06272\
97 & 1.12754 & 1.11829 & 1.11021 & 1.09980 & 1.07716 & 1.05707\
98 & 1.12349 & 1.11454 & 1.10641 & 1.09652 & 1.07434 & 1.05269\
99 & 1.11887 & 1.10978 & 1.10183 & 1.09193 & 1.07091 & 1.05387\
100 & 1.11470 & 1.10580 & 1.09740 & 1.08760 & 1.06480 & 1.04480\
101 & 1.11229 & 1.10332 & 1.09536 & 1.08528 & 1.06346 & 1.04473\
102 & 1.10803 & 1.09910 & 1.09122 & 1.08117 & 1.05942 & 1.04081\
103 & 1.10377 & 1.09488 & 1.08707 & 1.07706 & 1.05538 & 1.03689\
104 & 1.09951 & 1.09066 & 1.08292 & 1.07295 & 1.05134 & 1.03298\
105 & 1.09525 & 1.08644 & 1.07878 & 1.06884 & 1.04730 & 1.02906\
106 & 1.09099 & 1.08222 & 1.07463 & 1.06473 & 1.04326 & 1.02514\
107 & 1.08673 & 1.07800 & 1.07048 & 1.06062 & 1.03922 & 1.02122\
108 & 1.13400 & 1.12543 & 1.11806 & 1.10897 & 1.08813 & 1.07073\
109 & 1.13053 & 1.12200 & 1.11464 & 1.10560 & 1.08484 & 1.06766\
110 & 1.12706 & 1.11857 & 1.11123 & 1.10223 & 1.08154 & 1.06459\
111 & 1.12358 & 1.11514 & 1.10782 & 1.09886 & 1.07824 & 1.06152\
112 & 1.12011 & 1.11171 & 1.10441 & 1.09550 & 1.07495 & 1.05845\
113 & 1.11664 & 1.10828 & 1.10100 & 1.09213 & 1.07165 & 1.05537\
114 & 1.11316 & 1.10485 & 1.09759 & 1.08876 & 1.06835 & 1.05230\
115 & 1.10969 & 1.10143 & 1.09418 & 1.08539 & 1.06505 & 1.04923\
116 & 1.10622 & 1.09800 & 1.09077 & 1.08203 & 1.06176 & 1.04616\
117 & 1.10274 & 1.09457 & 1.08735 & 1.07866 & 1.05846 & 1.04309\
118 & 1.09927 & 1.09114 & 1.08394 & 1.07529 & 1.05516 & 1.04002\
119 & 1.09580 & 1.08771 & 1.08053 & 1.07192 & 1.05187 & 1.03695\
120 & 1.09233 & 1.08428 & 1.07712 & 1.06856 & 1.04857 & 1.03388\
121 & 1.08885 & 1.08085 & 1.07371 & 1.06519 & 1.04527 & 1.03081\
122 & 1.08538 & 1.07742 & 1.07030 & 1.06182 & 1.04197 & 1.02773\
123 & 1.08191 & 1.07399 & 1.06689 & 1.05845 & 1.03868 & 1.02466\
124 & 1.07843 & 1.07056 & 1.06348 & 1.05509 & 1.03538 & 1.02159\
125 & 1.12051 & 1.11296 & 1.10630 & 1.09811 & 1.07776 & 1.05946\
126 & 1.11767 & 1.11013 & 1.10349 & 1.09531 & 1.07514 & 1.05713\
127 & 1.11482 & 1.10731 & 1.10068 & 1.09251 & 1.07252 & 1.05480\
128 & 1.11198 & 1.10448 & 1.09786 & 1.08971 & 1.06990 & 1.05247\
129 & 1.10913 & 1.10166 & 1.09505 & 1.08690 & 1.06728 & 1.05014\
130 & 1.10628 & 1.09883 & 1.09223 & 1.08410 & 1.06466 & 1.04782\
131 & 1.10344 & 1.09601 & 1.08942 & 1.08130 & 1.06204 & 1.04549\
132 & 1.10059 & 1.09318 & 1.08661 & 1.07850 & 1.05942 & 1.04316\
133 & 1.09775 & 1.09036 & 1.08379 & 1.07570 & 1.05681 & 1.04083\
134 & 1.09490 & 1.08753 & 1.08098 & 1.07290 & 1.05419 & 1.03851\
135 & 1.09206 & 1.08471 & 1.07816 & 1.07009 & 1.05157 & 1.03618\
136 & 1.08921 & 1.08188 & 1.07535 & 1.06729 & 1.04895 & 1.03385\
137 & 1.08637 & 1.07906 & 1.07254 & 1.06449 & 1.04633 & 1.03152\
138 & 1.08352 & 1.07623 & 1.06972 & 1.06169 & 1.04371 & 1.02919\
139 & 1.08068 & 1.07341 & 1.06691 & 1.05889 & 1.04109 & 1.02687\
140 & 1.07783 & 1.07058 & 1.06409 & 1.05608 & 1.03847 & 1.02454\
141 & 1.07499 & 1.06776 & 1.06128 & 1.05328 & 1.03585 & 1.02221\
142 & 1.07214 & 1.06493 & 1.05846 & 1.05048 & 1.03323 & 1.01988\
143 & 1.06929 & 1.06211 & 1.05565 & 1.04768 & 1.03061 & 1.01756\
144 & 1.10745 & 1.10050 & 1.09418 & 1.08656 & 1.06964 & 1.05314\
145 & 1.10505 & 1.09812 & 1.09183 & 1.08422 & 1.06736 & 1.05108\
146 & 1.10266 & 1.09574 & 1.08948 & 1.08188 & 1.06508 & 1.04901\
147 & 1.10026 & 1.09336 & 1.08712 & 1.07954 & 1.06281 & 1.04694\
148 & 1.09787 & 1.09098 & 1.08477 & 1.07720 & 1.06053 & 1.04487\
149 & 1.09548 & 1.08860 & 1.08242 & 1.07486 & 1.05825 & 1.04281\
150 & 1.09308 & 1.08622 & 1.08006 & 1.07252 & 1.05598 & 1.04074\
151 & 1.09069 & 1.08384 & 1.07771 & 1.07018 & 1.05370 & 1.03867\
152 & 1.08830 & 1.08146 & 1.07536 & 1.06784 & 1.05142 & 1.03660\
153 & 1.08590 & 1.07908 & 1.07300 & 1.06550 & 1.04915 & 1.03454\
154 & 1.08351 & 1.07670 & 1.07065 & 1.06316 & 1.04687 & 1.03247\
155 & 1.08111 & 1.07432 & 1.06830 & 1.06081 & 1.04459 & 1.03040\
156 & 1.07872 & 1.07193 & 1.06594 & 1.05847 & 1.04232 & 1.02833\
157 & 1.07633 & 1.06955 & 1.06359 & 1.05613 & 1.04004 & 1.02627\
158 & 1.07393 & 1.06717 & 1.06123 & 1.05379 & 1.03776 & 1.02420\
159 & 1.07154 & 1.06479 & 1.05888 & 1.05145 & 1.03548 & 1.02213\
160 & 1.06915 & 1.06241 & 1.05653 & 1.04911 & 1.03321 & 1.02006\
161 & 1.06675 & 1.06003 & 1.05417 & 1.04677 & 1.03093 & 1.01800\
162 & 1.06436 & 1.05765 & 1.05182 & 1.04443 & 1.02865 & 1.01593\
163 & 1.09895 & 1.09248 & 1.08657 & 1.07949 & 1.06421 & 1.05123\
164 & 1.09692 & 1.09046 & 1.08456 & 1.07749 & 1.06225 & 1.04937\
165 & 1.09488 & 1.08844 & 1.08255 & 1.07549 & 1.06030 & 1.04752\
166 & 1.09285 & 1.08643 & 1.08054 & 1.07349 & 1.05834 & 1.04566\
167 & 1.09081 & 1.08441 & 1.07853 & 1.07149 & 1.05638 & 1.04381\
168 & 1.08878 & 1.08239 & 1.07652 & 1.06949 & 1.05443 & 1.04196\
169 & 1.08674 & 1.08037 & 1.07451 & 1.06749 & 1.05247 & 1.04010\
170 & 1.08471 & 1.07836 & 1.07250 & 1.06549 & 1.05052 & 1.03825\
171 & 1.08267 & 1.07634 & 1.07049 & 1.06348 & 1.04856 & 1.03640\
172 & 1.08064 & 1.07432 & 1.06848 & 1.06148 & 1.04660 & 1.03454\
173 & 1.07860 & 1.07231 & 1.06647 & 1.05948 & 1.04465 & 1.03269\
174 & 1.07657 & 1.07029 & 1.06446 & 1.05748 & 1.04269 & 1.03084\
175 & 1.07453 & 1.06827 & 1.06245 & 1.05548 & 1.04073 & 1.02898\
176 & 1.07250 & 1.06626 & 1.06044 & 1.05348 & 1.03878 & 1.02713\
177 & 1.07047 & 1.06424 & 1.05843 & 1.05148 & 1.03682 & 1.02528\
178 & 1.06843 & 1.06222 & 1.05642 & 1.04948 & 1.03486 & 1.02342\
179 & 1.06640 & 1.06020 & 1.05441 & 1.04748 & 1.03291 & 1.02157\
180 & 1.06436 & 1.05819 & 1.05240 & 1.04548 & 1.03095 & 1.01971\
181 & 1.06233 & 1.05617 & 1.05039 & 1.04348 & 1.02900 & 1.01786\
182 & 1.06029 & 1.05415 & 1.04838 & 1.04148 & 1.02704 & 1.01601\
183 & 1.05826 & 1.05214 & 1.04637 & 1.03948 & 1.02508 & 1.01415\
184 & 1.08994 & 1.08388 & 1.07872 & 1.07242 & 1.05801 & 1.04182\
185 & 1.08820 & 1.08216 & 1.07699 & 1.07071 & 1.05634 & 1.04025\
186 & 1.08647 & 1.08044 & 1.07526 & 1.06899 & 1.05468 & 1.03868\
187 & 1.08473 & 1.07872 & 1.07353 & 1.06728 & 1.05301 & 1.03711\
188 & 1.08300 & 1.07700 & 1.07181 & 1.06556 & 1.05134 & 1.03554\
189 & 1.08126 & 1.07527 & 1.07008 & 1.06384 & 1.04968 & 1.03397\
190 & 1.07952 & 1.07355 & 1.06835 & 1.06213 & 1.04801 & 1.03240\
191 & 1.07779 & 1.07183 & 1.06663 & 1.06041 & 1.04635 & 1.03083\
192 & 1.07605 & 1.07011 & 1.06490 & 1.05869 & 1.04468 & 1.02926\
193 & 1.07432 & 1.06839 & 1.06317 & 1.05698 & 1.04301 & 1.02769\
194 & 1.07258 & 1.06666 & 1.06145 & 1.05526 & 1.04135 & 1.02612\
195 & 1.07084 & 1.06494 & 1.05972 & 1.05354 & 1.03968 & 1.02455\
196 & 1.06911 & 1.06322 & 1.05799 & 1.05183 & 1.03801 & 1.02298\
197 & 1.06737 & 1.06150 & 1.05626 & 1.05011 & 1.03635 & 1.02141\
198 & 1.06564 & 1.05977 & 1.05454 & 1.04840 & 1.03468 & 1.01984\
199 & 1.06390 & 1.05805 & 1.05281 & 1.04668 & 1.03301 & 1.01827\
200 & 1.06216 & 1.05633 & 1.05108 & 1.04496 & 1.03135 & 1.01670\
201 & 1.06043 & 1.05461 & 1.04936 & 1.04325 & 1.02968 & 1.01513\
202 & 1.05869 & 1.05289 & 1.04763 & 1.04153 & 1.02802 & 1.01356\
203 & 1.05696 & 1.05116 & 1.04590 & 1.03981 & 1.02635 & 1.01199\
204 & 1.05522 & 1.04944 & 1.04417 & 1.03810 & 1.02468 & 1.01042\
205 & 1.05348 & 1.04772 & 1.04245 & 1.03638 & 1.02302 & 1.00885\
206 & 1.05175 & 1.04600 & 1.04072 & 1.03466 & 1.02135 & 1.00728\
207 & 1.08098 & 1.07536 & 1.07029 & 1.06418 & 1.05092 & 1.03650\
208 & 1.07947 & 1.07385 & 1.06879 & 1.06268 & 1.04946 & 1.03508\
209 & 1.07795 & 1.07234 & 1.06729 & 1.06119 & 1.04800 & 1.03365\
210 & 1.07643 & 1.07083 & 1.06579 & 1.05969 & 1.04654 & 1.03223\
211 & 1.07492 & 1.06932 & 1.06429 & 1.05820 & 1.04508 & 1.03081\
212 & 1.07340 & 1.06781 & 1.06279 & 1.05670 & 1.04361 & 1.02938\
213 & 1.07189 & 1.06630 & 1.06129 & 1.05521 & 1.04215 & 1.02796\
214 & 1.07037 & 1.06479 & 1.05979 & 1.05371 & 1.04069 & 1.02654\
215 & 1.06886 & 1.06328 & 1.05829 & 1.05222 & 1.03923 & 1.02511\
216 & 1.06734 & 1.06177 & 1.05679 & 1.05073 & 1.03777 & 1.02369\
217 & 1.06582 & 1.06026 & 1.05529 & 1.04923 & 1.03631 & 1.02227\
218 & 1.06431 & 1.05875 & 1.05379 & 1.04774 & 1.03484 & 1.02084\
219 & 1.06279 & 1.05724 & 1.05229 & 1.04624 & 1.03338 & 1.01942\
220 & 1.06128 & 1.05573 & 1.05078 & 1.04475 & 1.03192 & 1.01800\
221 & 1.05976 & 1.05422 & 1.04928 & 1.04325 & 1.03046 & 1.01657\
222 & 1.05825 & 1.05271 & 1.04778 & 1.04176 & 1.02900 & 1.01515\
223 & 1.05673 & 1.05120 & 1.04628 & 1.04026 & 1.02753 & 1.01373\
224 & 1.05521 & 1.04969 & 1.04478 & 1.03877 & 1.02607 & 1.01230\
225 & 1.05370 & 1.04818 & 1.04328 & 1.03727 & 1.02461 & 1.01088\
226 & 1.05218 & 1.04667 & 1.04178 & 1.03578 & 1.02315 & 1.00946\
227 & 1.05067 & 1.04516 & 1.04028 & 1.03428 & 1.02169 & 1.00803\
228 & 1.04915 & 1.04365 & 1.03878 & 1.03279 & 1.02023 & 1.00661\
229 & 1.04763 & 1.04214 & 1.03728 & 1.03129 & 1.01876 & 1.00519\
230 & 1.07488 & 1.06968 & 1.06493 & 1.05941 & 1.04715 & 1.03673\
231 & 1.07355 & 1.06836 & 1.06362 & 1.05811 & 1.04591 & 1.03557\
232 & 1.07222 & 1.06704 & 1.06231 & 1.05680 & 1.04466 & 1.03441\
233 & 1.07090 & 1.06571 & 1.06100 & 1.05550 & 1.04342 & 1.03325\
234 & 1.06957 & 1.06439 & 1.05969 & 1.05419 & 1.04218 & 1.03210\
235 & 1.06824 & 1.06307 & 1.05838 & 1.05289 & 1.04094 & 1.03094\
236 & 1.06691 & 1.06175 & 1.05707 & 1.05158 & 1.03970 & 1.02978\
237 & 1.06559 & 1.06043 & 1.05576 & 1.05028 & 1.03846 & 1.02862\
238 & 1.06426 & 1.05911 & 1.05444 & 1.04897 & 1.03722 & 1.02746\
239 & 1.06293 & 1.05778 & 1.05313 & 1.04767 & 1.03598 & 1.02630\
240 & 1.06161 & 1.05646 & 1.05182 & 1.04636 & 1.03474 & 1.02515\
241 & 1.06028 & 1.05514 & 1.05051 & 1.04506 & 1.03350 & 1.02399\
242 & 1.05895 & 1.05382 & 1.04920 & 1.04375 & 1.03226 & 1.02283\
243 & 1.05763 & 1.05250 & 1.04789 & 1.04244 & 1.03102 & 1.02167\
244 & 1.05630 & 1.05118 & 1.04658 & 1.04114 & 1.02978 & 1.02051\
245 & 1.05497 & 1.04985 & 1.04527 & 1.03983 & 1.02854 & 1.01935\
246 & 1.05364 & 1.04853 & 1.04395 & 1.03853 & 1.02730 & 1.01820\
247 & 1.05232 & 1.04721 & 1.04264 & 1.03722 & 1.02606 & 1.01704\
248 & 1.05099 & 1.04589 & 1.04133 & 1.03592 & 1.02482 & 1.01588\
249 & 1.04966 & 1.04457 & 1.04002 & 1.03461 & 1.02358 & 1.01472\
250 & 1.04834 & 1.04325 & 1.03871 & 1.03331 & 1.02234 & 1.01356\
251 & 1.04701 & 1.04192 & 1.03740 & 1.03200 & 1.02110 & 1.01240\
252 & 1.04568 & 1.04060 & 1.03609 & 1.03070 & 1.01986 & 1.01124\
253 & 1.04436 & 1.03928 & 1.03478 & 1.02939 & 1.01862 & 1.01009\
254 & 1.04303 & 1.03796 & 1.03346 & 1.02809 & 1.01738 & 1.00893\
255 & 1.06817 & 1.06341 & 1.05913 & 1.05390 & 1.04269 & 1.03194\
256 & 1.06702 & 1.06225 & 1.05796 & 1.05274 & 1.04155 & 1.03085\
257 & 1.06587 & 1.06109 & 1.05680 & 1.05159 & 1.04041 & 1.02975\
258 & 1.06471 & 1.05993 & 1.05563 & 1.05043 & 1.03926 & 1.02865\
259 & 1.06356 & 1.05877 & 1.05447 & 1.04927 & 1.03812 & 1.02756\
260 & 1.06241 & 1.05761 & 1.05330 & 1.04811 & 1.03698 & 1.02646\
261 & 1.06126 & 1.05645 & 1.05213 & 1.04695 & 1.03583 & 1.02537\
262 & 1.06010 & 1.05529 & 1.05097 & 1.04580 & 1.03469 & 1.02427\
263 & 1.05895 & 1.05413 & 1.04980 & 1.04464 & 1.03355 & 1.02318\
264 & 1.05780 & 1.05297 & 1.04863 & 1.04348 & 1.03241 & 1.02208\
265 & 1.05664 & 1.05181 & 1.04747 & 1.04232 & 1.03126 & 1.02099\
266 & 1.05549 & 1.05066 & 1.04630 & 1.04116 & 1.03012 & 1.01989\
267 & 1.05434 & 1.04950 & 1.04514 & 1.04000 & 1.02898 & 1.01880\
268 & 1.05319 & 1.04834 & 1.04397 & 1.03885 & 1.02783 & 1.01770\
269 & 1.05203 & 1.04718 & 1.04280 & 1.03769 & 1.02669 & 1.01660\
270 & 1.05088 & 1.04602 & 1.04164 & 1.03653 & 1.02555 & 1.01551\
271 & 1.04973 & 1.04486 & 1.04047 & 1.03537 & 1.02441 & 1.01441\
272 & 1.04857 & 1.04370 & 1.03930 & 1.03421 & 1.02326 & 1.01332\
273 & 1.04742 & 1.04254 & 1.03814 & 1.03306 & 1.02212 & 1.01222\
274 & 1.04627 & 1.04138 & 1.03697 & 1.03190 & 1.02098 & 1.01113\
275 & 1.04511 & 1.04022 & 1.03581 & 1.03074 & 1.01983 & 1.01003\
276 & 1.04396 & 1.03906 & 1.03464 & 1.02958 & 1.01869 & 1.00894\
277 & 1.04281 & 1.03790 & 1.03347 & 1.02842 & 1.01755 & 1.00784\
278 & 1.04166 & 1.03674 & 1.03231 & 1.02727 & 1.01641 & 1.00674\
279 & 1.04050 & 1.03559 & 1.03114 & 1.02611 & 1.01526 & 1.00565\
280 & 1.03935 & 1.03443 & 1.02997 & 1.02495 & 1.01412 & 1.00455\
281 & 1.06263 & 1.05794 & 1.05390 & 1.04900 & 1.03793 & 1.02781\
282 & 1.06161 & 1.05693 & 1.05288 & 1.04798 & 1.03696 & 1.02694\
283 & 1.06059 & 1.05591 & 1.05187 & 1.04696 & 1.03600 & 1.02607\
284 & 1.05957 & 1.05490 & 1.05085 & 1.04594 & 1.03503 & 1.02520\
285 & 1.05855 & 1.05389 & 1.04983 & 1.04492 & 1.03406 & 1.02433\
286 & 1.05753 & 1.05288 & 1.04882 & 1.04390 & 1.03310 & 1.02347\
287 & 1.05652 & 1.05186 & 1.04780 & 1.04288 & 1.03213 & 1.02260\
288 & 1.05550 & 1.05085 & 1.04679 & 1.04186 & 1.03117 & 1.02173\
289 & 1.05448 & 1.04984 & 1.04577 & 1.04084 & 1.03020 & 1.02086\
290 & 1.05346 & 1.04883 & 1.04475 & 1.03982 & 1.02923 & 1.01999\
291 & 1.05244 & 1.04781 & 1.04374 & 1.03880 & 1.02827 & 1.01912\
292 & 1.05143 & 1.04680 & 1.04272 & 1.03778 & 1.02730 & 1.01825\
293 & 1.05041 & 1.04579 & 1.04170 & 1.03676 & 1.02633 & 1.01738\
294 & 1.04939 & 1.04478 & 1.04069 & 1.03574 & 1.02537 & 1.01651\
295 & 1.04837 & 1.04376 & 1.03967 & 1.03472 & 1.02440 & 1.01564\
296 & 1.04735 & 1.04275 & 1.03865 & 1.03369 & 1.02343 & 1.01477\
297 & 1.04633 & 1.04174 & 1.03764 & 1.03267 & 1.02247 & 1.01390\
298 & 1.04532 & 1.04073 & 1.03662 & 1.03165 & 1.02150 & 1.01303\
299 & 1.04430 & 1.03971 & 1.03561 & 1.03063 & 1.02054 & 1.01217\
300 & 1.04328 & 1.03870 & 1.03459 & 1.02961 & 1.01957 & 1.01130\
301 & 1.04226 & 1.03769 & 1.03357 & 1.02859 & 1.01860 & 1.01043\
302 & 1.04124 & 1.03668 & 1.03256 & 1.02757 & 1.01764 & 1.00956\
303 & 1.04023 & 1.03566 & 1.03154 & 1.02655 & 1.01667 & 1.00869\
304 & 1.03921 & 1.03465 & 1.03052 & 1.02553 & 1.01570 & 1.00782\
305 & 1.03819 & 1.03364 & 1.02951 & 1.02451 & 1.01474 & 1.00695\
306 & 1.03717 & 1.03263 & 1.02849 & 1.02349 & 1.01377 & 1.00608\
307 & 1.03615 & 1.03161 & 1.02748 & 1.02247 & 1.01280 & 1.00521\
308 & 1.03513 & 1.03060 & 1.02646 & 1.02145 & 1.01184 & 1.00434\
309 & 1.05700 & 1.05279 & 1.04895 & 1.04417 & 1.03414 & 1.02581\
310 & 1.05608 & 1.05187 & 1.04804 & 1.04325 & 1.03324 & 1.02492\
311 & 1.05517 & 1.05096 & 1.04712 & 1.04234 & 1.03233 & 1.02402\
312 & 1.05426 & 1.05004 & 1.04621 & 1.04143 & 1.03143 & 1.02313\
313 & 1.05335 & 1.04913 & 1.04529 & 1.04052 & 1.03052 & 1.02224\
314 & 1.05244 & 1.04821 & 1.04438 & 1.03960 & 1.02962 & 1.02135\
315 & 1.05153 & 1.04730 & 1.04347 & 1.03869 & 1.02871 & 1.02046\
316 & 1.05062 & 1.04638 & 1.04255 & 1.03778 & 1.02781 & 1.01957\
317 & 1.04971 & 1.04547 & 1.04164 & 1.03687 & 1.02691 & 1.01868\
318 & 1.04880 & 1.04456 & 1.04072 & 1.03595 & 1.02600 & 1.01779\
319 & 1.04789 & 1.04364 & 1.03981 & 1.03504 & 1.02510 & 1.01690\
320 & 1.04698 & 1.04273 & 1.03889 & 1.03413 & 1.02419 & 1.01601\
321 & 1.04606 & 1.04181 & 1.03798 & 1.03322 & 1.02329 & 1.01512\
322 & 1.04515 & 1.04090 & 1.03706 & 1.03230 & 1.02238 & 1.01423\
323 & 1.04424 & 1.03998 & 1.03615 & 1.03139 & 1.02148 & 1.01333\
324 & 1.04333 & 1.03907 & 1.03523 & 1.03048 & 1.02057 & 1.01244\
325 & 1.04242 & 1.03816 & 1.03432 & 1.02957 & 1.01967 & 1.01155\
326 & 1.04151 & 1.03724 & 1.03341 & 1.02866 & 1.01876 & 1.01066\
327 & 1.04060 & 1.03633 & 1.03249 & 1.02774 & 1.01786 & 1.00977\
328 & 1.03969 & 1.03541 & 1.03158 & 1.02683 & 1.01695 & 1.00888\
329 & 1.03878 & 1.03450 & 1.03066 & 1.02592 & 1.01605 & 1.00799\
330 & 1.03787 & 1.03358 & 1.02975 & 1.02501 & 1.01514 & 1.00710\
331 & 1.03696 & 1.03267 & 1.02883 & 1.02409 & 1.01424 & 1.00621\
332 & 1.03605 & 1.03175 & 1.02792 & 1.02318 & 1.01334 & 1.00532\
333 & 1.03513 & 1.03084 & 1.02700 & 1.02227 & 1.01243 & 1.00443\
334 & 1.03422 & 1.02993 & 1.02609 & 1.02136 & 1.01153 & 1.00354\
335 & 1.03331 & 1.02901 & 1.02518 & 1.02044 & 1.01062 & 1.00264\
336 & 1.03240 & 1.02810 & 1.02426 & 1.01953 & 1.00972 & 1.00175\
337 & 1.05310 & 1.04897 & 1.04528 & 1.04088 & 1.03216 & 1.02235\
338 & 1.05227 & 1.04815 & 1.04446 & 1.04007 & 1.03131 & 1.02156\
339 & 1.05145 & 1.04733 & 1.04364 & 1.03926 & 1.03046 & 1.02077\
340 & 1.05063 & 1.04651 & 1.04282 & 1.03844 & 1.02961 & 1.01999\
341 & 1.04980 & 1.04569 & 1.04200 & 1.03763 & 1.02876 & 1.01920\
342 & 1.04898 & 1.04487 & 1.04118 & 1.03682 & 1.02791 & 1.01841\
343 & 1.04816 & 1.04405 & 1.04037 & 1.03601 & 1.02706 & 1.01763\
344 & 1.04734 & 1.04323 & 1.03955 & 1.03519 & 1.02622 & 1.01684\
345 & 1.04651 & 1.04241 & 1.03873 & 1.03438 & 1.02537 & 1.01605\
346 & 1.04569 & 1.04159 & 1.03791 & 1.03357 & 1.02452 & 1.01526\
347 & 1.04487 & 1.04077 & 1.03709 & 1.03275 & 1.02367 & 1.01448\
348 & 1.04404 & 1.03995 & 1.03627 & 1.03194 & 1.02282 & 1.01369\
349 & 1.04322 & 1.03913 & 1.03546 & 1.03113 & 1.02197 & 1.01290\
350 & 1.04240 & 1.03830 & 1.03464 & 1.03032 & 1.02112 & 1.01212\
351 & 1.04158 & 1.03748 & 1.03382 & 1.02950 & 1.02028 & 1.01133\
352 & 1.04075 & 1.03666 & 1.03300 & 1.02869 & 1.01943 & 1.01054\
353 & 1.03993 & 1.03584 & 1.03218 & 1.02788 & 1.01858 & 1.00975\
354 & 1.03911 & 1.03502 & 1.03136 & 1.02707 & 1.01773 & 1.00897\
355 & 1.03828 & 1.03420 & 1.03055 & 1.02625 & 1.01688 & 1.00818\
356 & 1.03746 & 1.03338 & 1.02973 & 1.02544 & 1.01603 & 1.00739\
357 & 1.03664 & 1.03256 & 1.02891 & 1.02463 & 1.01518 & 1.00661\
358 & 1.03582 & 1.03174 & 1.02809 & 1.02381 & 1.01434 & 1.00582\
359 & 1.03499 & 1.03092 & 1.02727 & 1.02300 & 1.01349 & 1.00503\
360 & 1.03417 & 1.03010 & 1.02645 & 1.02219 & 1.01264 & 1.00424\
361 & 1.03335 & 1.02928 & 1.02564 & 1.02138 & 1.01179 & 1.00346\
362 & 1.03252 & 1.02846 & 1.02482 & 1.02056 & 1.01094 & 1.00267\
363 & 1.03170 & 1.02763 & 1.02400 & 1.01975 & 1.01009 & 1.00188\
364 & 1.03088 & 1.02681 & 1.02318 & 1.01894 & 1.00924 & 1.00110\
365 & 1.03006 & 1.02599 & 1.02236 & 1.01812 & 1.00840 & 1.00031\
366 & 1.02923 & 1.02517 & 1.02154 & 1.01731 & 1.00755 & 0.99952\
367 & 1.04869 & 1.04481 & 1.04132 & 1.03715 & 1.02815 & 1.01890\
368 & 1.04795 & 1.04407 & 1.04059 & 1.03642 & 1.02743 & 1.01826\
369 & 1.04722 & 1.04333 & 1.03986 & 1.03569 & 1.02670 & 1.01762\
370 & 1.04648 & 1.04260 & 1.03912 & 1.03496 & 1.02598 & 1.01699\
371 & 1.04575 & 1.04186 & 1.03839 & 1.03423 & 1.02525 & 1.01635\
372 & 1.04501 & 1.04113 & 1.03765 & 1.03350 & 1.02453 & 1.01571\
373 & 1.04428 & 1.04039 & 1.03692 & 1.03277 & 1.02380 & 1.01507\
374 & 1.04354 & 1.03965 & 1.03619 & 1.03204 & 1.02308 & 1.01443\
375 & 1.04281 & 1.03892 & 1.03545 & 1.03131 & 1.02235 & 1.01379\
376 & 1.04207 & 1.03818 & 1.03472 & 1.03058 & 1.02162 & 1.01315\
377 & 1.04134 & 1.03744 & 1.03398 & 1.02985 & 1.02090 & 1.01251\
378 & 1.04060 & 1.03671 & 1.03325 & 1.02912 & 1.02017 & 1.01187\
379 & 1.03987 & 1.03597 & 1.03252 & 1.02839 & 1.01945 & 1.01123\
380 & 1.03913 & 1.03524 & 1.03178 & 1.02766 & 1.01872 & 1.01059\
381 & 1.03840 & 1.03450 & 1.03105 & 1.02693 & 1.01800 & 1.00995\
382 & 1.03766 & 1.03376 & 1.03031 & 1.02620 & 1.01727 & 1.00931\
383 & 1.03693 & 1.03303 & 1.02958 & 1.02547 & 1.01655 & 1.00868\
384 & 1.03619 & 1.03229 & 1.02884 & 1.02474 & 1.01582 & 1.00804\
385 & 1.03546 & 1.03156 & 1.02811 & 1.02401 & 1.01510 & 1.00740\
386 & 1.03472 & 1.03082 & 1.02738 & 1.02328 & 1.01437 & 1.00676\
387 & 1.03399 & 1.03008 & 1.02664 & 1.02255 & 1.01364 & 1.00612\
388 & 1.03325 & 1.02935 & 1.02591 & 1.02182 & 1.01292 & 1.00548\
389 & 1.03252 & 1.02861 & 1.02517 & 1.02109 & 1.01219 & 1.00484\
390 & 1.03178 & 1.02787 & 1.02444 & 1.02036 & 1.01147 & 1.00420\
391 & 1.03105 & 1.02714 & 1.02371 & 1.01963 & 1.01074 & 1.00356\
392 & 1.03031 & 1.02640 & 1.02297 & 1.01890 & 1.01002 & 1.00292\
393 & 1.02958 & 1.02567 & 1.02224 & 1.01817 & 1.00929 & 1.00228\
394 & 1.02884 & 1.02493 & 1.02150 & 1.01744 & 1.00857 & 1.00164\
395 & 1.02811 & 1.02419 & 1.02077 & 1.01671 & 1.00784 & 1.00101\
396 & 1.02737 & 1.02346 & 1.02004 & 1.01598 & 1.00711 & 1.00037\
397 & 1.02664 & 1.02272 & 1.01930 & 1.01525 & 1.00639 & 0.99973\
398 & 1.04487 & 1.04108 & 1.03781 & 1.03391 & 1.02503 & 1.01735\
399 & 1.04421 & 1.04042 & 1.03715 & 1.03325 & 1.02437 & 1.01671\
400 & 1.04354 & 1.03976 & 1.03648 & 1.03258 & 1.02371 & 1.01607\
401 & 1.04288 & 1.03910 & 1.03582 & 1.03192 & 1.02305 & 1.01543\
402 & 1.04221 & 1.03844 & 1.03516 & 1.03126 & 1.02239 & 1.01479\
403 & 1.04155 & 1.03778 & 1.03449 & 1.03059 & 1.02173 & 1.01415\
404 & 1.04088 & 1.03711 & 1.03383 & 1.02993 & 1.02107 & 1.01351\
405 & 1.04022 & 1.03645 & 1.03317 & 1.02926 & 1.02041 & 1.01287\
406 & 1.03955 & 1.03579 & 1.03250 & 1.02860 & 1.01975 & 1.01223\
407 & 1.03889 & 1.03513 & 1.03184 & 1.02793 & 1.01909 & 1.01159\
408 & 1.03822 & 1.03447 & 1.03118 & 1.02727 & 1.01843 & 1.01095\
409 & 1.03756 & 1.03380 & 1.03051 & 1.02660 & 1.01777 & 1.01031\
410 & 1.03689 & 1.03314 & 1.02985 & 1.02594 & 1.01710 & 1.00967\
411 & 1.03623 & 1.03248 & 1.02919 & 1.02528 & 1.01644 & 1.00903\
412 & 1.03556 & 1.03182 & 1.02852 & 1.02461 & 1.01578 & 1.00839\
413 & 1.03490 & 1.03116 & 1.02786 & 1.02395 & 1.01512 & 1.00775\
414 & 1.03423 & 1.03049 & 1.02720 & 1.02328 & 1.01446 & 1.00711\
415 & 1.03357 & 1.02983 & 1.02653 & 1.02262 & 1.01380 & 1.00647\
416 & 1.03290 & 1.02917 & 1.02587 & 1.02195 & 1.01314 & 1.00583\
417 & 1.03224 & 1.02851 & 1.02521 & 1.02129 & 1.01248 & 1.00519\
418 & 1.03158 & 1.02785 & 1.02454 & 1.02062 & 1.01182 & 1.00455\
419 & 1.03091 & 1.02718 & 1.02388 & 1.01996 & 1.01116 & 1.00391\
420 & 1.03025 & 1.02652 & 1.02322 & 1.01929 & 1.01050 & 1.00327\
421 & 1.02958 & 1.02586 & 1.02255 & 1.01863 & 1.00984 & 1.00263\
422 & 1.02892 & 1.02520 & 1.02189 & 1.01797 & 1.00918 & 1.00199\
423 & 1.02825 & 1.02454 & 1.02123 & 1.01730 & 1.00852 & 1.00135\
424 & 1.02759 & 1.02387 & 1.02056 & 1.01664 & 1.00786 & 1.00071\
425 & 1.02692 & 1.02321 & 1.01990 & 1.01597 & 1.00720 & 1.00007\
426 & 1.02626 & 1.02255 & 1.01924 & 1.01531 & 1.00654 & 0.99943\
427 & 1.02559 & 1.02189 & 1.01857 & 1.01464 & 1.00588 & 0.99879\
428 & 1.02493 & 1.02123 & 1.01791 & 1.01398 & 1.00522 & 0.99815\
429 & 1.02426 & 1.02056 & 1.01725 & 1.01331 & 1.00455 & 0.99751\
430 & 1.02360 & 1.01990 & 1.01658 & 1.01265 & 1.00389 & 0.99687\
431 & 1.04110 & 1.03745 & 1.03418 & 1.03052 & 1.02236 & 1.01457\
432 & 1.04050 & 1.03685 & 1.03359 & 1.02991 & 1.02176 & 1.01399\
433 & 1.03990 & 1.03625 & 1.03299 & 1.02931 & 1.02116 & 1.01342\
434 & 1.03929 & 1.03565 & 1.03239 & 1.02871 & 1.02056 & 1.01285\
435 & 1.03869 & 1.03505 & 1.03179 & 1.02810 & 1.01995 & 1.01227\
436 & 1.03809 & 1.03445 & 1.03119 & 1.02750 & 1.01935 & 1.01170\
437 & 1.03749 & 1.03385 & 1.03059 & 1.02690 & 1.01875 & 1.01112\
438 & 1.03688 & 1.03325 & 1.02999 & 1.02630 & 1.01815 & 1.01055\
439 & 1.03628 & 1.03265 & 1.02940 & 1.02569 & 1.01755 & 1.00997\
440 & 1.03568 & 1.03205 & 1.02880 & 1.02509 & 1.01694 & 1.00940\
441 & 1.03507 & 1.03145 & 1.02820 & 1.02449 & 1.01634 & 1.00883\
442 & 1.03447 & 1.03085 & 1.02760 & 1.02388 & 1.01574 & 1.00825\
443 & 1.03387 & 1.03025 & 1.02700 & 1.02328 & 1.01514 & 1.00768\
444 & 1.03327 & 1.02965 & 1.02640 & 1.02268 & 1.01454 & 1.00710\
445 & 1.03266 & 1.02905 & 1.02580 & 1.02207 & 1.01393 & 1.00653\
446 & 1.03206 & 1.02845 & 1.02521 & 1.02147 & 1.01333 & 1.00596\
447 & 1.03146 & 1.02785 & 1.02461 & 1.02087 & 1.01273 & 1.00538\
448 & 1.03086 & 1.02725 & 1.02401 & 1.02026 & 1.01213 & 1.00481\
449 & 1.03025 & 1.02665 & 1.02341 & 1.01966 & 1.01152 & 1.00423\
450 & 1.02965 & 1.02605 & 1.02281 & 1.01906 & 1.01092 & 1.00366\
451 & 1.02905 & 1.02545 & 1.02221 & 1.01845 & 1.01032 & 1.00309\
452 & 1.02844 & 1.02485 & 1.02161 & 1.01785 & 1.00972 & 1.00251\
453 & 1.02784 & 1.02425 & 1.02102 & 1.01725 & 1.00912 & 1.00194\
454 & 1.02724 & 1.02365 & 1.02042 & 1.01664 & 1.00851 & 1.00136\
455 & 1.02664 & 1.02305 & 1.01982 & 1.01604 & 1.00791 & 1.00079\
456 & 1.02603 & 1.02245 & 1.01922 & 1.01544 & 1.00731 & 1.00022\
457 & 1.02543 & 1.02185 & 1.01862 & 1.01484 & 1.00671 & 0.99964\
458 & 1.02483 & 1.02125 & 1.01802 & 1.01423 & 1.00611 & 0.99907\
459 & 1.02423 & 1.02065 & 1.01742 & 1.01363 & 1.00550 & 0.99849\
460 & 1.02362 & 1.02005 & 1.01683 & 1.01303 & 1.00490 & 0.99792\
461 & 1.02302 & 1.01945 & 1.01623 & 1.01242 & 1.00430 & 0.99735\
462 & 1.02242 & 1.01885 & 1.01563 & 1.01182 & 1.00370 & 0.99677\
463 & 1.02182 & 1.01825 & 1.01503 & 1.01122 & 1.00310 & 0.99620\
464 & 1.02121 & 1.01765 & 1.01443 & 1.01061 & 1.00249 & 0.99562\
465 & 1.03769 & 1.03430 & 1.03122 & 1.02757 & 1.01995 & 1.01307\
466 & 1.03714 & 1.03375 & 1.03067 & 1.02702 & 1.01939 & 1.01253\
467 & 1.03659 & 1.03319 & 1.03012 & 1.02647 & 1.01884 & 1.01199\
468 & 1.03604 & 1.03264 & 1.02957 & 1.02592 & 1.01828 & 1.01146\
469 & 1.03549 & 1.03209 & 1.02902 & 1.02537 & 1.01773 & 1.01092\
470 & 1.03494 & 1.03154 & 1.02847 & 1.02482 & 1.01717 & 1.01038\
471 & 1.03439 & 1.03099 & 1.02791 & 1.02427 & 1.01662 & 1.00984\
472 & 1.03384 & 1.03043 & 1.02736 & 1.02372 & 1.01606 & 1.00931\
473 & 1.03329 & 1.02988 & 1.02681 & 1.02317 & 1.01551 & 1.00877\
474 & 1.03274 & 1.02933 & 1.02626 & 1.02262 & 1.01495 & 1.00823\
475 & 1.03219 & 1.02878 & 1.02571 & 1.02207 & 1.01440 & 1.00769\
476 & 1.03164 & 1.02823 & 1.02516 & 1.02152 & 1.01384 & 1.00716\
477 & 1.03109 & 1.02767 & 1.02461 & 1.02097 & 1.01329 & 1.00662\
478 & 1.03054 & 1.02712 & 1.02406 & 1.02042 & 1.01273 & 1.00608\
479 & 1.02999 & 1.02657 & 1.02351 & 1.01987 & 1.01218 & 1.00554\
480 & 1.02944 & 1.02602 & 1.02296 & 1.01932 & 1.01163 & 1.00501\
481 & 1.02889 & 1.02547 & 1.02241 & 1.01876 & 1.01107 & 1.00447\
482 & 1.02834 & 1.02491 & 1.02186 & 1.01821 & 1.01052 & 1.00393\
483 & 1.02779 & 1.02436 & 1.02131 & 1.01766 & 1.00996 & 1.00339\
484 & 1.02724 & 1.02381 & 1.02076 & 1.01711 & 1.00941 & 1.00286\
485 & 1.02669 & 1.02326 & 1.02021 & 1.01656 & 1.00885 & 1.00232\
486 & 1.02614 & 1.02271 & 1.01966 & 1.01601 & 1.00830 & 1.00178\
487 & 1.02559 & 1.02215 & 1.01911 & 1.01546 & 1.00774 & 1.00124\
488 & 1.02504 & 1.02160 & 1.01856 & 1.01491 & 1.00719 & 1.00071\
489 & 1.02449 & 1.02105 & 1.01801 & 1.01436 & 1.00663 & 1.00017\
490 & 1.02394 & 1.02050 & 1.01746 & 1.01381 & 1.00608 & 0.99963\
491 & 1.02340 & 1.01995 & 1.01691 & 1.01326 & 1.00552 & 0.99910\
492 & 1.02285 & 1.01939 & 1.01636 & 1.01271 & 1.00497 & 0.99856\
493 & 1.02230 & 1.01884 & 1.01580 & 1.01216 & 1.00441 & 0.99802\
494 & 1.02175 & 1.01829 & 1.01525 & 1.01161 & 1.00386 & 0.99748\
495 & 1.02120 & 1.01774 & 1.01470 & 1.01106 & 1.00330 & 0.99695\
496 & 1.02065 & 1.01719 & 1.01415 & 1.01051 & 1.00275 & 0.99641\
497 & 1.02010 & 1.01663 & 1.01360 & 1.00996 & 1.00219 & 0.99587\
498 & 1.01955 & 1.01608 & 1.01305 & 1.00941 & 1.00164 & 0.99533\
499 & 1.01900 & 1.01553 & 1.01250 & 1.00886 & 1.00108 & 0.99480\
500 & 1.03464 & 1.03135 & 1.02835 & 1.02489 & 1.01768 & 1.01220\
\[table:s2\]
[ccccccc]{}
\
Sample Sizes & p=0.1 & p=0.05 & p=0.025 & p=0.01 & p=0.001 & p=0.0001\
\
20 & 0.06682 & 0.03219 & 0.01591 & 0.00659 & 0.00065 & 0.00007\
21 & 0.06464 & 0.03226 & 0.01573 & 0.00632 & 0.00062 & 0.00005\
22 & 0.06512 & 0.03175 & 0.01531 & 0.00620 & 0.00062 & 0.00006\
23 & 0.06590 & 0.03182 & 0.01594 & 0.00618 & 0.00070 & 0.00007\
24 & 0.06512 & 0.03226 & 0.01601 & 0.00647 & 0.00062 & 0.00006\
25 & 0.06479 & 0.03165 & 0.01585 & 0.00622 & 0.00064 & 0.00006\
26 & 0.06357 & 0.03102 & 0.01566 & 0.00631 & 0.00063 & 0.00006\
27 & 0.06366 & 0.03112 & 0.01531 & 0.00607 & 0.00062 & 0.00007\
28 & 0.06309 & 0.03098 & 0.01512 & 0.00599 & 0.00061 & 0.00007\
29 & 0.06346 & 0.03038 & 0.01503 & 0.00600 & 0.00060 & 0.00006\
30 & 0.06293 & 0.03024 & 0.01495 & 0.00590 & 0.00057 & 0.00005\
31 & 0.06257 & 0.03042 & 0.01481 & 0.00586 & 0.00058 & 0.00007\
32 & 0.06206 & 0.03057 & 0.01498 & 0.00579 & 0.00058 & 0.00006\
33 & 0.06207 & 0.03016 & 0.01502 & 0.00606 & 0.00056 & 0.00005\
34 & 0.06169 & 0.03010 & 0.01497 & 0.00592 & 0.00058 & 0.00005\
35 & 0.06171 & 0.03003 & 0.01490 & 0.00584 & 0.00059 & 0.00005\
36 & 0.06153 & 0.03016 & 0.01465 & 0.00572 & 0.00058 & 0.00007\
37 & 0.06146 & 0.02965 & 0.01465 & 0.00574 & 0.00058 & 0.00005\
38 & 0.06027 & 0.02944 & 0.01454 & 0.00565 & 0.00057 & 0.00006\
39 & 0.06069 & 0.02942 & 0.01447 & 0.00566 & 0.00055 & 0.00005\
40 & 0.06032 & 0.02926 & 0.01420 & 0.00564 & 0.00056 & 0.00005\
41 & 0.06083 & 0.02960 & 0.01451 & 0.00572 & 0.00052 & 0.00006\
42 & 0.05989 & 0.02923 & 0.01446 & 0.00574 & 0.00055 & 0.00005\
43 & 0.06029 & 0.02928 & 0.01470 & 0.00577 & 0.00054 & 0.00005\
44 & 0.06045 & 0.02904 & 0.01431 & 0.00558 & 0.00055 & 0.00006\
45 & 0.05955 & 0.02907 & 0.01417 & 0.00570 & 0.00057 & 0.00005\
46 & 0.05950 & 0.02885 & 0.01419 & 0.00553 & 0.00054 & 0.00006\
47 & 0.05953 & 0.02891 & 0.01410 & 0.00558 & 0.00055 & 0.00006\
48 & 0.05962 & 0.02908 & 0.01415 & 0.00553 & 0.00055 & 0.00005\
49 & 0.05952 & 0.02874 & 0.01410 & 0.00549 & 0.00054 & 0.00005\
50 & 0.05903 & 0.02858 & 0.01417 & 0.00561 & 0.00053 & 0.00005\
51 & 0.05913 & 0.02857 & 0.01412 & 0.00558 & 0.00055 & 0.00006\
52 & 0.05933 & 0.02896 & 0.01406 & 0.00540 & 0.00053 & 0.00005\
53 & 0.05925 & 0.02884 & 0.01417 & 0.00557 & 0.00053 & 0.00004\
54 & 0.05918 & 0.02879 & 0.01408 & 0.00553 & 0.00052 & 0.00005\
55 & 0.05867 & 0.02861 & 0.01393 & 0.00550 & 0.00055 & 0.00005\
56 & 0.05871 & 0.02862 & 0.01423 & 0.00554 & 0.00052 & 0.00005\
57 & 0.05874 & 0.02838 & 0.01382 & 0.00552 & 0.00056 & 0.00005\
58 & 0.05874 & 0.02865 & 0.01394 & 0.00543 & 0.00053 & 0.00005\
59 & 0.05868 & 0.02845 & 0.01408 & 0.00550 & 0.00052 & 0.00006\
60 & 0.05843 & 0.02840 & 0.01378 & 0.00545 & 0.00053 & 0.00005\
61 & 0.05849 & 0.02840 & 0.01398 & 0.00547 & 0.00055 & 0.00006\
62 & 0.05806 & 0.02831 & 0.01390 & 0.00541 & 0.00052 & 0.00005\
63 & 0.05810 & 0.02812 & 0.01372 & 0.00546 & 0.00053 & 0.00005\
64 & 0.05832 & 0.02852 & 0.01391 & 0.00544 & 0.00053 & 0.00005\
65 & 0.05770 & 0.02831 & 0.01386 & 0.00543 & 0.00054 & 0.00006\
66 & 0.05831 & 0.02817 & 0.01378 & 0.00543 & 0.00052 & 0.00005\
67 & 0.05779 & 0.02805 & 0.01379 & 0.00539 & 0.00051 & 0.00005\
68 & 0.05776 & 0.02800 & 0.01379 & 0.00551 & 0.00054 & 0.00005\
69 & 0.05768 & 0.02780 & 0.01356 & 0.00534 & 0.00052 & 0.00005\
70 & 0.05776 & 0.02806 & 0.01382 & 0.00536 & 0.00054 & 0.00005\
71 & 0.05744 & 0.02778 & 0.01368 & 0.00536 & 0.00053 & 0.00005\
72 & 0.05746 & 0.02776 & 0.01363 & 0.00540 & 0.00053 & 0.00005\
73 & 0.05736 & 0.02798 & 0.01370 & 0.00544 & 0.00052 & 0.00005\
74 & 0.05709 & 0.02770 & 0.01358 & 0.00537 & 0.00053 & 0.00005\
75 & 0.05712 & 0.02766 & 0.01352 & 0.00526 & 0.00051 & 0.00005\
76 & 0.05675 & 0.02759 & 0.01362 & 0.00538 & 0.00053 & 0.00006\
77 & 0.05669 & 0.02769 & 0.01355 & 0.00531 & 0.00053 & 0.00005\
78 & 0.05698 & 0.02765 & 0.01361 & 0.00534 & 0.00052 & 0.00005\
79 & 0.05656 & 0.02766 & 0.01354 & 0.00538 & 0.00053 & 0.00005\
80 & 0.05704 & 0.02791 & 0.01359 & 0.00534 & 0.00053 & 0.00005\
81 & 0.05709 & 0.02756 & 0.01352 & 0.00530 & 0.00051 & 0.00005\
82 & 0.05706 & 0.02765 & 0.01360 & 0.00530 & 0.00051 & 0.00005\
83 & 0.05659 & 0.02730 & 0.01336 & 0.00527 & 0.00052 & 0.00005\
84 & 0.05707 & 0.02768 & 0.01363 & 0.00539 & 0.00051 & 0.00005\
85 & 0.05680 & 0.02755 & 0.01347 & 0.00526 & 0.00052 & 0.00005\
86 & 0.05686 & 0.02750 & 0.01362 & 0.00538 & 0.00054 & 0.00005\
87 & 0.05707 & 0.02766 & 0.01352 & 0.00535 & 0.00053 & 0.00006\
88 & 0.05655 & 0.02756 & 0.01347 & 0.00530 & 0.00051 & 0.00005\
89 & 0.05685 & 0.02762 & 0.01352 & 0.00532 & 0.00051 & 0.00005\
90 & 0.05624 & 0.02709 & 0.01332 & 0.00524 & 0.00053 & 0.00005\
91 & 0.05628 & 0.02728 & 0.01343 & 0.00527 & 0.00053 & 0.00005\
92 & 0.05639 & 0.02727 & 0.01346 & 0.00529 & 0.00051 & 0.00005\
93 & 0.05645 & 0.02736 & 0.01344 & 0.00530 & 0.00051 & 0.00005\
94 & 0.05645 & 0.02751 & 0.01347 & 0.00536 & 0.00051 & 0.00005\
95 & 0.05639 & 0.02755 & 0.01353 & 0.00530 & 0.00051 & 0.00005\
96 & 0.05637 & 0.02727 & 0.01341 & 0.00529 & 0.00051 & 0.00005\
97 & 0.05646 & 0.02716 & 0.01337 & 0.00528 & 0.00051 & 0.00005\
98 & 0.05668 & 0.02731 & 0.01350 & 0.00524 & 0.00050 & 0.00005\
99 & 0.05615 & 0.02736 & 0.01346 & 0.00527 & 0.00053 & 0.00006\
100 & 0.05616 & 0.02737 & 0.01343 & 0.00527 & 0.00052 & 0.00005\
\[table:s3\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Optical limiters transmit low-level radiation while blocking electromagnetic pulses with excessively high energy (energy limiters) or with excessively high peak intensity (power limiters). A typical optical limiter absorbs most of the high-level radiation which can cause its destruction via overheating. Here we introduce the novel concept of a reflective energy limiter which blocks electromagnetic pulses with excessively high total energy by reflecting them back to space, rather than absorbing them. The idea is to use a defect layer with temperature dependent loss tangent embedded in a low-loss photonic structure. The low energy pulses with central frequency close to that of the localized defect mode will pass through. But if the cumulative energy carried by the pulse exceeds certain level, the entire photonic structure reflects the incident light (and does not absorb it!) for a broad frequency window. The underlying physical mechanism is based on self-regulated impedance mismatch which increases dramatically with the cumulative energy carried by the pulse.'
address:
- 'Department of Physics, Wesleyan University, Middletown CT-06459, USA'
- 'The Air Force Research Laboratory, Sensors Directorate, Wright Patterson AFB, OH 45433 USA'
- 'Department of Physics, Wesleyan University, Middletown CT-06459, USA'
- 'The Air Force Research Laboratory, Sensors Directorate, Wright Patterson AFB, OH 45433 USA'
author:
- 'Eleana Makri, Tsampikos Kottos'
- Ilya Vitebskiy
title: Reflective Optical Limiter Based on Resonant Transmission
---
Introduction
============
The protection of photosensitive optical components from high incident radiation has applications to areas as diverse as microwave and optical communications to optical sensing [@ST91; @O97; @PHR98]. As a result, a considerable research effort has focused on developing novel protection schemes and materials that provide control of high-level optical and microwave radiation and prevent damages of optical sensors (including the human eye) and microwave antennas [@limiter1; @limiter2; @limiter3; @RC69; @BSMBS85; @DJ95]. Optical limiters constitute an important class of such protection devices. They are supposed to transmit low-level radiation, while blocking light pulses with high level of radiation. A typical passive optical limiter absorbs most of the high-level radiation, which can cause its destruction via overheating. The most common set-up of a passive optical limiter consists of a single protective layer with complex permittivity $\epsilon =\epsilon ^{\prime }+i\epsilon ^{\prime \prime }$, where the imaginary part $\epsilon ^{\prime \prime }$ increases sharply with the radiation level. For low-level radiation, the absorption is negligible, and the protective layer is transparent. An increase in the radiation level results in an increase in $\epsilon ^{\prime \prime }$, which renders the protective layer opaque. As a consequence, most of the high-level radiation will be absorbed by the limiter, which can cause its overheating and destruction. It turns out that if the same protective layer is incorporated into a certain photonic layered structure, the entire multilayer can become highly reflective for high-level radiation, while remaining transmissive at certain frequencies if the radiation level is low. Such a reflective limiter can be immune to overheating and destruction by high-level laser radiation, which is our main objective.
The physical reasons for the sharp increase in $\epsilon ^{\prime \prime }$ with the radiation level can be different. For instance, it can be photoconductivity, heating, two-photon absorption, or any combination of the these mechanisms. In our previous publication [@MRKV14] we considered the particular case of a strong non-linear dependence of $\epsilon ^{\prime
\prime }$ of the protective layer on light intensity. This can be attributed, for instance, to a two-photon absorption. We showed that incorporation of such a nonlinear layer in a properly designed low-loss layered structure makes the entire assembly act as a reflective power limiter. In this paper, we consider a more practical particular case where the increase in $\epsilon ^{\prime \prime }$ is due to heating of the protective layer. We show that, depending on the pulse duration as compared to the thermal relaxation time, the properly design layered structure incorporating such a protective layer can act as a reflective energy limiter, or as a reflective power limiter. Specifically, for short pulses, such a layered structure acts as an energy limiter, reflecting light pulses carrying excessively high energy. By comparison, for sufficiently long pulses, the same structure will act as a power limiter. In either case, most of the incident radiation will be reflected back to space, even though a stand-alone protective layer would act as an absorptive optical limiter.
The proposed architecture consists of a (protective) defect layer embedded in a low-loss Bragg grating . In contrast to the reflective power limiter introduced in [@MRKV14], the defect layer does not have to be nonlinear, but it must display strong temperature dependence $\epsilon ^{\prime \prime}(T)$ of the imaginary part of its permittivity. If the total energy carried by the pulse is low, $\epsilon ^{\prime \prime}(T)$ remains small enough to support a localized mode and the resonant transmittance associated with this mode. If, on the other hand, the energy carried by the pulse exceeds certain level, the defect layer becomes lossy enough to suppress the localized mode, along with the resonant transmittance. The entire stack turns highly reflective, which is consistent with our goal. We refer to this limiter as a *reflective energy limiter* in order to distinguish it from the nonlinear reflective power limiter introduced in [@MRKV14]. Finally, if the pulse duration significantly exceeds the thermal relaxation time of the defect layer, the entire layered structure will again act as a reflective power limiter with the cut-off light intensity determined by the thermal relaxation time of the defect layer – not by the nonlinearity in $\epsilon ^{\prime \prime }$, as was the case in [@MRKV14].
The organization of the paper is as follows. In Sec. \[model\], a conceptual design for the reflective energy limiter is presented, along with the mathematical formalism used in our calculations. In Sec. \[Theory\], we analyze the role of thermal conductivity. The latter plays an important role if the pulse duration is comparable or exceeds the thermal relaxation time of the defect layer.
![(Color online) A schematics of a reflective energy limiter. Two identical lossless Bragg reflectors are placed on the left and right of a lossy layer (green). The value of $\protect\epsilon ^{\prime \prime }$ in the defect layer is an increasing function of temperature. (a) Field distribution at the frequency of resonance transmission for an incident pulse with low energy – the field amplitude at the location of the defect layer is exponentially higher than that of the incident wave. (b) Transmittance vs. light wavelength for low incident light energy. (c) Field distribution at the frequency of maximum transmittance for an incident pulse with high energy –the amplitude of the suppressed localized mode is lower than that of the incident wave. (d) Transmittance vs. wavelength for an incident pulse with high energy.[]{data-label="fig1"}](fig1.pdf){width="1\linewidth"}
Physical Structure and mathematical model {#model}
=========================================
We consider two identical losses Bragg reflectors consisting of two alternating layers. Each mirror consists of forty layers which are placed at $-L\leq z
\leq 0$ and $d\leq z\leq L+d$. For the sake of the discussion we assume that the layers consist of Al$_2$O$_3$ and SiO$_2$ with corresponding permittivities $\epsilon_1=3.08$ and $\epsilon_2=2.1$. These values are typical for these materials at wavelengths $\lambda \sim 1 \mu m$. The width of layers is assumed to be $d_1=151nm$ and $d_2\approx 183nm$ respectively. At $0\leq z\leq d$ we introduce a defect lossy layer with complex permittivity $\epsilon_{\mathrm{d}} = \epsilon_{\mathrm{d}}^{\prime
}+i\epsilon_{\mathrm{d}}^{\prime \prime }$. We further assume that the imaginary part of the permittivity of the defect layer depends on the temperature $T$ i.e. $\epsilon_{\mathrm{d}}^{\prime \prime }=\epsilon_{\mathrm{d}}^{\prime \prime }(T)$. For simplicity, we assume linear dependence i.e. $\epsilon_{\mathrm{d}}^{\prime \prime }(T)=c_1+c_2 T$ where $c_1,c_2$ are some characteristic constants of the defect. Below we assume that $\epsilon_{\mathrm{d}}^{\prime }=12.11$ (which is a typical value for, say GaAs, at near infrared), $c_1=10^{-5}$ and $c_2= 1$ while the width of the defect layer is taken to be $d=151nm$.
The transmittance $\mathcal{T}$, reflectance $\mathcal{R}$ and absorption $\mathcal{A}$ of our set up, and the field profile at any frequency can be calculated via the transfer matrix approach. Specifically, a monochromatic electric field of frequency $\omega$ satisfies the Helmholtz equation: $${\frac{\partial ^{2}E(z)}{\partial z^{2}}}+{\frac{\omega ^{2}}{c^{2}}}\epsilon (z)E(z)=0\,\,\,.
\label{Helm}$$ At each layer inside the grating, Eq. (\[Helm\]) admits the solution $E^{(j)}= E_f^{(j)} \exp(in_j k z) + E_b^{(j)}\exp(-in_j k z)$, where $n_j=\sqrt{\epsilon_j}$ is the refraction index of the $j$-th layer and $k$ is the wave vector $k=\omega /n_{0}c$ ($c$ is the speed of light in the vacuum and $n_0$ is the refractive index of air). Imposing continuity of the field and its derivative at each layer interface, as well as taking into consideration the free propagation in each layer, we get the following iteration relation $$\begin{aligned}
\label{transfer0}
\left(\begin{array}{c}
E_{f}^{(j)} \\
E_{b}^{(j)}\end{array}
\right) & = & \mathcal{M}^{(j)} \left(\begin{array}{c}
E_{f}^{(j-1)} \\
E_{b}^{(j-1)}\end{array}
\right); \mathcal{M}^{(j)}= P_{R}^{(j)} Q^{(j)} K^{(j)} P_{L}^{(j)}.\end{aligned}$$ where $$\begin{aligned}
\label{sltm}
Q^{(j)}=&\left(\begin{array}{cc}
e^{ikn_jd_j} & 0 \\
0 & e^{-ikn_jd_j}
\end{array}
\right) \nonumber \\
K^{(j)}=& \left(\begin{array}{cc}
{\frac{n_j+n_{j-1}}{n_j}} & {\frac{n_j-n_{j-1}}{n_j}} \\
{\frac{n_j-n_{j-1}}{n_j}} & {\frac{n_j+n_{j-1}}{n_j}}
\end{array}
\right) \nonumber \\
P_{R}^{(j)}=&\left(
\begin{array}{cc}
e^{ikn_jz} & 0 \\
0 & e^{-ikn_jz}
\end{array}
\right) \nonumber \\
P_{L}^{(j)}=&\left(\begin{array}{cc}
e^{-ikn_{j-1}(z-d_j)} & 0 \\
0 & e^{ikn_{j-1}(z-d_j)}\end{array}
\right)\end{aligned}$$
At the same time the field outside the layered structured can be written as $E_{0}^{-}(z)=E_{f}^{-}
\exp(ikz)+E_{b}^{-}\exp (-ikz)$ for $z<-L$ and $E_{0}^{+}(z)=E_{f}^{+}\exp (ikz)+E_{b}^{+}\exp (-ikz)$ for $z>L+d$. The amplitudes of forward and backward propagating waves on the left $z<-L$ and right $z>L+d$ domains are related via the total transfer matrix $\mathcal{M}
=P_{R}^{(2N+2)} K^{(2N+2)}\Pi_j \mathcal{M}^{(j)}$ (where $N$ is the number of layers on each grating and $n_{2N+2}=n_{0}$): $$\begin{aligned}
\label{transfer}
\left(\begin{array}{c}
E_{f}^+ \\
E_{b}^+\end{array}
\right) & = & \left(
\begin{array}{cc}
\mathcal{M}_{11} & \mathcal{M}_{12} \\
\mathcal{M}_{21} & \mathcal{M}_{22}
\end{array}
\right) \left(
\begin{array}{c}
E_{f}^- \\
E_{b}^-
\end{array}
\right)\end{aligned}$$ The transmission and reflection coefficients and the field profile, say for a left incident wave, can be obtained by iterating backwards Eqs. (\[transfer0\],\[transfer\]) together with the boundary conditions $E_b^+=0$ and $\left|E_f^+\right|=1$ (due to the linearity of the equations, one can always impose a value for the outgoing field and calculate via a backward iteration of the transfer matrices the input field to which corresponds [@backward]). Specifically we have $\mathcal{T}\equiv|E_{f}^{+}/E_{f}^{-}|^{2}$; $\mathcal{R}\equiv
|E_{b}^{-}/E_{f}^{-}|^{2}$. These can be expressed in terms of the transfer matrix elements as $\mathcal{T}= \left|{\frac{1}{\mathcal{M}_{22}}}\right|^2;
\mathcal{R}= \left|{\frac{\mathcal{M}_{21}}{\mathcal{M}_{22}}}\right|^2$. The absorption coefficient $\mathcal{A}$ can then be evaluated in terms of transmittances and reflectances as $\mathcal{A}\equiv 1- \mathcal{T} -\mathcal{R}$.
Theoretical analysis {#Theory}
====================
In the case that the permittivity of the defect layer is replaced by $\epsilon _{d}=\epsilon _{1}$, the whole structure is periodic and displays a typical dispersion relation consisting of transparent frequency windows (bands) where light is transmitted with near-unity transmittance alternated with frequency windows (gaps) where the incident light is experiencing almost complete reflection.
When the defect is included in the middle of the grating, for zero temperature $T=0$ corresponding to permittivity $\epsilon_{ \mathrm{d}}\approx
\epsilon_{\mathrm{d}}^{\prime }$, the layered structure supports a localized resonant defect mode (see Fig. \[fig1\]a) with a frequency lying in a photonic band gap of the Bragg grating (see Figs. \[fig1\]b). For the specific set up that we consider here, we find that a resonant mode is located in the middle of the gap at wavelength $\lambda_r\approx 1060nm$. This defect mode is localized in the vicinity of the defect layer and decays exponentially away from the defect (see Fig. \[fig1\]a). In the vicinity of the localized mode frequency $\omega_r$, the entire layered structure displays a strong resonant transmission due to the excitation of the localized mode (see Fig. \[fig1\]b). In other words, the transmittance is $\mathcal{T} \approx 1$ while the reflectance and the absorption in the absence of any losses are $\mathcal{R}\approx 0$ and $\mathcal{A}\approx 0$ respectively. This picture is still applicable even in the presence of small (but non-zero) dissipative permittivity $\epsilon_{\mathrm{d}}^{\prime
\prime} \neq 0$ (see Fig. \[fig1\]a,b).
An alternative expression for the absorption coefficient $\mathcal{A}$ can be given in terms of the permittivity and the field intensity $|E(z)|^2$ inside the defect layer. The resulting expression is derived by subtracting the product of Eq. (\[Helm\]) with $E^*(z)$ from its complex conjugate form and then integrating the outcome over the interval $-L\leq z\leq L$. We get $$\label{A_wf}
\left(E^*{\frac{dE}{dz}} - E {\frac{dE^*}{dz}}\right)_{z=-L}^{z=L} +2 i k^2\int_{-L}^{L} \mathcal{I}m \epsilon(z) \left|E(z)\right|^2 dz =0.$$ Substituting in Eq. (\[A\_wf\]) the expressions of the electric field at $z=-L$ and $z=L$ respectively we get $$\label{Asfin}
\mathcal{A}\equiv 1-\mathcal{T} -\mathcal{R} = {\frac{k}{\left|E_f^-\right|^2}} \int_{-L}^{L} dz |E(z)|^2 \mathcal{I}m \epsilon(z).$$ Furthermore we assume that $\mathcal{I}m \epsilon (z)$ is zero everywhere inside the layered structure apart from the interval $0\leq z\leq d$ where the defect layer is placed. In this interval it takes a uniform value $\mathcal{I}m \epsilon (0\leq z\leq d) =
\epsilon_{\mathrm{d}}^{\prime \prime }(T)$. These simplifications allow us to express the absorption coefficient of Eq. (\[Asfin\]) in the following form $$\mathcal{A}(T)=\rho\left( T\right) \omega\epsilon_{\mathrm{d}
}^{\prime\prime}\left( T\right) \label{Afin}$$ where $\rho\left( T\right) =\mathcal{I}_{\mathrm{d}}/\left|E_f^{-}\right|^2$ is the ratio of the integral of light intensity $\mathcal{I}_{\mathrm{d}}
=\int_0^d dz \left|E(z)\right|^{2}$ at the lossy layer and the incident light intensity. It is obvious from Eq. (\[Afin\]) that $\mathcal{A}(T)$ depends on both the dissipative part of the permittivity and the value of the electric field inside the defect layer. Although the former increases monotonically with the temperature $T$ and thus with the duration time of the incident pulse, this is not true for $\rho(T)$. The latter, which is a unique function of the permittivity, remains approximately constant up to some value of $\epsilon_{\mathrm{d}}^{\prime \prime }$ above which it decreases, leading eventually to a total decrease of the absorption coefficient together with a simultaneous increase of the reflectivity of the structure. This is related to the fact that the increase of $\epsilon_{\mathrm{d}}^{\prime \prime }$ spoils the resonant localized mode (see Fig. \[fig1\]c) which is responsible for high transmittance. Specifically, when the losses due to $\epsilon_{\mathrm{d}}^{\prime \prime }$ overrun the losses due to leakage from the boundaries of the structure, the resonant mode cease to exist (see Fig. \[fig1\]c) and the structure becomes reflective, i.e. $\mathcal{R} \approx 1$, and $\mathcal{T} \approx 0$, see Fig. \[fig1\]d. As a consequence we have that $\mathcal{A}=1-\mathcal{T}-
\mathcal{R}\approx 0$ and the system does not absorb the high incident energy of the incoming light source but rather reflects it back in space.
In fact, the non-monotonic shape of the envelope of the scattering field in Fig. \[fig1\]c is a direct consequence of the fact that the structure becomes reflective $\mathcal{R}\approx 1; \mathcal{T}\approx 0$. One has to realize that in the case that both Bragg gratings on the left and right of the defect layer are finite, the field inside each half-space is written as a linear combination of two evanescent contributions with exponentially decreasing and exponentially increasing amplitudes. Their relative weight is determined by the boundary conditions $E(z=-L) =
E_0^-(-L)$ and $E(L)=E_0^{+}(L)= E_f^- \sqrt{\mathcal{T}}$ at the two outer interfaces of the layered structure. In the case of reflective structures these boundary conditions lead to the relation $E(-L)=E_f^-\sim \mathcal{O}(1)$ and $E(L)\approx 0$. It can be shown rigorously that in this case, the field on the left half-space of the structure is dominated originally by the exponentially decaying component while after some turning point $z_0$ the exponentially increasing component becomes dominant up to the defect layer. After that the field decays exponentially as in the resonant case. Similar scattering field profiles have been found in cases of active (gain) defects [@PACY10].
One can use a simple qualitative argument that allows to estimate the condition under which $\mathcal{A}(T)$ continues to increase. As we discuss previously, we assume that the electromagnetic energy losses occur in the lossy defect layer. The dissipated power can be estimated from Eq. (\[Afin\]) to be ${\dot Q}\propto \mathcal{A} \cdot \left|E_f^-\right|^2=\omega \epsilon_{\mathrm{d}}^{\prime
\prime }\mathcal{I}_{\mathrm{d}}$. Due to the energy conservation, the rate of energy dissipation cannot exceed the energy supply provided by the incident wave. The latter is $\mathcal{S}_{\mathrm{in}}\propto c \cdot \left|E_f^{-}\right|^2$. Taking this constraint into account we get the following upper limit on the field intensity at the defect layer location $$\label{ineq}
{\frac{c}{\omega \epsilon^{\prime \prime }(T) d}} \left|E_f^{-}\right|^2\geq
\left|E_{\mathrm{d}}\right|^2$$ Above we have made the additional approximation that $\mathcal{I}_{\mathrm{d}}\sim \left|E_{\mathrm{d}}\right|^2 \cdot d$, where $E_{\mathrm{d}}$ is a typical value of the field inside the defect layer.
Next we recall that a resonant mode with a frequency $\omega$ inside the band-gap has a Bloch wave number which is imaginary $k=
ik^{\prime \prime }$. The electric field inside the layered structure, can be expressed as a pair of evanescent modes, one of which is decaying with the distance $z$ and another one which is growing i.e. $E(z) = E_f \exp(-k^{\prime \prime }z) +E_b \exp(k^{\prime
\prime }z)$. To the left of the defect ($-L<z<0$), the electric field is dominated by the rising evanescent mode $E(z)\approx E_b
\exp(k^{\prime \prime }z)$ while to the right of the defect ($0<z<L$), the dominant contribution is provided by the decaying mode $E(z)\approx E_f \exp(-k^{\prime \prime }z)$ [@notefield].
The field $E_{\mathrm{d}}$ at the location of the defect layer is provided by the rising evanescent mode evaluated at $z=0$ i.e. $E_{\mathrm{d}}\sim E_b $. Therefore, the value of this evanescent mode at the left stack boundary at $z=-L$ is $$\label{equal}
E(-L) \propto E_{\mathrm{d}} \exp(-k^{\prime \prime }L)$$ Comparing (\[ineq\]) and (\[equal\]) we can conclude that if $$\label{ineq2}
{\frac{c}{\omega \epsilon^{\prime \prime }(T) d}} \exp(-2k^{\prime \prime}L)\ll 1$$ then the amplitude of rising evanescent mode $E(z=-L)$ at the left stack boundary is much less than amplitude of the incident wave $$\label{inequal}
\left|E(z=-L)\right|^2\ll \left|E_f^{-}\right|^2.$$ The latter condition Eq. (\[inequal\]), implies that the energy density inside the left grating is much smaller than the energy density of the incident wave, hence, only a small portion of the incident light energy $S_I\propto c \left|E_f^{-}\right|^2$ will cross the stack boundary at $z=-L$. In other word, the condition Eq. (\[ineq2\]) automatically implies high reflectivity at the stack interface. The condition Eq. (\[ineq2\]) for high stack reflectivity (and hence low transmittance and absorption) will always be satisfied if the loss tangent $\epsilon^{\prime \prime }(T)$ of the defect layer is large enough and/or if the number of layers in the Bragg grating is large enough.
![(a) The imaginary part $\protect\epsilon _{d}^{\prime \prime }$ of permittivity as a function of pulse duration $t_{f}$. The solid line corresponds to the layered structure in Fig. \[fig1\], while the dashed line corresponds to the stand-alone lossy layer. (b) The absorption $\mathcal{A}(t_{f})$ (black solid line), reflectance $\mathcal{R}(t_{f})$ (red solid lines) and transmission $\mathcal{T}(t_{f})$ (blue solid line) of the layered structure in Fig. \[fig1\] vs. pulse duration. For longer pulse duration (and larger cumulative energy of the pulse), the absorption $\mathcal{A}$ is suppressed and the set-up becomes highly reflective ($\mathcal{R}\approx 1$). The dashed lines show the respective values for the stand-alone lossy layer, in which case, the absorption for pulses with longer duration (and larger cumulative energy) is much higher, while the reflectivity is much lower than those of the layered structure in Fig. \[fig1\].[]{data-label="fig2"}](fig2.pdf){width="1\linewidth"}
![(Color online) The same as in Fig. \[fig2\] but now in the presence of thermal exchange between the defect layer and its surroundings ($\protect\kappa =0.05$). For longer pulse duration, a steady state regime is reached, which corresponds to a crossover from energy limiting regime to a power limiting regime.[]{data-label="fig3"}](fig3.pdf){width="1\linewidth"}
Next, we want to quantify the above arguments. To this end, we calculate explicitly the transport characteristics of our grating structure for an incident laser pulse. Although the analysis can be generalized for any incident pulse shape, in our numerical simulations below, we have assumed for simplicity that the incident laser pulse has a train-form [@notetrain] $$\begin{array}{cccc}
\mathcal{W}_I(t) & = 0 & \mathrm{for} & t\leq 0 \\
& =w_0 & \mathrm{for} & 0\leq t \leq t_f \\
& = 0 & \mathrm{for} & t\geq 0\nonumber
\end{array}$$ We want to calculate the total energy transmitted, reflected, and absorbed during the duration of the pulse. These can be expressed in terms of the time-dependent transmittance $\mathcal{T}(t)$, reflectance $\mathcal{R}(t)$ and absorption $\mathcal{A}(t)$ which are the main quantities that we analyze below. All other observables can be easily deduced from them. For example, the integrated (over the period of the pulse) absorption $\bar {\mathcal{A}}$ can be defined as $$\bar{\mathcal{A}}= {\frac{\int\limits_{-\infty}^{\infty} dt\mathcal{A}(t)
\mathcal{W}_{I}(t) }{\int\limits_{-\infty}^{\infty}dt \mathcal{W}_{I}(t)}};
\label{5}$$ while similar expressions can be used for calculating the total (over the period of the pulse) transmission $\bar {\mathcal{T}}$, and reflection $\bar {\mathcal{R}}$.
Our starting point is the “rate” equation $$\frac{d}{dt}T\left( t\right) =\frac{1}{C}\left(\mathcal{A}(T)\mathcal{W}_{I}\left( t\right)+\kappa (T_0-T)\right),
\label{heat1}$$ that describes the heating rate of the defect layer. Above, $C$ is the heat capacity, $\mathcal{W}_{I}\left( t\right)\equiv\left|
\mathcal{E}_{I}(t)\right|^{2}=\left|\int d\omega E(\omega) \exp(i\omega t)d\omega\right|^2$ is the incident light intensity, and $\kappa$ is the thermal conductance of the defect layer. The first term in Eq. (\[heat1\]) describes the heating process of the lossy layer while the second one corresponds to heat dissipation from the defect layer to the mirror (if any) or to the air. To further simplify our calculations, we assume that the temperature changes are within a domain where both thermal conductance and heat capacity are constants and independent of temperature changes.
Substitution of the absorption coefficient from Eq. (\[Afin\]) into Eq. (\[heat1\]) leads us to the following equation $$\frac{d}{dt}T\left( t\right) =\frac{1}{C} \left(\omega \varepsilon_{\mathrm{d}}^{\prime\prime}(T) \rho(T) \mathcal{W}_{I}
(t) +\kappa (T_0-T)\right)
\label{heat4}$$ which expresses the temporal behavior of the temperature $T(t)$ in terms of the given profile $\mathcal{W}_{I}(t)$ of the incident pulse. Everything else, e.g. $\epsilon_{\mathrm{d}}^{\prime \prime }(t)$, $\mathcal{A}(t)$, $\mathcal{T}(t)$ and $\mathcal{R}(t)$, can be directly and explicitly expressed in terms of $T(t)$.
In case that $\kappa=0$, one can further show that the outcomes can be written in terms of the total incident energy $U_f=\int_0^{t_f} \mathcal{W}_I(t) dt$. Furthermore, using Eq. (\[heat4\]) we get that $T_f=\int_0^{U_f}\mathcal{A}(U)
dU/C$. The associated total absorption is $\bar{\mathcal{A}} = \left(\int_0^{U_f} \mathcal{A}(U) dU\right)/U_f$, while similar expressions can be derived for the other transport characteristics.
In Fig. \[fig2\] we report the outcomes of a direct integration of Eq. (\[heat4\]) for the case of $\kappa=0$. In Fig. \[fig2\]a we report the temporal behavior of permittivity $\epsilon_{d}^{\prime \prime}$ as a function of the pulse duration $t_f$. Notice that for train pulses the pulse duration $t_f$ is directly analogous of the total incident energy $U_f$. We will therefore alternate, in our presentation below, the dependence of $\epsilon_{\mathrm{d}}^{\prime \prime}$, $\mathcal{T,R,A}$ from the pulse duration with the (more natural parameter for an energy limiter) total incident energy of the pulse.
Originally $\epsilon_{d}^{\prime \prime}$ is essentially unaffected by the incident energy and the same is true for the resonance mechanism (via the defect mode) that is responsible for high transmittance in the absence of losses. In this domain $\mathcal{T}\approx 1$, $\mathcal{R}\approx 0$ while there is a slow increase of the absorption $\mathcal{A}$, as it can be seen from Fig. \[fig2\]b (solid lines). Once the incident energy (pulse duration time) exceeds some critical value, there is a rather abrupt increase in $\epsilon_{\mathrm{d}}^{ \prime \prime}$ which results to the destruction of the resonance mode. Subsequently, the incident energy does not resonate into the structure, leading to a decaying absorption $\mathcal{A}\approx 0$, while the same is true for the transmittance $\mathcal{T}\approx 0$. At the same time, there is a noticeable growth of the reflectance which becomes approximately equal to unity $\mathcal{R}\approx 1$. For comparison we also plot at the same figure the results of the stand-alone layer. We find that for large incident energies (pulse durations $t_f$) the absorption $\mathcal{A}(t)$ is higher by more than two orders of magnitude as compared to the case of reflective energy limiter.
We have also performed the same analysis for the case where the thermal conductance $\kappa $ is different from zero. In Fig. \[fig3\] we report the results of the numerical integration of Eq. (\[heat4\]) in the presence of thermal conductivity. For long pulse duration we find a steady state behavior of the transport characteristics of the reflective energy limiter. The physical nature of the steady-state regime is quite obvious. It corresponds to the situation when the heat released in the defect layer is completely carried away by thermal conductivity. At this point, the temperature of the defect layer stabilizes and the time derivative $dT(t)/dt$ in Eqs. (\[heat1\],\[heat4\]) vanishes. The latter condition determines the steady-state values of the defect layer temperature as a function of the incident light amplitude. In this limiting case our structure acts as a power limiter. For comparison, the results of the stand-alone lossy layer are also reported in this figure. We find that in the steady-state regime our structure performs superbly resulting in absorption values which are more than two orders of magnitude smaller than the onces achieved by the stand-alone lossy layer.
Conclusions
===========
At infrared and optical frequencies, the reflectivity of known uniform materials is well below 90%, especially so when the incident light intensity is dangerously high. So, if we want to build a highly reflective optical limiter, we have to rely on photonic structures which would support some kind of low-intensity resonant transmission via slow or localized modes at photonic band-gap frequencies. If the incident light intensity increases, the respective localized mode must disappear, and the entire photonic structure will behave as a simple Bragg reflector. Here we considered the so-called “dissipative” mechanism of the localized mode suppression. At first glance, it seems counterintuitive, because the high reflectivity and low absorption are caused by the increase in the loss tangent of the defect layer in Fig. \[fig1\]. A qualitative explanation for such a phenomenon is that the large value of $\epsilon ^{\prime \prime }$ in the defect layer results in decoupling of the left and the right Bragg reflectors in Fig. \[fig1\]. Of course, there might be other ways to suppress resonant transmittance when the incident light intensity, or the total energy of the pulse, grow dangerously high. Still, the presented “dissipative” mechanism seems simple and practical.
*Acknowledgments -* This work is sponsored by the Air Force Office of Scientific Research LRIR09RY04COR and by an AFOSR MURI grant FA9550-14-1-0037.
[99]{} B. E. A. Saleh and M. C. Teich, *Fundamentals of Photonics* (Wiley, New York, 1991).
T. Ohtsuki, J. Lightwave Technol. **15**, 452 (1997)
N. S. Patel, K. L. Hall and K. A. Rauschenbach, Appl. Opt. **37**, 2831 (1998)
L. W. Tutt, T. F. Boggess, Prog. Quant. Electr. **17**, 299 (1993); A. E. Siegman, Appl. Opt. **1**, 739 (1962); J. E. Geusic, S. Singh, D. W. Tipping, and T. C. Rich, Phys. Rev. Lett. **19**, 1126 (1969).
Y. Zeng, X. Chen, W. Lu, J. Appl. Phys. **99**, 123107 (2006); M. Scalora, J. P. Bowling, C. M. Bowden, M. J. Bloemer, Phys. Rev. Lett. **73**, 1368 (1994)
S. Husaini, *et al.*, Appl. Phys. Lett., 191112 (2013); S. Pawar, *et al.*, J. Nonlin. Opt. Phys. & Mat. **21**, 1250017 (2012).
J. M. Ralston, R. K. Chang, Appl. Phys. Lett. **15**, 164 (1969); V. V. Arsenev, V. S. Dneprovksii, D. N. Klyshko, A. N. Penin, Sov. Phys. JETP **29**, 413 (1969).
T. F. Boggess, A. L. Smirl, S. C. Moss, I. W. Boyd and E. W. Van Stryland, IEEE J. Quantum Electron. **21**, 488 (1985); T. F. Boggess, S. C. Moss, I. W. Boyd, A. L. Smirl, Opt. Lett. **9**, 291 (1984).
M. D. Dvorak and B. L. Justus, Optics Communications **114**, 147 (1995)
E. Makri, H. Ramezani, T. Kottos, I. Vitebskiy, Phys. Rev. A **89**, 031802(R) (2014).
Macleod H. A. 2001 *Thin-Film Optical Filters* (Bristol and Philadelphia: Institute of Physics Publishing)
G. Tsironis and D. Hennig, Phys. Rep. **307**, 333 (1999).
B. Payne, J. Andreasen, H. Cao, A. Yamilov, Phys. Rev. B **82**, 104204 (2010).
We stress that in the case that both Bragg gratings on the left and right of the defect layer are finite, both evanescent contributions are present in either half-space, although, only one if them is dominant on either side. Furthermore one can show that the presence of both evanescent contributions on either side of the defect layer can provide an energy flux and, hence, a non-zero transmittance.
A. Figotin and I. Vitebskiy, Waves in Random and Complex Media **16**, 293 (2006).
We have checked that the same qualitative behaviour is obtained for other pulse shapes as well.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show how quantum coherence effects can be used to improve the resolution and the contrast of diffraction-limited images imprinted onto a probe field. The narrow and sharp spectral features generated by double-dark resonances (DDR) are exploited to control absorption, dispersion and diffraction properties of the medium. The spatial modulated control field can produce inhomogeneous susceptibility of the medium that encodes the spatial feature of the control image to probe field in the presence of DDR. The transmission of cloned image can be enhanced by use of incoherent pump field. We find that the feature size of cloned image is four times smaller than the initial characteristic size of the control image even though the control image is completely distorted after propagation through 3 cm long Rb vapour cell. We further discuss how spatial optical switching is possible by using of induced transparency and absorption of the medium.'
author:
- 'Onkar N.'
- 'Tarak N.'
title: 'Enhancement of image resolution beyond the diffraction-limit by interacting dark resonances'
---
Introduction
============
The ability to enhance spatial resolution of a Rayleigh or Sparrow limited image is one of the main challenges in optics [@Rayleigh]. Conventional optics has failed to resolve the characteristic size of an image beyond a value comparable to the wavelength of the probing light [@saleh]. Main constraint of high resolution imaging comes from the diffraction and the absorption. The diffraction of an image is inevitable due to its geometrical origin [@born]. The above obstacles can completely or partially be eliminated by use of quantum interference effects. Coherent electromagnetic fields interacting in a multilevel atomic system induce atomic coherence. The induced atomic coherence can be exploited to demonstrate many interesting phenomena such as coherent population trapping (CPT) [@Arimond_po_96], electromagnetically induced transparency (EIT) [@harris; @michael], lasing without inversion (LWI) [@olga_92] and saturated absorption techniques [@hansch_71; @agarwal_lpr_09]. A suitable spatially-dependent profile of the control field can produce a waveguide-like structure inside the medium which controls image propagation without diffraction[@truscott_99; @kapoor_00; @howell_09; @jorg_11]. This spatially varying refractive index can also guide focusing[@Moseley1; @Moseley2; @focusing3; @focusing4; @mitsunaga], de-focusing[@defocusing], self imaging[@cheng_07] and steering of the probe beam [@lida]. Most of the schemes employ a spatially inhomogeneous control field to protect the image from diffraction. In a different development, Firstenberg [*et al.*]{} theoretically and experimentally found that Dicke narrowing induced by atomic motion and velocity-changing collisions is useful to eliminate the diffraction of an arbitrary image [@Firstenberg1; @Firstenberg2; @Firstenberg3]. Tailoring the optical properties of the medium along the transverse direction can open up a new possibilities of transferring the characteristics features of the control field to the probe field. This is because the propagation dynamics of probe field is dependent on the diffraction and dispersive properties of the medium. The diffraction and dispersion characteristics of the atomic medium can be manipulated by using proper spatially inhomogeneous control field. This concept has been demonstrated in both experimentally [@Li] and theoretically [@om] in a CPT system where well resolved control field structure used for optical cloning. Further, the transmitted cloned image has feature size four times smaller as compared to the initial control image. However, all of these schemes suffer from strong absorption due to breaking of two-photon resonance condition. Hence the absorption based mechanism limits practical implementation. Therefore, one can take advantages of gain based schemes to generate high resolution cloned image. Resolution of cloned image can be improved by engineering the contrast of the refractive index of atomic waveguides of the gain medium. Quantum interference effects induced by interacting dark resonances have been shown to drastically increase the contrast of the refractive index profile [@chris; @Lukin]. In this paper, we have used interacting dark resonances to imprint the Rayleigh limited or Sparrow limited control image to probe field with high resolution and contrast. To facilitate these processes, we use four-level atomic system. A single dark state can be created by the control and the probe fields couple to the two arms of $\Lambda$-system. This interaction gives rise usual single transparency window. The double-dark states can generate by using a microwave or optical field which interacts with a magnetic or electric dipole moments of relevant atomic transitions [@chen; @yelin; @ye]. We find that the interference between two dark states results a new sharp absorption peak at line centre. The double dark resonance(DDR) spectra shows two transparency windows accompanied with one sharp absorption peak. Furthermore, we demonstrated that a very weak incoherent pump field is sufficient to turn the induced absorption dips to gain peaks. We exploit these sharp spectral features to write waveguide inside medium. We begin with Rayleigh limited control field structure and do a comparative study of inhomogeneous susceptibility for EIT, Microwave induced absorption (MIA), and LWI. The result shows that the presence of three fields with an incoherent pump provides a sharp contrast in refractive index from core to cladding than other two cases. We efficiently use this sharp refractive index contrast for cloning the Rayleigh limited control field image to the probe field with high resolution. Finally, we also show that Sparrow limited three modes of the control image can also be cast onto the probe field with appreciable resolution and high transmission. Later, we also use induced absorption and transparency mechanism to demonstrate the spatial switching (off or on) of probe beam. The spatial optical beam switching based on spatial phase modulation has been discussed recently in optical lattice [@hong]. The organisation of the paper is as follows. In the next section, we introduce our model configuration, discuss the equations of motion for four-level system, describe the perturbative analysis of linear susceptibility of the probe field and derive the beam propagation equations for both probe and control fields under paraxial approximations. In Sec. III, we present our results. First, we describe the linear response of the medium to the probe field under the action of the continuous wave(cw) as well as the spatially dependent control beam. We then employ the spatial dependent susceptibility to explain the basic principle of cloning of Rayleigh limited control image to the probe field with high resolution and high contrast. Next we provide numerical results on propagation dynamics of cloned images with different spatial structure of the control field for LWI, EIT and MIA cases. Sec. IV provides a summary and discussion of our results.
Theoretical Formulations
========================
Model configuration
-------------------
In this work, we consider a homogeneously broadened four level atomic system consisting of an excited state ${\ensuremath{|4\rangle}}$ and three metastable states ${\ensuremath{|1\rangle}}$, ${\ensuremath{|2\rangle}}$, and ${\ensuremath{|3\rangle}}$ interacting with two optical fields and one microwave field as shown in the Fig. \[fig:Fig1\]. The excited state ${\ensuremath{|4\rangle}}$ is coupled to two degenerate ground states ${\ensuremath{|1\rangle}}$, and ${\ensuremath{|3\rangle}}$ by two coherent fields, namely, a weak probe field with frequency $\omega_1$ and a control field with frequency $\omega_2$, respectively, which form a three level $\Lambda$-system. The ground state ${\ensuremath{|3\rangle}}$ is further coupled to the metastable state ${\ensuremath{|2\rangle}}$ by an additional microwave field with frequency $\omega_3$. We define two co-propagating optical fields along the $z$-axis as $$\label{field}
{\vec{E}_j}(\vec{r},t)= \hat{e}_{j}\mathcal{E}_{j}(\vec{r})~e^{- i\left(\omega_j t- k_j z\right )} + {c.c.}\,,$$ where, $\mathcal{E}_{j}(\vec{r})$ is the slowing varying envelope, $\hat{e}_{j}$ is the unit polarization vector, $\omega_j$ is the laser field frequency and $k_j$ is the wave number of field, respectively. The index $j\in \{1,2\}$ denotes the probe or control field, respectively. The microwave field is defined as $$\label{field}
{\vec{E}_3}({r},t)= \hat{e}_{3}\mathcal{E}_{3}(\vec{r})~e^{- i\left(\omega_3 t-k_3z \right)} + {c.c.}\,,$$ where, $\mathcal{E}_{3}(\vec{r})$ is constant amplitude, $\omega_3$ is the frequency of the microwave field.
![\[fig:Fig1\] (Color online) Schematic diagram of the four-level $^{87}$Rb atomic system. The atomic transition ${\ensuremath{|4\rangle}}\leftrightarrow{\ensuremath{|1\rangle}}$ is coupled by the coherent probe field $g$ and incoherent pump field $r$. The control field $G$ interacts to the atomic transition ${\ensuremath{|4\rangle}}\leftrightarrow{\ensuremath{|3\rangle}}$. A microwave field $\Omega$ acts on the transition ${\ensuremath{|3\rangle}}\leftrightarrow{\ensuremath{|2\rangle}}$ to produce the double dark resonance of the system.](Fig1.eps){width="0.8\columnwidth"}
In the presence of three coherent fields, the Hamiltonian of the system under the electric dipole and rotating-wave approximation can be expressed as,
\[Hschroed\] $$\begin{aligned}
H =& H_0 + H_I\,,\\
H_0 =& \hbar\omega_{43}{\ensuremath{|4\rangle}}{\ensuremath{\langle 4 |}} - \hbar\omega_{23} {\ensuremath{|2\rangle}}{\ensuremath{\langle 2 |}} - \hbar\omega_{13} {\ensuremath{|1\rangle}}{\ensuremath{\langle 1 |}}\,,\\
H_I =& - ( {\ensuremath{|4\rangle}}{\ensuremath{\langle 1 |}} {\boldsymbol{d}}_{41}\cdot\mathcal{E}_1e^{- i\left(\omega_1 t- k_1 z\right )} \nonumber \\
& + {\ensuremath{|4\rangle}}{\ensuremath{\langle 3 |}} {\boldsymbol{d}}_{43}\cdot\mathcal{E}_2e^{- i\left(\omega_2 t- k_2 z\right )}\nonumber\\
& + {\ensuremath{|3\rangle}}{\ensuremath{\langle 2 |}} {\boldsymbol{d}}_{32}\cdot\mathcal{E}_3e^{- i\left(\omega_3 t- k_3 z\right )}\,+\,\text{H.c.})\,,
\end{aligned}$$
The time dependent parts of the above Hamiltonian can be removed by use of unitary transformation,
$$\begin{aligned}
W&=e^{-\frac{i}{\hbar}U t} \,,\\
U&=\hbar \omega_2 {\ensuremath{|4\rangle}}{\ensuremath{\langle 4 |}} - \hbar \omega_3 {\ensuremath{|2\rangle}}{\ensuremath{\langle 2 |}} - \hbar (\omega_1-\omega_2){\ensuremath{|1\rangle}}{\ensuremath{\langle 1 |}}\,.\end{aligned}$$
Now, we can rewrite transformed Hamiltonian as $$\begin{aligned}
{V}/{\hbar} = & -{\Delta_2}{\ensuremath{|4\rangle}}{\ensuremath{\langle 4 |}} + {\Delta_3}{\ensuremath{|2\rangle}}{\ensuremath{\langle 2 |}} + ({\Delta_1}-{\Delta_2}){\ensuremath{|1\rangle}}{\ensuremath{\langle 1 |}} \notag \\
& - (g{\ensuremath{|4\rangle}}{\ensuremath{\langle 1 |}} + G \,{\ensuremath{|4\rangle}}{\ensuremath{\langle 3 |}} \,+ \Omega \,{\ensuremath{|3\rangle}}{\ensuremath{\langle 2 |}} \notag \\
& \,+\,\text{H.c.})\,,
\label{Heff}\end{aligned}$$ where ${\Delta_1}= \omega_1 - \omega_{41}$, ${\Delta_2}= \omega_2 - \omega_{43}$, ${\Delta_3}= \omega_2 - \omega_{32}$ are the single-photon detunings and $$\label{field}
g=\frac{\vec{d}_{41}\cdot\vec{\mathcal{E}}_{\rm{1}}e^{ik_1 z}}{\hbar},~~G=\frac{\vec{d}_{43}\cdot\vec{\mathcal{E}}_{\rm{2}}e^{ik_2 z}}{\hbar},~~ \Omega=\frac{\vec{d}_{32}\cdot\vec{\mathcal{E}}_{\rm{3}}e^{ik_3 z}}{\hbar}\nonumber$$ are the Rabi frequencies of the probe, control and the microwave fields, respectively. The atomic transition frequencies and the corresponding dipole moment matrix elements are denoted by $\omega_{ij}$ and ${\vec{d}}_{ij}$, respectively.
Dynamical equations
-------------------
We use Liouville equation to incorporate the coherent and incoherent processes of the atomic system. Thus the dynamics of the system is governed by the following Liouville equation $$\begin{aligned}
\label{master}
\dot{\rho}=-\frac{{i}}{\hbar}\left[V,\rho\right]+\mathcal{L}\rho\,.\end{aligned}$$ where the second term represents the incoherent processes that can be determined by $$\begin{aligned}
\mathcal{L}\rho = &\mathcal{L}_{\gamma}\rho+\mathcal{L}_{d}\rho+\mathcal{L}_{r}\rho\,,
\label{decay}\end{aligned}$$ with $$\begin{aligned}
\mathcal{L}_{\gamma}\rho = &-\sum\limits_{i=1}^3 \frac{\gamma_{i}}{2}\left({\ensuremath{|4\rangle}}{\ensuremath{\langle 4 |}}\rho-2{\ensuremath{|i\rangle}}{\ensuremath{\langle i |}}\rho_{44}+\rho{\ensuremath{|4\rangle}}{\ensuremath{\langle 4 |}}\right) \,,\nonumber\\
\mathcal{L}_{d}\rho = &-\sum\limits_{i=1}^3\sum\limits_{i\neq j=1}^3 \frac{\gamma_{c}}{2}\left({\ensuremath{|i\rangle}}{\ensuremath{\langle i |}}\rho-2{\ensuremath{|j\rangle}}{\ensuremath{\langle j |}}\rho_{ii}+\rho{\ensuremath{|i\rangle}}{\ensuremath{\langle i |}}\right) \,,\nonumber\\
\mathcal{L}_{r}\rho = &\mathcal{L}_{14}\rho + \mathcal{L}_{41}\rho \,,\nonumber\\
\mathcal{L}_{14}\rho = &- \frac{r}{2}\left({\ensuremath{|4\rangle}}{\ensuremath{\langle 4 |}}\rho-2{\ensuremath{|1\rangle}}{\ensuremath{\langle 1 |}}\rho_{44}+\rho{\ensuremath{|4\rangle}}{\ensuremath{\langle 4 |}}\right) \,,\nonumber\\
\mathcal{L}_{41}\rho = &- \frac{r}{2}\left({\ensuremath{|1\rangle}}{\ensuremath{\langle 1 |}}\rho-2{\ensuremath{|4\rangle}}{\ensuremath{\langle 4 |}}\rho_{11}+\rho{\ensuremath{|1\rangle}}{\ensuremath{\langle 1 |}}\right)\,.\nonumber
\label{idecay}\end{aligned}$$ The first term of Eq.(\[decay\]) refers to the radiative decay from excited state $|4\rangle$ to ground states $|j\rangle$ as labelled by $\gamma_{j}$. The second term, $\mathcal{L}_{d}\rho$, represents pure dephasing for the coherence $\rho_{ij}$ due to collision with rate $\gamma_{c}$. The incoherent pumping between levels ${\ensuremath{|1\rangle}}$ and ${\ensuremath{|4\rangle}}$ with rate $r$ is descryibed by $\mathcal{L}_{r}\rho$. The dynamics of the population and atomic coherences in the four level system can be described by the following set of density matrix equations
\[Full\_density\] $$\begin{aligned}
\dot{\rho}_{11}=&-r\rho_{11}+r\rho_{44}+\gamma_{1}\rho_{44}+ {i}g^* \rho_{41}
- {i}g \rho_{14} \,,\\
\dot{\rho}_{22}=& \gamma_{2}\rho_{44} + {i}\Omega^* \rho_{32} - {i}\Omega
\rho_{23}\,,\\
\dot{\rho}_{33}=& \gamma_{3}\rho_{44} + {i}\Omega \rho_{23} - {i}\Omega^*
\rho_{32}+ {i}G^*\rho_{43} - {i}G\rho_{34}\,,\\
\dot{\rho}_{44}=&-\dot{\rho}_{11}-\dot{\rho}_{22}-\dot{\rho}_{33} \,,\\
\dot{\rho}_{21}=&-\left[\frac{r}{2} + \gamma_{21} - {i}(\Delta_1-\Delta_2-\Delta_3)\right]\rho_{21} + {i}\Omega^* \rho_{31} \nonumber\\
&- {i}g \rho_{24} \,,\\
\dot{\rho}_{23}=&-\left[\gamma_{23} + {i}\Delta_3\right]\rho_{23}
- {i}G\rho_{24} + {i}\Omega^*(\rho_{33} - \rho_{22})\,,\\
\dot{\rho}_{24}=&-\left[ \gamma_{24} + {i}(\Delta_2+\Delta_3)\right]\rho_{24}
- {i}g^* \rho_{21}- {i}G^* \rho_{23} \nonumber\\
& + {i}\Omega^*\rho_{34} \,,\\
\dot{\rho}_{31}=&-\left[\frac{r}{2} + \gamma_{31} + {i}(\Delta_2-\Delta_1)\right]\rho_{31}
+ {i}\Omega\rho_{21}- {i}g^* \rho_{34} \nonumber\\
& + {i}G^*\rho_{41} \,,\\
\dot{\rho}_{34}=&-\left[ \gamma_{34} - {i}\Delta_2\right]\rho_{34}
- {i}g^*\rho_{31}+ {i}\Omega\rho_{24} \nonumber\\
& - {i}G^*(\rho_{33} - \rho_{44}) \,,\\
\dot{\rho}_{41}=&-\left[\frac{r}{2} + \gamma_{41} - {i}\Delta_1\right]\rho_{41}
+ {i}G \rho_{31} - {i}g(\rho_{11} - \rho_{44}) \,,\\
\dot{\rho}_{ij}=&\dot{\rho}_{ji}^*\,.
\end{aligned}$$
where, the overdots stand for time derivatives and $``*"$ denotes complex conjugate. The total dephasing rate of the atomic coherences is given by ${\gamma}_{ij} = {\gamma}_{c} +{{\gamma}_{i}}/{2}$.
Perturbative analysis
---------------------
We adopt steady state solutions of the master equations (\[Full\_density\]) to study the response of the medium. The equations (\[Full\_density\]) can be solved to all orders in the control and probe field provided both the fields have approximately equal amplitude [@om]. However, in the spirit of weak probe field limit, we calculate the coherences and populations to the first order in $g$ and to all order in control field $G$ and microwave field $\Omega$. Hence the steady state solutions of the density matrix equations can be written in the form of $$\rho_{_{ij}}=\rho_{_{ij}}^{(0)}+g\rho_{_{ij}}^{(+)}+g^*\rho_{_{ij}}^{(-)},$$ where, $\rho_{ij}^{(0)}$ describes the solution in the absence of the probe field. The second and third terms denote the solutions at positive and negative frequencies of the probe field. We now substitute the above expression in equations (\[Full\_density\]) and equate the coefficients of $g$, $g^{*}$ and the constant terms. Thus, we obtain a set of sixteen coupled simultaneous equations. The solutions of simultaneous equations which are relevant for susceptibility expression are given in Appendix \[app-A\]. Now, the steady state value of the atomic coherence ${\rho}_{41}^{(+)}$ will yield susceptibility $\chi_{41}$ at frequency $\omega_1$ $$\begin{aligned}
\label{mollow_chi}
{\rho}_{41}^{(+)} &=i\left(\frac{{(\Gamma_{21}\Gamma_{31}+\Omega^2)({\rho}_{11}^{(0)}-{\rho}_{44}^{(0)})}+A G^2}{\Gamma_{41}(\Gamma_{21}\Gamma_{31}+\Omega^2)+\Gamma_{21}G^2}\right)\,,
$$ with $$\begin{aligned}
{A} &= \frac{B({\rho}_{44}^{(0)} - {\rho}_{33}^{(0)})+C({\rho}_{33}^{(0)} - {\rho}_{22}^{(0)})}{(\Gamma_{23}(\Gamma_{24}\Gamma_{34}+\Omega^2)+\Gamma_{34}G^2)}\,,\nonumber \\
{B} &=(\Gamma_{21}(\Gamma_{23}\Gamma_{24}+G^2)-\Gamma_{23}\Omega^2)\,,\nonumber \\
{C} &= (\Gamma_{21}+\Gamma_{34})\Omega^2\,.\nonumber
$$ where ${\Gamma}_{21}=\left[r/2 + \gamma_{21} - {i}(\Delta_1-\Delta_2-\Delta_3)\right]$, ${\Gamma}_{23}=\left[\gamma_{23} + {i}\Delta_3\right]$, ${\Gamma}_{24}=\left[ \gamma_{24} + {i}(\Delta_2+\Delta_3)\right]$, ${\Gamma}_{31}=\left[r/2 + \gamma_{31} + {i}(\Delta_2-\Delta_1)\right]$, ${\Gamma}_{34}=\left[ \gamma_{34} - {i}\Delta_2\right]$, and ${\Gamma}_{41}=\left[r/2 + \gamma_{41} - {i}\Delta_1\right]$. For the simplicity, we have assumed equal decay rates from excited state, $\gamma_{1}=\gamma_{2}=\gamma_{3}=\gamma$ and coherence dephasing rates $\gamma_{41}=\gamma_{24}=\gamma_{34}\approx\gamma$, $\gamma_{21}=\gamma_{31}=\gamma_{23}\approx\gamma_{c}=\Gamma$. We now express the macroscopic polarization of the medium in terms of both the atomic coherences as well as the susceptibility as $$\begin{aligned}
\label{polarization}
\vec{\mathcal{P}}_1&=\mathcal{N}\left(\vec{d}_{41}\rho_{41}^{(+)}e^{-{i}\omega_1 t}+c.c.\right)\notag\\
&=\left(\chi_{41} \hat{e}_{1}\mathcal{E}_1e^{-{i}\omega_1 t}+c.c.\right)\,,\end{aligned}$$ where $\mathcal{N}$ is the density of the atomic medium. Now Eq. (\[mollow\_chi\]) and (\[polarization\]), will yield the linear response of the medium as $$\begin{aligned}
\label{chi_41}
{\chi}_{41}({\Delta_1})&=\frac{\mathcal{N}|d_{41}|^2}{{\hbar}}{\rho}_{41}^{(+)} .\end{aligned}$$ The real and imaginary parts of the susceptibility ${\chi}_{41}$ in Eq. (\[chi\_41\]) gives the dispersion and absorption of the medium respectively. The optical properties of the medium can be manipulated coherently by proper consideration of spatial shape and intensity of the different applied fields. The effect of different fields such as optical, microwave and incoherent pump field on the medium properties are in sequence in the results and discussions section.
Beam propagation equation with paraxial approximation
-----------------------------------------------------
The spatial dynamics of the probe and control fields along the $z$-direction of the medium is governed by the Maxwell’s wave equations. The wave equation under slowly varying envelope and paraxial wave approximations can result the beam propagation equation. The spatial evolution equations for the probe and control fields are obtained as
$$\begin{aligned}
\frac{\partial g}{\partial z}
&= \frac{{i}}{2{k_1}} \left( \frac{\partial^2 }{\partial x^2}
+ \frac{\partial^2 }{\partial y^2} \right) g + 2i{\pi}k_1{\chi}_{41}\,{g} \,,\label{probe} \\
\frac{\partial G}{\partial z}
&= \frac{i}{2{k_2}} \left( \frac{\partial^2 }{\partial x^2}
+ \frac{\partial^2 }{\partial y^2} \right) G \,\label{control}.\end{aligned}$$
The terms within the parentheses on the right hand side of Eq. (\[probe\]) and Eq. (\[control\]) are related with transverse variation of the laser beam. These terms account for the diffraction either in free space or in the medium. The second term on the right hand side of Eq. (\[probe\]) is responsible for the dispersion and absorption or gain of the probe beam. Note that the effects of the atomic coherences on the control beam propagation are very negligible under the weak probe field [@om]. Therefore, we study the effect of both diffraction and dispersion for the spatial evolution of the probe beam where we include only the effect of diffraction for the control beam dynamics.
Results and Discussions
========================
Susceptibility with continuous wave fields
------------------------------------------
![\[fig:Fig2\] (Color online) The variations of the imaginary part of the probe susceptibility with the detuning ${\Delta_1}$ in the presence and absence of both microwave field and incoherent pump is plotted. The zoomed part of the absorption spectrum corresponds to medium loss, gain or transparency at the line center is shown in the inset. The corresponding parameters for these regimes are: ${\Omega}=0.01{\gamma}$, $r=0$ (red dot-dashed line), ${\Omega}=0.01{\gamma}$, $r=0.0005\gamma$ (green dashed line), and ${\Omega}=0{\gamma}$, $r=0$ (black solid line). The common parameters are $G=1.0{\gamma}$, ${\Delta_2}={\Delta_3}=0$, $\Gamma=0.0001\gamma$, $\gamma=3\pi\times10^6$ rad/sec, and $N=5{\times}10^{11}$ atoms/cm$^{3}$.](Fig2.eps){width="1.0\columnwidth"}
We first study the atomic coherences by using cw optical and microwave fields at steady-state condition. The quantum interference of atomic coherences induces EIT, MIA and LWI in our system. The characteristic of these quantum interference phenomena is illustrated in Fig. \[fig:Fig2\]. In Fig. \[fig:Fig2\] we have plotted the variations of the imaginary part of the probe susceptibility with probe field detuning ${\Delta_1}$ in the presence and absence of both microwave and incoherent pump fields. In the absence of both microwave and incoherent pump fields four-level system reduces to three-level $\Lambda$ system with a weak probe and a strong control field. The probability amplitudes of two arms of the $\Lambda$ system leads to destructive interference. This interference enable us to cancellation of absorption of probe field provided two-photon resonance condition is fulfilled as shown in Fig. \[fig:Fig2\]. This phenomenon is known as EIT. In EIT, a single transparency window is accompanied by two absorptive peaks which originates from the strong control field. Now this single transparency window can be split into double transparency windows by the use of the microwave field. It is clear from Fig. \[fig:Fig2\] that the double transparency window is accompanied with very narrow absorption peak. This peak occurs due to the double dark states formed by microwave field at three-photon resonance condition. Furthermore, the position and width of these two transparency windows strongly depend on the intensity of microwave field. Now a relatively weak incoherent pump acting along the probe transition can switch the absorption peak to the gain dip. The second term in the numerator of Eq. (\[mollow\_chi\]) is responsible for gain around line center. This gain characteristic is illustrated by green dashed line line in Fig. \[fig:Fig2\]. At three photon resonance the second term is negative and is lager than the first term which changes the properties of the medium from absorption into gain. Thus the presence of both weak microwave and incoherent pump fields is able to produce a gain window for the medium.
![ \[fig:Fig3\] (Color online) Spatial intensity variation of the control image is plotted against the transverse axis $x$ with $y=0$ at entry face of the vapor cell. The Rayleigh limited and Sparrow limited control image are formed by choosing $a_1=-a_2=0.01$cm and $a_1=-a_2=0.009$cm, respectively . The individual peaks can be well resolved by changing $a_1=-a_2=0.02$cm. The common parameters of two graphs are $G_{0}=1\gamma$, and ${w_c}=100{\mu}$m.](Fig3.eps){width="1.0\columnwidth"}
![\[fig:Fig4\] (Color online) The Spatial variation of the real (Re$[\chi_{41}]$) and imaginary (Im$[\chi_{41}]$) parts of $\chi_{41}$. The plots are shown against the transverse axis coordinate $x$ of the control beam for $y=0$ plane. The different curves are for three different set of parameters: ${\Omega}=0.015{\gamma}$, $r=0$, ${\Delta_1}=0.001{\gamma}$ (red long dashed, and dot-dashed lines); ${\Omega}=0.015{\gamma}$, $r=0.0005\gamma$, ${\Delta_1}=0.001{\gamma}$ (blue dashed double-dot, and dot double-dashed lines), and ${\Omega}=0$, $r=0$, ${\Delta_1}=-0.001{\gamma}$ (black solid, and short dashed lines). The control beam parameters are $G_{0}=1\gamma$, ${w_c}=100{\mu}$m, and $a_1=-a_2=0.012$cm.](Fig4.eps){width="1.0\columnwidth"}
Susceptibility with inhomogeneous control field
-----------------------------------------------
In this section, we discuss the effect of spatial inhomogeneous field on linear susceptibility given in Eq. (\[chi\_41\]). For this purpose, we change the control field profile from cw to spatially inhomogeneous field while keeping rest of the fields as cw for further study. The spatially inhomogeneous transverse profile of the control field is a combination of more than one Gaussian peak. At $z=0$, the control beam can be written as, $$\begin{aligned}
\label{shape_drive1}
G(x,y)=&{G}_0\sum\limits_{i=1}^n \:e^{-\frac{\left [(x-a_i)^2+y^2\right]}{w_c^{2}}}\,,\end{aligned}$$ where, $G_0$ is initial peak amplitude, $w_c$ is beam width and $a_i$ are the individual peak position. The full-width-at-half-maximum (FWHM) of individual peak is $\sqrt{2ln2}w_c$. Figure [\[fig:Fig3\]]{} shows the intensity distribution of the control field against radial position $x$ at the entry face of the medium. The overlapping of two peaks gives rise to a central minimum with non-zero intensity as shown in Fig.\[fig:Fig3\]. The Rayleigh-limited or Sparrow limited control images can be formed when the intensity of the peak normalized central minimum is $I_{min}\sim0.5$ or $\sim 0.7$, respectively. The resolution of the diffraction limited images can be improved by reducing the central minimum intensity to zero. Thus, by increasing the peak separation or by decreasing the width of the individual peak can able to create high resolution image.
The spatially modulated control field perturbs the probe beam susceptibility along the transverse direction as shown in Fig. \[fig:Fig4\]. Fig. \[fig:Fig4\] illustrates the spatial variation of the real and imaginary parts of $\chi_{41}$ as a function of the transverse axis $x$ for $y=0$ plane. The very special inhomogeneous character of dispersion Re$[\chi_{41}]$ and absorption Im$[\chi_{41}]$ causes the spatial modulation in phase and amplitude for the probe field, respectively. Since the phase of probe beam is influenced by the co-propagating control beam, therefore, this phase modulation is termed as cross phase modulation (XPM) [@agrawal]. The mutual coupling between the optical beams is attributed to XPM which causes focusing to the probe beam. The amplitude modulation results in attenuation or gain to the probe beam.
The curves of Fig. \[fig:Fig4\] represent three different cases of EIT, MIA, and LWI, respectively. It is clear from Fig. \[fig:Fig4\] that for MIA and LWI cases two transparency windows are formed at higher intensity regions whereas absorption occurs in relatively low intensity regions of control field $G$ defined by two Gaussian modes using Eq. (\[shape\_drive1\]). The real part of the susceptibility is maximized at these higher intensity regions. This resembles two parallel waveguide like structures with claddings (0.0075 cm$ \gtrsim |x| \gtrsim $ 0.0175 cm) and cores (0.0175 cm$ \gtrsim |x| \gtrsim $ 0.0075 cm). In order to have a perfect wave-guiding, there should be a high contrast between core and cladding. In case of EIT, it is evident from Fig. \[fig:Fig4\] that a single transparency window is formed and the variation in refractive index around $x=0$ is very small. Therefore, the single transparency window is failed to create two parallel waveguide. As a result, EIT is not suitable to separate out the modes with high resolution. However, in the case of MIA, one can see a sharp variation in refractive index (red long dashed line) around $x=0$, with a rapid increase in contrast from core to cladding. But there is reasonable increase in absorption in the region between 0.0175 cm$ \gtrsim |x| \gtrsim $ 0.0075 cm of the doublet compared to EIT. This increment will reduce transmission of the probe beam and therefore, its visibility seems to be restricted.
![ \[fig:Fig6\] (Color online) A comparison study of EIT and DDR with incoherent pump for cloning of the just resolved control images onto the probe beam at the output of the vapor cell with length $L=2.5$ cm. The parameters are same as in Fig.(\[fig:Fig4\])](Fig6.eps){width="1.0\columnwidth"}
Interestingly, in case of LWI, the refractive index contrast between core and cladding is higher than the other two cases. This contrast enhancement causes strong focusing of the probe beam towards the center of the two peaks of the control field. As a result the width of the probe beam becomes narrow which can improve the contrast of the cloned image on the probe field. Also the two deeps of the doublet changes from absorption into gain can produce the enhancement of the cloned beam transmission. Hence the weak probe beam is not only guided or focused but also amplified in order to preserve the information during the propagation through the optical medium. This is the key mechanism of cloning the un-resolvable or just-resolvable control field profile to the probe field with high resolution. In the following, we use the inhomogeneous susceptibility for LWI case to illustrate the improvement of the resolution of the cloned of images of the control field onto the probe field.
Beam propagation dynamics
-------------------------
We numerically integrate the paraxial wave equations (\[probe\]) and (\[control\]) by using a higher order split operator method [@Shen] to study the propagation dynamics of both control and probe beams. First we explore the cloning of Rayleigh limited control beam onto the probe beam in presence of both microwave and incoherent pump fields. For this purpose, we set ${w_c}=100{\mu}$m and $a_1=-a_2=0.01$cm in the Eq.(\[shape\_drive1\]). The results for the spatial evolution of the control and the probe profiles throughout the medium are shown in Fig. \[fig:Fig5\]. It is clear from Fig. \[fig:Fig5a\] that within a very short distance, the control field structure is mapped on to the probe with central minimum reduced to zero. As a result, the finesse, which is the ratio of the spacing between peaks to the width of peaks of the transmitted probe beam at $z=2.5$ cm, is 4 times smaller than initial control beam finesse. We also find that the integrated transmission of the output probe beam at $z=2.5$ cm is about $98\%$. The probe beam transmission can be changed by changing the incoherent pump field rate $r$. Figure \[fig:Fig5b\] depicts the intensity profile of the control beam at different propagation distances $z$. We find that the the shape of the control beam is gradually distorted as it propagates through the medium due to diffraction. As a consequence, control beam induced waveguide structure in the medium is modified. Accordingly the shape of the cloned beam starts experiencing diffraction after $z=2.5$ cm propagation distance as shown in Fig.\[fig:Fig5a\]. Long distance diffractionless cloned image propagation can be achieved by considering tightly focused control beam [@howell_09] or self-reconstructing Bessel control beam[@Fahrbach].
Figure \[fig:Fig6\] compares the cloning mechanism in presence and absence of both microwave and incoherent pump fields. The Rayleigh limited control field structure generated a double transparency window and a single transparency window for DDR and EIT system, respectively. It is clear from Fig. \[fig:Fig6\] that the double transparency window enable to perfectly clone the control image whereas single transparency window failed to clone the control image to the transmitted probe beam. We also notice that the DDR induced waveguide structure can support the propagation of cloned probe beam without any diffraction. In contrast, for EIT case, the transmitted probe beam suffers severe distortion due to lack of parallel waveguide like structure inside the medium. Hence EIT based mechanism has limitation to clone unresolved or just resolved control image onto probe beam without loss of generality.
Next, we demonstrate how the microwave and incoherent pump fields offer the unprecedented control over the image cloning for unresolved images. For this purpose, we consider more complex structure of control beam consisting of three Gaussian peaks. Fig. \[fig:Fig7\] shows the radial distribution of the input Sparrow limited control beam (at z = 0) and output probe beam at z = 1 cm. As in Fig. \[fig:Fig7b\] it can be seen that the cloned probe images contains three distinguishable peaks even though the control beam profile is unresolved. Surprisingly, the integrated transmission intensity of the cloned probe image is approximately 74%. Thus microwave and incoherent pump fields allow one to cloned the diffraction limited control field image onto the probe beam with improved spatial resolution and high transmission. We also verified that the resolution enhancement of cloned images can be possible even for Rayleigh limited control images with the Bessel as well as non-Gaussian shape. These studies may be useful for practical applicability such as optical microscope, quantum metrology and quantum imaging[@application].
Spatial optical switchting
--------------------------
Here, we show how the propagation dynamics of the probe beam can be controlled by switching the microwave field on and off. The well resolved control beam image is being considered for this demonstration. The individual peak has width 100$\mu$m correspond to 4 cm Rayleigh length. The spatially dependent control field assisted atomic waveguide can protect the feature of the cloned beam in a 4 cm long medium. Fig.\[fig:Fig8a\] illustrates that the nondiffracting cloned probe beam propagation is possible inside the medium in both EIT as well as LWI system. We found that the width and the transmission of the cloned beam at $z=3$ cm are 25$\mu$m (100 $\mu$m) and 60% (5%) for LWI (EIT) mechanism. Therefore, the precise control of finesse and the contrast of the output cloned probe beam can be achieved by application of coherent fields and incoherent pump field interacting in a four level atomic medium. Fig.\[fig:Fig8b\] shows how the microwave induced absorption can be utilized to attenuate the probe beam gradually inside the medium in the absence of incoherent pump field. Thus, microwave field which connects the lower level metastable states of four level system can switch off the probe beam propagation inside the medium. This investigation can be applicable for all optical switching and logic gates [@hong; @nie].
\[Conclusion\]Conclusion
========================
In conclusion, we have revealed a scheme to improve the resolution of the cloned image based on the quantum interference effects induced by interacting dark resonances. For this purpose, we have used four levels atomic system interacting with three coherent fields and an incoherent pump field. An atomic wave-guide structure is formed inside the medium by using a spatially modulated control field. The refractive-index contrast between core and cladding of the atomic waveguide can be increased by use of sharp absorption peak associated with double dark resonances. The high contrast atomic wave-guide enables us to imprint the Rayleigh or Sparrow limited control images to probe field with high resolution. The transverse feature of control image is efficiently cast on to the probe field even though the control image suffers distortion due to the diffraction during the propagation. Our numerical result show that the propagation of high resolution cloned image is possible until the feature of the control image lost completely. We use incoherent pump field in order to increase the transmission of the cloned probe image. Finally, we have also demonstrated that spatial optical switching is possible by use of EIT, LWI and MIA mechanism.
[One of the authors (T.N.D) gratefully acknowledge funding by the Science and Engineering Board(SR/S2/LOP-0033/2010).]{}
\[app-A\]Coefficients for susceptibility
========================================
$$\begin{aligned}
{\rho}_{11}^{(0)}=&\frac{(2(r+\gamma)|G|^2\Omega^2({\gamma}({\Gamma}{\gamma}+|G|^2)+{\Gamma}(({{\Delta_2}}+{{\Delta_3}})^2+{\Omega}^2)))}{D}\\
{\rho}_{22}^{(0)}=&(r(|G|^6+G^4(2\Gamma\gamma^2-2\gamma{\Delta_2}{\Delta_3}-2\gamma{\Delta_3}^3+\Gamma\Omega^2-\gamma\Omega^2)+2\Gamma\Omega^2({\Delta_2}^4+2{\Delta_2}^3{\Delta_3}+\gamma^2{\Delta_2}^3+{\Delta_2}^2(2\gamma^2+{\Delta_3}^2-2\Omega^2)\notag\\
&+2{\Delta_2}{\Delta_3}(\gamma^2-\Omega^2)+(\gamma^2+\Omega^2)^2)+|G|^2(\gamma{\Delta_3}^4+(\gamma^2+\Omega^2)({\Gamma}^2\gamma-(\Gamma-2\gamma)\Omega^2)+{\Delta_3}^2(\gamma(\Gamma^2+\gamma^2)\notag\\
&+2(\Gamma+2{\gamma})\Omega^2)+{\Delta_2}^2({\Gamma}^2\gamma+\gamma{\Delta_3}^2+(5\Gamma+2\gamma)\Omega^2)+{\Delta_2}{\Delta_3}(2{\Gamma}^2\gamma+2\gamma{\Delta_3}^2+(7\Gamma+6\gamma)\Omega^2))))/{D}\\
{\rho}_{33}^{(0)}=&(r\Omega^2(2{\Gamma}{\Delta_2}^4+4\Gamma{\Delta_2}^3{\Delta_3}+{\Delta_3}^2(2\Gamma{\gamma}^2+(2\Gamma+\gamma)|G|^2)+{\Delta_2}^2(2\Gamma{\Delta_3}^2+(3\Gamma+2\gamma)G^2+4\Gamma(\gamma^2-\Omega^2))\notag\\
&+{\Delta_2}{\Delta_3}((5\Gamma+2\gamma)|G|^2+4\Gamma(\gamma^2-\Omega^2))+(|G|^2+2({\gamma}^2+\Omega^2))(\gamma|G|^2+\Gamma({\gamma}^2+\Omega^2))))/{D}\\
{\rho}_{44}^{(0)}=&\frac{(2r|G|^2\Omega^2({\gamma}({\Gamma}{\gamma}+|G|^2)+{\Gamma}(({{\Delta_2}}+{{\Delta_3}})^2+{\Omega}^2)))}{D}\\
D=&r\gamma |G|^6 +|G|^4\left(2r \gamma(\Gamma \gamma- {\Delta_3}({\Delta_2}+{\Delta_3}))+(2 \gamma^2+r(\Gamma +4 \gamma))\Omega^2\right)\notag\\
&+4r\Gamma \Omega^2 \left((\gamma^2+{\Delta_2}^2)(\gamma^2 +({\Delta_2}+{\Delta_3})^2)+2(\gamma^2 -{\Delta_2}({\Delta_2}+{\Delta_3}))\Omega^2+\Omega^4\right)\notag\\
&+|G|^2 \left(r\gamma(\Gamma^2+{\Delta_3}^2\right)\left(\gamma^2+({\Delta_2}+{\Delta_3})^2\right)\notag\\
&+(\gamma(2\Gamma\gamma^2 +r(\Gamma+2\gamma)^2)+2 \left(\Gamma\gamma+2r(3\Gamma+\gamma)) {\Delta_2}^2+4(5r \Gamma+2r \gamma + \Gamma \gamma \right){\Delta_2}{\Delta_3}\notag\\
&+(8r\Gamma+5r \gamma +2\Gamma \gamma) {\Delta_3}^2 )\Omega^2+2 (\Gamma\gamma +2r( \gamma +\Gamma)\Omega^4))\,.\end{aligned}$$
[99]{}
Lord Rayleigh, Philos. Mag. [**8**]{}, 261 (1879).
B. A. A. Saleh and M. C. Teich, [*Fundamentals of Photonics*]{} (John Wiley & Sons, New York, 1991).
M. Born and E. Wolf, [*[Principles of Optics]{}*]{} (Cambridge University Press, Cambridge, England, 1999).
E. Arimondo, Progress in Optics, [**35**]{}, 257(1996).
S. E. Harris, Phys. Today [**50(7)**]{}, 36 (1997).
M. Fleischhauer, and A. Imamoglu, and J. P. Marangos, , 663 (2005).
O. Kocharovskaya, Phys. Rep. 219, 175 (1992).
T. W. Hansch, M. D. Levenson, and A. L. Schawlow, Phys. Rev. Lett. 26, 946 (1971).
G. S. Agarwal and T. N. Dey, Laser & Photonics Reviews [**3**]{}, 287 (2009).
A. G. Truscott, M. E. J. Friese, N. R. Heckenberg, and H. Rubinsztein-Dunlop, , 1438 (1999).
R. Kapoor and G. S. Agarwal, , 053818-1-053818-4 (2000).
P. K. Vudyasetu, D. J. Starling, and J. C. Howell, , 123602 (2009).
T. N. Dey, and J. Evers, , 043842 (2011).
R. R. Moseley, S. Shepherd, D. J. Fulton, B. D. Sinclair, and M. H. Dunn, , 670 (1995).
R. R. Moseley, S. Shepherd, D. J. Fulton, B. D. Sinclair, and M. H. Dunn, , 408 (1996).
D. R. Walker, D. D. Yavuz, M. Y. Shverdin, G. Y. Yin, A. V. Sokolov, and S. E. Harris, Opt. Lett. [**27**]{}, 2094 (2002)
N. A. Proite, B. E. Unks, J. T. Green, and D. D. Yavuz, Phys. Rev. A [**77**]{}, 023819 (2008).
M. Mitsunaga, M. Yamashita, and H. Inoue , 013817 (2000).
D. Bortman-Arbiv, A. D. Wilson-Gordon, and H. Friedmann, , 031801(R) (2001).
J. Cheng and S. Han, Opt. Lett. 32, 1162 (2007).
L. Zhang, T. N. Dey, and J. Evers, , 043842 (2013).
O. Firstenberg, M. Shuker, N. Davidson, and A. Ron, , 043601 (2009).
O. Firstenberg, P. London, M. Shuker, A. Ron, and N. Davidson, Nature Phys. [**5**]{}, 665 (2009).
O. Firstenberg, M. Shuker, A. Ron, and N. Davidson, , 941 (2013).
H. Li, V. A. Sautenkov, M. M. Kash, A. V. Sokolov, G. R. Welch, Y. V. Rostovtsev, M. S. Zubairy, and M. O. Scully, , 013803 (2008).
O. N. Verma, L. Zhang, J. Evers, and T. N. Dey, , 013810 (2013).
C. O’Brien, and O. Kocharovskaya, , 137401 (2011).
M. D. Lukin, S. F. Yelin, M. Fleischhauer, and M. O. Scully, , 3225 (1999).
Y. C. Chen, Y. A. Liao, H. Y. Chiu, J. J. Su, and I. A. Yu, , 053806 (2001).
S. F. Yelin, V. A. Sautenkov, M. M. Kash, G. R. Welch, and M. D. Lukin, , 063801 (2003).
C. Y. Ye, A. S. Zibrov, Y. V. Rostovtsev, and M. O. Scully, , 043805 (2002).
H. Wang, and X. Peng, , 429 (2005).
G. P. Agrawal, , 2487 (1990).
A. D. Bandrauk, H. Shen, Journal of Physics A: Mathematical and General [**27**]{}, 7747 (1994).
F. O. Fahrbach and A. Rohrbach, Nature Communications [**3**]{}, 632 (2012)
C. A. Mack, Fundamental Principles of Optical Lithography: the Science of Microfabrication (Wiley, West Sussex, England, 2007); R. T. Glasser, H. Cable, and J. P. Dowling, F. De Martini, F. Sciarrino, and C. Vitelli, , 012339 (2008); V. Giovannetti, S. Lloyd, L. Maccone, and J. H. Shapiro, , 013827 (2009).
Z. Nie, H. Zheng, Y. Zhang, Y. Zhao, C. Zuo, C. Li, H. Chang, and M. Xiao, Opt. Express [**18**]{}, 899 (2010).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper proposes a generic formulation that significantly expedites the training and deployment of image classification models, particularly under the scenarios of many image categories and high feature dimensions. As a defining property, our method represents both the images and learned classifiers using binary hash codes, which are simultaneously learned from the training data. Classifying an image thereby reduces to computing the Hamming distance between the binary codes of the image and classifiers and selecting the class with minimal Hamming distance. Conventionally, compact hash codes are primarily used for accelerating image search. Our work is first of its kind to represent classifiers using binary codes. Specifically, we formulate multi-class image classification as an optimization problem over binary variables. The optimization alternatively proceeds over the binary classifiers and image hash codes. Profiting from the special property of binary codes, we show that the sub-problems can be efficiently solved through either a binary quadratic program (BQP) or linear program. In particular, for attacking the BQP problem, we propose a novel bit-flipping procedure which enjoys high efficacy and local optimality guarantee. Our formulation supports a large family of empirical loss functions and is here instantiated by exponential / hinge losses. Comprehensive evaluations are conducted on several representative image benchmarks. The experiments consistently observe reduced complexities of model training and deployment, without sacrifice of accuracies.'
author:
- |
Fumin Shen$^{\dagger}$, Yadong Mu$^{\ddagger}$, Wei Liu$^{\sharp}$, Yang Yang$^{\dagger}$, Heng Tao Shen$^{\dagger\natural}$\
$^\dagger$ University of Electronic Science and Technology of China\
$^\ddagger$ AT&T Labs Research $^\sharp$ Didi Research, Beijing China $^\natural$ The University of Queensland
bibliography:
- 'FSRef.bib'
title: Learning Binary Codes and Binary Weights for Efficient Classification
---
Introduction
============
In recent years, large-scale visual recognition problem has attracted tremendous research enthusiasm from both academia and industry owing to the explosive increase of data size and feature dimensionality. Classifying an image into thousands of categories often entails heavy computations by using a conventional classifier, exemplified by $k$ nearest neighbor ($k$-NN) and support vector machines (SVM), on a commodity computer. For the image recognition problem with many categories, the computational and memory overhead primarily stems from the large number of classifiers to be learned. The complexities can be high at the stages of both training and deploying these classifier. Considering a classification task with $C$ different classes and $D$-dimensional feature representation, even the simplest linear models are comprised of $D \times C$ parameters. As an inspiring example to our work in this paper, the ImageNet dataset [@deng2009imagenet] contains annotated images from 21,841 classes in total. When experimenting with some state-of-the-art visual features (, 4096-dimensional deep neural networks feature), a huge number of 80 million parameters need to be learned and stored, which clearly indicates slow training and low efficacy at the deployment phase. Real-world applications (such as industrial image search engine) often require near-real-time response. The conventional ways of training multi-class image classifiers thus have much space to be improved.
Compact binary hash codes [@LSH99] have demonstrated notable empirical success in facilitating large-scale similarity-based image search, referred to as image hashing in the literature. In a typical setting of supervised learning, the hash codes are optimized to ensure smaller Hamming distances between images of the same semantic kind. In practice, image hashing techniques have been widely utilized owing to its low memory footprint and theoretically-guaranteed scalability to large data.
Though the hashing techniques for image search has been a well-explored research area, its application on large-scale optimization still remains a nascent topic in the fields of machine learning and computer vision. Intuitively, one can harness the hash codes for the image classification task through naive methods such as $k$-NN voting. Both the training and testing images are indexed with identical hashing functions. A new image is categorized by the majority semantic label within the hashing bucket where it is projected into. However, since the hash codes are initially optimized for image search purpose, such a naive scheme does not guarantee high accuracy for image recognition.
The most relevant works to ours are approximately solving non-linear kernel SVM via hashing-based data representation [@li2011hashing; @li2013sign; @mu2014hash]. These methods first designate a set of hashing functions that transform the original features into binary codes. The original non-linear kernels (, RBF kernel) are theoretically proved to be approximated by the inner product between binary hash bits. Prominent advantages of such a treatment are two-folds: the required hash bits only weakly hinge on the original feature dimensionality, and meanwhile the non-linear optimization problem is converted into a linear alternative. As a major drawback, these works still rely on the regular real-valued based classifiers upon the binary features. Though it enables the direct application of linear solvers, the potential of binary codes is not fully utilized.
Our work is a non-trivial extension of the aforementioned line of research. We further elevate the efficacy of classification models by binarizing both the features and the weights of classifiers. In other words, our goal is to develop a generic multi-class classification framework. The classifier weights and image codes are simultaneously learned. Importantly, both of them are represented by binary hash bits. This way the classification problem is transformed to an equivalent and simpler operation, namely searching the minimal Hamming distance between the query and the $C$ binary weight vectors. This can be extremely fast by using the built-in XOR and popcount operations in modern CPUs. We implement this idea by formulating the problem of minimizing the empirical classification error with purely binary variables.
The major technical contributions of this work are summarized as below:
1. We define a novel problem by binarizing both classifiers and image features and simultaneously learning them in a unified formulation. The prominent goal is to accelerate large-scale image recognition. Our work represents an unexplored research direction, namely extending hashing techniques from fast image search to the new topic of hashing-accelerated image classification.
2. An efficient solver is proposed for the binary optimization problem. We decouple two groups of binary variables (image codes and binary classifier weights) and adopt an alternating-minimizing style procedure. Particularly, we show that the sub-problems are in the form of either binary quadratic program (BQP) or linear program. An efficient bit-flipping scheme is designed for the BQP sub-problem. Profiting from the special traits of binary variables, we are able to specify the local optimality condition of the BQP.
3. Our formulation supports a large family of empirical loss functions and is here instantiated by exponential / hinge losses. In our quantitative evaluations, both variants are compared with key competing algorithms, particulary a highly-optimized implementation of the SVM algorithm known as LibLinear [@liblinear08]. Our proposed method demonstrates significant superiority in terms of train/test CPU time and the classification accuracy, as briefly depicted by Figure \[fig:acc\_sun\].
[0.23]{} ![Comparison of our method with the LibLinear implementation of Linear SVM for classification on the SUN dataset with 108K images from 397 scene categories. By coding both the image feature and learned classifiers with a small number of hash bits, our method achieves better results than Linear SVM, even with much smaller model training complexity. []{data-label="fig:acc_sun"}](ac.pdf "fig:"){width="\textwidth"}
[0.23]{} ![Comparison of our method with the LibLinear implementation of Linear SVM for classification on the SUN dataset with 108K images from 397 scene categories. By coding both the image feature and learned classifiers with a small number of hash bits, our method achieves better results than Linear SVM, even with much smaller model training complexity. []{data-label="fig:acc_sun"}](tm.pdf "fig:"){width="\textwidth"}
Related work {#Sec:related}
============
Let us first review the related works which strongly motivate ours. They can be roughly cast into two categories:
**Hashing for fast image search and beyond**: learning compact hash codes [@ICML13SHEN; @SH08] recently becomes a hot research topic in the computer vision community. The proliferation of digital photography has made billion-scale image collections a reality, and efficiently searching a similar image to the query is a critical operation in such image collections.
The seminal work of LSH [@LSH99] sheds a light on fast image search with theoretic guarantee. In a typical pipeline of hash code based image search [@SSH2012; @xia2015sparse], a set of hashing functions are generated either in an unsupervised manner or learned by perfectly separating similar / dissimilar data pairs on the training set. The latter is also called supervised hashing since it often judges the similarity of two samples according to their semantic labels. For an unseen image, it finds the most similar images in the database by efficiently comparing the corresponding hash codes. It can be accomplished in sub-linear time using hash buckets [@LSH99]. Representative methods in this line include Binary Reconstructive Embedding (BRE [@BRE2009]), Minimal Loss Hashing (MLH [@MLH2011]), Kernel-Based Supervised Hashing (KSH [@KSH2012]), CCA based Iterative Quantization (CCA-ITQ [@gong2013iterative]), FastHash [@lin2014fast], Graph Cuts Coding (GCC [@ge2014graph]), .
The success of hashing technique is indeed beyond fast image search. For example, Dean et al. [@Dean13] used hash tables to accelerate the dot-product convolutional kernel operator in large-scale object detection, achieving a speed-up of approximately 20,000 times when evaluating 100,000 deformable-part models.
**Hashing for large-scale optimization**: Noting the Hamming distance is capable of faithfully preserve data similarity, it becomes a natural thought to extend it for approximating non-linear kernels. Optimizing with non-linear kernels generally require more space to store the entire kernel matrix, which prohibits its scalability to large data. Real vectors based explicit feature mapping [@vedaldi2012efficient] partially remedies above issue, approximating kernel functions by the inner product between real vectors. However, they typically require high dimension towards an accurate approximation, and is thus beyond the scope of most practitioners. A more recent strand of research instead approximates non-linear kernel with binary bits, of which the prime examples can be found in [@mu2014hash; @li2013sign; @li2011hashing]. In particular, Mu et al. developed a random subspace projection which transforms original data into compact hash bits. The inner product of hash code essentially plays the role of kernel functions. Consequently, the non-linear kernel SVM as their problem of interest can be converted into a linear SVM and resorts to efficient linear solvers like LibLinear [@liblinear08].
The philosophy underlying all aforementioned works can be summarized as binarizing the features and harnessing the compactness of binary code. We here argue that the potential of hashing technique in the context of large-scale classification has not been fully explored yet. Related research is still in its embryonic stage. For example, a recent work by Shen et al. [@SDH15] proposed a Supervised Discrete Hashing (SDH) method under the assumption that good hash codes were optimal for linear classification. However, similar to other methods, SDH still classified the learned binary codes by real-valued weights. Thus the test efficiency for binary codes is still not improved compared to real-valued features.
After surveying related literature, we are motivated to advocate in this paper an extreme binary learning model, wherein both image features and classifiers shall be represented by binary bits. This way the learned models get rid of real-valued weight vectors and can fully benefit from high-optimized hash bit operators such as XOR.
The proposed model {#Sec:prop}
==================
Suppose that we have generated a set of binary codes ${{\mathbf{B}}}= \{{{\mathbf{b}}}_i\}_{i=1}^n \in \{-1,1\}^{r \times n}$, where ${{\mathbf{b}}}_i$ is the $r$-bit binary code for original data ${{\mathbf{x}}}_i$ from the training set ${{\mathbf{X}}}= \{{{\mathbf{x}}}_i\}_{i=1}^n$. For simplicity, we assume a linear hash function $$\begin{aligned}
h({{\mathbf{x}}}) = {{\mathrm{sgn}}}({{\mathbf{P}}}^\top {{\mathbf{x}}}),\end{aligned}$$ where ${{\mathbf{P}}}\in \mathbb{R}^{d \times r}$.
In the context of linear classification, the binary codes ${{\mathbf{b}}}$ is classified according to the maximum of the score vector $${{\mathbf{W}}}^{\T}{{\mathbf{b}}}= [{{\mathbf{w}}}_1^{\T}{{\mathbf{b}}}, \cdots,
{{\mathbf{w}}}_C^{\T}{{\mathbf{b}}}]^{\T},$$ where ${{\mathbf{w}}}_c \in \{-1,1\}^{r}$ is the binary parameter vector for class $c \in [1, \cdots, C]$. Taking advantage of the binary nature of both ${{\mathbf{w}}}_c$ and ${{\mathbf{b}}}$, the inner product ${{\mathbf{w}}}_c^\top {{\mathbf{b}}}$ can be efficiently computed by $r - 2\mathbb{D_H}({{\mathbf{w}}}_c,{{\mathbf{b}}})$, where $\mathbb{D_H}(\cdot,\cdot)$ is the Hamming distance. Thereby the standard classification problem is transformed to searching the minimum from $C$ Hamming distances (or equivalently the maximum of binary code inner products).
Following above intuition, this paper proposes a multi-class classification framework, simultaneously learning the binary feature codes and classifier. Suppose ${{\mathbf{b}}}_i$ is the binary code of sample ${{\mathbf{x}}}_i$ and it shall be categorized as class $c_i$. Ideally, it expects the smallest Hamming distance (or largest inner product) to ${{\mathbf{w}}}_{c_i}$, in comparison with other classifier ${{\mathbf{w}}}_c$, $c \neq c_i$. An intuitive way of achieving this is through optimizing the inter-class “margin". Formally, we can minimize the loss $\ell \big( -({{\mathbf{w}}}_{c_i}^\top {{\mathbf{b}}}_i - {{\mathbf{w}}}_c^\top {{\mathbf{b}}}_i) \big)$, $\forall c$, where $\ell(\cdot)$ is a generic loss function. We re-formulate multi-class classification problem as below: $$\begin{aligned}
\label{eqn:obj_general}
\min_{{{\mathbf{W}}}, {{\mathbf{B}}}} & \quad \sum_{i=1}^n \sum_{c=1}^C \ell \big( -({{\mathbf{w}}}_{c_i}^\top {{\mathbf{b}}}_i - {{\mathbf{w}}}_c^\top {{\mathbf{b}}}_i) \big) \\
{{\,\,\mathrm{s.t.\,\,}}}& \quad {{\mathbf{b}}}_i \in \{-1,1\}^r, ~ \forall i, \;{{\mathbf{w}}}_c \in \{-1,1\}^r, ~ \forall c. \nonumber\end{aligned}$$
We instantiate $\ell(\cdot)$ with the exponential loss (Section \[Sec:exploss\]) and hinge loss (Section \[Sec:hingeloss\]). In fact, the loss function in Problem can be broadly defined. Any proper loss function $\ell(\cdot)$ can be applied as long as it is monotonically increasing.
Learning with exponential loss {#Sec:exploss}
------------------------------
Using the exponential loss function, we have the formulation below: $$\begin{aligned}
\label{eqn:obj}
\min_{{{\mathbf{W}}}, {{\mathbf{B}}}} & \sum_{i=1}^n \sum_{c=1}^C \exp \left[ -({{\mathbf{w}}}_{c_i}^\top \mathbf{b}_i - {{\mathbf{w}}}_c^\top \mathbf{{{\mathbf{b}}}}_i) \right] \\
{{\,\,\mathrm{s.t.\,\,}}}& \mathbf{{{\mathbf{b}}}}_i \in \{-1,1\}^r,~ \forall i,\;
{{\mathbf{w}}}_c \in \{-1,1\}^r, ~\forall c. \nonumber\end{aligned}$$
We tackle problem by alternatively solving the two sub-problems with ${{\mathbf{W}}}$ and ${{\mathbf{B}}}$, respectively.
### Classifying binary codes with binary Weights {#SEC:W}
Assume ${{\mathbf{B}}}$ is known. We iteratively update ${{\mathbf{W}}}$ row by row, i.e., one bit each time for ${{\mathbf{w}}}_c, c = 1, \cdots, C$, while keep all other $r-1$ bits fixed. Let ${{\mathbf{w}}}(k)$ denote the $k_{\mathrm{th}}$ entry of ${{\mathbf{w}}}$ and ${{\mathbf{w}}}(\setminus k)$ the vector which zeros its $k$-th element. We then have $$\begin{aligned}
\label{EQ:decom}
\exp({{\mathbf{w}}}^\top {{\mathbf{b}}}) =& \exp\left[ {{\mathbf{w}}}(\setminus k)^\top {{\mathbf{b}}}+ {{\mathbf{w}}}(k){{\mathbf{b}}}(k)\right]\nonumber\\
=& \exp\left[ {{\mathbf{w}}}(\setminus k)^\top {{\mathbf{b}}}\right] \cdot \exp\left[{{\mathbf{w}}}(k){{\mathbf{b}}}(k)\right].\end{aligned}$$ It can be verified that $$\begin{aligned}
\label{EQ:linear1}
&& \exp\left[{{\mathbf{w}}}(k){{\mathbf{b}}}(k)\right] \nonumber \\
&=& \left\{ \begin{array}{rc}
\frac{e^{-1}+e^1}{2} + \frac{e^{-1}-e^1}{2} \cdot {{\mathbf{w}}}(k), & {{\mathbf{b}}}(k) = -1\\
\frac{e^{-1}+e^1}{2} - \frac{e^{-1}-e^1}{2} \cdot {{\mathbf{w}}}(k), &{{\mathbf{b}}}(k) = 1.
\end{array}
\right.\end{aligned}$$
Denote $u = \frac{e^{-1}+e^1}{2}$ and $v = \frac{e^{-1}-e^1}{2}$. Equation can be simplified by the sign of ${{\mathbf{b}}}(k)$ as $$\label{EQ:linear2}
\exp\left[{{\mathbf{w}}}(k){{\mathbf{b}}}(k)\right] = u - v \cdot {{\mathbf{b}}}(k){{\mathbf{w}}}(k).$$ Equation clearly shows the exponential function of the product of two binary bits equals to a linear function of the product. By applying and , we write the loss term in as follows. $$\begin{aligned}
\label{EQ:loss}
&\exp \left[ -({{\mathbf{w}}}_{c_i}^\top \mathbf{b}_i - {{\mathbf{w}}}_c^\top \mathbf{{{\mathbf{b}}}}_i) \right] \\
=& \exp({{\mathbf{w}}}_c^\top \mathbf{{{\mathbf{b}}}}_i) \cdot \exp(-{{\mathbf{w}}}_{c_i}^\top \mathbf{b}_i) \nonumber\\
=& \gamma_{ick} \cdot \left[u - v \cdot {{\mathbf{b}}}_i(k){{\mathbf{w}}}_c(k))\right] \cdot \left[u + v \cdot {{\mathbf{b}}}_i(k){{\mathbf{w}}}_{c_i}(k))\right],\nonumber\end{aligned}$$ where the constant $\gamma_{ick} = \exp\left[ \big({{\mathbf{w}}}_c(\setminus k) - {{\mathbf{w}}}_{c_i}(\setminus k)\big)^\top {{\mathbf{b}}}_i \right]$. Clearly, the non-linear exponential loss becomes a quadratic polynomial with regard to ${{\mathbf{w}}}_c(k)$.
After merging terms with the same orders, optimizing problem with regard to ${{\mathbf{w}}}^k= [{{\mathbf{w}}}_1(k); \cdots; {{\mathbf{w}}}_C(k)]$ becomes $$\begin{aligned}
\label{BQ}
{{\mathbf{w}}}^k \leftarrow & \arg \min_{{{\mathbf{w}}}^k} \; \frac{1}{2} {{\mathbf{w}}}^{k^\top } {{\mathbf{H}}}^k {{\mathbf{w}}}^k + {{\mathbf{w}}}^{k\top}{{\mathbf{g}}}^k,\end{aligned}$$ where $$\begin{array}{rcl}
{{\mathbf{H}}}^k &=& -2v^2 {{\mathbf{Y}}}\mathbf{\Gamma}^k, \\
{{\mathbf{g}}}^k &=& uv {{\mathbf{Y}}}({{\mathbf{b}}}^k \odot \mathbf{\Gamma} {{\mathbf{1}}}) - uv \Gamma^\top {{\mathbf{b}}}^k.
\end{array}$$ Here $\mathbf{\Gamma}^{k} \in \mathbb{R}^{n \times C}$ includes its entries $\gamma_{ick}$, ${{\mathbf{b}}}^k$ is the $n$-dimension vector including the $k_{\mathrm{th}}$ binary bits of training data. ${{\mathbf{Y}}}\in \mathbb{R}^{C \times n}$ is the label matrix whose entry $y_{ci}$ at coordinate $(c,i)$ equals to 1 if sample ${{\mathbf{x}}}_i$ belongs to class $c$ and 0 otherwise. $\odot$ denotes the element-wise product. ${{\mathbf{1}}}$ is the vector with all ones.
Problem is a binary quadratic program (BQP) and can be efficiently solved by sequential bit flipping operation. A local optimum can be guaranteed. We solve problem by investigating the local optimality condition. Intuitively, for any local optimum, flipping any of its hash bits will not decrease the objective value of problem . Let ${{\mathbf{H}}}_{*,c}, {{\mathbf{H}}}_{c,*}$ denote the column or row vector indexed by $c$ respectively. ${{\mathbf{g}}}(c)$ and ${{\mathbf{w}}}(c)$ represents the $c_{\mathrm{th}}$ element of ${{\mathbf{g}}}$ and ${{\mathbf{w}}}$, respectively. In problem , collecting all terms pertaining to ${{\mathbf{w}}}(c)$ obtains $$\begin{aligned}
f({{\mathbf{w}}}^k(c)) &= \frac{1}{2} \left( {{\mathbf{H}}}^{k\top}_{*,c} + {{\mathbf{H}}}^{k}_{c,*} \right) {{\mathbf{w}}}^k \cdot {{\mathbf{w}}}^k(c) + {{\mathbf{g}}}^k(c) \cdot {{\mathbf{w}}}^k(c). \nonumber \end{aligned}$$ By flipping ${{\mathbf{w}}}(c)$, the gain of the objective function of problem is $$\begin{aligned}
\Delta_{{{\mathbf{w}}}^k(c)\rightarrow -{{\mathbf{w}}}^k(c)} = f(-{{\mathbf{w}}}^k(c)) - f({{\mathbf{w}}}^k(c)).\end{aligned}$$
Regarding the local optimality condition of problem , we have the observation below:
\[thm:localopt\] **(Local optimality condition)**: Let ${{\mathbf{w}}}^\ast$ be a solution of problem . ${{\mathbf{w}}}^\ast$ is a local optimum when the condition $\Delta_{{{\mathbf{w}}}(c)\rightarrow -{{\mathbf{w}}}(c)} \ge 0$ holds for $c=1,\ldots,C$.
The conclusion holds by a simple application of proof of contradiction. Recall that ${{\mathbf{w}}}$ is a binary vector. Flipping any bit of ${{\mathbf{w}}}$ will incur specific change of the objective function as described by $\Delta_{{{\mathbf{w}}}(i)\rightarrow -{{\mathbf{w}}}(i)}$. When the changes incurred by all these flipping operations are all non-negative, ${{\mathbf{w}}}^\ast$ is supposed to be locally optimal. Otherwise, we can flip the bit with negative $\Delta_{{{\mathbf{w}}}(i)\rightarrow -{{\mathbf{w}}}(i)}$ to further optimize the objective function.
With the above analysis, we summarize our algorithm for updating the $C$-bit ${{\mathbf{w}}}^k$ as in Algorithm \[alg:bitflip\].
Calculate the bit-flipping gain $\Delta_{{{\mathbf{w}}}(c)\rightarrow -\w(c)}$ for $c=1,\ldots,C$; Select $\hat c = \arg \min_c \Delta_{{{\mathbf{w}}}(c)\rightarrow -{{\mathbf{w}}}(c)}$ and $\Delta_{min} = \min_c \Delta_{{{\mathbf{w}}}(k)\rightarrow -{{\mathbf{w}}}(c)}$; Set ${{\mathbf{w}}}(c) \leftarrow -{{\mathbf{w}}}(c)$; Exit;
### Binary code learning
Similar as the optimization procedure for ${{\mathbf{W}}}$, we solve for ${{\mathbf{B}}}$ by a coordinate descent scheme. In particular, at each iteration, all the rest $r-1$ hash bits are fixed except for the $k$-th hash bit ${{\mathbf{b}}}^k = \left[{{\mathbf{b}}}_1(k); \cdots; {{\mathbf{b}}}_n(k)\right]$. Let ${{\mathbf{b}}}_i(\setminus k)$ denote the vector which zeros its $k$-th element ${{\mathbf{b}}}_i(k)$. We rewrite equation w.r.t ${{\mathbf{b}}}_i(k)$ as $$\begin{aligned}
& \exp \left[ -(\w_{c_i}^\top \mathbf{b}_i - \w_c^\top \mathbf{b}_i)\right] \\
=& z_{ick} \cdot \exp \left[ ( \w_c(k) - \w_{c_i}(k)) \mathbf{b}_i(k) \right] \nonumber \end{aligned}$$ where $z_{ick} = \exp \left[ (\w_c - \w_{c_i})^\top \mathbf{b}_i(\setminus k) \right] $.
Similar as , we have $$\begin{aligned}
&& \exp \left[ ( \w_c(k) - \w_{c_i}(k)) \mathbf{b}_i(k) \right] \nonumber \\
&=& \left\{ \begin{array}{rc}
0, & \w_{c_i}(k) = \w_c(k) \\
\frac{e^{-2}+e^2}{2} + \frac{e^{-2}-e^2}{2} \cdot \mathbf{b}_i(k), & \w_{c_i}(k) = 1,\w_c(k)=-1 \\
\frac{e^{-2}+e^2}{2} - \frac{e^{-2}-e^2}{2} \cdot \mathbf{b}_i(k), & \w_{c_i}(k) = -1,\w_c(k)=1.
\end{array}
\right. \nonumber\end{aligned}$$ We can see that, the non-linear exponential loss term becomes either a constant or linear function with regard to $\mathbf{b}_i(k)$.
Denote $u' = \frac{e^{-2}+e^2}{2}$, $v' = \frac{e^{-2}-e^2}{2}$. Let matrix $\mathbf{Z}^k \in \mathbb{R}^{n \times C}$ include its entry at the coordinate $(i,c)$ as $z_{ick}$ if $\w_{c_i}(k) = 1,\w_c(k)=-1$ and 0 otherwise; similarly let matrix $\mathbf{\bar{Z}}^k \in \mathbb{R}^{n \times C}$ include its entry at the coordinate $(i,c)$ as $z_{ick}$ if $\w_{c_i}(k) = -1,\w_c(k)=1$ and 0 otherwise. Then the loss in can be written as w.r.t ${{\mathbf{b}}}^k$ $$\begin{aligned}
\label{P}
&\sum_{i=1}^n \sum_{c=1}^C \exp \left[ -({{\mathbf{w}}}_{c_i}^\top \mathbf{b}_i - {{\mathbf{w}}}_c^\top \mathbf{{{\mathbf{b}}}}_i) \right] \\
=&\sum_{i=1}^n \sum_{c=1}^C \mathbf{Z}^k (i,c) \cdot (u'+v'{{\mathbf{b}}}_i(k)) + \mathbf{\bar{Z}}^k (i,c) \cdot (u'-v'{{\mathbf{b}}}_i(k))\nonumber\\
=& \sum_{i=1}^n \mathbf{z}^k(i) \cdot (u'+v'{{\mathbf{b}}}_i(k)) + \mathbf{\bar{z}}^k(i)\cdot(u'-v'{{\mathbf{b}}}_i(k)),\end{aligned}$$ where ${{\mathbf{z}}}^k(i) = \sum_{c=1}^C {{\mathbf{Z}}}^k (i,c)$ and $\bar{{{\mathbf{z}}}}^k(i) = \sum_{c=1}^C \bar{{{\mathbf{Z}}}}^k (i,c)$.
Then we have the following optimization problem $$\begin{aligned}
\min_{{{\mathbf{b}}}^k} & \; v' ({{\mathbf{z}}}^k - \bar{{{\mathbf{z}}}}^k)^\top {{\mathbf{b}}}^k \nonumber\\
{{\,\,\mathrm{s.t.\,\,}}}& \; {{\mathbf{b}}}^k \in \{-1,1\}^r,\end{aligned}$$ which has a optimal solution ${{\mathbf{b}}}^k = -{{\mathrm{sgn}}}(v' ({{\mathbf{z}}}^k - \bar{{{\mathbf{z}}}}^k))$.
[0.23]{} ![Object value as a function of bit updating iteration $k$ in optimizing ${{\mathbf{W}}}$ and ${{\mathbf{B}}}$ on the dataset of SUN397.[]{data-label="fig:obj"}](obj_w.pdf "fig:"){width="100.00000%"}
[0.23]{} ![Object value as a function of bit updating iteration $k$ in optimizing ${{\mathbf{W}}}$ and ${{\mathbf{B}}}$ on the dataset of SUN397.[]{data-label="fig:obj"}](obj_b.pdf "fig:"){width="100.00000%"}
Figure \[fig:obj\] shows the objective value as a function of the bit updating iteration number. As can be seen, with the proposed coordinate descent optimizing procedure for both ${{\mathbf{W}}}$ and ${{\mathbf{B}}}$, the object value consistently decreases as updating the hash bits in each sub-problem. The optimization for the original problem typically converges in less than 3 iterations of alternatively optimizing ${{\mathbf{W}}}$ and ${{\mathbf{B}}}$.
Learning with the hinge loss {#Sec:hingeloss}
----------------------------
With the widely used hinge loss, we obtain $$\begin{aligned}
\label{EQ:svm}
\min_{{{\mathbf{B}}},{{\mathbf{W}}},\mathbb{\xi}} &\quad \lambda ||{{\mathbf{W}}}||^2 + \sum_{i=1}^n \sum_{c=1}^C \xi_{i} \\
{{\,\,\mathrm{s.t.\,\,}}}& \quad \forall i,c \quad {{\mathbf{w}}}_{c_i}^{\T}{{\mathbf{b}}}_i + y_{ci} - {{\mathbf{w}}}_c^{\T}{{\mathbf{b}}}_i \geq 1 - \xi_{i},\nonumber\\
&\quad {{\mathbf{b}}}_i = h({{\mathbf{x}}}_i), ~ \forall i, \;\w_c \in \{-1,1\}^r, ~ \forall c. \nonumber\end{aligned}$$ Here $\xi_{i} \geq 0$ is the slack variable; ${{\mathbf{I}}}_r$ is the $r \times r$ identity matrix. We denote $y_{ci} = 1$ if ${{\mathbf{x}}}_i$ belongs to class $c$ and $-1$ otherwise. The non-negative constraint $\xi_{i} \ge 0$ always holds: If $\xi_{i} < 0$ for a particular ${{\mathbf{b}}}_i$, the first constraint in will be violated when $y_{ci} = 1$. In other words, the constraint corresponding to the case the label of ${{\mathbf{b}}}_i$ equals to $c$ ensures the non-negativeness of $\xi_{i}$. This formulation is similar to the multi-class SVM [@crammer2002algorithmic] except that it is exposed to the binary constraints of both input features and classification matrix.
Similar as with the exponential loss, we tackle problem by alternatively solving the two sub-problems regarding to ${{\mathbf{W}}}$ and ${{\mathbf{B}}}$, respectively. Observing $||{{\mathbf{W}}}||^2$ is constant, we can write problem as w.r.t. ${{\mathbf{W}}}$,
$$\begin{aligned}
\label{EQ:W1}
\min_{{{\mathbf{W}}}} &\quad \sum_{i=1}^n \sum_{c=1}^C ( {{\mathbf{w}}}_c^{\T}{{\mathbf{b}}}_i -{{\mathbf{w}}}_{c_i}^{\T}{{\mathbf{b}}}_i)\\
{{\,\,\mathrm{s.t.\,\,}}}&\quad {{\mathbf{w}}}_c \in \{-1,1\}^r, \forall c. \nonumber\end{aligned}$$
Collecting all terms with ${{\mathbf{w}}}_c,\;\forall c$, problem writes $$\begin{aligned}
\label{EQ:Hinge3}
\min_{{{\mathbf{W}}}} &\quad \sum_{c=1}^C {{\mathbf{w}}}_c^{\T}\big(\sum_{i=1}^n{{\mathbf{b}}}_i - C\sum_{i=1, c_i=c}^n{{\mathbf{b}}}_i\big) \\
{{\,\,\mathrm{s.t.\,\,}}}& \quad {{\mathbf{w}}}_c \in \{-1,1\}^r, \forall c, \nonumber\end{aligned}$$
which has optimal solution $$\begin{aligned}
\label{EQ:W}
{{\mathbf{w}}}_c &= {{\mathrm{sgn}}}\big( C \sum_{i=1, c_i=c}^n{{\mathbf{b}}}_i - \sum_{i=1}^n{{\mathbf{b}}}_i \big), \forall c.\end{aligned}$$
For the sub-problem regarding to ${{\mathbf{B}}}$, we first let matrix ${{\mathbf{W}}}^o$ of size $r \times n$ include its $i_{\mathrm{th}}$ column as ${{\mathbf{w}}}^o_i = \sum_{c=1}^C {{\mathbf{w}}}_c - C{{\mathbf{w}}}_{c_i}$. Problem writes w.r.t. ${{\mathbf{B}}}$ as $$\begin{aligned}
\label{EQ:B1}
\min_{{{\mathbf{B}}}} &\; {{\mathrm{trace}}}({{\mathbf{W}}}^{o\T}{{\mathbf{B}}})\\
{{\,\,\mathrm{s.t.\,\,}}}& \; {{\mathbf{B}}}\in \{-1,1\}^{r \times n}. \nonumber\end{aligned}$$ ${{\mathbf{B}}}$ can be efficiently computed by $-{{\mathrm{sgn}}}({{\mathbf{W}}}^{o})$. It is clear that, both of the two sub-problems associated with ${{\mathbf{B}}}$ and ${{\mathbf{W}}}$ have closed-form solutions, which significantly reduce the computation overhead of classifier training and binary code learning.
Binary code prediction
----------------------
With the binary codes ${{\mathbf{B}}}$ for training data ${{\mathbf{X}}}$ obtained, the hash function $h({{\mathbf{x}}}) = {{\mathrm{sgn}}}({{\mathbf{P}}}^{\T}{{\mathbf{x}}})$ is obtained by solving a simple linear regression system $${{\mathbf{P}}}= ({{\mathbf{X}}}^{\T}{{\mathbf{X}}})^{-1}{{\mathbf{X}}}^{\T}{{\mathbf{B}}}.$$ Then for a new sample ${{\mathbf{x}}}$, the classification is conducted by searching the minimum of $C$ Hamming distances: $$c^{\ast} = \arg \min_c \{\mathbb{D_H}\big({{\mathbf{w}}}_c, h({{\mathbf{x}}})\big)\},\; c = 1, \cdots, C.$$
The binary coding step occupies the main computation in the testing stage, which is $O(dr)$ in time complexity.
Experiments {#Sec:exp}
===========
------------------ ----------- ------------ ----------- ----------- ------------ ----------- ---------- ------------ -----------
acc (%) train time test time acc (%) train time test time acc (%) train time test time
OVA-SVM **77.39** 818.87 1.55e-5 **79.84** 151.02 1.15e-5 57.1 55.17 4.02e-7
Multi-SVM 75.28 380.94 1.01e-5 79.48 93.12 1.21e-5 [57.7]{} 35.55 4.27e-7
LSH 54.11 417.42 7.75e-6 58.16 107.41 1.32e-5 39.0 39.64 2.26e-6
CCA-ITQ 69.33 452.34 8.78e-6 76.30 142.95 1.25e-5 56.4 47.23 2.35e-6
SDH 72.56 2522.33 7.43e-6 76.64 1102.21 1.43e-5 55.3 115.63 2.32e-6
Ours-Exponential 75.44 772.11 3.67e-6 **79.04** 245.14 6.54e-6 **59.2** 16.01 2.01e-6
Ours-Hinge **76.56** **16.45** 3.86e-6 77.88 **35.16** 6.86e-6 54.2 **2.31** 2.13e-6
------------------ ----------- ------------ ----------- ----------- ------------ ----------- ---------- ------------ -----------
\[tab:3data\]
In this section, we evaluate the proposed two methods on three large-scale datasets: CIFAR-10[^1], SUN397 [@xiao2010sun] and ImageNet [@deng2009imagenet]. As a subset of the well-known 80M tiny image collection [@80Mtiny2008], the CIFAR-10 dataset consists of 60,000 images which are manually labelled as 10 classes with $6,000$ samples for each class. We represent each image in this dataset by a GIST feature vector [@GIST2001] of dimension $512$. The whole dataset is split into a test set with $1,000$ samples and a training set with all remaining samples. SUN397 [@xiao2010sun] contains about 108K images from 397 scene categories, where each image is represented by a 1,600-dimensional feature vector extracted by PCA from 12,288-dimensional Deep Convolutional Activation Features [@gong2014multi]. We use a subset of this dataset including 42 categories with each containing more than 500 images; 100 images are sampled uniformly randomly from each category to form a test set of 4,200 images. As a subset of ImageNet [@deng2009imagenet], the large dataset ILSVRC 2012 contains over 1.2 million images of totally 1,000 categories. We form the evaluation database by the 100 largest classes with total 128K images from training set, and 50,000 images from validation set as test set. We use the 4096-dimensional features extracted by the convolution neural networks (CNN) model [@krizhevsky2012imagenet].
The proposed methods (denoted by Ours-Exponential and Ours-Hinge for the exponential and hinge loss, respectively) are extensively compared with two popular linear classifiers: one-vs-all linear SVM (OVA-SVM) and multi-class SVM (Multi-SVM [@crammer2002algorithmic]), both of which are implemented using the LibLinear software package [@liblinear08]. For these two methods, we tune the parameter $c$ from the range \[1e-3, 1e3\] and the best results are reported. We also compare our methods against several state-of-the-art binary code learning methods including Locality Sensitive Hashing (LSH) implemented by signed random projections, CCA-ITQ [@gong2013iterative], and SDH [@SDH15]. The classification results of these hashing methods are obtained by performing the multi-class linear SVM over the predicted binary codes by the corresponding hash functions. We use the public codes and suggested parameters of these methods from the authors. We extensively evaluate these compared methods in terms of storage memory overhead, classification accuracy and computation time.
Accuracy and computational efficiency
-------------------------------------
In this part, we extensively evaluate the proposed two methods with the compared algorithms in both classification accuracy and computation time. We use 128-bit for our method and the three hashing algorithms LSH, CCA-ITQ and SDH.
We report the results on [SUN397]{}, [ImageNet]{} and [CIFAR-10]{} in Table \[tab:3data\]. Between the algorithms with different loss functions of our proposed model, we can see that the method with the exponential loss performs slightly better than that with the hinge loss, while the latter one benefits from a much more efficient training. This is not surprising because Ours-Hinge solves two alternating sub-problems both with closed-form solutions.
Compared to other methods, the results clearly show that, our methods achieve competitive (or even better) classification accuracies on all three large-scale datasets with the state-of-the-art linear SVMs. In the meanwhile, even being constrained to learn binary classification weights, our methods obtain much better results than the compared hashing algorithms. Specifically, Ours-Exponential outperforms the best results obtained the hashing algorithms by 2.88%, 2.4% and 2.8% on SUN397, ImageNet and CIFAR-10, respectively.
In terms of training time, we can see that our method with the hinge loss runs way faster than all other methods on all the evaluated three datasets. In particular, on SUN397, Ours-Hinge is $50\times$ and $23\times$ faster than the LibLinear implementations of one-vs-all SVM and mult-class SVM. Compared with the hashing algorithm, our method runs over $28\times$ faster than the fastest LSH followed by Liblinear. For testing time, the benefit of binary dimension reduction for our methods together with other three hashing algorithms is clearly shown on the SUN397 dataset with a large number of categories. Our methods requires less testing time than the hashing based classification methods, which is due to the extremely fast classification implemented by searching in the Hamming space.
We also evaluate the compared methods with different code lengths, where the detailed results on SUN397 and ImageNet are shown in Figure \[fig:acc\_two\]. From Figure \[fig:acc\_two\], it is clear that with a relatively small number of bits (, 256 bits), our method can achieve close classification performance to or even better than the real-valued linear SVM with real-valued features. We can also observe that our method consistently outperforms other hashing algorithms by noticeable gaps at all code lengths on SUN397. On the ImageNet dataset, our method achieves marginally better results than SDH and CCA-ITQ, while much better than the random LSH algorithm.
[0.3]{} ![Comparative results of various algorithms in classification accuracy on SUN397 and ImageNet.[]{data-label="fig:acc_two"}](acc_sun397.pdf "fig:"){width="100.00000%"}
[0.3]{} ![Comparative results of various algorithms in classification accuracy on SUN397 and ImageNet.[]{data-label="fig:acc_two"}](acc_imagenet100.pdf "fig:"){width="100.00000%"}
[0.3]{} ![Comparative results of various algorithms in training time on SUN397 and ImageNet.[]{data-label="fig:time_two"}](traintime_sun3971.pdf "fig:"){width="100.00000%"}
[0.3]{} ![Comparative results of various algorithms in training time on SUN397 and ImageNet.[]{data-label="fig:time_two"}](traintime_imagenet1001.pdf "fig:"){width="100.00000%"}
Figure \[fig:time\_two\] demonstrates the consumed training time by our method and Linear SVM by the Liblinear solver on two large-scale datasets. The computation efficiency advantage of our method is clearly shown. Our method has a nearly linear training time complexity with the code length, which can facilitate its potentially applications in high dimensional binary vector learning.
---------- ------------ ------- ------------ ------
Linear SVM Ours Linear SVM Ours
ImageNet 3943.91 24.37 30.86 9.92
SUN397 1283.83 2.01 4.79 1.54
---------- ------------ ------- ------------ ------
: Memory overhead (MB) to store the training features and classification model using Linear SVM and our method (128-bit). Note that, for our method, the trained model includes the real-valued hashing matrix (${{\mathbf{P}}}$) and the binary classification weight matrix ${{\mathbf{W}}}$.
\[tab:mem\]
Running memory overhead
-----------------------
In this subsection, we compare our method with linear SVM in term of running memory overhead for loading training features in the training stage and storing classification models in the testing stage. For our method, the trained model includes the real-valued hash function matrix (${{\mathbf{P}}}$) and the binary classification weight matrix. The results are reported in Table \[tab:mem\]. It is clearly shown that our approach requires much less memory than Linear SVM for loading both the training features and classification models. Taking the ImageNet database for example, Linear SVM needs over 150 times more RAM than our method for the training data, and over 3 times more RAM for the trained models.
Conclusions {#Sec:conc}
===========
This work proposed a novel classification framework, by which classification was equivalently transformed to searching the nearest binary weight code in the Hamming space. Different from previous methods, both the feature and classifier weight vectors were simultaneously learned with binary hash codes. Our framework could employ a large family of empirical loss functions, and we here especially studied the representative exponential and hinge loss. For the two sub-problems regarding to the binary classifier and image hash codes, a binary quadratic program (BQP) and linear program was formulated, respectively. In particular, for the BQP problem, a novel bit-flipping procedure which enjoys high efficacy and local optimality guarantee was presented. The two methods with exponential loss and hinge loss were extensively evaluated on several large-scale image datasets. Significant computation overhead reduction of model training and deployment were obtained, while without sacrifice of classification accuracies.
[^1]: <http://www.cs.toronto.edu/~kriz/cifar.html>.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give a dynamical characterisation of odd-dimensional balls within the class of all contact manifolds whose boundary is a standard even-dimensional sphere. The characterisation is in terms of the non-existence of short periodic Reeb orbits.'
address:
- 'Mathematisches Institut, Universität zu Köln, Weyertal 86–90, 50931 Köln, Germany'
- 'Mathematisches Institut, WWU Münster, Einsteinstrae 62, 48149 Münster, Germany'
author:
- Hansjörg Geiges
- Kai Zehmisch
title: Reeb dynamics detects odd balls
---
[^1]
Introduction
============
Definitions and the main result
-------------------------------
Let $(M,\alpha)$ be a compact, connected contact manifold (with a fixed choice of contact form $\alpha$) of dimension $2n+1$, $n\in{\mathbb{N}}$, whose boundary $\partial M$ is diffeomorphic to $S^{2n}$.
We write $\inf_0(\alpha)$ for the infimum of all positive periods of *contractible* closed orbits of the Reeb vector field $R_{\alpha}$. When there are no closed contractible Reeb orbits, we have $\inf_0(\alpha)=\infty$, otherwise $\inf_0(\alpha)$ is a minimum and in particular positive.
Our main result will be a criterion for $M$ to be diffeomorphic to a ball in terms of $\inf_0(\alpha)$ and an embeddability condition on $\partial M$. To formulate this condition, we introduce the following terminology.
\(a) Write $D$ for the closed unit disc in ${\mathbb{R}}^2$. The $(2n+1)$-dimensional manifold (with boundary) $$Z:={\mathbb{R}}\times D\times{\mathbb{C}}^{n-1}$$ with contact form $${\alpha_{\mathrm{cyl}}}:={\mathrm{d}}b+\frac{1}{2}(x_0\,{\mathrm{d}}y_0-y_0\,{\mathrm{d}}x_0)
-\sum_{j=1}^{n-1}y_j\,{\mathrm{d}}x_j$$ (with the obvious denomination of cartesian coordinates) will be referred to as the **contact cylinder**.
\(b) We say that $\partial M$ admits a **contact embedding** into the contact cylinder $Z$ if there is an embedding $\varphi$ of a collar neighbourhood of $\partial M\subset M$ into $\operatorname{Int}(Z)$ with $\varphi^*{\alpha_{\mathrm{cyl}}}=\alpha$ and with the image of the collar under $\varphi$ contained in the interior of $\varphi (\partial M)$.
\[thm:main\] Assume that the boundary $\partial M\cong S^{2n}$ of a contact manifold $(M,\alpha)$ as above admits a contact embedding into the contact cylinder, and $\inf_0(\alpha)\geq\pi$. Then $M$ is diffeomorphic to a ball.
This theorem has been proved for $\dim M=3$ by Eliashberg and Hofer [@elho94]. In that paper, they also announced the theorem for the higher-dimensional case, but a proof has never been published. They formulated the higher-dimensional case under the additional homological assumption $H_2(M;{\mathbb{R}})=0$; this condition, as we shall see, is superfluous.
For simplicity, we shall assume throughout that $n\geq 2$, although a large part of our argument also works for $n=1$. Our proof shows that $(M,\alpha)$ is diffeomorphic to a ball whenever $\partial M\cong S^{2n}$ admits a contact embedding into the cylinder $Z_r:={\mathbb{R}}\times D^2_r\times{\mathbb{C}}^{n-1}$ of radius $r$, and $\inf_0(\alpha)>
\pi r^2$. Given a contact embedding into $Z=Z_1$, it may be regarded as an embedding into a cylinder of slightly smaller radius. Hence, even though the proof below will be based on the assumption $\inf_0(\alpha)>\pi$, the result holds under the weaker assumption $\inf_0(\alpha)\geq\pi$.
Idea of the proof {#subsection:idea}
-----------------
The contact embedding $\varphi$ of $\partial M$ into the contact cylinder $Z$ allows us to form a new contact manifold ${\widehat{{\mathbb{R}}}^{2n+1}}$ by removing the bounded component of ${\mathbb{R}}^{2n+1}\setminus\varphi(\partial M)$ and gluing in $M$ instead. Similarly, we write ${\widehat{Z}}$ for the cylinder $Z$ with $M$ glued in. We shall be studying the moduli space ${\mathcal{W}}$ of holomorphic discs $u=(a,f)\co {\mathbb{D}}\rightarrow {\mathbb{R}}\times{\widehat{{\mathbb{R}}}^{2n+1}}=:W$ in the symplectisation $W$ of ${\widehat{{\mathbb{R}}}^{2n+1}}$, where the discs are subject to certain boundary and homological conditions. (We always write ${\mathbb{D}}$ for the closed unit disc in ${\mathbb{C}}$ when regarded as the domain of definition of our holomorphic discs.) It will turn out that $f({\mathbb{D}})$ is always contained in ${\widehat{Z}}$. We then have the following dichotomy. Either the evaluation map $$\begin{array}{rccc}
\operatorname{ev}\co & {\mathcal{W}}\times{\mathbb{D}}& \longrightarrow & {\widehat{Z}}\\
& \bigl( (a,f),z\bigr) & \longmapsto & f(z)
\end{array}$$ is proper and surjective, i.e. gives a filling, in which case topological arguments involving the $h$-cobordism theorem can be used to show that $M$ must be a ball. Otherwise there will be breaking of holomorphic discs, which entails the existence of short contractible periodic Reeb orbits as in Hofer’s paper [@hofe93].
Remarks
-------
\(1) The bound $\pi$ in Theorem \[thm:main\] is optimal. Inside $Z$ one can form the connected sum as described by Weinstein [@wein91], cf. [@geig08 Section 6.2], with any contact manifold, producing a belt sphere of radius $r_0$ smaller than, but arbitrarily close to $1$. Inside this belt sphere one finds a periodic orbit of length $\pi r_0^2$.
\(2) In the $3$-dimensional case, Theorem \[thm:main\] can be strengthened. If $\inf_0(\alpha)\geq\pi$, then there are in fact no closed Reeb orbits at all. (This was part of the formulation of the theorem in [@elho94].) In this $3$-dimensional case, the holomorphic discs project to embedded discs in ${\widehat{Z}}$, where they produce a foliation by discs transverse to the Reeb direction, see [@elho94 Section 2]. This precludes closed orbits.
\(3) The existence of a foliation by discs as in (2) implies that there cannot even be trapped Reeb orbits, i.e. orbits that are bounded in forward or backward time. In [@grz] we show in joint work with Nena Röttgen that this is a purely $3$-dimensional phenomenon. In higher dimensions it is possible to have a Reeb dynamics on Euclidean space, standard outside a compact set, with trapped orbits but no periodic ones.
\(4) One may consider manifolds $M$ with disconnected boundary (and boundary components different from $S^{2n}$). The requirement of a contact embedding into the contact cylinder $Z$ is made for each component of $\partial M$ individually. By translating the images of these components in the ${\mathbb{R}}$-direction one may then assume without loss of generality that they are not nested. The collection $\varphi (\partial M)$ of these images is contained in a large ellipsoid $E$ inside $\operatorname{Int}(Z)$. The manifold obtained from $E$ by removing the interiors of the components of $\varphi(\partial M)$ and gluing in $M$ instead has non-trivial fundamental group: by taking a path in $M$ joining two boundary components, and a second path joining these two boundary points in the exterior of $\varphi(\partial M)\subset E$, one creates an essential loop. It follows that this manifold contains a contractible Reeb orbit of period smaller than $\pi$. This orbit must in fact be contained entirely in $M$, since the Reeb flow on $Z$ is positively transverse to any hypersurface $\{b\}\times D^{2n}$.
In other words, Theorem \[thm:main\] provides a means of detecting contractible periodic orbits on non-compact manifolds or manifolds with boundary. See [@bpv09; @bprv; @suze] for related work.
Symplectisations of contactisations
===================================
The contact cylinder $Z$ may be regarded as the contactisation of the exact symplectic manifold $D\times{\mathbb{C}}^{n-1}\subset{\mathbb{C}}^n$. In the latter, we have the obvious holomorphic discs $D\times\{*\}$. In order to lift these to holomorphic discs in the symplectisation of $Z$, it is advantageous to proceed in two steps: first lift them to holomorphic discs in ${\mathbb{C}}\times D\times{\mathbb{C}}^{n-1}$, and then transform them to holomorphic discs in the symplectisation ${\mathbb{R}}\times Z$ using an explicit biholomorphism $$\Phi\co {\mathbb{R}}\times{\mathbb{R}}\times D\times{\mathbb{C}}^{n-1}\longrightarrow
{\mathbb{C}}\times D\times{\mathbb{C}}^{n-1}.$$ The desired boundary condition for the holomorphic discs on the left-hand side gives us the boundary conditions for the holomorphic discs on the right.
This allows one to transform a Cauchy–Riemann problem on the left with respect to a ‘twisted’ almost complex structure (which preserves the contact hyperplanes and pairs the Reeb with the symplectisation direction) into a Poisson problem on a single real-valued function.
This idea is implicit in [@elho94 p. 1320] and has also been used in [@nied06 Proposition 5]. Before we turn to our specific situation, we discuss this transformation in slightly greater generality.
Lifting holomorphic discs {#subsection:lifting}
-------------------------
Let $(V,J_V)$ be a Stein manifold of complex dimension $n$. We write $\psi$ for a plurisubharmonic potential on $V$, so that $\omega_V:=-{\mathrm{d}}({\mathrm{d}}\psi\circ J_V)$ is a Kähler form on $V$. In fact, what is really relevant for the following discussion is the existence of such a potential, not the integrability of $J_V$, cf. [@geze12 Section 3.1]. Write $\lambda:=-{\mathrm{d}}\psi
\circ J_V$ for the primitive $1$-form of the symplectic form $\omega_V$.
The contactisation of $V$ is $({\mathbb{R}}\times V,\alpha:={\mathrm{d}}b+\lambda)$, where $b$ denotes the ${\mathbb{R}}$-coordinate. Notice that $\partial_b$ is the Reeb vector field of the contact form $\alpha$. A symplectisation of this manifold is $$\bigl({\mathbb{R}}\times{\mathbb{R}}\times V,\omega:={\mathrm{d}}(\tau\alpha)\bigr),$$ where $\tau$ is a strictly increasing smooth positive function on the first ${\mathbb{R}}$-factor (whose coordinate we shall denote by $a$). A compatible almost complex structure $J$ on this symplectic manifold, which in addition preserves the contact hyperplanes $$\ker\alpha=\{ v-\lambda(v)\partial_b\co v\in TV\}$$ on $\{ a\}\times{\mathbb{R}}\times V$, is given by $$J(\partial_a)=\partial_b\;\;\;\text{and}\;\;\;
J(v-\lambda(v)\partial_b)=J_V v-\lambda(J_V v)\partial_b.$$ If $J_V$ is not integrable, then $J$ may only be tamed by $\omega$.
A straightforward calculation gives the following generalisation of [@nied06 Proposition 5]:
\[prop:biholo\] The map $$\begin{array}{rccc}
\Phi\co & ({\mathbb{R}}\times{\mathbb{R}}\times V,J) & \longrightarrow &
({\mathbb{C}}\times V,{\mathrm{i}}\oplus J_V)\\
& (a,b,{\mathbf{z}}) & \longmapsto &
(a-\psi({\mathbf{z}})+{\mathrm{i}}b,{\mathbf{z}})
\end{array}$$ is a biholomorphism.
Given a holomorphic disc ${\mathbb{D}}\ni z\mapsto h(z)\in V$, we want to lift this to a holomorphic disc $${\mathbb{D}}\ni z\longmapsto (a(z), b(z),h(z))$$ in the symplectisation, with boundary in the zero level of the symplectisation, i.e. $a|_{\partial{\mathbb{D}}}\equiv 0$. By Proposition \[prop:biholo\], the functions $a$ and $b$ are found as follows. Let $a\co{\mathbb{D}}\rightarrow{\mathbb{R}}$ be the unique solution, smooth up to the boundary, of the Poisson problem $$\left\{\begin{array}{rcll}
\Delta a & = & \Delta (\psi\circ h) & \text{on $\operatorname{Int}({\mathbb{D}})$},\\
a & = & 0 & \text{on $\partial{\mathbb{D}}$}.
\end{array}\right.$$ Then $a-\psi\circ h$ is harmonic, and we may choose the function $b$ (unique up to adding a constant) such that $a-\psi\circ h+{\mathrm{i}}b$ is holomorphic. Notice that the function $a$ is subharmonic.
Examples {#subsection:ex-disc}
--------
\(1) Our first example shows how to derive the set-up of [@elho94] in this general context. We take $V={\mathbb{C}}$ with plurisubharmonic potential $\psi(x+{\mathrm{i}}y)=x^2/2$. This yields the contact form ${\mathrm{d}}b+x\,{\mathrm{d}}y$ on ${\mathbb{R}}\times{\mathbb{C}}$. Start with the holomorphic disc $h\co{\mathbb{D}}\rightarrow {\mathbb{C}}$ given by inclusion. The solution $a$ of the corresponding Poisson problem — this is equation (52) in [@elho94] — is given by $a(x,y)=(x^2+y^2-1)/4$. For $b$ one obtains $b(x,y)=b_0-xy/2$. Notice that $a-\psi\circ h+{\mathrm{i}}b$ is the holomorphic function $z\mapsto
-(z^2+1)/4+{\mathrm{i}}b_0$.
\(2) For our second example we take $V={\mathbb{C}}$ with plurisubharmonic potential $\psi(z)=|z|^2/4$. This gives rise to the contact form ${\mathrm{d}}b+(x\,{\mathrm{d}}y-y\,{\mathrm{d}}x)/2$ on ${\mathbb{R}}\times{\mathbb{C}}$. The solution $a$ of the Poisson problem is unchanged, but $b$ is now simply a constant function. The example in [@nied06] is obtained by crossing this $V$ with a cotangent bundle $T^*Q$, on which one takes the plurisubharmonic potential $\|\mathbf{p}\|^2/2$, with $\mathbf{p}$ denoting the fibre coordinate, corresponding to the canonical Liouville $1$-form on $T^*Q$.
The contact cylinder {#subsection:cylinder}
--------------------
The contact form ${\alpha_{\mathrm{cyl}}}$ on the contact cylinder $Z={\mathbb{R}}\times D\times{\mathbb{C}}^{n-1}$ derives from the plurisubharmonic potential $$\psi (z_0;z_1,\ldots,z_{n-1}):=\frac{1}{4}|z_0|^2+
\frac{1}{2}\sum_{j=1}^{n-1}y_j^2$$ on $D\times{\mathbb{C}}^{n-1}$, where $z_j=x_j+{\mathrm{i}}y_j$, $j=0,1,\ldots,n-1$.
Similar to Example \[subsection:ex-disc\] (2), for any choice of parameters $b\in{\mathbb{R}}$, ${\mathbf{s}},{\mathbf{t}}\in{\mathbb{R}}^{n-1}$, we have the holomorphic discs $$\begin{array}{rccc}
u^{{\mathbf{t}}}_{{\mathbf{s}},b}\co & {\mathbb{D}}& \longrightarrow & {\mathbb{R}}\times{\mathbb{R}}\times D^{2n}\\
& z & \longmapsto & \bigl(\frac{1}{4}
(|z|^2-1),b,z,{\mathbf{s}}+{\mathrm{i}}{\mathbf{t}}\bigr),
\end{array}$$ lifting the obvious holomorphic discs in $D\times{\mathbb{C}}^{n-1}$. The disc $u^{{\mathbf{t}}}_{{\mathbf{s}},b}$ has boundary on the Lagrangian cylinder $$L^{{\mathbf{t}}}:=\{0\}\times{\mathbb{R}}\times S^1\times{\mathbb{R}}^{n-1}\times\{{\mathbf{t}}\}$$ in ${\mathbb{R}}\times Z$. These Lagrangian cylinders foliate $\partial(\{0\}\times Z)$.
The moduli space of holomorphic discs
=====================================
We now form the contact manifold $({\widehat{{\mathbb{R}}}^{2n+1}},{\hat{\alpha}})$ as explained in Section \[subsection:idea\]. Let $$\bigl(W:={\mathbb{R}}\times{\widehat{{\mathbb{R}}}^{2n+1}},\omega:={\mathrm{d}}(\tau{\hat{\alpha}})\bigr)$$ be its symplectisation, where $\tau\co{\mathbb{R}}\rightarrow{\mathbb{R}}^+$ is a smooth function with $\tau'>0$ and $\tau(a)={\mathrm{e}}^a$ for $a\geq 0$. The freedom of choosing $\tau$ on $\{ a<0\}$ is required for the asymptotic analysis cited in Section \[section:compactness\].
The almost complex structure {#subsection:complex}
----------------------------
Choose $b_0,r,R\in{\mathbb{R}}^+$ with $r<1$ such that $\varphi(\partial M)$ is contained in the interior of the box $$B:=[-b_0,b_0]\times D^2_r\times D^{2n-2}_R\subset Z,$$ where $D_{\rho}^{2k}\subset{\mathbb{C}}^k$ denotes a closed $2k$-disc of radius $\rho$. We write ${\widehat{B}}$ for the result of gluing $M$ into this box, in other words, $${\widehat{{\mathbb{R}}}^{2n+1}}= {\widehat{B}}\cup_{\partial B}\bigl(({\mathbb{R}}\times{\mathbb{C}}\times{\mathbb{C}}^{n-1})
\setminus\operatorname{Int}(B)\bigr).$$ We shall also have occasion to use the notation ${\widehat{Z}}$ for the cylinder $Z$ with $M$ glued in, that is, $${\widehat{Z}}={\widehat{{\mathbb{R}}}^{2n+1}}\setminus \bigl( {\mathbb{R}}\times ({\mathbb{C}}\setminus\operatorname{Int}(D))\times
{\mathbb{C}}^{n-1}\bigr).$$
On the symplectic manifold $(W,\omega)$ we choose an almost complex structure $J$ compatible with $\omega$ subject to the following conditions:
- On the complement of ${\mathbb{R}}\times\operatorname{Int}({\widehat{B}})$, the almost complex structure $J$ equals the one described in Section \[subsection:lifting\].
- On ${\mathbb{R}}\times\operatorname{Int}({\widehat{B}})$, we make a generic choice (in a sense explained in Section \[subsection:regular\]) of an ${\mathbb{R}}$-invariant almost complex structure $J$ preserving $\ker{\hat{\alpha}}$ and satisfying $J(\partial_a)=R_{{\hat{\alpha}}}$.
Condition (J1) will allow us to prove that holomorphic discs in the relevant region are standard. Condition (J2) implies that the breaking of holomorphic discs corresponds to cylindrical ends asymptotic to Reeb orbits.
The moduli space {#subsection:moduli}
----------------
We now consider holomorphic discs (smooth up to the boundary) of the form $$u=(a,f)\co ({\mathbb{D}},\partial{\mathbb{D}})\longrightarrow (W={\mathbb{R}}\times{\widehat{{\mathbb{R}}}^{2n+1}},L^{{\mathbf{t}}}),$$ i.e. with Lagrangian boundary condition, where ${\mathbf{t}}$ is allowed to vary over ${\mathbb{R}}^{n-1}$. We shall call the value of ${\mathbf{t}}$ corresponding to a given $u$ the ‘boundary level’ of the holomorphic disc.
We define ${\mathcal{W}}$ to be the moduli space of such discs $u$, which are supposed to satisfy the following conditions:
- The relative homology class $[u]\in H_2(W,L^{{\mathbf{t}}})$, with ${\mathbf{t}}$ equal to the boundary level of $u$, equals that of ${u^{{\mathbf{t}}}_{{\mathbf{s}},b}}$ for some $b\in{\mathbb{R}}$, ${\mathbf{s}}\in{\mathbb{R}}^{n-1}$, where $|b|,|{\mathbf{s}}|$ are large (such that ${u^{{\mathbf{t}}}_{{\mathbf{s}},b}}$ may be regarded as a holomorphic disc in $W$).
- For $k=0,1,2$ we have $u({\mathrm{i}}^k)\in L^{{\mathbf{t}}}\cap\{z_0={\mathrm{i}}^k\}$.
Let $u=(a,f)$ be a holomorphic disc satisfying (M1). By the maximum principle, $f({\mathbb{D}})$ is contained in ${\widehat{Z}}$, see Lemma \[lem:bounds\]. By the boundary lemma of E. Hopf, applied to a small disc in ${\mathbb{D}}$ touching a given boundary point and mapping to the complement of ${\mathbb{R}}\times\operatorname{Int}({\widehat{B}})$, so that the $z_0$-component of $u$ is defined and holomorphic on that small disc, the boundary $u(\partial{\mathbb{D}})$ is transverse to $$\{0\}\times{\mathbb{R}}\times\{{\mathrm{e}}^{{\mathrm{i}}\theta}\}\times{\mathbb{R}}^{n-1}
\times\{{\mathbf{t}}\}\subset L^{{\mathbf{t}}}$$ for each ${\mathrm{e}}^{{\mathrm{i}}\theta}\in S^1$ and, by (M1), in fact positively transverse. Thus, condition (M2) fixes a parametrisation of $u$.
Properties of the holomorphic discs
-----------------------------------
Here we collect some basic properties of the discs $u\in{\mathcal{W}}$.
\[lem:Maslov\] The Maslov index $\mu$ of any disc $u\in{\mathcal{W}}$, i.e.the index of the bundle pair $(u^*TW,(u|_{\partial{\mathbb{D}}})^*TL^{{\mathbf{t}}})$, equals $2$.
We appeal to the axiomatic definition of the Maslov index in [@mcsa04 Section C.3]. For the disc $u_0:=u^{{\mathbf{t}}}_{{\mathbf 0},0}$ in ${\mathbb{R}}\times{\mathbb{R}}^{2n+1}$, the bundle $u_0^*T({\mathbb{R}}\times{\mathbb{R}}^{2n+1})$ is a trivial ${\mathbb{C}}^{n+1}$-bundle. The fibre of the totally real subbundle $(u_0|_{\partial{\mathbb{D}}})^*TL^{{\mathbf{t}}}$ over ${\mathrm{e}}^{{\mathrm{i}}\theta}\in\partial{\mathbb{D}}$ is given by ${\mathbb{R}}{\mathrm{i}}\oplus{\mathbb{R}}{\mathrm{i}}{\mathrm{e}}^{{\mathrm{i}}\theta}\oplus{\mathbb{R}}^{n-1}$. So the normalisation property of the Maslov index implies $\mu (u^{{\mathbf{t}}}_{{\mathbf 0},0})=2$.
By the homotopy invariance of the Maslov index, we have $\mu(u^{{\mathbf{t}}}_{{\mathbf{s}},b})=2$ for all standard discs $u^{{\mathbf{t}}}_{{\mathbf{s}},b}$ in $W$. Finally, given any $u\in{\mathcal{W}}$, we may choose $u^{{\mathbf{t}}}_{{\mathbf{s}},b}$ in the same relative homology class, so that $u-u^{{\mathbf{t}}}_{{\mathbf{s}},b}$ is a boundary. This implies $\mu(u)=2$.
\[lem:energy\] Each disc $u\in{\mathcal{W}}$ has symplectic energy $\int_{{\mathbb{D}}}u^*\omega$ equal to $\pi$.
Choose a standard disc ${u^{{\mathbf{t}}}_{{\mathbf{s}},b}}$ in the same relative class in $H_2(W,L^{{\mathbf{t}}})$ as $u$. Then in particular $[\partial u]=
[\partial{u^{{\mathbf{t}}}_{{\mathbf{s}},b}}]$ in $H_1(L^{{\mathbf{t}}})$. Since $L^{{\mathbf{t}}}$ is Lagrangian, the pull-back of the $1$-form ${\hat{\alpha}}$ to $L^{{\mathbf{t}}}$ is closed, and hence $$\int_{\partial u}{\hat{\alpha}}=\int_{\partial{u^{{\mathbf{t}}}_{{\mathbf{s}},b}}}{\hat{\alpha}}.$$ One then computes $$\int_u\omega=\int_{\partial u}{\hat{\alpha}}=\int_{\partial{u^{{\mathbf{t}}}_{{\mathbf{s}},b}}}{\hat{\alpha}}=\int_{\partial{u^{{\mathbf{t}}}_{{\mathbf{s}},b}}}{\alpha_{\mathrm{cyl}}}=\pi.\qedhere$$
\[rem:energy\] By the same argument we see that any non-constant holomorphic disc in $W$ with boundary on $L^{{\mathbf{t}}}$ has symplectic energy in $\pi{\mathbb{N}}$.
\[lem:simple\] All discs $u\in{\mathcal{W}}$ are simple.
According to [@lazz11 Theorem A], the homology class $[u]\in H_2(W,L^{{\mathbf{t}}})$ of a holomorphic disc with totally real boundary condition can be decomposed into positive multiples of homology classes represented by simple discs, which are obtained from a decomposition of ${\mathbb{D}}$. Since the class $[u]=[{u^{{\mathbf{t}}}_{{\mathbf{s}},b}}]\in H_2(W,L^{{\mathbf{t}}})$ is indecomposable by Lemma \[lem:energy\] and Remark \[rem:energy\], the disc $u$ itself must be simple.
Simplicity of the discs $u=(a,f)$ will not be quite enough for our purposes. We shall also need simplicity of $f$ in the sense of the following lemma, cf. [@hwz99 Theorem 1.14]. Here $\pi$ denotes the projection of $TM$ onto $\ker{\hat{\alpha}}$ along the Reeb vector field $R_{{\hat{\alpha}}}$.
\[lem:f-injective\] For each $u=(a,f)\in{\mathcal{W}}$, the set $${F_{\mathrm{inj}}}:=\bigl\{ z\in{\mathbb{D}}\co \pi\circ T_zf\neq 0,\
f^{-1}(f(z))=\{ z\}\bigr\}$$ of ‘$f$-injective points’ is open and dense in ${\mathbb{D}}$.
The combination of defining conditions for ${F_{\mathrm{inj}}}$ is open, so we need only show that ${F_{\mathrm{inj}}}$ is dense in ${\mathbb{D}}$. We begin with three observations about the behaviour of the holomorphic discs $u$.
First of all, in a neighbourhood of the boundary $\partial{\mathbb{D}}\subset{\mathbb{D}}$ we can write $f$ in components as $f=(b,{\mathbf{h}})=(b,h_0,\ldots,h_{n-1})$ with each $h_j$ holomorphic. By the comment in Section \[subsection:moduli\], $h_0|_{\partial{\mathbb{D}}}$ is an immersion, hence $\pi\circ Tf|_{\partial{\mathbb{D}}}\neq 0$. Moreover, a variant of the Carleman similarity principle [@elho94 pp. 1315/6] implies that the set $\{z\in{\mathbb{D}}\co
\pi\circ T_zf=0\}$ is finite.
Secondly, the boundary $\partial{\mathbb{D}}$ maps under $f$ to ${\mathbb{R}}\times S^1\times{\mathbb{C}}^{n-1}$. Near any point in $\operatorname{Int}({\mathbb{D}})$ that putatively maps to ${\mathbb{R}}\times ({\mathbb{C}}\setminus \operatorname{Int}(D))\times{\mathbb{C}}^{n-1}$, we could write $f=(b,{\mathbf{h}})$ as above, and we would find that $h_0$ violates the maximum principle. We conclude in particular that there are no mixed intersections of the holomorphic disc $u$, i.e. pairs of an interior and a boundary point with the same image.
Thirdly, from the work in [@zehm13] it follows that the immersion $u|_{\partial{\mathbb{D}}}= (0,f|_{\partial{\mathbb{D}}})$ has at most finitely many double points. Otherwise the respective preimages would accumulate in two separate points — for in a common limit point the differential $Tu$ would be singular — and [@zehm13 Lemma 4.2] would imply that the differentials $Tu$ in the two limit points are collinear over ${\mathbb{R}}$. Furthermore, by Lemma \[lem:bounds\] (i) below, the collinearity factor would have to be positive. Then [@zehm13 Lemma 4.3] would imply that $u$ is not simple, contradicting the preceding lemma.
From these last two observations we infer that ${F_{\mathrm{inj}}}$ contains $\partial{\mathbb{D}}$ with the exception of at most finitely many points, and in particular is non-empty.
Now we prove that ${F_{\mathrm{inj}}}$ is dense, arguing by contradiction. If ${F_{\mathrm{inj}}}$ were not dense, the set $\operatorname{Int}({\mathbb{D}})\setminus{F_{\mathrm{inj}}}$ would have non-empty interior. By the preceding observations we can find an open subset $U\subset\operatorname{Int}({\mathbb{D}})$ such that for each $z\in U$ the set $f^{-1}(f(z))\subset\operatorname{Int}({\mathbb{D}})$ contains more than just the point $z$, and such that $\pi\circ T_wf\neq 0$ in all points $w\in f^{-1}(f(U))$. The latter implies that the points in $f^{-1}(f(z))$ are isolated, and hence finite in number.
What follows is an explication of an argument in [@hwz99 p. 459]. Fix a point $z_0\in U$ and write $f^{-1}(f(z_0))=\{z_0,z_1,\ldots, z_N\}$. Choose pairwise disjoint (and disjoint from $U$) open neighbourhoods $U_k\subset\operatorname{Int}({\mathbb{D}})$ of $z_k$, $k=1,\ldots, N$, such that $f|_{U_k}$ is an embedding. By a compactness argument, $U$ can be chosen so small that $$f(U)\subset \bigcup_{k=1}^N f(U_k),$$ and such that $f|_{U}$ is likewise an embedding. Choose relatively compact neighbourhoods $U_k'\subset U_k$ of $z_k$, $k=1,\ldots, N$. By shrinking $U$ to a smaller neighbourhood of $z_0$, we can ensure that $$f(U)\subset \bigcup_{k=1}^N f(U_k').$$ Set $A_k:=(f|_U)^{-1}(f(U_k'))\subset U$. If $A_1$ has non-empty interior, we can shrink $U$ such that $f(U)\subset f(U_1')$ (but $U$ need no longer be a neighbourhood of $z_0$). The argument then concludes as in [@hwz99 pp. 459/60], leading to a contradiction to $u$ being simple. If $A_1$ has empty interior, so that $U\setminus A_1$ is dense in $U$, we find that $$f(U)\subset \bigcup_{k=2}^N f(\overline{U'_k})\subset
\bigcup_{k=2}^N f(U_j).$$ The argument concludes inductively.
Bounds on the holomorphic discs
-------------------------------
In the next lemma we collect some restrictions on the image $u({\mathbb{D}})$ of the holomorphic discs $u\in{\mathcal{W}}$.
\[lem:bounds\] For $u=(a,f)\in{\mathcal{W}}$ we have:
- $a<0$ on $\operatorname{Int}({\mathbb{D}})$.
- $f(\operatorname{Int}({\mathbb{D}}))$ is contained in the interior of ${\widehat{Z}}$, i.e.$$f(\operatorname{Int}({\mathbb{D}}))\cap\bigl({\mathbb{R}}\times({\mathbb{C}}\setminus\operatorname{Int}(D))
\times{\mathbb{C}}^{n-1}\bigr)=\emptyset.$$
\(i) The holomorphicity of $u=(a,f)$ (with respect to an almost complex structure preserving $\ker{\hat{\alpha}}$ and satisfying $J(\partial_a)=R_{{\hat{\alpha}}}$) implies $f^*{\hat{\alpha}}=-{\mathrm{d}}a\circ{\mathrm{i}}$, so $a$ is subharmonic. We have $a|_{\partial{\mathbb{D}}}\equiv 0$, but $a$ cannot be identically zero on all of ${\mathbb{D}}$, for otherwise we would have $f^*\alpha\equiv 0$ and $f^*{\mathrm{d}}\alpha\equiv 0$, which would imply that $u$ has zero symplectic energy density and hence is constant, contradicting (M1). The strong maximum principle for $a$ then implies the claim.
\(ii) Near the points of ${\mathbb{D}}$ mapping under $f$ to ${\mathbb{R}}\times({\mathbb{C}}\setminus\operatorname{Int}(D))\times{\mathbb{C}}^{n-1}$ we can write this map in components as $f=(b,{\mathbf{h}})$. If $f(\operatorname{Int}({\mathbb{D}}))$ were not contained in $\operatorname{Int}(Z)$, we would find that the map $h_0$ is defined and locally constant on a non-empty open and closed subset of ${\mathbb{D}}$, and hence on all of ${\mathbb{D}}$, contradicting the homological assumption (M1).
Since a generic choice of the almost complex structure $J$ is only allowed on ${\mathbb{R}}\times\operatorname{Int}({\widehat{B}})$, this can be used to guarantee regularity in the sense of [@mcsa04 Definition 3.1.4] only for those holomorphic discs that pass through this ‘perturbation domain’, see [@mcsa04 Remark 3.2.3]. We therefore want to show that all other discs belong to the standard family ${u^{{\mathbf{t}}}_{{\mathbf{s}},b}}$, where transversality is obvious. This will be used below to show that ${\mathcal{W}}$ is actually a manifold.
\[lem:standard1\] Let $u=(a,f)\in{\mathcal{W}}$. If $f({\mathbb{D}})\subset{\widehat{{\mathbb{R}}}^{2n+1}}\setminus\operatorname{Int}({\widehat{B}})$, then $u={u^{{\mathbf{t}}}_{{\mathbf{s}},b}}$ for some ${\mathbf{s}}\in{\mathbb{R}}^{n-1}$, $b\in{\mathbb{R}}$, and ${\mathbf{t}}$ equal to the boundary level of $u$.
Since $f$ maps to the complement of $\operatorname{Int}({\widehat{B}})$, we can write it *globally* as $f=(b,{\mathbf{h}})$, with every component $h_j$ of ${\mathbf{h}}$ a holomorphic map ${\mathbb{D}}\rightarrow{\mathbb{C}}$. The boundary condition for $u$ means that for $j=1,\ldots ,n-1$ we have $\operatorname{Im}h_j=t_j$ on $\partial{\mathbb{D}}$. The minimum and maximum principle for harmonic functions implies that $\operatorname{Im}h_j=t_j$ on all of ${\mathbb{D}}$. Hence, by the open mapping theorem, $\operatorname{Re}h_j=:s_j$ is likewise constant on ${\mathbb{D}}$ for $j=1,\ldots,n-1$.
The component $h_0$ is a holomorphic disc in ${\mathbb{C}}$ with $h_0|_{\partial{\mathbb{D}}}$ an orientation-preserving diffeomorphism of $\partial{\mathbb{D}}$, cf. the comment after condition (M2). The argument principle implies that $h_0$ is an orientation-preserving automorphism of ${\mathbb{D}}$, and then (M2) forces $h_0=\operatorname{id}_{{\mathbb{D}}}$.
By Proposition \[prop:biholo\], the function $$z\longmapsto a(z)-\frac{1}{4}|h_0(z)|^2-\frac{1}{2}\sum_{j=1}^{n-1}
(\operatorname{Im}h_j)^2
=a(z)-\frac{1}{4}|z|^2-\frac{1}{2}\sum_{j=1}^{n-1}
(\operatorname{Im}h_j)^2$$ on ${\mathbb{D}}$ is harmonic, taking the constant value $-1/4-|{\mathbf{t}}|^2/2$ on $\partial{\mathbb{D}}$, hence it is constant on ${\mathbb{D}}$. This means that the imaginary part $b$ that makes this into a holomorphic function must also be constant. Solving for $a(z)$ we get $$a(z)=\frac{1}{4}(|z|^2-1)\;\;\text{on ${\mathbb{D}}$},$$ i.e. $u={u^{{\mathbf{t}}}_{{\mathbf{s}},b}}$.
The next lemma will allow us to control the degree of the evaluation map $\operatorname{ev}$. It says hat non-standard disc can never reach $b$-levels with $|b|>b_0$. This is also relevant for compactness.
\[lem:standard2\] Let $u=(a,f)\in{\mathcal{W}}$. On the closed set $A:=
f^{-1}({\widehat{{\mathbb{R}}}^{2n+1}}\setminus\operatorname{Int}({\widehat{B}}))\subset{\mathbb{D}}$, which includes the whole boundary $\partial{\mathbb{D}}$ in its interior, we write $f=(b,{\mathbf{h}})$. If the function $b$ takes values outside $[-b_0,b_0]$, then $f$ maps to a $b$-level set $\{b_1\}\times D\times{\mathbb{C}}^{n-1}$ with $|b_1|>b_0$ and hence, by the preceding lemma, the holomorphic curve $u$ equals $u_{{\mathbf{s}},b_1}^{{\mathbf{t}}}$ for some ${\mathbf{s}}$ and ${\mathbf{t}}$.
Choose $z_*\in A$ with $b_*:=b(z_*)$ of maximal absolute value. Notice that $z_*$ is an interior point of $A$. By Proposition \[prop:biholo\], the function $$g:= a-\frac{1}{4}|h_0|^2-\frac{1}{2}\sum_{j=1}^{n-1}(\operatorname{Im}h_j)^2+{\mathrm{i}}b$$ is holomorphic on $\operatorname{Int}(A)$. We should now like to argue with the maximum principle that the imaginary part $b$ of $g$ has to be constant equal to $b_*$ on an open and closed subset of ${\mathbb{D}}$. If $z_*\in\operatorname{Int}({\mathbb{D}})$, this inference is indeed conclusive, just as in part (ii) of Lemma \[lem:bounds\]. If $z_*\in\partial{\mathbb{D}}$, we reason as follows.
The real part of the holomorphic function $g$ takes the constant value $a_{\partial}:=-1/4-|{\mathbf{t}}|^2/2$ on $\partial{\mathbb{D}}\subset\operatorname{Int}(A)$. It follows that the function can be extended by Schwarz reflection to the complementary set $\overline{A}$ of $A$ in $\hat{{\mathbb{C}}}\setminus{\mathbb{D}}$, with $\hat{{\mathbb{C}}}$ denoting the Riemann sphere. Indeed, the holomorphic function ${\mathrm{i}}(g-a_{\partial})$ takes real values on $\partial{\mathbb{D}}$, so the Schwarz reflection principle applies to this function, and we simply transform the extension via the map $w\mapsto -{\mathrm{i}}w+a_{\partial}$ to a holomorphic extension of $g$. Now $z_*$ is an interior point of $A\cup\overline{A}$, and we conclude as before with the maximum principle.
Finally, we establish a $C^0$-bound in the ${\mathbb{C}}^{n-1}$-direction on non-standard discs.
\[lem:standard3\] Let $u=(a,f)\in{\mathcal{W}}$. If $f({\mathbb{D}})$ intersects $${\mathbb{R}}\times{\mathbb{C}}\times \bigl({\mathbb{C}}^{n-1}\setminus D^{2n-2}_{R+\sqrt{2}}\bigr),$$ then $u$ equals one of the standard discs ${u^{{\mathbf{t}}}_{{\mathbf{s}},b}}$.
Consider the open subset $$G:= f^{-1}\bigl({\mathbb{R}}\times{\mathbb{C}}\times\bigl({\mathbb{C}}^{n-1}\setminus D^{2n-2}_R
\bigr)\bigr)\subset{\mathbb{D}},$$ which will be non-empty under the assumption on $f$ in the lemma. On the closure ${\overline{G}}$ of $G$ we write $f=(b,{\mathbf{h}})$ as before and consider the subharmonic function $h:=|h_1|^2+\cdots+|h_{n-1}|^2$.
Write $\partial G$ for the topological boundary of $G$ in ${\mathbb{D}}$. We have $h|_{\partial G}\equiv R^2$, so the maximum of $h$ on ${\overline{G}}$ must be attained at a point in $G\cap\partial{\mathbb{D}}$ (in particular, this intersection must be non-empty).
If ${\overline{G}}={\mathbb{D}}$, we are done by Lemma \[lem:standard1\]. Otherwise, we perform Schwarz reflection on the holomorphic function $h_j-{\mathrm{i}}t_j$, which is possible since $\operatorname{Im}h_j\equiv t_j$ on $\partial{\mathbb{D}}$. To the extended function we add ${\mathrm{i}}t_j$ again to obtain the extension of $h_j$. Geometrically, this corresponds to a reflection of $h_j(G)$ in the line $\{z_j={\mathrm{i}}t_j\}\subset{\mathbb{C}}$.
Write $S$ for the compact subset of the Riemann sphere given as the union of ${\overline{G}}$ and its reflected copy, and continue to write $h$ for the extension of the plurisubharmonic function to $S$. Beware that $h$ may take larger values on $S$ than on ${\overline{G}}$.
Choose a point $s_0\in G\cap\partial D$ where $h|_{{\overline{G}}}$ attains its maximum $(R+\delta)^2$. Now consider an open $\delta$-ball $B_{\delta}$ about the point ${\mathbf{h}}(s_0)\in{\mathbb{C}}^{n-1}$. Then ${\mathbf{h}}(\partial G)$ is contained in the complement of $B_{\delta}$, and since the extension of ${\mathbf{h}}$ to $S$ was obtained by Schwarz reflection along $\partial{\mathbb{D}}\ni s_0$, the full boundary ${\mathbf{h}}(\partial S)$ after reflection will likewise be contained in the complement of $B_{\delta}$.
This allows us to apply the monotonicity lemma [@humm97 Theorem 1.3], which tells us that the area of ${\mathbf{h}}(S)\cap\overline{B}_{\delta}$ is bounded from below by $\pi\delta^2$. (In [@humm97] the estimate is given in the form $\text{const.}\cdot\delta^2$; in the present Euclidean setting the constant $\pi$ comes from the classical isoperimetric inequality.) So the area of ${\mathbf{h}}({\overline{G}})\cap\overline{B}_{\delta}$ is bounded from below by $\pi\delta^2/2$, and from above by the energy $\pi$ of $u$. This implies $\delta\leq\sqrt{2}$.
To sum up: Any holomorphic disc $u$ whose ${\mathbf{h}}$-component stays outside $D^{2n-2}_R$ is standard; for all other discs the ${\mathbf{h}}$-component stays inside $D^{2n-2}_{R+\sqrt{2}}$.
Compactness {#section:compactness}
===========
In this section we establish, under the assumption $\inf_0(\alpha)>\pi$, compactness of the *truncated* moduli space $${\mathcal{W}}':=\bigl\{ u=(a,f)\in{\mathcal{W}}\co f({\mathbb{D}})\subset
[-b_0,b_0]\times D\times D^{2n-2}_{R+\sqrt{2}}\bigr\},$$ i.e. the space obtained from ${\mathcal{W}}$ by cutting off ends containing standard discs only.
Variable boundary condition {#subsection:variable}
---------------------------
The holomorphic discs $u\in{\mathcal{W}}$ have boundary on the Lagrangian cylinder $L^{{\mathbf{t}}}$, which varies with the parameter ${\mathbf{t}}\in{\mathbb{R}}^{n-1}$. It is possible to fix the boundary condition, at the cost of allowing the almost complex structure to vary. This is done with the help of a flow that enables us to identify different copies of $L^{{\mathbf{t}}}$. That flow will also provide explicit charts when we discuss transversality.
Start with a constant vector field ${\mathbf{v}}$ on the space $\operatorname{Im}{\mathbb{C}}^{n-1}$ of ${\mathbf{t}}$-coordinates, and regard this as a vector field on ${\mathbb{R}}\times{\mathbb{R}}\times{\mathbb{C}}\times{\mathbb{C}}^{n-1}$. Cut this off with a bump function supported near $$\{0\}\times [-b_0,b_0]\times S^1\times{\mathbb{C}}^{n-1}$$ and identically $1$ in a smaller neighbourhood of that set. Then write $\psi_t^{{\mathbf{v}}}$ for the flow of this vector field.
For a sequence $u_{\nu}$ of holomorphic discs of level ${\mathbf{t}}_{\nu}\rightarrow {\mathbf{t}}_0$, we can then use the maps $\psi_1^{{\mathbf{t}}_{\nu}-{\mathbf{t}}_0}$ to pull back the $u_{\nu}$ to $J_{\nu}$-holomorphic discs of level ${\mathbf{t}}_0$, where $J_{\nu}:=(\psi_1^{{\mathbf{t}}_{\nu}-{\mathbf{t}}_0})^*J$ is $C^{\infty}$-convergent to $J$ and coincides with $J$ outside the neighbourhood described in the preceding paragraph.
Proof of compactness
--------------------
Now we apply this construction to the truncated moduli space ${\mathcal{W}}'$. Consider a sequence $(u_{\nu})$ of holomorphic discs $u_{\nu}=(a_{\nu},f_{\nu})\in{\mathcal{W}}'$. Then, in particular the levels ${\mathbf{t}}_{\nu}$ will be contained in the compact set $D^{n-1}_{R+\sqrt{2}}$. Hence, after passing to a subsequence, we may assume that ${\mathbf{t}}_{\nu}\rightarrow
{\mathbf{t}}_0$ for some ${\mathbf{t}}_0\in D^{n-1}_{R+\sqrt{2}}$. With the construction from the preceding section we may take the $u_{\nu}$ to be $J_{\nu}$-holomorphic discs of fixed boundary level ${\mathbf{t}}_0$. The almost complex structures $J_{\nu}$ equal $J$ outside a neighbourhood of $\{0\}\times [-b_0,b_0]\times S^1\times{\mathbb{C}}^{n-1}$ and converge to $J$ in the $C^{\infty}$-topology. By Lemma \[lem:energy\], all discs $u_{\nu}$ have symplectic energy equal to $\pi$.
We claim that there is a uniform bound on $\max_{{\mathbb{D}}}|\nabla u_{\nu}|$. Here $|\,.\,|$ denotes the norm corresponding to an ${\mathbb{R}}$-invariant metric on $W$ of the form ${\mathrm{d}}a^2+g_{{\widehat{{\mathbb{R}}}^{2n+1}}}$, with $g_{{\widehat{{\mathbb{R}}}^{2n+1}}}$ any Riemannian metric on ${\widehat{{\mathbb{R}}}^{2n+1}}$. The mean value theorem then gives a uniform $C^0$-bound on $(a_{\nu})$, and compactness follows as in [@geze13] with [@mcsa04 Theorem B.4.2].
Bubbling off analysis as in [@geze13 Section 6] shows that, *a priori*, the following phenomena might occur:
- bubbling of spheres
- bubbling of finite energy planes
- breaking
- bubbling of discs (this can only happen at boundary points).
The first is impossible in an exact symplectic manifold. The second and third phenomenon are precluded by the assumption $\inf_0(\alpha)>\pi$ and the energy estimate from Lemma \[lem:energy\], cf. [@geze13 p. 584], since a finite energy plane in a symplectisation is asymptotic to a *contractible* Reeb orbit. Notice that this rules out any kind of bubbling at interior points.
This leaves the bubbling of discs at boundary points. By Remark \[rem:energy\], there could be at best a single bubble disc at the boundary, taking away the full energy $\pi$, cf. [@mcsa04 Theorem 4.6.1]. But the $C^{\infty}_{{\mathrm{loc}}}$ convergence on the complement of the bubble point, together with condition (M2), is incompatible with a ghost disc.
Transversality {#section:transversality}
==============
The purpose of this section is to show that the truncated moduli space ${\mathcal{W}}'$ is a smooth, oriented manifold with boundary. As usual, this is achieved by proving transversality results in the setting of $W^{1,p}$-maps for some $p>2$. Smoothness of the holomorphic discs is then implied by elliptic regularity.
Let ${\mathcal{B}}$ denote the space of $W^{1,p}$-maps $$u\co ({\mathbb{D}},\partial{\mathbb{D}})\longrightarrow (W,\{0\}\times{\widehat{{\mathbb{R}}}^{2n+1}}),$$ where $u(\partial{\mathbb{D}})$ is supposed to be contained in $L^{{\mathbf{t}}}$ for some ${\mathbf{t}}\in{\mathbb{R}}^{n-1}$, and $u$ is required to satisfy the homological condition (M1) from Section \[subsection:moduli\]. Write ${\mathcal{B}}^{{\mathbf{t}}}\subset{\mathcal{B}}$ for the subspace of discs corresponding to a fixed boundary level ${\mathbf{t}}$.
The space ${\mathcal{B}}^{{\mathbf{t}}}$ is a (separable) Banach manifold modelled on the Banach space of $W^{1,p}$-sections of $u^*(TW,TL^{{\mathbf{t}}})$ (i.e. vector fields along $u$ that are tangent to $L^{{\mathbf{t}}}$ along the boundary); charts are obtained from such vector fields along $u$ by choosing a metric for which the submanifold $L^{{\mathbf{t}}}$ is totally geodesic and then applying the exponential map, see [@elia67]. The construction from Section \[subsection:variable\] shows that the map sending a disc $u\in{\mathcal{B}}$ to its level ${\mathbf{t}}$ gives ${\mathcal{B}}$ the structure of a locally trivial fibration over ${\mathbb{R}}^{n-1}$ with fibre ${\mathcal{B}}^{{\mathbf{t}}}$. Tangent vectors at $u\in{\mathcal{B}}$ can be written uniquely as ${\mathfrak{u}}+{\mathfrak{v}}|_u$, where ${\mathfrak{u}}\in T_u{\mathcal{B}}^{{\mathbf{t}}}$, and ${\mathfrak{v}}$ is a vector field as in Section \[subsection:variable\] coming from a constant vector field ${\overline{\mathfrak{v}}}$ on $\operatorname{Im}{\mathbb{C}}^{n-1}$.
The linearised Cauchy–Riemann operator
--------------------------------------
Over ${\mathcal{B}}$ we have a Banach space bundle ${\mathcal{E}}$ whose fibre over the point $u\in{\mathcal{B}}$ is the space $L^p(u^*TW)$ of $L^p$-vector fields along $u$; see for instance [@abklr94 Proposition 6.13] for the construction of the bundle structure. This bundle inherits the local product structure from ${\mathcal{B}}$.
Fix an almost complex structure $J$ on $W$ satisfying the conditions (J1) and (J2). The Cauchy–Riemann operator $u\mapsto u_x+J(u)u_y$ defines a section of ${\mathcal{E}}$. In order to discuss transversality, we need to compute the vertical differential $D_u$ of this section at $u\in{\mathcal{B}}$. To this end, consider a path of holomorphic curves $$u^s:=\psi_1^{s{\mathfrak{v}}}\circ\exp_u(s{\mathfrak{u}})$$ for $s$ in some small interval around $0$, where $\psi$ denotes the flow as in Section \[subsection:variable\]. This path is tangent to ${\mathfrak{u}}+{\mathfrak{v}}|_u$ in $s=0$. Let $\nabla$ be a torsion-free connection on $TW$. Write $$\nabla_s=\bigl(\nabla_{\partial u^s/\partial s}\bigr)|_{s=0},\;\;\;
\nabla_x=\bigl(\nabla_{\partial u^s/\partial x}\bigr)|_{s=0},$$ and likewise $\nabla_y$. Since the torsion of $\nabla$ vanishes, we have $$\nabla_s\frac{\partial u^s}{\partial x}=
\nabla_x\frac{\partial u^s}{\partial s}=\nabla_x({\mathfrak{u}}+{\mathfrak{v}}),$$ and similarly for $\partial u^s/\partial y$. Hence $$\begin{aligned}
D_u({\mathfrak{u}}+{\mathfrak{v}}|_u) & = & \nabla_s(u^s_x+J(u^s)u^s_y)\\
& = & \nabla_x({\mathfrak{u}}+{\mathfrak{v}})+J(u)\nabla_y({\mathfrak{u}}+{\mathfrak{v}})
+\bigl(\nabla_{{\mathfrak{u}}+{\mathfrak{v}}}J\bigr)(u)\, u_y\\
& = & D_u^{{\mathbf{t}}}{\mathfrak{u}}+K_u{\overline{\mathfrak{v}}},\end{aligned}$$ where $$\begin{aligned}
D_u^{{\mathbf{t}}}{\mathfrak{u}}& := & \nabla_x{\mathfrak{u}}+J(u)\nabla_y{\mathfrak{u}}+
\bigl(\nabla_{{\mathfrak{u}}}J\bigr)(u)u_y,\\
K_u{\overline{\mathfrak{v}}}& := & \nabla_x{\mathfrak{v}}+J(u)\nabla_y{\mathfrak{v}}+
\bigl(\nabla_{{\mathfrak{v}}}J\bigr)(u)u_y.\end{aligned}$$ The operator ${\overline{\mathfrak{v}}}\mapsto K_u{\overline{\mathfrak{v}}}$ is linear of order $0$ in ${\overline{\mathfrak{v}}}$, and hence a compact operator. The restriction of $D_u$ to the subspace $T_u{\mathcal{B}}^{{\mathbf{t}}}$ equals $D_u^{{\mathbf{t}}}$, which is a Fredholm operator of index $$\operatorname{index}(D_u^{{\mathbf{t}}})=\mu+n+1=n+3$$ by the index formula [@mcsa04 Theorem C.1.10] and Lemma \[lem:Maslov\]. The subspace ${\mathfrak{V}_u}\subset T_u{\mathcal{B}}$ made up of vectors of the form ${\mathfrak{v}}|_{u}$ is $(n-1)$-dimensional, and it is contained in the kernel of $D_u^{{\mathbf{t}}}$. Hence, by the invariance under compact perturbations of both the Fredholm property and the index, see [@mcsa04 Theorem A.1.4], we have — writing ${\mathbb{O}}$ for the zero operator — $$\operatorname{index}(D_u)=\operatorname{index}(D_u^{{\mathbf{t}}}+{\mathbb{O}}_{{\mathfrak{V}_u}})=
\operatorname{index}(D_u^{{\mathbf{t}}})+n-1=2n+2.$$
Regular almost complex structures {#subsection:regular}
---------------------------------
Given an almost complex structure $J$ on $W$ subject to the constraints (J1) and (J2), write ${\widetilde{\mathcal{W}}}$ for the space of *holomorphic* discs $u\in{\mathcal{B}}$, i.e.those $u$ with $u_x+J(u)u_y=0$. In other words, these are holomorphic discs satisfying condition (M1).
The almost complex structure $J$ is called *regular* if two conditions are satisfied:
- $D_u$ is onto for all $u\in{\widetilde{\mathcal{W}}}$.
- The evaluation map $$\begin{array}{ccc}
{\widetilde{\mathcal{W}}}& \longrightarrow & L^{{\mathbf{t}}}\times L^{{\mathbf{t}}}\times L^{{\mathbf{t}}}\\
u=(a,f) & \longmapsto & (f(1),f({\mathrm{i}}),f(-1))
\end{array}$$ is transverse to $({\mathbb{R}}\times L^{{\mathbf{t}}}_1)\times({\mathbb{R}}\times L^{{\mathbf{t}}}_{{\mathrm{i}}})
\times({\mathbb{R}}\times L^{{\mathbf{t}}}_{-1})$, where $L^{{\mathbf{t}}}_{e^{{\mathrm{i}}\theta}}:=
L^{{\mathbf{t}}}\cap\{z_0=e^{{\mathrm{i}}\theta}\}$.
If the first condition is satisfied, ${\widetilde{\mathcal{W}}}$ will be a manifold of the expected dimension $2n+2$; if in addition (ii) holds, then ${\mathcal{W}}$ will be a manifold of dimension $2n-1$.
The proof that the set of regular $J$ is non-empty, in fact of second Baire category, follows the standard line of reasoning as in the proof of Theorems 3.1.5 and 3.4.1 of [@mcsa04]. Selecting such a regular $J$ is the generic choice we make in (J2). For the standard discs $u^{{\mathbf{t}}}_{{\mathbf{s}},b}$, transversality is obvious. By Lemma \[lem:standard1\], all discs that are not standard pass through the region where $J$ may be chosen generically, which is sufficient to achieve transversality by [@mcsa04 Remark 3.2.3]. In contrast with the set-up in [@mcsa04], we are only allowed to perturb $J$ along $\xi$, keeping it compatible with ${\mathrm{d}}{\hat{\alpha}}$. But this is exactly the situation dealt with by Bourgeois in the appendix of [@bour06]. The proof given there carries over to our situation; the essential ingredient of Bourgeois’s argument is that the set of $f$-injective points is open and dense, which is precisely our Lemma \[lem:f-injective\].
Orientation
-----------
In order to speak of the degree of the evaluation map $\operatorname{ev}$ on ${\mathcal{W}}\times{\mathbb{D}}$, we need to put an orientation on the moduli space ${\mathcal{W}}$. Given the relation between ${\mathcal{W}}$ and ${\widetilde{\mathcal{W}}}$ described in the preceding section, it suffices to orient ${\widetilde{\mathcal{W}}}$, and that in turn amounts to showing that the determinant line bundle $\det D$ over ${\widetilde{\mathcal{W}}}$ is oriented, since $\ker D_u=T_u{\widetilde{\mathcal{W}}}$.
Recall that the determinant line $\det F$ is defined for any Fredholm operator $F$ as $\det F=\det\ker F\otimes
(\det\operatorname{\mathrm{coker}}F)^*$. Since $D_u$ is surjective for all $u\in{\widetilde{\mathcal{W}}}$, the determinant line bundle is simply $\det\ker D=\bigwedge^{2n+2}\ker D$. In the arguments that follow, however, we use deformations through not necessarily surjective Fredholm operators, so we need to work with determinant lines, in general.
As we have seen, the operator $D_u$ splits (by slight abuse of notation) as $D_u=D_u^{{\mathbf{t}}}+K_u$. The linear interpolation of $D_u$ to $D_u^{{\mathbf{t}}}+{\mathbb{O}}_{{\mathfrak{V}_u}}$ is via Fredholm operators, since $K_u$ is compact. It follows that $\det(D_u)=\det(D_u^{{\mathbf{t}}}+{\mathbb{O}}_{{\mathfrak{V}_u}})$, see [@fooo09 p. 680], whence $$\det(D_u)=\det D_u^{{\mathbf{t}}}\otimes\det{\mathfrak{V}_u}.$$ The second factor inherits a natural orientation from the orientation of ${\mathbb{R}}^{n-1}$. The first factor is naturally oriented by the construction in [@fooo09 Section 8.1]. Our situation is a particularly simple one, since $TL^{{\mathbf{t}}}$ is a trivial bundle. This implies that any bundle pair $(u^*TW,(u|_{\partial{\mathbb{D}}})^*TL^{{\mathbf{t}}})$ comes with a natural trivialisation of the boundary bundle, and this suffices for the construction of a natural orientation of the determinant line bundle.
Proof of Theorem \[thm:main\]
=============================
By Sections \[section:compactness\] and \[section:transversality\] (notably Lemma \[lem:standard2\]), the assumption $\inf_0(\alpha)>\pi$ of Theorem \[thm:main\] implies that the evaluation map $$\begin{array}{rccc}
\operatorname{ev}\co & {\mathcal{W}}\times{\mathbb{D}}& \longrightarrow & {\widehat{Z}}\\
& \bigl( (a,f),z\bigr) & \longmapsto & f(z)
\end{array}$$ is a proper map of degree $1$. By Lemmata \[lem:standard1\] and \[lem:standard3\], we may pretend that ${\mathcal{W}}\times{\mathbb{D}}$ and ${\widehat{Z}}$ are — after smoothing corners — compact, oriented manifolds with boundary, without changing the homotopy type of these spaces, and that $\operatorname{ev}$ is a smooth degree $1$ map between these manifolds.
Homotopical and homological arguments similar to the ones that follow were used by Eliashberg–Floer–McDuff, see [@mcdu91].
\[prop:pi1\] The manifold ${\widehat{Z}}$ is simply connected.
Given a loop in ${\widehat{Z}}$, we homotope it to an embedded circle $C$ inside $\operatorname{Int}({\widehat{Z}})$ that intersects the complement of ${\widehat{B}}$, in other words, such that it passes through the region where all holomorphic discs (more precisely, their $f$-components) are standard. We can make the evaluation map $$\begin{array}{ccc}
{\mathcal{W}}\times{\mathbb{D}}& \longrightarrow & {\widehat{Z}}\\
\bigl( (a,f),z\bigr) & \longmapsto & f(z)
\end{array}$$ transverse to $C$ by a perturbation compactly supported in $\operatorname{Int}({\widehat{B}})$. The preimage of $C$ under this perturbed map will then be a single circle $C'\subset{\mathcal{W}}\times{\mathbb{D}}$ mapping with degree 1 onto $C$. The homotopy of $C'$ to a loop in ${\mathcal{W}}\times\{1\}$ induces a homotopy of $C$ to a loop in the cell ${\mathbb{R}}\times\{1\}\times{\mathbb{C}}^{n-1}$.
Let $\phi\co (P,\partial P)\rightarrow (Q,\partial Q)$ be a degree $1$ map between compact, oriented $m$-dimensional manifolds with boundary. Then the induced homomorphism $\phi_*\co H_k(P;{\mathbb{F}})\rightarrow
H_k(Q;{\mathbb{F}})$ in singular homology with coefficients in a field ${\mathbb{F}}$ is surjective in each degree $k\in{\mathbb{N}}_0$.
Over a field, the Kronecker pairing between homology and cohomology is non-degenerate, so equivalently we need to show injectivity of the induced homomorphism $\phi^*$ in cohomology.
Given a non-zero class $\beta\in H^k(Q)$, Poincaré duality allows us to find a class $\gamma\in H^{m-k}(Q,\partial Q)$ such that $\beta\cup\gamma$ is the orientation generator of $H^m(Q,\partial Q)$. Since $\phi$ is of degree $1$, we have $$0\neq\phi^*(\beta\cup\gamma)=\phi^*\beta\cup\phi^*\gamma,$$ which forces $\phi^*$ to be injective on $H^k(Q)$.
\[prop:H\*\] The manifold ${\widehat{Z}}$ has the integral homology of a point.
With the preceding lemma this follows with an argument completely analogous to the proof of Proposition \[prop:pi1\].
Since $2n\neq 3$, the smooth Schoenflies theorem tells us that the subset of ${\widehat{Z}}$ bounded by $\varphi(\partial M)$ and a standard ellipsoid surrounding $\varphi(\partial M)$ is diffeomorphic to a collar of $\partial M$. Hence $M$ is a strong deformation retract of ${\widehat{Z}}$. So by Propositions \[prop:pi1\] and \[prop:H\*\], the manifold $M$ is a simply connected homology ball with boundary diffeomorphic to $S^{2n}$. It follows that $M$ is diffeomorphic to a ball: for $n\geq 3$ we appeal to Proposition A on page 108 of Milnor’s lectures [@miln65]; for $n=2$, to Proposition C on page 110.
We thank Peter Albers for useful conversations about compactness questions, and Chris Wendl for drawing our attention to Frédéric Bourgeois’s work on transversality in the setting of symplectisations. A part of the work on this paper was done during the workshop on Legendrian submanifolds, holomorphic curves and generating families at the Académie Royale de Belgique, August 2013, organised by Frédéric Bourgeois.
[10]{} , *Symplectic Geometry*, Progr. Math. **124**, Birkhäuser, Basel (1994). , Closed characteristics on non-compact hypersurfaces in ${\mathbb{R}}^{2n}$, *Math. Ann.* **343** (2009), 247–284. , Closed characteristics on non-compact mechanical contact manifolds, preprint (2013), [arXiv: 1303.6461]{}. , Contact homology and homotopy groups of the space of contact structures, *Math. Res. Lett.* **13** (2006), 71–85. , A Hamiltonian characterization of the three-ball, *Differential Integral Equations* **7** (1994), 1303–1324. , Geometry of manifolds of maps, *J. Differential Geometry* **1** (1967), 169–194. , *Lagrangian Intersection Floer Theory: Anomaly and Obstruction, Part II*, AMS/IP Stud. Adv. Math. **46.2**, American Mathematical Society, Providence, RI (2009). , *An Introduction to Contact Topology*, Cambridge Stud. Adv. Math. **109**, Cambridge University Press (2008). , Trapped Reeb orbits do not imply periodic ones, *Invent. Math.*, to appear. , Symplectic cobordisms and the strong Weinstein conjecture, *Math. Proc. Cambridge Philos. Soc.* **153** (2012), 261–279. , How to recognize a $4$-ball when you see one, *Münster J. Math.* [**6**]{} (2013), 525–554; erratum: pp. 555–556. , Pseudoholomorphic curves in symplectizations with applications to the Weinstein conjecture in dimension three, *Invent. Math.* **114** (1993), 515–563. , Properties of pseudoholomorphic curves in symplectizations III: Fredholm theory, in *Topics in Nonlinear Analysis*, Progr. Nonlinear Differential Equations Appl. [**35**]{}, Birkhäuser, Basel (1999), 381–475. , *Gromov’s Compactness Theorem for Pseudo-holomorphic Curves*, Progr. Math. [**151**]{}, Birkhäuser, Basel (1997). , Relative frames on $J$-holomorphic curves, *J. Fixed Point Theory Appl.* **9** (2011), 213–256. , Symplectic manifolds with contact type boundaries, *Invent. Math.* **103** (1991), 651–671. , *$J$-holomorphic Curves and Symplectic Topology*, Amer. Math. Soc. Colloq. Publ. **52**, American Mathematical Society, Providence, RI (2004). , *Lectures on the $h$-Cobordism Theorem*, Princeton University Press, Princeton, NJ (1965). , The plastikstufe – a generalization of the overtwisted disk to higher dimensions, *Algebr. Geom. Topol.* **6** (2006), 2473–2508. , Linking and closed orbits, preprint (2013), [arXiv: 1305.2799]{}. , Contact surgery and symplectic handlebodies, *Hokkaido Math. J.* **20** (1991), 241–251. , The annulus property of simple holomorphic discs, *J. Symplectic Geom.* [**11**]{} (2013), 135–161.
[^1]: H. G. and K. Z. are partially supported by DFG grants GE 1245/2-1 and ZE 992/1-1, respectively.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Landau-Zener-Stückelberg (LZS) interference of continuously driven superconducting qubits is studied. Going beyond the second order perturbation expansion, we find a time dependent stationary population evolution as well as unsymmetrical microwave driven Landau-Zener transitions, resulting from the nonresonant terms which are neglected in rotating-wave approximation. For the low-frequency driving, the qubit population at equilibrium is a periodical function of time, owing to the contribution of the nonresonant terms. In order to obtain the average population, it is found that the average approximation based on the perturbation approach can be applied to the low-frequency region. For the extremely low frequency which is much smaller than the decoherence rate, we develop noncoherence approximation by dividing the evolution into discrete time steps during which the coherence is lost totally. These approximations present comprehensive analytical descriptions of LZS interference in most of parameter space of frequency and decoherence rate, agreeing well with those of the numerical simulations and providing a simple but integrated understanding to system dynamics. The application of our models to microwave cooling can obtain the minimal frequency to realize effective microwave cooling.'
address: 'National Laboratory of Solid State Microstructures and Department of Physics, Nanjing University, Nanjing 210093, China'
author:
- 'Lingjie Du,Yang Yu'
title: 'Low-frequency Landau-Zener-Stückelberg interference in dissipative superconducting qubits'
---
Introduction
============
Superconducting devises based on Josephson tunnel junctions have great importance in demonstrating quantum phenomena at macroscopic scales and hold promise for applications in quantum information. [@makhlin; @nc; @nori; @mooij; @clark] In the studies of these systems, crossovers between energy levels are usually associated with a wide variety of interesting phenomena, particularly in connection with Landau-Zener-Stückelberg (LZS) interference. [@sn; @shytove; @landau]
Recently, a series of great progresses [oliver,sillanpaa,naka,ssaito,lzm1,wilson,bernnature,sun,jb,oliver1,bernsprl,rudner]{} have been implemented to study the coherent dynamics of qubits in the regime dominated by LZS interference. The strong high-frequency microwave generates controllable population transitions, allowing for fast and reliable control of quantum systems. The delicate LZS interference patterns also provide a way to calibrate the parameters of the qubit and its interaction with the environment. On the other hand, through continuous Landau-Zener (LZ) transitions, the low-frequency microwave can be used to achieve an effective temperature of qubit up to two orders of magnitude lower than the bath temperature through an analogue of optical sideband cooling. [@va; @du2] This microwave cooling provides an effective means for preparing qubit state with high fidelity and suppressing decoherence. Therefore, understanding dynamics of LZS interference in strongly driven superconducting qubits has both fundamental and practical significances. A number of theoretical investigations [sn,wilson,goy,gri,goo,du,du1,son,fol,hau,ashhab,liu,wilson1,green,sn1,sn2,wub,zq,ksaito,pegg]{} have been developed in driven quantum two-level systems (TLS). These works generally addressed the coherent phenomena in the high-frequency driving in superconducting qubits, as reviewed recently in Ref. 6. In the realistic situation, the influence of environment always needs to be considered. Compared with natural atoms, the solid state systems contain macroscopic degrees of freedom strongly coupled to environment bathes that decohere the quantum states to be controlled. To address the driven dissipative TLS, two typical methods usually taken into account are rotating-wave approximation (RWA) [@ashhab; @pegg] and perturbation-RWA. [@bernsprl; @du1]
RWA provides a simple physical picture and good approximation for the stationary population. However, it limits at the high-frequency region, i.e. $\omega \gg \Gamma _{2},\Delta J_{n}\left( A/\omega \right) $, where $n$ is the closest integer to $\varepsilon _{0}/\omega $ and $\varepsilon _{0}$ is the dc detuning. The other method, perturbation-RWA, applies the second order perturbation approximation based on RWA and can be used in the parameter region with the high frequency and strong decoherence, i.e. $\omega >\Delta ,\Gamma _{2}$ and $\Gamma _{2}\gg $ $W_{01}$, where $W_{01}$ is microwave driven Landau-Zener (MDLZ) transition rate from $|0\rangle $ and $|1\rangle $. However in the low-frequency driving where microwave cooling takes effects, both methods would be inappropriate.
In this article, we focus on analyzing the stationary state which is useful in spectroscopy experiments [@oliver; @sillanpaa; @ssaito; @bernnature] and microwave cooling [@va] in the presence of decoherence. Going beyond the second order perturbation and focusing on the nonresonant terms which are dropped in RWA, we present the real-time population evolution as well as unsymmetrical MDLZ transitions in a strongly driven TLS of the long-time dynamics. In the high-frequency case, i.e. $\omega >\Delta $, the result of perturbation approach can fully cover those of RWA and perturbation-RWA. In addition, the perturbation approach can be applied to larger parameters space, at which other methods are inappropriate.
Then we move on lower frequency region, i.e. $\omega <\Delta $. In this parameter region, microwave cooling begins to take effects, motivating us to give a thorough understanding of the dynamics. The steady dynamics shows a periodic time dependence of qubit population, which is from the nonresonant terms and cannot be explained by perturbation-RWA. Then we focus on the average stationary population. Although one can use the perturbation approach, it is cumbersome thus difficult to get the final solutions. Therefore, we use the average perturbation approximation (APA), which is based on perturbation approach through the average over one driving period. Interestingly, the analytical expression in APA has the same form with that of perturbation-RWA. If the frequency further decreases, APA cannot correctly give the qubit population. We thereby propose the noncoherence approximation (NCA). The incoherent evolution is divided into discrete incoherent time steps, from which we can obtain the analytical solution. It is found that when the frequency is much smaller than the decoherence rate, the time dependent and average population calculated by NCA agree quantitatively with those of the numerical simulation. At last, we extend the analytical results of APA and NCA to the multi-level system. We discuss microwave cooling [@va] and obtain the minimal frequency of the effective microwave cooling.
Through these approximations of perturbation approach, we present simple and comprehensive descriptions in the average stationary population in most of parameter space of frequency and decoherence rate. The agreement with numerical results indicates that the picture captures the underlying physics, thus giving a better understanding of the behavior of system dynamics. The models and methods we used here are also valid for other TLSs with similar structures.
The article is organized as follows. In section. 2, we introduce the basic model and approximations. In section. 3, assuming high order perturbation in the long-time dynamics, we present the perturbation approach which considers the nonresonant terms. Then with this approach we discuss the high-frequency driving case. In section. 4, We discuss the average stationary population in the low frequency. Through APA, we obtain the average population in the steady case under the low-frequency driving. In section. 5, for the extremely low frequency, we demonstrate that NCA can be used to get the analytical solution, which is useful to describe the system dynamics. In section. 6, we extend our model to multi-level flux qubit and discuss the frequency range of effective microwave cooling.
Basic model
===========
{width="7.5cm"}
For a superconducting flux qubit [@mooij2; @chio2] with three Josephson tunnel junctions, if the external flux bias $\Phi \approx 0.5\Phi _{0}$, where $\Phi _{0}=h/2e$ is the flux quantum, a double well landscape of the potential energy parameterized by the dc flux detuning $\delta \Phi
_{dc}=\Phi -0.5\Phi _{0}$ exists as shown in Fig. 1. In the millikelvin temperature $\sim 10$ mK, a series of diabatic states are localized in different wells. States $|0\rangle $ and $|1\rangle $ are the lowest levels in the left and right wells respectively. They are characterized by different flowing directions of the persistent currents, namely clockwise and anticlockwise. The states in each well are separated by large energy spacing on the order of 20 GHz, [@bernnature] while the coupling strength between two wells is small compared with the energy spacing. As the qubit is driven with a microwave $\Phi _{ac}=\Phi _{rf}\sin \omega t$, where $\Phi _{rf}$ is the microwave amplitude, the time dependent flux detuning is$$\begin{aligned}
\delta \Phi (t)=\delta \Phi _{dc}+\Phi _{ac}=\delta \Phi _{dc}+\Phi
_{rf}\sin \omega t. \nonumber\end{aligned}$$Then the time dependent energy detuning of states $|0\rangle $ and $|1\rangle $ can be described as$$\begin{aligned}
\varepsilon (t)=\varepsilon _{0}+A\sin \omega t, \nonumber\end{aligned}$$where $\varepsilon _{0}=(|m_{0}|+|m_{1}|)\delta \Phi _{dc}$ is the dc energy detuning of states $|0\rangle $ and $|1\rangle $, $A=(|m_{0}|+|m_{1}|)\Phi
_{rf}$ is the microwave energy amplitude, and $m_{i}=dE_{i}(\Phi )/d\Phi $ is the diabatic energy-level slope of state $|i\rangle $ in units of frequency per flux. For the initial time $t=0$, the qubit is at the detuning $\varepsilon (0)=\varepsilon _{0}$.
We start from a periodically driven TLS coupled to an environment bath. Such a process can be described by Hamiltonian (we set $\hbar =k_{B}=1$) [du,du1,ksaito,amin,Amin2]{}$$\begin{aligned}
H=-\frac{\Delta }{2}\sigma _{x}-\frac{\varepsilon (t)}{2}\sigma
_{z}+H_{int}+H_{B},\end{aligned}$$where $\sigma _{x}$ and $\sigma _{z}$ are Pauli matrices, $\sigma
_{z}=|1\rangle \langle 1|-|0\rangle \langle 0|$ and $\sigma _{x}=|1\rangle
\langle 0|-|0\rangle \langle 1|$, $H_{int}$ is the system-environment interaction Hamiltonian, and $H_{B}$ is the environment Hamiltonian. We assume the qubit is embedded in the bath in equilibrium at temperature $T$ with the density matrix of environment $\rho _{B}=e^{-H_{B}/T}$. The system-environment interaction Hamiltonian is written as $$\begin{aligned}
H_{int}=-Q\sigma _{z}/2,\end{aligned}$$where $Q$ is an operator acting on the environment.
{width="7.5cm"}
Before investigating the system dynamics in the presence of the environment bath, it is worth recalling main features of its free time evolution which is described by Hamiltonian Eq. (1) without $H_{int}$ and $H_{B}$. By a gauge transformation, the Hamiltonian is brought to the form$$\begin{aligned}
H_{1}=-\frac{\Delta }{2}e^{\phi (t)}|1\rangle \langle 0|+H.c.\end{aligned}$$where $\phi (t)=-i\int_{0}^{t}\varepsilon (\tau )d\tau $. Using Bessel functions, we have $e^{\phi (t)}=e^{-i\varepsilon _{0}t-iA/\omega }\sum
J_{n}(A/\omega )e^{in(\omega t+\pi /2)}$ and find $$\begin{aligned}
H_{2}=-\frac{\Delta }{2}\sum J_{n}\left( \frac{A}{\omega }\right)
e^{-i(\varepsilon _{0}t-n\omega t)+in\pi /2-iA/\omega }|1\rangle \langle
0|+H.c.\end{aligned}$$With the relation $e^{iA/\omega }=\sum J_{n}(A/\omega )e^{in\pi /2}$, we consider parameters close to the $n$-photon resonance, where $n\omega
=\varepsilon _{0}$, and rely on RWA to neglect the fast oscillating terms with $k\neq n$ for the high frequency $\omega $. Therefore, when the system is initially in state $|1\rangle $ we have the population in state $|0\rangle $ [@sn; @oliver; @ashhab] $$\begin{aligned}
\fl\rho _{00}=\frac{1}{2}\sum \frac{\Delta ^{2}J_{n}^{2}(A/\omega )}{\Delta
^{2}J_{n}^{2}(A/\omega )+(\varepsilon _{0}-n\omega )^{2}}(1-\cos \sqrt{\Delta ^{2}J_{n}^{2}(A/\omega )+(\varepsilon _{0}-n\omega )^{2}}t).\end{aligned}$$
This approximation neglects all the nonresonant terms which oscillate very fast in the time scale of system dynamics, so the high frequency is needed to keep the validity of RWA, i.e. $\omega >\Delta \gg $ $\Delta J_{n}\left(
A/\omega \right) $. [@sn; @son; @ashhab] We compare the real-time numerical solution with the analytical results in Eq. (5), shown in Fig. 2. RWA has the high accuracy under the high-frequency driving. However, when the frequency is low, i.e. $\omega <\Delta $, the difference between numerical and RWA results are not negligible. It should be noted that in Fig. 2 (a) there exist two kinds of oscillations. One agrees with Eq. (5) while another as shown in the panel in Fig. 2 (a) is from the nonresonant terms in Eq. (4). When the frequency is small just as shown in Fig. 2 (b), the effect of nonresonant terms could not be neglected. In next sections, we will further show the influence of these nonresonant terms on system dynamics.
In the realistic solid-state system, the decoherence which is always strong needs to be taken into account. In this case, the Bloch equations are usually used to provide the numerical stimulation, which read $$\begin{aligned}
\dot{\rho}=-i[H_{1},\rho ]+\Gamma \lbrack \rho ], \label{6}\end{aligned}$$here $\Gamma \lbrack \rho ]$ is the dissipative term [@averin; @abr]$$\begin{aligned}
\Gamma \lbrack \rho ]=\left[
\begin{array}{cc}
-\Gamma _{10}\rho _{11}+\Gamma _{01}\rho _{00} & -\Gamma _{2}\rho _{01} \\
-\Gamma _{2}\rho _{10} & \Gamma _{10}\rho _{11}-\Gamma _{01}\rho _{00}\end{array}\right] ,\end{aligned}$$where $\Gamma _{2}=1/T_{2}$ is the decoherence rate, $T_{2}$ is the decoherence time, $\Gamma _{ij}$ is the interwell relaxation rate from state $|i\rangle $ to $|j\rangle $, $\Gamma _{10}=\Gamma _{01}e^{-\varepsilon
_{0}/T}$, $1/T_{1}=\max \{\Gamma _{10},\Gamma _{01}\}$, and $T_{1}$ is the interwell relaxation time. After the long-time system dynamical evolution, whatever the initial state is, the temporal oscillation in Fig. 2 is eliminated by the decoherence and the qubit population will converge to a stationary solution. We make RWA with the equation Eq. (6), neglect the fast oscillating terms, and obtain the probability in state $|0\rangle $ [@sn]
$$\begin{aligned}
\rho _{00}=\sum_{n}\frac{\frac{\Delta ^{2}J_{n}^{2}(A/\omega )\Gamma _{2}}{2[\Gamma _{2}^{2}+(\varepsilon _{0}-n\omega )^{2}]}+\Gamma _{10}}{\frac{\Delta ^{2}J_{n}^{2}(A/\omega )\Gamma _{2}}{\Gamma _{2}^{2}+(\varepsilon
_{0}-n\omega )^{2}}+\Gamma _{01}+\Gamma _{10}}. \label{8}\end{aligned}$$
To ensure Eq. (8) is valid, we need not only the high frequency, i.e. $\omega >\Delta \gg $ $\Delta J_{n}\left( \frac{A}{\omega }\right) $, but also the high coherence, i.e. $\omega \gg \Gamma _{2}$. If the decoherence is strong $\omega \sim \Gamma _{2}$, the population $\rho _{00}$ may even be larger than 1 and this stationary solution is obviously not reasonable.
In order to describe the dynamics in the strong decoherence, Berns *et al.* [@bernsprl] proposed the perturbation-rotating-wave approximation. Under the Gaussian white low-frequency noise, based on the second order expansion perturbation, we have MDLZ transition rate from state $|0\rangle $ to $|1\rangle $ $$\begin{aligned}
W_{01}=\frac{\Delta ^{2}}{2}\sum_{n}\frac{\Gamma _{2}J_{n}^{2}(A/\omega )}{(\varepsilon _{0}-n\omega )^{2}+\Gamma _{2}^{2}}.\end{aligned}$$Correspondingly, $W_{10}(\varepsilon _{0})=W_{01}(-\varepsilon
_{0})=W_{01}(\varepsilon _{0})$. Then for the long-time dynamics, we employ the rate equations, in which the qubit population in state $|0\rangle $ and $|1\rangle $, i.e. $\rho _{00}$ and $\rho _{11}$, follow $$\begin{aligned}
\dot{\rho}_{00}=-(W_{01}+\Gamma _{01})\rho _{00}+(W_{10}+\Gamma _{10})\rho
_{11} \nonumber \\
1=\rho _{00}+\rho _{11}.\end{aligned}$$Since MDLZ transition rate is constant in the stationary case, the equations would lead to the constant population, i.e. $\dot{\rho}_{00}=\dot{\rho}_{11}=0$, and the population in state $|0\rangle $ becomes $$\begin{aligned}
\rho _{00}=\sum\limits_{n}\frac{\frac{\Delta ^{2}J_{n}^{2}(A/\omega )\Gamma
_{2}}{2[\Gamma _{2}^{2}+(\varepsilon _{0}-n\omega )^{2}]}+\Gamma _{10}}{\sum\limits_{n}\frac{\Delta ^{2}J_{n}^{2}(A/\omega )\Gamma _{2}}{\Gamma
_{2}^{2}+(\varepsilon _{0}-n\omega )^{2}}+\Gamma _{01}+\Gamma _{10}}.\end{aligned}$$
{width="15cm"}
In order to ensure the validity of this approach,[@bernsprl; @du1] the high frequency and strong decoherence, i.e. $\omega >\Delta ,\Gamma _{2}$ and $\Gamma _{2}\gg $ $W_{01}$, are necessary, which greatly restrict the application of this approximation. In next sections, we will introduce approximations to extend the quantitative analysis to larger parameters space.
Perturbation approach
=====================
In order to get analytical results in larger parameters space, we go back to Hamiltonian (1) and take the perturbation approach. First of all, we follow the process in Ref. 27 and 44, and the complete time evolution operator in the laboratory picture can be expressed as$$\begin{aligned}
U(t_{1},t_{2})=e^{-iH_{B}t_{1}}U_{0}(t_{1})U_{1}(t_{1})U_{V}(t_{1},t_{2})U_{1}^{\dag }(t_{2})U_{0}^{\dag }(t_{2})e^{iH_{B}t_{2}},\end{aligned}$$where $U_{0}(t)=\mathcal{T}\exp [\frac{i}{2}\int_{0}^{t}d\tau \varepsilon
(\tau )\sigma _{z}]$, $U_{1}(t)=\mathcal{T}\exp [\frac{i}{2}\int_{0}^{t}d\tau Q(\tau )\sigma _{z}]$, $Q(t)=e^{iH_{B}t}Qe^{-iH_{B}t},U_{V}(t_{1},t_{2})=\mathcal{T}\exp
[-i\int_{t_{2}}^{t_{1}}d\tau V(\tau )]$, $V(t)=-\frac{\Delta }{2}\sum\limits_{n}J_{n}(A/\omega )\exp [-i(\varepsilon _{0}t-n\omega t)+in\pi
/2-iA/\omega ]U_{1}^{\dag }(t)|1\rangle \langle 0|U_{1}(t)+H.c.$, and $\mathcal{T}$ is the time ordering operator. Let $\rho _{SB}$ express the total density matrix of the system plus environment bath, which can be described as $\rho _{SB}(t)=U(t,0)\rho _{SB}(0)U^{\dag }(t,0)$. Then the system reduced density matrix is defined as $\rho (t)=$ Tr$_{B}\{\rho
_{SB}(t)\}$, where Tr$_{B}\{...\}$ denotes the trace over the environmental degrees of freedom. We assume at the initial time the density matrix is separable, i.e. $\rho _{SB}(0)=\rho (0)\otimes \rho _{B}$, and the system is initialized in state $|0\rangle $, i.e. $\rho (0)=|0\rangle \langle 0|$.
We perform a perturbation expansion of $U_{V}$ in $V(\tau )$ , which reads
$$\begin{aligned}
\fl U_{V}(t,0)\approx 1-i\int_{0}^{t}d\tau V(\tau )-\int_{0}^{t}d\tau V(\tau
)\int_{0}^{\tau }d\tau ^{\prime }V(\tau ^{\prime }) \nonumber \\ \fl
+i\int_{0}^{t}d\tau
V(\tau )\int_{0}^{\tau }d\tau ^{\prime }V(\tau ^{\prime })\int_{0}^{\tau
^{\prime }}d\tau ^{\prime \prime }V(\tau ^{\prime \prime })+...\end{aligned}$$
For the strong decoherence, i.e. $\Gamma _{2}\gg $ $\Delta J_{n}\left(
\frac{A}{\omega }\right) \sim W_{01}$, we can consider the expansion to the second order in a time interval $1/\Gamma _{2}\ll t<1/[\Delta J_{n}(A/\omega
)]$. [@du1; @Amin2] However, in the high coherence region where $\Gamma
_{2}<\Delta J_{n}(A/\omega )$, instead of considering the time interval $t<1/[\Delta J_{n}(A/\omega )]$, we need take account of a time interval $t$ much larger than the decoherence time. Hence this time interval $t$ is also much larger than $1/[\Delta J_{n}(A/\omega )]$. Moreover, for the low frequency $\Delta J_{n}(A/\omega )\lesssim\omega <\Delta $, in order to calculate MDLZ transition in the stationary case, even if $1/[\Delta
J_{n}(A/\omega )]$ is much longer than the decoherence time, the time interval $t<1/[\Delta J_{n}(A/\omega )]$ is still not sufficiently long relative to the driving period. As a result we consider $t$ $\gg 1/\omega
,1/[\Delta J_{n}(A/\omega )],1/\Gamma _{2}$. It is clear that the second order perturbation expansion would be insufficient. [@du1; @Amin2] Considering high order expansion terms, we obtain the population in state $|1\rangle $ (see appendix)
$$\begin{aligned}
\fl\rho _{11}(t)\approx \int_{0}^{t}\int_{0}^{t}d\tau d\tau ^{\prime
}Tr_{B}\{\langle 1|V(\tau )\rho _{SB}(0)V^{\dag }(\tau ^{\prime })|1\rangle
\} \nonumber \\\fl
-2\int_{0}^{t}\int_{0}^{t_{1}}dt_{1}dt_{2}Tr_{B}\{\langle
1|[\int_{0}^{t}\int_{0}^{t}dt_{3}dt_{4}V(t_{4})\rho _{SB}(0)V^{\dag
}(t_{3})]V^{\dag }(t_{2})V^{\dag }(t_{1})|1\rangle \} \nonumber \\ \fl
-Tr_{B}\{\int_{0}^{t}\int_{0}^{t}dt_{1}dt_{4}\langle
1|V(t_{4})[2\int_{0}^{t}\int_{0}^{t_{2}}dt_{2}dt_{3}\rho _{SB}(0)V^{\dag
}(t_{3})V^{\dag }(t_{2})]V^{\dag }(t_{1})|1\rangle \}+... \label{14}\end{aligned}$$
For the short time interval, i.e. $t<1/[\Delta J_{n}(A/\omega )]$, we use the second order perturbation expansion and can keep only the first term in Eq. (14) expressing the population in state $|1\rangle $ transferred from $|0\rangle $, just as discussed in Ref. 27. As the time goes on, the population is not completely in state $|0\rangle $. The first term does not consider the effect of the reduced population in state $|0\rangle $ such that it cannot characterize the population transferred from state $|0\rangle
$ to $|1\rangle $ any more. Higher order terms in $U_{V}(t,0)$ need to be considered, which result into the second as well as third term in Eq. (14).In the second term, $\int_{0}^{t}\int_{0}^{t}dt_{3}dt_{4}V(t_{4})\rho
_{SB}(0)V^{\dag }(t_{3})$ describes the population transferred from state $|0\rangle $ to $|1\rangle $ in the time interval $t$. Then the second term characterizes the transition of population from state $|1\rangle $ to $|0\rangle $ based on this part of population in $|1\rangle $. Similarly, $2\int_{0}^{t}\int_{0}^{t_{2}}dt_{2}dt_{3}\rho _{SB}(0)V^{\dag
}(t_{3})V^{\dag }(t_{2})$ in the third term describes the reduced population in state $|0\rangle $ in the time interval $t$. The third term describes the decrease of population transferred from state $|0\rangle $ to $|1\rangle $ because of the reduced population in state $|0\rangle $. When the time interval $t$ becomes further larger, more terms will appear in Eq. (14) as a ripple effect and we can also extend the integration limit $t$ to $\infty $.
Then we only discuss the first term in Eq. (14) as an example, and other terms can be treated with a similar procedure. We rewrite the first term of Eq. (14) to $$\begin{aligned}
y(t)=\frac{\Delta ^{2}}{4}\int_{0}^{t}dt_{1}\int_{0}^{t}dt_{2}K\langle
U_{-}^{+}(t_{1})U_{+}(t_{1})U_{+}^{+}(t_{2})U_{-}(t_{2})\rangle \\
K=\sum_{n,m}J_{n}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })e^{i(\varepsilon _{0}-m\omega )t_{1}-i(\varepsilon _{0}-n\omega
)t_{2}+in\pi /2-im\pi /2}, \nonumber\end{aligned}$$where $U_{\pm }(t)=\mathcal{T}\exp [\pm (i/2)\int_{0}^{t}Q(\tau )d\tau ]$. For the low frequency, i.e. $\Delta J_{n}(A/\omega )\lesssim \omega <\Delta $, the terms in $K$ with $n\neq m$ would oscillate not fast enough, leading the invalidity of RWA. Assuming the white Gaussian low-frequency noise, we rewrite Eq. (15) to$$\begin{aligned}
y(t)=\frac{\Delta ^{2}}{4}\int_{0}^{t}dt_{1}\int_{0}^{t}dt_{2}\sum_{m}e^{i(\varepsilon _{0}-m\omega )(t_{1}-t_{2})-\Gamma _{2}|t_{1}-t_{2}|} \nonumber \\
\{J_{m}^{2}(\frac{A}{\omega })+\sum_{n\neq 0}J_{n+m}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })e^{in\omega t_{2}+in\pi /2}\}.\end{aligned}$$For the long-time dynamics, we extend the integration limits of Eq. (16) to $\infty $, and have [@du1] $$\begin{aligned}
\fl\dot{y}(t)=\frac{\Delta ^{2}}{2}\sum_{n}\frac{\Gamma _{2}J_{n}^{2}(A/\omega )}{(\varepsilon _{0}-n\omega )^{2}+\Gamma _{2}^{2}}+f(t) \label{17} \\\fl
f(t)=\frac{\Delta ^{2}}{4}\sum_{n\neq 0,m}\lim_{t\rightarrow \infty }\frac{d}{dt}\int_{0}^{t}dt_{1}\int_{0}^{t}dt_{2}J_{n+m}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })e^{i(\varepsilon _{0}-m\omega )\tau +in\omega t_{2}+in\pi
/2-\Gamma _{2}|\tau |}, \nonumber\end{aligned}$$where $\tau =t_{1}-t_{2}$. The first term in Eq. (17) is the familiar result in perturbation-RWA. For convenience we change the form of $f(t)$ to$$\begin{aligned}
\fl f(t)=\frac{\Delta ^{2}}{4}\sum_{n\neq 0,m}J_{n+m}(\frac{A}{\omega })J_{m}(\frac{A}{\omega }) \nonumber \\ \fl
\lim_{t\rightarrow \infty }\frac{d}{dt}\int_{0}^{t}d\tau
^{\prime }\int_{-\tilde{t}}^{\tilde{t}}d\tau e^{i(\varepsilon _{0}-m\omega
)\tau +in\omega (\tau ^{\prime }-\tau /2)+in\pi /2-\Gamma _{2}|\tau |},\end{aligned}$$where $\tau ^{\prime }=(t_{1}+t_{2})/2$, and $\tilde{t}=\min \{2\tau
^{\prime },2t-2\tau ^{\prime }\}$. Therefore we have the leading term given by $$\begin{aligned}
f(t)\approx \frac{\Delta ^{2}}{2}\sum_{n\neq 0,m}J_{n+m}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })\frac{\Gamma _{2}\cos n(\omega t+\pi /2)}{(\varepsilon _{0}-m\omega )^{2}+\Gamma _{2}^{2}}.\end{aligned}$$Then we consider other terms in Eq. (14) and follow the above process, obtaining $$\begin{aligned}
\dot{\rho}_{11}(t) &\approx &\dot{y}(t,\varepsilon _{0})-\dot{y}(t,\varepsilon _{0})\rho _{11}(t)-\dot{y}(t,-\varepsilon _{0})\rho _{11}(t)
\nonumber \\
&=&W_{01}(t)\rho _{00}(t)-W_{10}(t)\rho _{11}(t),\end{aligned}$$where $W_{01}$ ($W_{10}$) is MDLZ transition rate from state $|0\rangle $ to $|1\rangle $ ($|1\rangle $ to $|0\rangle $). We emphasize that in the derivation of Eq. (20), all the expansion terms in Eq. (14), not only the notable terms, have the contribution. Eq. (20) gives the evolution of the diagonal part of the system reduced density matrix under the “Markovian” approximation.(see appendix) It should be noted that we take the relaxation phenomenonally and hence it does not emerge in Hamiltonian and rate equations, while in the calculation of population we will add it. Furthermore, MDLZ transition rates obtained are the combination of Lorentzian line shapes and time dependent, $$\begin{aligned}
W_{01}(t)=\frac{\Delta ^{2}}{2}\sum_{n,m}J_{n+m}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })\frac{\Gamma _{2}\cos n(\omega t+\pi /2)}{(\varepsilon
_{0}-m\omega )^{2}+\Gamma _{2}^{2}} \nonumber \\
W_{10}(t)=\frac{\Delta ^{2}}{2}\sum_{n,m}(-1)^{n}J_{n+m}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })\frac{\Gamma _{2}\cos n(\omega t+\pi /2)}{(\varepsilon _{0}-m\omega )^{2}+\Gamma _{2}^{2}},\end{aligned}$$where we use the relation $J_{m}(x)=(-1)^{m}J_{-m}(x)$. Different from Eq. (9), MDLZ transition rates in Eq. (21) are unsymmetrical. Comparing Eq. (21) and Eq. (9), we find that their difference is the additional time-dependent oscillating terms $f(t)$, which comes from the nonresonant terms. The nonresonant terms also correspond to the weak oscillation in Fig. 2 (b), due to the coherence evolution. It is clear that the coherence in the nonresonant terms results into unsymmetrical MDLZ transitions, i.e. $W_{01}(t)\neq W_{10}(t)$. In addition, although the coherence is preserved in the nonresonant terms, nonresonant terms with different oscillation frequencies counteract the resonant terms which would become less obvious. Therefore, not only the decoherence but also the low-frequency driving would eliminate the multiphoton resonant fringes.
Then we have the population in state $|0\rangle $ $$\begin{aligned}
\fl \rho _{00}=e^{\int_{0}^{t}dt^{\prime }[-W_{01}(t^{\prime })-W_{10}(t^{\prime
})-\Gamma _{01}-\Gamma _{10}]}\{\int_{0}^{t}dt^{\prime }[W_{10}(t^{\prime
})+\Gamma _{10}] \nonumber \\
\fl e^{\int_{0}^{t^{\prime }}dt^{\prime \prime
}[-W_{01}(t^{\prime \prime })-W_{10}(t^{\prime \prime })-\Gamma _{01}-\Gamma
_{10}]}+1\}. \label{22}\end{aligned}$$
Similarly, by following the above procedure, we can extend perturbation approach to higher levels and different noise sources, which were previously explored under high-frequency driving. [@du; @du1] Especially, MDLZ transition between higher levels can be used to analyze microwave cooling with the maximal efficiency. [@du2]
Under the large amplitude driving, the coherence of LZS interference comes from two aspects. The first one is traditionally from the prepared initial states. The off diagonal elements of the density matrix describe this part of coherence, which is usually needed to be preserved for the application of quantum computation. The second one is the coherence induced by the crossover which acts as a beam splitter. [@oliver; @sun] The microwave drives the state through the crossover, splitting the state into the superposition of states $|0\rangle $ and $|1\rangle ,$ and the system gains the coherence again. For a stationary case, the first kind has disappeared fully while the second one is created incessantly. However, it should also be noted that usually the coherence induced by the crossover only exists between two consecutive passing through the crossover within one period. There is no coherence for the processes in arbitrary periods.
To make this method reasonable, as discussed in appendix, the coherent evolution produced by the crossover cannot produce remarkable population change in the stationary case, which means $\Delta J_{n}(A/\omega )$ has little effect in one driving period. If the tunnel coupling strength $\Delta
$ is large and the decoherence is weak, the coherence will persist for a long time, producing remarkable coherent population change. Therefore, the validity of this method depends on two factors. The first one is the high frequency, i.e. $\omega $ $\gg $ $\Delta J_{n}(A/\omega )$, which makes coherent dynamics not have enough time to cause population change. The second one is the strong decoherence, which makes the coherence disappear quickly and produce little effect on the population, i.e. $\Gamma
_{2}>\Delta J_{n}(A/\omega )$. If one of factors is satisfied, the method would be available.
For the high frequency, i.e. $\omega $ $>\Delta \gg $ $\Delta J_{n}(A/\omega
)$, the time-dependent terms in Eq. (21) oscillate fast and can be neglected just like that in RWA. Therefore MDLZ transition Eq. (21) returns to symmetrical perturbation-RWA result Eq. (9) and is constant. In addition, as shown in the derivation, Eq. (9) can be further extended to the incoherent region $\omega <\Gamma _{2}$ and the high coherence $\Gamma _{2}<\Delta
J_{n}(A/\omega )$. Furthermore, in the condition, i.e. $\omega \gg \Gamma
_{2},\Delta J_{n}(A/\omega )$, the stationary population Eq. (11) in perturbation-RWA is equivalent to Eq. (8) in RWA. To validate the perturbation-RWA result in the two regions, we have numerically solved the Bloch equations to compare with the analytical results obtained by Eq. (9) on the steady occupation probability in state $|0\rangle $, as illustrated in Fig. 3. Here we consider the frequency $\omega /2\pi =90$ MHz with different decoherence rates. The analytical results agree well with the numerical simulations.
For the low frequency, i.e. $\omega <$ $\Delta $, the nonresonant terms will be comparable with the oscillating terms and cannot be neglected. In this case, perturbation-RWA result is not available and Eq. (21) must be needed. As discussed above, if the decoherence rate is small, e.t. $\Gamma
_{2}<\Delta J_{n}(A/\omega )\sim W_{01}$, the strong coherent population evolution may break perturbation approach. Therefore in this region, for smaller frequency, perturbation approach can be used in stronger decoherence. When the decoherence is enough strong, i.e. $\Gamma _{2}>\Delta
J_{n}(A/\omega )$, the coherence generated by the crossover would have no effects..
When the state is driven with the low frequency through the crossover $\Delta $, the small sweeping velocity makes the effect of nonresonant terms released. Nonresonant terms generate the periodic change of MDLZ transition rate with the same period of the driving, which makes the interwell relaxation participate in the dynamics and cause the change of the qubit population, i.e. $\dot{\rho}_{00}=\dot{\rho}_{11}\neq 0$, in the stationary case. Then the population would also oscillate with the frequency $\omega $. As shown in Fig. 4, in the steady case with large amplitude $A>\varepsilon _{0}$, for the driving frequencies 10 MHz and 1 MHz, the population in state $|0\rangle
$ exhibits a periodic change \[Fig. 4 (a) and (b)\] while for the frequency 90 MHz the population is always constant \[Fig. 4 (c)\]. On the other hand, at the small amplitude, the curve becomes smooth again \[Fig. 4 (d)\]. We can approximate $\Delta J_{n}(A/\omega )=(\Delta /n!)(A/2\omega )^{n}$. Small amplitude makes $\Delta J_{n}(A/\omega )$ much smaller than $\omega $ such that the high-frequency case Eq. (9) can be used. This case can also be understood simply: when the amplitude is small as illustrated, the state does not pass through the crossover and LZ transition would not occur, resulting little change of population in the steady case.
{width="7.5cm"}
{width="15cm"}
{width="15cm"}
Average perturbation approximation
==================================
As discussed in Ref. 21 and 22, the low frequency $\omega $, i.e. $\Delta
J_{n}(A/\omega )\lesssim \omega <\Delta $, can produce the optimal microwave cooling which is used to realize the active cooling of qubit and other artificial devises through MDLZ transitions. Although the low frequency makes the qubit population time-dependent in the stationary case, we can consider the average population over time. For the high frequency, the average population is just the steady population described by Eq. (11) while for the low frequency, the average population can be obtained by perturbation approach. However, it is complicate and not convenient to calculate the average population with Eq. (21). Therefore, we will employ average perturbation approximation (APA) based on perturbation approach in this section.
Then, we go back to the expression of $f(t)$, Eq. (19). In the stationary case, considering the time in one period, i.e. $kT<t<kT+T$, for the term $M(n,m,t)$ in $f(t)$, we have $$\begin{aligned}
M(n,m,t)=J_{n+m}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })\frac{\Gamma
_{2}\cos n(\omega t+\pi /2)}{(\varepsilon _{0}-m\omega )^{2}+\Gamma _{2}^{2}},n\neq 0,\end{aligned}$$where $M(n,m,kT+t^{\prime })=M(n,m,kT+T/2n+t^{\prime })$ with an even number $n$ and $M(n,m,kT+t^{\prime })=M(n,m,kT+T-t^{\prime })$ with an odd number $n$. Therefore, in one period of time, [the above relation]{} leads to $\lim\limits_{kT\rightarrow \infty }\int_{kT}^{kT+T}dt^{\prime }f(t^{\prime
})=0$, and we obtain the average MDLZ transition rates by averaging Eq. (21) over one period in the stationary case $$\begin{aligned}
\langle W_{10}\rangle =\langle W_{01}\rangle \approx \frac{1}{T}\lim_{kT\rightarrow \infty }\int_{kT}^{kT+T}dt^{\prime }W_{01}(t^{\prime })
\nonumber \\
=\frac{\Delta ^{2}}{2}\sum_{n}\frac{\Gamma _{2}J_{n}^{2}(A/\omega )}{(\varepsilon _{0}-n\omega )^{2}+\Gamma _{2}^{2}}. \label{24}\end{aligned}$$
It should be mentioned that although Eq. (24) coincides with the transition rate Eq. (9) in perturbation-RWA, the latter is the stationary transition rate under the high-frequency driving while the other is the average transition rate through averaging over the driving period under the low-frequency driving. Having obtained the transition rate, we can describe the population of system dynamics using the rate equations, where the average population $\langle \rho _{ii}\rangle $ obey$$\begin{aligned}
\langle \dot{\rho}_{00}\rangle =-(\langle W_{01}\rangle +\Gamma
_{01})\langle \rho _{00}\rangle +(\langle W_{10}\rangle +\Gamma
_{10})\langle \rho _{11}\rangle \nonumber \\
1=\langle \rho _{00}\rangle +\langle \rho _{11}\rangle\end{aligned}$$
After the long-time system evolution, we have $\langle \dot{\rho}_{00}\rangle =\langle \dot{\rho}_{11}\rangle =0$, and the average qubit population in state $|0\rangle $ can be easily solved from Eq. (25) $$\begin{aligned}
\langle \rho _{00}\rangle =(\langle W_{01}\rangle +\Gamma _{10})/(2\langle
W_{01}\rangle +\Gamma _{10}+\Gamma _{01}).\end{aligned}$$
The rate equations used here are different from Eq. (20). Under the low-frequency driving, Eq. (20) describes the real-time dynamics with the periodic change of population. However, Eq. (26) neglects the population change within one period as shown in Figs. 4 (a) and (b) and focuses on the scale of one period, in which the average population would be constant. If APA is valid, the change of qubit population must be slow on the scale of the driving period, i.e. $\omega >\langle W_{01}\rangle $. If $\omega $ is larger than $\langle W_{01}\rangle $, the population change in one period is so small that $\int_{t^{\prime }}^{t^{\prime }+T}W_{01}\rho _{00}dt/T\approx
\langle \rho _{00}\rangle \int_{t^{\prime }}^{t^{\prime
}+T}W_{01}dt/T=\langle \rho _{00}\rangle \langle W_{01}\rangle $ and the rate equations can be changed to the form of Eq. (25).
In order to show the region of validity for APA, we compare the numerical solutions with analytical results. The average transition rate $\langle
W_{01}\rangle $ is probably less than $5\times 2\pi $ MHz and generally the same order with $\Delta J_{n}(A/\omega )$. For several decoherence rates, we select the frequency 1 MHz and obtain the average steady population, shown in Fig. 5. The deviation between the numerical and analytical results cannot be neglected and APA is not proper. Fig. 4 (b) shows the real-time population in state $|0\rangle $ with the numerical stimulation under the frequency 1 MHz. Comparing with the result of frequency 10 MHz in Fig. 4 (a), due to $\omega <\langle W_{01}\rangle $ the population change in Fig. 4 (b) is large, demonstrating the invalidity of APA. (Fig. 5)
Moreover, as shown in Fig. 5 (a-c), at the frequency 1 MHz, when the amplitude is less than the dc detuning, which means the state does not reach the crossover, APA exhibits the good agreement with the numerical results. This is more clearly seen in Fig. 5 (d-f) with a small amplitude $A/2\pi =10$ MHz. Since the population change in the steady case is little \[Fig. 4 (d)\], APA is still available. In Fig. 6, for the frequency 10 MHz, the analytical results show very good agreement for several decoherence rates, changing from the rate less than $\Delta J_{n}(A/\omega )$ to the one much larger than $\Delta $. When the amplitude is larger than the dc detuning, the population is near 0.5, which also provides a way to determine the location of crossover with the low frequency. \[Fig. 6 (a-d)\] It should be noted that with the decoherence $\Gamma _{2}/2\pi =$ 3 MHz and the frequency 10 MHz, the agreement is not very good at the detuning $\varepsilon _{0}\approx A$. The high coherence makes the average transition rate large. As discussed in Ref. 22, at the detuning $\varepsilon _{0}\approx A$, the transition rate reaches maximum. Therefore the frequency would be less than the average transition rate, i.e. $\omega <\langle W_{01}\rangle $, and results into the disagreement.
Noncoherence approximation
==========================
For the extremely low frequency, i.e. $\omega <\Delta J_{n}(A/\omega )\sim
\langle W_{01}\rangle $, APA becomes insufficient just as shown in Fig. 5. Therefore in order to analyze this region conveniently, here we develop an analytical method based on noncoherence approximation (NCA), which will give an intuitive and clear physical picture.
When the driving frequency is high and coherent, LZS interference and multiphoton effect are clear and Bessel functions can give a convenient picture. However, the decrease of frequency makes LZS interference gradually disappear. Therefore the description with Bessel functions would not be proper for the extremely low frequency and we return back to the sinusoidal description.
We go back to Eq. (16) and have$$\begin{aligned}
\fl \dot{y}(t)=\frac{\Delta ^{2}}{4}\frac{d}{dt}\int_{0}^{t}d\tau ^{\prime
}\int_{-\tilde{t}}^{\tilde{t}}d\tau \exp \{i\int_{\tau ^{\prime }-\tau
/2}^{\tau ^{\prime }+\tau /2}dt^{\prime }(\varepsilon _{0}+A\sin \omega
t^{\prime })-\Gamma _{2}|\tau |\} \nonumber \\ \fl
=\frac{\Delta ^{2}}{4}\int_{0}^{t}d\tau \exp \{i\int_{t}^{t+\tau }dt^{\prime
}(\varepsilon _{0}+A\sin \omega t^{\prime })-\Gamma _{2}|\tau |\} \nonumber \\ \fl
+\frac{\Delta ^{2}}{4}\int_{-t}^{0}d\tau \exp \{i\int_{t-\tau }^{t}dt^{\prime
}(\varepsilon _{0}+A\sin \omega t^{\prime })-\Gamma _{2}|\tau |\}.
\label{27}\end{aligned}$$
Because of the strong decoherence, the extreme low frequency usually is much less than the decoherence rate. The decoherence makes the qubit lose coherence quickly, even though the state is split by the crossover and obtains the coherence again in each period. Hence we can approximately replace $\sin \omega t^{\prime }$ to $\sin \omega t$. Then we follow the procedure in Sec. 3, and have MDLZ transition rates
$$\begin{aligned}
W_{10}(t)=W_{01}(t)=\frac{\Delta ^{2}}{2}\frac{\Gamma _{2}}{(\varepsilon
_{0}+A\sin \omega t)^{2}+\Gamma _{2}^{2}}.\end{aligned}$$
Eq. (28) is actually the result of Eq. (21) in NCA limit. Interestingly, MDLZ transition rates become symmetrical. The expression can also be understood as macroscopic resonant tunneling (MRT) [@amin; @averin; @han] model under the sinusoidal driving with the extremely low frequency. When the state is driven slowly to pass through each detuning, we divide the evolution into discrete time steps within which the coherence lost totally at the corresponding detunings. [@js] Furthermore there is no coherence to connect time steps and the phase accumulation [@oliver] has not existed. Since each time step is much larger than the decoherence time, we can approximately consider that MRT occurs at the detuning with the tunneling transition rate Eq. (28). Hence, the dynamics is changed from LZS interference to the periodically driven MRT and we have the rate equations $$\begin{aligned}
\dot{\rho}_{00}=-[W_{01}(t)+\Gamma _{01}]\rho _{00}+[W_{10}(t)+\Gamma
_{10}]\rho _{11} \nonumber \\
1=\rho _{00}+\rho _{11},\end{aligned}$$where $W_{01}$ is defined by Eq. (28) and $\Gamma _{01}(\Gamma _{10})$ has the same definition with that in Eq. (7). In the evolution equations Eq. (29), the periodic change makes the previous treatment $\dot{\rho}_{00}=\dot{\rho}_{11}=0$ incorrect. When the initial state is $|0\rangle $, we have the population in state $|0\rangle $ $$\begin{aligned}
\fl \rho _{00}=e^{\int_{0}^{t}dt^{\prime }[-2W_{01}(t^{\prime })-\Gamma
_{01}-\Gamma _{10}]}\{\int_{0}^{t}dt^{\prime }[W_{10}(t^{\prime })+\Gamma
_{10}]e^{\int_{0}^{t^{\prime }}dt^{\prime \prime }[-2W_{01}(t^{\prime \prime
})-\Gamma _{01}-\Gamma _{10}]}+1\}. \end{aligned}$$The limit of perturbation approach also has function on NCA, which requires the strong decoherence, $\Gamma _{2}>\Delta J_{n}(A/\omega )\sim \langle
W_{01}\rangle $, where $\langle W_{01}\rangle $ is defined by Eq. (24). Moreover NCA needs the extremely low frequency compared with the strong decoherence. Therefore, in this region, this method is reasonable with the condition $\omega <\langle W_{01}\rangle <\Gamma _{2}$. In the frequencies where the average perturbation theory fails, we use Eq. (30) to reconsider the average population in the steady situation as shown in Fig. 7. At the decoherence rate $\Gamma _{2}/2\pi =3$ MHz, the agreement is not good, because the decoherence rate is comparable with the driving frequency. However, for other decoherence rates, the numerical stimulations agree well with the analytical results. Furthermore, we consider the real-time evolution and compare it with the numerical stimulation (Fig. 8). As the frequency is comparable or larger than the decoherence rate, NCA will be insufficient. \[Fig. 8 (b)\]
{width="7.5cm"}
{width="7.5cm"}
{width="15cm"}
From perturbation-RWA to NCA, we can find the transition from quantum to classic effect of the microwave with the frequency decreasing. For the high-frequency driving, being in a coherent state and with high photon numbers, the microwave behaves as the quantized electromagnetic field exchanging photons with the qubit. At the same time, the large microwave amplitude forms LZS interferometry through the crossover, modulating the intensity of the n-photon resonance. In this case MDLZ transition Eq. (9) is based on perturbation-RWA and the nonresonant terms in the transition rate can be neglected. With the frequency further decreasing, the decoherence becomes stronger relative to the driving frequency and the nonresonant terms are comparable with the resonant ones, which makes the photons resonances as well as the modulation from LZS interferometry vanish gradually. Then we have to use Eq. (21) to describe MDLZ transition. For the low frequency, in the scale of one driving period, the resonant information can still be notable and we can use APA to obtain the average population. For the extremely low frequency compared with the decoherence, the microwave can finally be treated as a classic field and the dynamic process is changed to the periodically driven MRT as shown in Eq. (30).
To sum up, we plot the average stationary population obtained from analytical calculation and numerical stimulation as shown in Fig. 9 (a). The different approximations of perturbation approach cover most of parameter space. Fig. 9 (b) is the detail of Fig. 9 (a) at the decoherence rate $\Gamma _{2}/2\pi =110$ MHz. We show the applicable regions for these approximations in the average stationary population in Fig. 9 (c). The circles mark where perturbation-RWA holds. In the red shaded region I, a high-frequency coherent field is used to realize multi-photon resonance. [@oliver; @bernsprl] With the frequency further decreasing, we enter $\langle W_{01}\rangle \lesssim \omega <\Delta .$ In the stationary state the qubit population exhibits the periodic oscillation which cannot be explained using perturbation-RWA. Now APA is applicable, marked with blue squares in Fig. 9 (c). In the green shaded region II, the low-frequency driving is able to generate microwave cooling and the optimal microwave cooling was realized by Valenzuela $et$ $al.$.[@va] Then for the extremely low frequency $\omega <\langle W_{01}\rangle <\Gamma _{2}$, APA is also unsuccessful. NCA can describe the behavior, as marked by the black ellipses. In the blue shaded region III, which is not covered by any before mentioned approximation, the high coherence and low frequency make it possible and convenient to realize adiabatic LZ process or LZS interference by a single crossing [@sun; @gos; @kal] with large amplitude. It should be emphasized that the region in Fig. 9 (c) would change with the amplitude just as shown in Fig. 5. For the amplitude larger than the dc detuning, the small frequency locates in NCA region while for the amplitude less than the dc detuning, this frequency may move to APA region.
Multi-level system
==================
Our approach gives a simple but effective physical picture, which enables us easily to extend the discussion to a multi-level flux qubit \[Fig. 10 (a)\].[@bernnature; @du1] In this section, based on the approximations we will discuss the lower limit of driving frequency in effective microwave driven cooling.
![(Color online). (a). Schematic energy diagram of a multi-level flux qubit. Red solid curve represents the microwave $\Phi _{ac}$. The dashed black line marks a particular static flux detuning $\protect\delta \Phi
_{dc} $. State $|0\rangle $ is in the left well; $|1\rangle $ and $|2\rangle
$ are in the right well. The red solid path describes the cooling with the population in state $|0\rangle $ transferred to $|1\rangle $. The locations of the crossovers $\Delta _{01}$, and $\Delta _{20}$ are 0 m$\Phi _{0}$, and 8.4 m$\Phi _{0}$, respectively. The diabatic energy-level slope $|m_{0}|$($|m_{1}|$) =$2\protect\pi \times 1.44$ GHz/m$\Phi _{0}$, and $|m_{2}|$ =$2\protect\pi \times 1.09$ GHz/m$\Phi _{0}$. $\Delta _{01}/2\protect\pi =0.013$ GHz, $\Delta _{02}/2\protect\pi =0.09$ GHz. The parameters used above are from the experiment.[@bernnature] In calculation, $\Gamma _{2}/2\protect\pi =0.06$ GHz, $\Gamma _{21}/2\protect\pi =0.1$ GHz,[du1]{} and the interwell relaxation time $T_{1}=1/(2\protect\pi \times
0.00005)$ ns. (b). The yellow surface with solid line shows the population in state $|0\rangle $ versus microwave frequency and amplitude at dc flux detuning $\protect\delta \Phi _{dc}=0.05$ m$\Phi _{0}$. The temperature is 50 mK. The dashed line indicates the population in equilibrium at this detuning. ](10.eps){width="7.5cm"}
In Ref. 21, it was found that a lower driving frequency realized better cooling effect and at 5 MHz the optimal cooling appeared. The driving microwave frequency used is so low that one does not need the high-frequency microwave generator and microwave line to realize active cooling. On the other hand, generally, if the microwave frequency is too low, there would not be the cooling effect on the qubit. Therefore, there exists a minimum frequency to produce the effective microwave cooling. Considering the three-level system consisting of an “internal” qubit ($|1\rangle $ and $|0\rangle $) and an “internal” oscillator like state $|2\rangle $, as shown in Fig. 10 (a), we rewrite the rate equations Eq. (29) to three-level form $$\begin{aligned}
\dot{\rho}_{00}=-[W_{01}(t)+\Gamma _{01}+W_{02}(t)]\rho
_{00}+[W_{10}(t)+\Gamma _{10}]\rho _{11}+W_{20}(t)\rho _{22} \nonumber \\
\dot{\rho}_{11}=[W_{01}(t)+\Gamma _{01}]\rho _{00}-[W_{10}(t)+\Gamma
_{10}]\rho _{11}+\Gamma _{21}\rho _{22} \nonumber \\
1=\rho _{00}+\rho _{11}+\rho _{22}.\end{aligned}$$where $\Gamma _{21}$ is the intrawell relaxation from state $|2\rangle $ to $|1\rangle $.
For the low frequency as discussed in Sec. 4, APA can be applicable and we have$$\begin{aligned}
W_{20}(t)=W_{02}(t)=\frac{\Delta _{20}^{2}}{2}\sum_{n}\frac{(\Gamma
_{2}+\Gamma _{21}/2)J_{n}^{2}(A/\omega )}{(\varepsilon _{0}^{\prime
}-n\omega )^{2}+(\Gamma _{2}+\Gamma _{21}/2)^{2}},\end{aligned}$$and,$$\begin{aligned}
W_{01}(t)=W_{10}(t)=\frac{\Delta _{01}^{2}}{2}\sum_{n}\frac{\Gamma
_{2}J_{n}^{2}(A/\omega )}{(\varepsilon _{0}-n\omega )^{2}+\Gamma _{2}^{2}},\end{aligned}$$where $\varepsilon _{0}^{\prime }=(|m_{0}|+|m_{2}|)(\delta \Phi _{dc}-\Phi
_{20})$, $\varepsilon _{0}=(|m_{0}|+|m_{1}|)\delta \Phi _{dc}$, and $\Phi
_{20}$ is the flux detuning where the crossover $\Delta _{20}$ is reached. It should be mentioned that Eqs. (32) and (33) actually express the average transition rate. Then for the extremely low frequency as discussed in Sec. 5, we have$$\begin{aligned}
W_{20}(t)=W_{02}(t)=\frac{\Delta _{20}^{2}}{2}\frac{\Gamma _{2}+\Gamma
_{21}/2}{\varepsilon ^{\ast }(t){}^{2}+(\Gamma _{2}+\Gamma _{21}/2)^{2}},\end{aligned}$$and, $$\begin{aligned}
W_{01}(t)=W_{10}(t)=\frac{\Delta _{01}^{2}}{2}\frac{\Gamma _{2}}{\varepsilon
(t){}^{2}+\Gamma _{2}^{2}}.\end{aligned}$$where $\varepsilon (t)=(|m_{0}|+|m_{1}|)\delta \Phi (t)$, and $\varepsilon
^{\ast }(t){}=(|m_{0}|+|m_{2}|)[\delta \Phi (t)-\Phi _{20}]$. Here, we focus the extremely low frequency region. Since the analytical solution is too complex to extract a clear physical picture, we do not write it out explicitly here. Fig. 10 (b) shows the population in state $|0\rangle $ at the dc flux detuning $\delta \Phi _{dc}$ = 0.05 m$\Phi _{0}$ as functions of frequency and amplitude. The thick dashed line marks the population at equilibrium with the temperature 50 mK. It can be seen that when the frequency is less than 0.1 MHz, microwave cooling would be ineffective.
In the extremely low frequency, the interwell relaxation changes the qubit population within one period as shown in Eq. (31). When the state is driven through the crossover $\Delta _{20}$, LZ transition would transfer the population from state $|0\rangle $ to $|2\rangle $, which then quickly relaxes into state $|1\rangle $. As the states is driven back to the initial detuning, the interwell relaxation warms the qubit again if the driving frequency is less than the repopulation rate of the warm process, i.e. $\Gamma _{10}+\Gamma _{01}$, which is approximately $2\Gamma _{10}$ at the small dc detuning. The analysis agrees with the result in Fig. 10 (b).
Conclusion
==========
In this article we have studied the population evolution of a strongly driven superconducting qubit of the long-time dynamics. First of all, we use perturbation approach, which considers high order perturbation expansion and nonresonant terms, to obtain the time dependent unsymmetrical MDLZ transitions. Moreover, we point out that this approach has a high validity in the case where the influence of coherence resulted from $\Delta $ is small with the high frequency or strong decoherence. Then for different frequencies, we apply several approximations to discuss the system dynamics.
For the high frequency, i.e. $\omega >\Delta $, MDLZ transition rate in perturbation approach is constant and leads to perturbation-RWA result. In previous works [@oliver; @bernsprl] perturbation-RWA has been proven successful in the coherent region under the strong decoherence, i.e. $\omega
>\Delta ,\Gamma _{2}$ and $\Gamma _{2}\gg $ $W_{01}$. Besides this region, we show perturbation-RWA result can also be used in the incoherent case where the frequency is smaller than the decoherence rate, $\Gamma
_{2}>\omega $, and even the high coherence region $\Gamma _{2}<$ $W_{01}$.
With the frequency further decreasing to the region of the optimal microwave cooling, $\langle W_{10}\rangle \lesssim \omega <\Delta $, since the coherence has effect on the nonresonant terms and nonresonant terms are comparable with the resonant terms, the qubit population exhibits periodic oscillation which cannot be addressed by perturbation-RWA. In this case, we study the average stationary population in the low-frequency driving. Then we employ APA by averaging MDLZ transition rate of perturbation approach with time. Although the analytical expression is the same with that of perturbation-RWA, they have different physical pictures. In perturbation-RWA, nonresonant terms have no effects and can be neglected, while in APA the nonresonant terms result oscillations and can be averaged over to obtain the average population. Therefore in APA, we do not take account into the real-time dynamics, but consider one period as the unit time. Then in the extremely low frequency, i.e. $\omega <\langle
W_{10}\rangle <\Gamma _{2}$, where APA has failed, we propose NCA by dividing the evolution into discrete time steps during which the coherence is lost totally. We show the dynamic process equivalent to the periodically driven MRT, giving a convenient and intuitive picture. In addition, in the multi-level system, based on this approximation we examine the possibility of employing the extremely low frequency driving as a useful tool to realize microwave cooling and prove that there exists a lower limit of frequency for effective microwave cooling. The different approximations of perturbation approach cover most of parameter space when describing the average population. \[Fig. 9 (c)\]
Our theory offers the effective analytical description of the driven dissipative TLS in the stationary case, including real-time and average situation. The good agreement of our theory with the numerical results shows the validity of the approximations used. By shifting these approximations, we also show the transition from quantum regime to classic regime, with dynamics changing from LZS interference to periodically driven MRT. This could enhance the understanding to the influence of decoherence on the qubit. Furthermore, our results can be applied to understand the behavior of other systems with similar structure. Especially, our theory in the low frequency can be used to analyze the optimal microwave cooling of qubit as well as other artificial device such as nano mechanical resonators.
Acknowledgement
===============
Thanks to S. O. Valenzuela, S. N. Shevchenko and M. H. S. Amin for useful discussions. This work was supported in part by the State Key Program for Basic Researches of China (2011CB922104, 2011CBA00205), the NSFC (10725415, 91021003), and the Natural Science Foundation of Jiangsu Province (BK2010012).
Derivation of Eq. (14)
======================
In this appendix, we will accept the straightforward physical picture for intuitive understanding. Assuming that the system is initialized in state $|0\rangle $, first of all, we try to calculate the population in state $|0\rangle $, i.e. $\rho _{00}(t)$. In a short time interval, i.e. $t<1/[\Delta J_{n}(A/\omega )]$, considering the perturbation expansion of $U_{V}$ in $V(\tau )$ up to the second order term, we have
$$\begin{aligned}
\fl \rho _{00}(t) =Tr_{B}\{\langle 0|\rho _{SB}(t)|0\rangle \}=Tr_{B}\{\langle
0|U_{V}(t,0)\rho _{SB}(0)U_{V}^{\dag }(t,0)|0\rangle \} \nonumber \\\fl=Tr_{B}\{\langle 0|\rho _{SB}(0)|0\rangle
\}-Tr_{B}\{\int_{0}^{t}\int_{0}^{\tau }d\tau d\tau ^{\prime }\langle
0|V(\tau )V(\tau ^{\prime })\rho _{SB}(0)|0\rangle \} \nonumber \\\fl
-Tr_{B}\{\int_{0}^{t}\int_{0}^{\tau }d\tau d\tau ^{\prime }\langle 0|\rho
_{SB}(0)V^{\dag }(\tau ^{\prime })V^{\dag }(\tau )|0\rangle \}. \label{A(1)}\end{aligned}$$
We substitute $V(t)$ into Eq. (A.1) and have$$\begin{aligned}
\fl \rho _{00}(t) =1-\int_{0}^{t}\int_{0}^{\tau }d\tau d\tau ^{\prime
}K_{1}\langle U_{-}^{\dag }(\tau )U_{+}(\tau )U_{+}^{\dag }(\tau ^{\prime
})U_{-}(\tau ^{\prime })\rangle \nonumber \\\fl
-\int_{0}^{t}\int_{0}^{\tau }d\tau d\tau ^{\prime }K_{2}\langle
U_{-}^{\dag }(\tau ^{\prime })U_{+}(\tau ^{\prime })U_{+}^{\dag }(\tau
)U_{-}(\tau )\rangle \nonumber \\\fl
=1-2\int_{0}^{t}\int_{0}^{\tau }d\tau d\tau ^{\prime }K_{1}\langle
U_{-}^{\dag }(\tau )U_{+}(\tau )U_{+}^{\dag }(\tau ^{\prime })U_{-}(\tau
^{\prime })\rangle , \label{A(2)}\end{aligned}$$where$$K_{1}=\sum\limits_{n,m}J_{n}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })e^{i(\varepsilon _{0}-m\omega )\tau -i(\varepsilon _{0}-n\omega )\tau
^{\prime }+in\pi /2-im\pi /2},$$and$$K_{2}=\sum\limits_{n,m}J_{n}(\frac{A}{\omega })J_{m}(\frac{A}{\omega })e^{-i(\varepsilon _{0}-m\omega )\tau +i(\varepsilon _{0}-n\omega )\tau
^{\prime }-in\pi /2+im\pi /2}.$$In the above derivation, we also obtain $Tr_{B}\{\langle
0|\int_{0}^{t}\int_{0}^{\tau }d\tau d\tau ^{\prime }V(\tau )V(\tau ^{\prime
})\rho _{SB}(0)|0\rangle \}=Tr_{B}\{\langle 0|\int_{0}^{t}\int_{0}^{\tau
}d\tau d\tau ^{\prime }\rho _{SB}(0)V^{\dag }(\tau ^{\prime })V^{\dag }(\tau
)|0\rangle \}$. Moreover $2Tr_{B}\{\langle 0|\int_{0}^{t}\int_{0}^{\tau
}d\tau d\tau ^{\prime }V(\tau )V(\tau ^{\prime })\rho _{SB}(0)|0\rangle \}$ just describes the reduced population in state $|0\rangle $ in the short time interval.
Then we turn to high order expansion. It should be mentioned for simpleness in the below derivation we only consider the first four terms in Eq. (13) and higher order terms can be treated with the similar process. We have
$$\begin{aligned}
\fl \rho _{11}(t) =Tr_{B}\{\langle 1|U_{V}(t,0)\rho _{SB}(0)U_{V}^{\dag
}(t,0)|1\rangle \} \nonumber \\
\fl =\int_{0}^{t}\int_{0}^{t}d\tau d\tau ^{\prime }Tr_{B}\{\langle 1|V(\tau
)\rho _{SB}(0)V^{\dag }(\tau ^{\prime })|1\rangle \} \nonumber \\
\fl -\int_{0}^{t}\int_{0}^{t}\int_{0}^{t_{1}}\int_{0}^{t_{2}}dt_{4}dt_{1}dt_{2}dt_{3}Tr_{B}\{\langle 1|V(t_{4})\rho
_{SB}(0)V^{\dag }(t_{3})V^{\dag }(t_{2})V^{\dag }(t_{1})|1\rangle \}
\nonumber \\
\fl -\int_{0}^{t}\int_{0}^{t_{1}}\int_{0}^{t_{2}}\int_{0}^{t}dt_{1}dt_{2}dt_{3}dt_{4}Tr_{B}\{\langle
1|V(t_{1})V(t_{2})V(t_{3})\rho _{SB}(0)V^{\dag }(t_{4})|1\rangle \}+...
\label{A(3)}\end{aligned}$$
Considering that the last two terms in Eq. (A.3) is equivalent, we can write Eq. (A.3) as $$\begin{aligned}
\fl \rho _{11}(t) =\int_{0}^{t}\int_{0}^{t}d\tau d\tau ^{\prime
}Tr_{B}\{\langle 1|V(\tau )\rho _{SB}(0)V^{\dag }(\tau ^{\prime })|1\rangle
\} \nonumber \\
\fl -2\int_{0}^{t}\int_{0}^{t}\int_{0}^{t_{1}}\int_{0}^{t_{2}}dt_{4}dt_{1}dt_{2}dt_{3}Tr_{B}\{\langle 1|V(t_{4})\rho
_{SB}(0)V^{\dag }(t_{3})V^{\dag }(t_{2})V^{\dag }(t_{1})|1\rangle \}
\nonumber \\
\fl=\int_{0}^{t}\int_{0}^{t}d\tau d\tau ^{\prime }Tr_{B}\{\langle 1|V(\tau
)\rho _{SB}(0)V^{\dag }(\tau ^{\prime })|1\rangle \} \nonumber \\
\fl-2\int_{0}^{t}\int_{0}^{t_{1}}\int_{0}^{t_{2}}\int_{0}^{t_{2}}dt_{1}dt_{2}dt_{3}dt_{4}Tr_{B}\{\langle 1|V(t_{4})\rho
_{SB}(0)V^{\dag }(t_{3})V^{\dag }(t_{2})V^{\dag }(t_{1})|1\rangle \}
\nonumber \\
\fl-2\int_{0}^{t}\int_{t_{2}}^{t}\int_{0}^{t_{1}}\int_{0}^{t_{2}}dt_{1}dt_{4}dt_{2}dt_{3}Tr_{B}\{\langle 1|V(t_{4})\rho
_{SB}(0)V^{\dag }(t_{3})V^{\dag }(t_{2})V^{\dag }(t_{1})|1\rangle \}+...
\label{A(4)}\end{aligned}$$Then we find that the system population rate change is given by$$\begin{aligned}
\fl \dot{\rho}_{11}(t) =\int_{0}^{t}d\tau ^{\prime }Tr_{B}\{\langle 1|V(t)\rho
_{SB}(0)V^{\dag }(\tau ^{\prime })|1\rangle \}+\int_{0}^{t}d\tau
Tr_{B}\{\langle 1|V(\tau )\rho _{SB}(0)V^{\dag }(t)|1\rangle \} \nonumber \\
\fl-2\int_{0}^{t}\int_{0}^{t_{2}}\int_{0}^{t_{2}}dt_{2}dt_{3}dt_{4}Tr_{B}\{\langle 1|V(t_{4})\rho _{SB}(0)V^{\dag }(t_{3})V^{\dag }(t_{2})V^{\dag
}(t)|1\rangle \} \nonumber \\
\fl-2\int_{0}^{t}\int_{0}^{t_{2}}\int_{t_{2}}^{t}dt_{2}dt_{3}dt_{4}Tr_{B}\{\langle 1|V(t_{4})\rho _{SB}(0)V^{\dag }(t_{3})V^{\dag }(t_{2})V^{\dag
}(t)|1\rangle \} \nonumber \\
\fl-2\int_{0}^{t}\int_{0}^{t_{1}}\int_{0}^{t_{2}}dt_{1}dt_{2}dt_{3}Tr_{B}\{\langle 1|V(t)\rho _{SB}(0)V^{\dag }(t_{3})V^{\dag }(t_{2})V^{\dag
}(t_{1})|1\rangle \}+... \label{A(5)}\end{aligned}$$Combining the equivalent terms in Eq. (A.5), we have$$\begin{aligned}
\fl \dot{\rho}_{11}(t) =Tr_{B}\{2\int_{0}^{t}d\tau \langle 1|V(t)\rho
_{SB}(0)V^{\dag }(\tau )|1\rangle \} \nonumber \\
\fl-Tr_{B}\{2\int_{0}^{t}dt_{2}\langle
1|[\int_{0}^{t_{2}}\int_{0}^{t_{2}}dt_{3}dt_{4}V(t_{4})\rho _{SB}(0)V^{\dag
}(t_{3})]V^{\dag }(t_{2})V^{\dag }(t)|1\rangle \} \nonumber \\
\fl-Tr_{B}\{2\int_{0}^{t}dt_{1}\langle
1|V(t)[2\int_{0}^{t_{1}}\int_{0}^{t_{2}}dt_{2}dt_{3}\rho _{SB}(0)V^{\dag
}(t_{3})V^{\dag }(t_{2})]V^{\dag }(t_{1})|1\rangle \}+... \label{A(6)}\end{aligned}$$where $\int_{0}^{t_{2}}\int_{0}^{t_{2}}dt_{3}dt_{4}V(t_{4})\rho
_{SB}(0)V^{\dag }(t_{3})$ describes the population transferred from state $|0\rangle $ to $|1\rangle $ in the time interval $t_{2}$. Comparing with Eqs. (A.1) and (A.2), we can find that the second term in Eq. (A.6) characterizes the population rate transferred from state $|1\rangle $ to $|0\rangle $ based on the population transferred from state $|0\rangle $. Similarly $2\int_{0}^{t_{1}}\int_{0}^{t_{2}}dt_{2}dt_{3}\rho _{SB}(0)V^{\dag
}(t_{3})V^{\dag }(t_{2})$ describes the reduced population in state $|0\rangle $ in the time interval $t_{1}$ and the third term in Eq. (A.6) describes the reduced population rate from state $|0\rangle $ to $|1\rangle $ due to the decrease of population in state $|0\rangle $. Then we replace the integration limit $t_{2}$ in Eq. (A.6) to $t$, which is similar with the widely known Markovian approximation dealing with the population transition caused by the coupling to the bath [@scu]. Here, different from the Markovian approximation, the approximation is used in the population transition caused by the tunnel coupling $\Delta $. To make this approximation valid, the system should be independent on the past, which means the coherent population change resulted from $\Delta $ should be little. Integrating with Eq. (A.6) we can obtain Eq. (14).
References {#references .unnumbered}
==========
[99]{} Makhlin Y, Schön G, and Shnirman A, Rev. Mod. Phys. 73, 357 (2001).
Nielsen M A and Chuang I L, *Quantum Computation and Quantum Information*, (Cambridge University Press, Cambridge, 2000).
You J Q and Nori F, Phys. Today 58, No. 11, 42 (2005);
Mooij J E, Science **307**,1210 (2005)
Clarke J and Wilhelm F K, Nature(London) **453**, 1031(2008)
Shevchenko S N, Ashhab S, and Nori F, Physics Reports **492**, 1 (2010).
Shytov A V, Ivanov D A, and Feigel’man M V, Eur. Phys. J. B 36, 263 (2003).
Landau L D, Phys. Z. Sowjetunion **2**, 46 (1932); Zener G, Proc. R. Soc. London. Ser. A **137,** 696 (1932); Stückelberg E C G, Helv. Phys. Acta **5,** 369 (1932).
Oliver W D, Yu Y, Lee J C, Berggren K K, Levitov L S, and Orlando T P, Science **310**, 1653(2005)
Sillanpää M, Lehtinen T, Paila A, Makhlin Y, and Hakonen P, Phys. Rev. Lett. **96**,187002(2006)
Nakamura Y, Pashkin Y A, and Tsai J S, Phys. Rev. Lett. 87, 246601 (2001)
Saito S, Thorwart M, Tanaka H, Ueda M, Nakano H, Semba K, and Takayanagi H, Phys. Rev. Lett. 93, 037001 (2004)
Izmalkov A, van der Ploeg S H W, Shevchenko S N, Grajcar M, Il’ichev E, Hübner U, Omelyanchouk A N, and Meyer H G, Phys. Rev. Lett. 101, 017003 (2008).
Wilson C M, Duty T, Persson F, Sandberg M, Johansson G, and Delsing P, Phys. Rev. Lett. 98, 257003 (2007)
Berns D M, Rudner M S, Valenzuela S O, Berggren K K, Oliver W D, Levitov L S, and Orlando T P, Nature(London) **455**, 51(2008)
Sun G Z, Wen X D, Mao B, Chen J, Yu Y, Wu P H, and Han S Y, Nature. Commun. 1:51 doi: 10.1038 / ncomms1050 (2010).
Bylander J, Rudner M S, Shytov A V, Valenzuela S O, Berns D M, Berggren K K, Levitov L S, and Oliver W D, Phys. Rev. B 80, 220506 (2009)
Oliver W D, Valenzuela S O, Quant. Info. Proc. 8, 261 (2009).
Berns D M, Oliver W D, Valenzuela S O,Shytov A V, Berggren K K, Levitov L S, and Orlando T P, Phys. Rev. Lett. **97** ,150520(2006)
Rudner M S, Shytov A V, Levitov L S, Berns D M, Oliver W D, Valenzuela S O, and Orlando T P, Phys. Rev. Lett. 101, 190502 (2008).
Valenzuela S O, Oliver W D, Berns D M, Berggren K K, Levitov L S, Orlando T P, Science 314, 1589 (2006)
Du L J, and Yu Y, Phys. Rev. B submitted (2010)
Goychuk I and Hänggi P, Adv. Phys. 54, 525 (2005).
Grifoni M and Hänggi P, Phys. Rep. 304, 229 (1998).
Goorden M C, and Wilhelm F K, Phys. Rev. B 68, 012508 (2003)
Du L J, Wang M J, and Yu Y, Phys. Rev. B 82, 045128 (2010)
Du L J, and Yu Y, Phys. Rev. B 82, 144524 (2010)
Son S K, Han S Y and Chu S I, Phys. Rev. A 79, 032301 (2009)
Földi P, Benedict M G, and Peeters F M, Phys. Rev. A 77, 013406 (2008)
Hausinger J and Grifoni M, Phys. Rev. A 81, 022117 (2010)
Ashhab S, Johansson J R, Zagoskin A M, and Nori F, Phys. Rev. A 75, 063414 (2007).
Liu Y, Sun C P, and Nori F, Phys. Rev. A 74, 052321 (2006)
Wilson C M, Johansson G, Duty T, Persson F, Sandberg M, and Delsing P, Phys. Rev. B 81, 024520 (2010)
Greenberg Y S, Phys. Rev. B 76, 104520 (2007)
Shevchenko S N, Kiyko A S, Omelyanchouk A N, and Krech W, Low Temp. Phys. 31, 569(2005)
Shevchenko S N, van der Ploeg S H W, Grajcar M, Il’ichev E, Omelyanchouk A N, Meyer H G, Phys. Rev. B 78, 174527 (2008)
Wubs M, Saito K, Kohler S, Kayanuma Y, and Hänggi P, New. J. Phys. 7, 218(2005).
Zhang Q, Hänggi P and Gong J B, New. J. Phys. 10 073008(2008).
Saito K, Wubs M, Kohler S, Kayanuma Y, and Hänggi P, Phys. Rev. B 75, 214308(2007).
Pegg D T, and Series G W, Proc. R. Soc. A 332, 281(1973)
Mooij J E, Orlando T P, Levitov L, Tian L, van der Wal C H, and Lloyd S, Science 285, 1036 (1999).
Chiorescu I, Nakamura Y, Harmans C J P M, and Mooij J E, Science 299, 1869 (2003).
Amin M H S, and Averin D V, Phys. Rev. Lett. **100**, 197001(2008)
Amin M H S, and Brito F, Phys. Rev. B 80, 214302 (2009)
Averin D V, Friedman J R, and Lukens J E, Phys. Rev. B 62, 11802 (2000)
Abragam A, The Principles of Nuclear Magnetism, (Clarendon, Oxford, 1961).
Rouse R, Han S Y, and Lukens J E, Phys. Rev. Lett. 75, 1614 (1995).
Johansson J, Amin M H S, Berkley A J, Bunyk P, Choi V, Harris R, Johnson M W, Lanting T M, Lloyd S, and Rose G, Phys. Rev. B 80, 012507 (2009)
Goswami D, Phys. Rep. 374, (2003)
Kral P, Thanopulos I, and Shapiro M, Rev. Mod. Phys. 79, 53(2007)
Scully M O, Zubairy M S, Quantum Optics (Cambridge University Press, New York, 1997)
| {
"pile_set_name": "ArXiv"
} |
\
T.O. Vulfs ^1^, E.I. Guendelman ^1,2,3^\
^1^ Department of Physics, Ben-Gurion University of the Negev, Beer-Sheva, Israel\
^2^ Frankfurt Institute for Advanced Studies, Giersch Science Center, Campus Riedberg, Frankfurt am Main, Germany\
^3^ Bahamas Advanced Study Institute and Conferences, 4A Ocean Heights, Hill View Circle, Stella Maris, Long Island, The Bahamas\
E-mail: [email protected], [email protected]
Starting with a simple two scalar field system coupled to a modified measure that is independent of the metric, we, first, find a Born-Infeld dynamics sector of the theory for a scalar field and second, show that the initial scale invariance of the action is dynamically broken and leads to a scale charge nonconservation, although there is still a conserved dilatation current.
Introduction
============
Every physical quantity is defined by its transformation properties. A scalar field, $\phi$, being a single real function of spacetime behaves as a scalar under Lorentz transformations: $\phi(x) \rightarrow \phi'(x')$. Its dynamics is determined by the kinetic and potential energy densities. In our paper we show, first, how this simplest scalar field can be transformed to a Born-Infeld scalar and second, how the dynamically broken scale invariance leads to the nonconservation of a scale charge. It all becomes possible when we use a modified measure in the action instead of a standard one.\
The notion of a measure is usually associated with the theories of gravity. There $\sqrt{-g}$, where $g$ is the determinant of the metric, is included to the action, i.e. $S = \int L \sqrt{-g} dx$, to make the volume element invariant. However, such choice for the measure is not unique. The only requirement that it must be a density under diffeomorphic transformations can be fulfilled in other ways. Our new measure is
$$\label{eq:0}
\Phi = \epsilon^{\mu_1\mu_2\ldots\mu_D}\epsilon_{a_1a_2\ldots a_D}\partial_{\mu_1}\varphi_{a_1}\ldots\partial_{\mu_D}\varphi_{a_D},$$
where $\epsilon^{\mu_1\mu_2\ldots\mu_D}$ and $\epsilon_{a_1a_2\ldots a_D}$ are Levi-Civita symbols and $\varphi_{a_1}\ldots\varphi_{a_D}$ are additional scalar fields that have nothing to do with the original $\phi$. This modified measure is already applied in gravity as first proposed in [@d; @e]. Different from (\[eq:0\]), the Galileon measure is considered in String Theory in [@e1; @e2].\
We assume that our spacetime is two-dimensional to avoid any unnecessary complications. Then
$$\Phi = \epsilon^{\mu\nu}\epsilon_{ab} \partial_\mu \varphi_a \partial_{\nu} \varphi_b.$$
We choose this particular realization for $\Phi$ because it is appropriate for our goals. We are able to do it because
$$\Phi \rightarrow \det(\frac{\partial x^{\mu'}}{\partial x^{\mu}})^{-1} \Phi, \quad d^2 x \rightarrow \det(\frac{\partial x^{\mu'}}{\partial x^{\mu}}) d^2 x.$$
Therefore,
$$\Phi d^2 x \rightarrow \Phi d^2 x.$$
The general form of the action is
$$\label{eq:5}
S = \int \Phi L d^D x.$$
To avoid a confusion in the terminology: the lagrangian is $\Phi L$, let’s call it $L_{full}$ and by $L$ we mean the part of $L_{full}$ without the measure $\Phi$. Notice that when the measure appears only linearly, as in (\[eq:5\]) and the measure fields $\varphi_a$ do not enter in $L$, there is an infinite dimensional symmetry
$$\varphi_a \rightarrow \varphi_a + f_a(L),$$
as has been discussed in [@d].\
As the background is clear, let’s check what exactly the goals are.\
The essence of the Born-Infeld theory is the requirement of finitness of a physical quantities. Originated as a specific theory of nonlinear electrodynamics in [@c], it put limitations on the self-energy of a point charge. Later it reappeared in string theory to describe the electromagnetic fields on the world-volumes of D-branes as it guarantees that the energy of the string is finite in [@g; @f]. Recently, to bring limits on scalar fields in cosmology, the Born-Infeld scalar was considered in [@h; @i]. This integration was developed later in [@i1; @i2; @i3; @i4; @i5; @i6; @i7; @i8]. Our first aim is a naturally arising restraints on our scalar field.\
The essence of the scale invariance is the requirement that physics must be the same at all scales, i.e. the system must be invariant under the global scale transformations ($\omega$ is a constant):
$$\label{eq:11}
g_{\mu\nu} \rightarrow \omega g_{\mu\nu}.$$
However, the physical universe definitely does not have such property, and different scales show different behavior. Then to approach reality the scale invariance must be broken. Our second aim is a naturally arising dynamical violation of the scale invariance. In this paper we are going to work with a fixed background metric, so the transformation (\[eq:11\]) will not be used, instead the fields will transform and a volume element independent of the metric will be allowed to transform.\
Moreover, according to Noether’s theorem, the symmetries and conservation laws are tightly connected. Then our third aim is to show that despite being a symmetry, the scale invariance does not lead to the conservation of the scale charge.\
The anomalous infrared behavior of the conserved chiral current in the presence of instantons was discussed in [@b]. The conclusion was made that in this case there was no conserved $U(1)$ charge and Goldstone’s theorem therefore failed, solving the $U(1)$ problem in QCD. The case of global scale invariance in the presence of a modified measure was considered in [@a; @a1; @a2; @a3; @a4; @a5; @a6; @a7; @a8] and the dilatation currents were calculated in a special model in [@a0], where the current was shown to be singular in the infrared. Here also, the resulting scale current produces a nonzero flux of the dilatation current to infinity, so once again, although there is a conserved current, there is no conserved scalar charge.\
Section 2 is devoted to the preparations for the later sections: the guiding principles are considered in more details and the lagrangian is provided. In Section 3 we arrive at the Born-Infeld scalar dynamics. In Section 4 we show how the asymptotic behavior of the conserved current leads to the nonconservation of a scale charge. The conclusions are given in Section 5.\
General Considerations
======================
The modified measure results in the dynamical violation of the scale invariance. To see that, we consider the variation of (\[eq:5\]) with respect to $\varphi_a$:
$$A^{\mu_1}_{a_1} \partial_{\mu_1} L = 0,$$
where $A^{\mu_1}_{a_1} = \epsilon^{\mu_1\mu_2\ldots\mu_D}\epsilon_{a_1a_2\ldots a_D}\partial_{\mu_2}\varphi_{a_2}\ldots\partial_{\mu_D}\varphi_{a_D}$ and we assume that $L$ is independent of $\varphi_a$’s.\
If $\det(A^{\mu_i}_{a_j}) \sim \Phi$ is non-trivial, then the solution is
$$\label{eq:4}
L = M = constant.$$
The appearance of the constant in (\[eq:4\]) breaks the scale invariance.\
To break the scale invariance in a consequence, the action must be scale invariant initially. Then
$$S = \frac12 \int \Phi \partial_{\mu} \phi \partial_{\nu}\phi g^{\mu\nu} d^2 x.$$
Without loss of generality, we assume that the scalar field possesses only the kinetic energy.\
The theory has the scale invariance with the following choice of the rescaling of the fields:
$$\label{eq:20}
\phi \rightarrow \lambda^{-\frac12} \phi, \quad \varphi_a \rightarrow \lambda^{\frac12} \varphi_a,$$
where $\lambda$ is the rescaling parameter that applies to the scalar fields and the measure only (and the metric remains invariant).\
However, it turns out that this model is not able to bring the enviable results. We must add one more scalar to the lagrangian.\
The final lagrangian is
$$L = \frac12 (\partial_{\mu} \phi_1\partial_{\nu}\phi_1 g^{\mu\nu} + \partial_{\mu} \phi_2\partial_{\nu}\phi_2 g^{\mu\nu}),$$
where $\phi_1$ is the former scalar field $\phi$ and $\phi_2$ is the supplemented one.\
So that the final action is
$$\label{eq:111}
S = \frac12 \int \Phi (\partial_{\mu} \phi_1\partial_{\nu}\phi_1 g^{\mu\nu} + \partial_{\mu} \phi_2\partial_{\nu}\phi_2 g^{\mu\nu}) d^2 x,$$
where for $S$ to be scale invariant we choose the rescaling of the additional field as
$$\label{eq:21}
\phi_2 \rightarrow \lambda^{-\frac12} \phi_2.$$
So at that level the dynamics of the initial scalar field $\phi$ is defined by the action (\[eq:111\]). To achieve our aims we added to that action three more scalar fields: $\phi_2$ is physically equivalent to $\phi$ and enter the lagrangian in the same footing, $\varphi_{a, a=1,2}$ are the base for the newly constructed modified measure in 2D. In the following sections we show how such complexity leads to the solutions.\
A step further is to obtain the equations of motion, i.e. the variations of $S$ with respect to the dynamical variables.\
For simplicity we consider flat Minkowski 2D spacetime so that
$$g_{\mu\nu} = \eta_{\mu\nu},$$
with the signature $(- +)$.\
The Appearance of a Born-Infeld Scalar Sector.
==============================================
The dynamical variables of (\[eq:111\]) are $\phi_1$, $\phi_2$, $\varphi_a$.\
The variation with respect to $\varphi_a$ is
$$\epsilon^{\mu\nu} \epsilon_{ab}\partial_{\nu} \varphi_{b} \partial_{\mu} L = 0.$$
If $\Phi$ is non-trivial then $\epsilon^{\mu\nu} \epsilon_{ab}\partial_{\nu} \phi_{b}$ is non-trivial. Then we can obtain
$$\label{eq:7}
L = \frac12 (\partial_{\mu} \phi_1\partial_{\nu}\phi_1 g^{\mu\nu} + \partial_{\mu} \phi_2\partial_{\nu}\phi_2 g^{\mu\nu}) = M.$$
Note that we can see (even before studying the Born-Infeld scalar sector) that for static case the gradients of the two scalar fields are bounded by $\sqrt{2M}$.\
(\[eq:7\]) is rewritten to give
$$\label{eq:10}
(\partial_{\mu} \phi_1)^2 + (\partial_{\mu} \phi_2)^2 = 2M.$$
It is interesting to notice that this is a kind of “nonlinear gradient $\sigma$ model”.\
The variation with respect to $\phi_2$ is
$$\label{eq:99}
\partial_{\mu} (\Phi \partial^{\mu} \phi_2) = 0.$$
We assume that $\phi_2$ depends only on the spatial coordinate $x$, $\phi_2 = \phi_2(x)$. Then (\[eq:99\]) becomes
$$\partial_1 (\Phi \partial_1 \phi_2) = 0,$$
which can be integrated to give
$$\Phi \partial_1 \phi_2 = J = constant.$$
We observe that the action has the additional shift symmetry:
$$\label{eq:9}
\phi_2 \rightarrow \phi_2 + c_2,$$
where $c_2$ is a constant.\
This symmetry leads to the conservation law. Then $J$ has the interpretation of a constant current flowing in the $x$-direction. Then
$$\partial_1 \phi_2 = \frac{J}{\Phi}.$$
Inserting this into (\[eq:7\]), we get
$$\partial_{\mu} \phi_1 \partial^{\mu} \phi_1 + \frac{J^2}{\Phi^2} = 2M,$$
which can be used to solve for the measure $\Phi$, giving
$$\label{eq:22}
\Phi = \frac{J}{\sqrt{2M - \partial_{\mu}\phi_1\partial^{\mu}\phi_1}}.$$
The variation with respect to $\phi_1$ is
$$\label{eq:23}
\partial_{\mu} (\Phi \partial^{\mu} \phi_1) = 0.$$
Making the same assumptions as for the $\phi_2$, namely $\phi_1 = \phi_1(x)$ and
$$\label{eq:8}
\phi_1 \rightarrow \phi_1 + c_1,$$
where $c_1$ is a constant,
we obtain
$$\partial_{\mu}(\Phi \partial^{\mu} \phi_1) = 0.$$
Inserting (\[eq:22\]) into (\[eq:23\]), we get the Born-Infeld scalar equation (for $M >0$):
$$\label{eq:100}
J \partial_{\mu}(\frac{\partial^{\mu}\phi_1}{\sqrt{2M - \partial_{\mu}\phi_1\partial^{\mu}\phi_1}}) = 0,$$
which is also obtained from the effective Born-Infeld action for this kind of solutions.
$$\label{eq:101}
S_{eff} = \int \sqrt{2M - \partial_{\mu}\phi_1\partial^{\mu}\phi_1} d^2 x.$$
The dynamics of $\phi_1$ defined by the equation (\[eq:100\]) is the same as the dynamics of $\phi_1$ derived from the variation of (\[eq:101\]). This means that $\partial_{\mu}\phi_1\partial^{\mu}\phi_1$ is bounded in this sector of the theory (the Born-Infeld scalar sector) Notice, however, that we are now considering only a sector of the theory.\
The Breaking of Charge Conservation.
====================================
The action (\[eq:111\]) is invariant under the scale transformations (\[eq:20\]) and (\[eq:21\]). By the Noether’s theorem a conservation quantity must appear, namely, the scale charge, $Q$. Then the continuity equation must be satisfied:\
$$\frac{\partial \rho}{\partial t} + \nabla j = 0,$$
where $\rho$ is a density of $Q$ and $j$ is the flux of $Q$.\
By the the integration
$$\int_{x_1}^{x^2} \frac{\partial \rho}{\partial t} d^2 x + \int_{x_1}^{x^2} \frac{\partial j^1}{\partial x} d^2 x = 0$$
we obtain
$$\frac{d Q}{dt} + j^1(x_2) - j^1(x_1)=0.$$
Therefore, the conservation of the total charge requires
$$j^1(x_1 \rightarrow -\infty)-j^1(x_2 \rightarrow +\infty) = 0.$$
However, it does not always happen. Our case is one of the exceptions.\
The conserved current is given by\
$$\label{eq:1}
j^{\mu} = \frac{\partial L_{full}}{\partial(\partial_{\mu}\varphi_a)}\delta \varphi_a + \frac{\partial L_{full}}{\partial(\partial_{\mu}\phi_1)}\delta \phi_1 + \frac{\partial L_{full}}{\partial(\partial_{\mu}\phi_2)}\delta \phi_2.$$
We consider a scale transformations infinitesimally closed to the identity: $\lambda = (1 + \theta)$, so that (\[eq:20\]) and (\[eq:21\]) turn into
$$\varphi_a \rightarrow (1 + \theta)^{\frac12}\varphi_a \simeq \varphi_a + \frac{\theta}{2}\varphi_a,$$
$$\phi_1 \rightarrow (1 + \theta)^{-\frac12}\phi_1 \simeq \phi_1 - \frac{\theta}{2}\phi_1,$$
$$\phi_2 \rightarrow (1 + \theta)^{-\frac12}\phi_2 \simeq \phi_2 - \frac{\theta}{2}\phi_2.$$
Therefore,
$$\delta\varphi_a = \frac{\theta}{2}\varphi_a, \quad \delta\phi_1 = -\frac{\theta}{2}\phi_1, \quad \delta\phi_1 = -\frac{\theta}{2}\phi_2.$$
Then (\[eq:1\]) becomes
$$j^{\mu} = M\frac{\theta}{2}\epsilon^{\mu\nu}\epsilon_{ab}\varphi_a\partial_{\nu}\varphi_b - \frac{\theta}{2}\Phi\partial^{\mu}\phi_1 - \frac{\theta}{2}\Phi\partial^{\mu}\phi_2.$$
Let’s go back to (\[eq:100\]) and find the static solutions ($\partial_0 \phi_1 = 0$):
$$\partial_1 (\frac{\partial_1 \phi_1}{\sqrt{2M - (\partial_1 \phi_1)^2}}) = 0.$$
By integration we get
$$\frac{\partial_1 \phi_1}{\sqrt{2M - (\partial_1 \phi_1)^2}} = c_3,$$
where $c_3$ is a constant.\
Then
$$\phi_1 = \frac{\sqrt{2M} |c_3|}{\sqrt{1 + |c_3|^2}} (x_2-x_1).$$
We have done all the calculations for $\phi_1$. However, the same is relevant for $\phi_2$. So that
$$\phi_2 = \frac{\sqrt{2M} |c_4|}{\sqrt{1 + |c_4|^2}} (x_2-x_1),$$
where $c_4$ is a constant.\
Inserting this solution to (\[eq:22\]), we obtain
$$\Phi = J \sqrt{\frac{1 + c_3^2}{(1 - 2M)c_3^2 + 1}}.$$
Then we see
$$\Phi = \Phi_0 = constant.$$
It is satisfied for
$$\varphi_1 = c_5 t, \quad \varphi_2 = c_6 x,$$
where $c_5$ and $c_6$ are constants.\
Then indeed
$$\Phi = c_5 c_6.$$
Now by inserting the solutions for $\varphi_a$, $\phi_1$ and $\phi_2$ into (\[eq:1\]), we obtain for $j^1$
$$j^1 = \frac{\theta}{2}M c_5 c_6 x - \frac{\theta}{2}c_5 c_6\frac{\sqrt{2M}|c_3|}{\sqrt{1+|c_3|^2}} - \frac{\theta}{2}c_5 c_6\frac{\sqrt{2M}|c_4|}{\sqrt{1+|c_4|^2}}.$$
We see that $j^1$ is a constant plus a term proportional to $x$ and therefore, $j^1(\infty)-j^1(-\infty) \ne 0$ and in fact diverges. Therefore, $Q$ is not conserved.\
Let’s calculate $j^0$ explicitly.
$$j^0 = -\frac{\theta}{2}Mc_5 c_6 t.$$
Then we see that
$$Q = \int_{x_1}^{x^2} j^0 dx = -(x_2 - x_1)\frac{\theta}{2}M c_5 c_6 t.$$
We checked that $Q$ is not conserved.\
Conclusions
===========
In this paper we start with the scalar field, surround it with three supplementary scalar fields and investigate the resulting action, (\[eq:111\]). One scalar field is physically equivalent to the former scalar field. However, the new measure of integration is constructed from the other two scalars. The source of the following findings is this modified measure. First, we show that the gradient of this initial scalar field is finite and in particular there is a sector which can be presented in the form of the Born-Infeld scalar. Second, the initial action is scale invariant, however, the invariance gets spontaneously broken. In addition to having spontaneous symmetry breaking, our physical system serves as an example of a system with the symmetry that does not lead to the conserved charge.\
A consideration of a scale invariance in cosmology started in [@v1; @v2] and was continued in [@a10; @a11]. In this case when the scale symmetry is spontaneously broken, there is a conserved current and since no singular behavior of the conserved current is obtained, so there is a conserved scale charge and the Goldstone theorem holds.\
Note that in our case of a scalar field there remains a massless field which is a Goldstone Boson of the shift symmetry $(\phi_1 \rightarrow \phi_1 + constant)$, not of the scale symmetry, because for the scale symmetry that theorem cannot be applied since the dilatation charge is not conserved.\
**Acknowledgments** TV acknowledges support by the Ministry of Aliyah and Integration (IL). EG is supported by the Foundational Questions Institute and COST actions CA15117, CA16104, CA18108. We thank Emil Nissimov and Svetlana Pacheva for interesting discussions.
[28]{} E.I. Guendelman, A.B. Kaganovich, Phys.Rev.D53, 7020, (1996), \[gr-qc/9605026\] E.I. Guendelman, A.B. Kaganovich, Phys.Rev.D55, 5970, (1997), \[gr-qc/9611046\] T.O. Vulfs, E.I. Guendelman, Mod.Phys.Lett. A32, no.38, 1750211, (2017), \[arXiv:1708.00458\] T.O. Vulfs, E.I. Guendelman, Annals Phys. 398 (2018) 138-145, (2018), \[arXiv:1709.01326\] M. Born, L. Infeld, Proc.Roy.Soc.Lond. A144, no.852, 425-451, (1934) C.G. Callan, J.M. Maldacena, Nucl.Phys. B513, 198-212, (1998), \[hep-th/9708147\] G.W. Gibbons, Nucl.Phys. B514, 603-639, (1998), \[hep-th/9709027\] J.A. Feigenbaum, P.G.O. Freund, M. Pigli, Phys.Rev. D57, 4738-4744, (1998), \[hep-th/9709196\] S. Deser, G.W. Gibbons, Class.Quant.Grav. 15, L35-L39, (1998), \[hep-th/9803049\] G.N. Felder, L. Kofman, , A. Starobinsky, JHEP 0209, 026, (2002), \[hep-th/0208019\] D.N. Vollick, Gen.Rel.Grav. 35, 1511-1516, (2003), \[hep-th/0102187\] Jian-gang Hao, Xin-zhou Li, Phys.Rev. D68, 043501, (2003), \[hep-th/0305207\] Dan N. Vollick, Phys.Rev. D72, 084026, (2005), \[gr-qc/0506091\] W. Fang, H.Q. Lu, Z.G. Huang, K.F. Zhang, Int.J.Mod.Phys. D15, 199-214, (2006), \[hep-th/0409080\] S. Jana, S. Kar, Phys.Rev. D94, no.6, 064016, (2016), \[arXiv:1605.00820\] V.I. Afonso, G.J. Olmo, D. Rubiera-Garcia, JCAP 1708, no.08, 031, (2017), \[arXiv:1705.01065\] S. Jana, S. Kar, Phys.Rev. D96, no.2, 024050, (2017), \[arXiv:1706.03209\] G. ’t Hooft, Phys.Rept. 142, 357-387, (1986) E.I. Guendelman, Class.Quant.Grav. 17, 361-372, (2000), \[gr-qc/9906025\] E.I. Guendelman, R. Herrera, P. Labrana, E. Nissimov, S. Pacheva, Gen.Rel.Grav. 47, no.2, 10, (2015), \[arXiv:1408.5344\] E.I. Guendelman, H. Nishino, S. Rajpoot, Phys.Lett. B732, 156-160, (2014), \[arXiv:1403.4199\] E.I. Guendelman, E. Nissimov, S. Pacheva, M. Vasihoun, Bulg.J.Phys. 40, 121-126, (2013), \[arXiv:1310.2772\] S. del Campo, E.I. Guendelman, A.B. Kaganovich, R. Herrera, P. Labrana, Phys.Lett. B699, 211-216, (2011), \[arXiv:1105.0651\] S. del Campo, E.I. Guendelman, R. Herrera, P. Labrana, JCAP 1006, 026, (2010), \[arXiv:1006.5734\] E.I. Guendelman, A.B. Kaganovich, Annals Phys. 323, 866-882, (2008), \[arXiv:0704.1998\] E.I. Guendelman, A.B. Kaganovich, Phys.Rev. D75, 083505, (2007), \[gr-qc/0607111\] E.I. Guendelman, O. Katz, Class.Quant.Grav. 20, 1715-1728, (2003), \[gr-qc/0211095\] E.I. Guendelman, Mod.Phys.Lett. A14, 1043-1052, (1999), \[gr-qc/9901017\] J. Garcia-Bellido, J. Rubio, M. Shaposhnikov, D. Zenhausern, Phys.Rev. D84, 123504, (2011), \[arXiv:1107.2163\] F. Bezrukov, G.K. Karananas, J. Rubio, M. Shaposhnikov, Phys.Rev. D87 (2013) no.9, 096001, (2013), \[arXiv:1212.4148\] P.G. Ferreira, C.T. Hill, G.G. Ross, Phys.Rev. D98, no.11, 116012, (2018), \[arXiv:1801.07676\] P. G. Ferreira, C.T. Hill, J. Noller, G.G. Ross, Phys.Rev. D97, no.12, 123516, (2018), \[arXiv:1802.06069\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Spin-only approaches to anisotropic effects in strongly interacting materials are often insufficient for systems close to the Mott regime. Within a model context, here the consequences of the low-symmetry relevant Dzyaloshinskii-Moriya (DM) interaction are studied for strongly correlated, but overall itinerant, systems. Namely, we investigate the Hubbard bilayer model supplemented by a DM term at half filling and in the hole-doped regime. As an add-on, further results for the two- impurity Anderson model with DM interaction are also provided. The model Hamiltonians are treated by means of the rotational invariant slave boson technique at saddle point within a (cellular) cluster approach. Already small values of the anisotropic interaction prove to have a strong influence on the phases and correlation functions with increasing $U$. An intriguing metallic spin-flop phase is found in the doped bilayer model and a reduction of the RKKY exchange in the two-impurity model.'
author:
- Sergej Schuwalow
- Christoph Piefke
- Frank Lechermann
bibliography:
- 'bibextra.bib'
title: 'Impact of the Dzyaloshinskii-Moriya interaction in strongly correlated itinerant systems'
---
Introduction
============
The effect of anisotropic magnetic exchange on the atomistic level has been recently brought back to a centre of interest in condensed matter physics due to its intriguing importance in e.g. the search for multiferroic materials [@ede05; @kho09], the understanding of complex metallic magnetic surface structures [@bod07] or the phenomenology of topological insulators. [@zhu11] A hallmark step in this research topic has already been performed some fifty years ago by Dzyaloshinskii [@dzy58] and Moriya, [@mor60] who derived an effective spin-spin interaction term from the spin-orbit coupling in low-symmetry cases, the so-called Dzyaloshinskii-Moriya (DM) interaction. The DM term or more generic magnetic anisotropies are nowadays believed to play furthermore a prominent role in many strongly correlated materials. However nearly exclusively, theoretical studies in this context were in the past bound to pure spin models without itinerancy, leaving the impact of charge fluctuations aside. Yet the latter are surely important, e.g. close to the Mott-critical regime of the metal-insulator transition. Allying the Hubbard model with spin-orbit terms has just recently gained rising interest. [@pes10; @men10]
In the present work we aim at a minimal modeling of the influence of the DM interaction in the strongly correlated metallic regime. There are many specific materials problems motivating such a case study, namely the complex magnetic behavior of doped cuprate systems, [@thi88; @cof91; @jur06] manganites, [@hir98; @mit01] and mono-oxides [@sol98] as well as anisotropic magnetic effects close to the metal-insulator transition in low-dimensional organic compounds [@kag08] or in the context of transition-metal impurities on metallic surfaces. [@zho10] While standard direct and indirect exchange processes favor collinear alignment of the local spins generated in the strongly correlated metallic regime, the DM interaction tends to align the spins in a perpendicular fashion. Thus the competition between the former conventional exchange processes and the DM interaction within an itinerant system shall give rise to nontrivial physics resulting in sophisticated spin arrangements/orderings.
To keep things simple and to build up on a somewhat canonical approach, we rely on two basic models, namely the bilayer model of two coupled single-band Hubbard planes [@zie96; @moe99; @fuh06; @kan07; @lec07; @bou08; @haf09; @yos09] and the two-impurity Anderson model. [@jay82; @fye87; @jonkot89; @fye89; @schi96; @nis06; @ferr09] The former Hamiltonian allows for a DM coupling between two lattice planes in the thermodynamic limit, whereas the latter one provides the possiblity to study the DM term within a local perspective via interacting impurities coupled to the same bath. Both setups render it possible to investigate nearest-neighbor (NN) correlation functions between sites in an itinerant background. Of course, such modelings are not sufficient to grasp the very details of the above named materials problems, yet it will be shown that the computed phenomenology is far from trivial and may apply to generic realistic phenomena. One key focus in the context of the Hubbard bilayer lattice is thereby on the competition between the antiferromagnetic (AFM) tendencies driven by direct exchange and the DM term within the metallic state. It will become clear that already rather small values of the DM integral may have a significant influence on the magnetic ordering tendencies in the larger Hubbard $U$ range, i.e., the AFM state is rather sensitive to only minor DM perturbations. A rich phase diagram results from the interplay of kinetic energy, onsite Coulomb and DM interaction. The latter also has important consequences in the two-impurity model, where its favor for perpendicular spin arrangement severely affects the local-limit competition between singlet-forming Kondo-screening and triplet-forming Ruderman-Kittel-Kasuya-Yosida (RKKY) interaction.
In the following we define the model Hamiltonians as well as our mean-field approach in section \[sec:mod\]. The results for the Hubbard bilayer at half filling and in the hole-doped case are discussed in section \[sec:hubi\]. Some basic observations retrieved from the studies on the two-impurity Anderson model with DM interaction will be presented in section \[sec:tiam\].
Hamiltonians and Theoretical Approach\[sec:mod\]
================================================
The first problem addressed here consists of two coupled two-dimensional infinite square-lattice planes with one orbital per site each facing an on-site Coulomb repulsion $U$ (see Fig. \[fig:mod\]). In both planes the electron dispersion is defined by identical simple NN hopping $t$. The inter-plane coupling is realized via a perpendicular hopping $t_\perp$ as well as a DM interaction mediated by the vector integral ${\bf D}$. The model Hamiltonian is accordingly written as $$\begin{aligned}
\label{eq:biham}
H_{\rm BL}=&&-t\sum_{\stackrel{\alpha\sigma}{\langle i,j\rangle}}
(c^\dagger_{\alpha i\sigma}c^{\hfill}_{\alpha j\sigma}+{\rm h.c.})+
t_{\perp}\sum_{i\sigma}(c^\dagger_{1i\sigma}c^{\hfill}_{2i\sigma}
+{\rm h.c.})\nonumber\\
&&+\;U\sum_{\alpha i}n_{\alpha i\uparrow}n_{\alpha i\downarrow}+
\sum_i{\bf D}\cdot({\bf S}_{1i}\times{\bf S}_{2i})\quad,\end{aligned}$$ where $c^{(\dagger)}_{\alpha i\sigma}$ creates/annihilates an electron in layer $\alpha$=1,2 at lattice site $i$ with spin projection $\sigma$=$\uparrow,\downarrow$. The $\nu$=$x,y,z$ component of the spin operator at each site $i$ of an individual layer $\alpha$ is provided by $S^{(\nu)}_{\alpha i}$=$1/2\,c^\dagger_{\alpha i\sigma}\,\tau^{(\nu)}_{\sigma\sigma'}\,c^{\hfill}_{\alpha i\sigma'}$ with the Pauli matrices $\tau^{(\nu)}$. In general, the vector interaction ${\bf D}$ is defined perpendicular to the bond between the involved lattice sites. [@dzy58; @mor60] Since otherwise there is a freedom of choice for the explicit direction, we pick ${\bf D}$ to point along the $y$ axis, i.e. ${\bf D}$=$D\,{\bf e}_y$. Note that the DM interaction may only occur if the inversion symmetry is broken. To facilitate this in the present case, one could e.g. think of an inter-layer coupling originally established via oxygen with an angle deviating from 180$^\circ$.
![(Color online) Bilayer model with DM interaction. The vertical blue line is a representant of the inter-layer hopping $t_\perp$, which of course is applied at every lattice point, and the ellipse marks the two-site cluster. The DM integral vector ${\bf D}$ is chosen to point in $y$ direction. \[fig:mod\]](model-struc.eps){width="7cm"}
In the smaller second part of this paper, we take the opportunity to also briefly discuss the well-known two-impurity Anderson model (TIAM) supplemented by a DM interaction between the impurities. We write that model in the form $$\begin{aligned}
\label{eq:impham}
H_{\rm TIAM}=&&\sum_{{\mathbf{k}}\sigma}\varepsilon_{{\mathbf{k}}}
c^\dagger_{{\mathbf{k}}\sigma}c^{\hfill}_{{\mathbf{k}}\sigma}+
\varepsilon_d\sum_{i\sigma}n_{i\sigma}\nonumber\\
&&+\,V\sum_{{\mathbf{k}}i\sigma}(c^\dagger_{{\mathbf{k}}\sigma}d^{\hfill}_{i\sigma}+{\rm h.c.})
+\;t_{12}\sum_{\sigma}(d^\dagger_{1\sigma}d^{\hfill}_{2\sigma}
+{\rm h.c.})\nonumber\\
&&+\,U\sum_{i}n_{i\uparrow}n_{i\downarrow}+
{\bf D}\cdot({\bf S}_{1}\times{\bf S}_{2})\quad,\end{aligned}$$ with the impurity-electron operators $d^{(\dagger)}_{i\sigma}$ ($i$=1,2), the impurity-level energy $\varepsilon_d$ and the impurity-impurity hopping $t_{12}$. The bath has associated operators $c^{(\dagger)}_{i\sigma}$ and a dispersion $\varepsilon_{{\mathbf{k}}}$. The impurity-bath coupling is denoted by $V$ and the Hubbard $U$ is located on the impurities with $n_{i\sigma}$=$d^{\dagger}_{i\sigma}d^{\hfill}_{i\sigma}$. In the present approach the bath is treated explicitly through a three-dimensional simple-cubic dispersion with bandwidth $W$=$12t$, choosing $t$=0.5. For the direction of ${\bf D}$ again the $y$ axis is selected. The impurities have a common bath, yet $V$ is assumed here to be $k$-independent and the explicit impurity-impurity distance is formally set to zero. A constant value of $V$=$-0.5$ is chosen in the present work. Hence only the local part of the RKKY interaction is accessible. Such a modeling is e.g. important for understanding the local spin interactions between correlated atoms on metallic surfaces, [@zho10] where there is indeed an intriguing interplay between conventional direct exchange, RKKY interaction, Kondo effect and anisotropic exchange.
For the numerical solution of the model Hamiltonians discussed here, the rotationally invariant slave-boson (RISB) formalism [@li89; @lec07] in the saddle-point approximation, similar to the generalized Gutzwiller approach, [@bue98] is employed. The RISB methodology amounts to a decomposition of an electron operator $a^{\hfill}_{\nu\sigma}$ with generic orbital/site index $\mu$ via $\underline{a}_{\mu\sigma}$=$\hat{R}[\phi]^{\sigma\sigma'}_{\mu\mu'}f_{\mu'\sigma'}$ into its quasiparticle (QP) part $f_{\mu\sigma}$ and the remaining high-energy excitations carried by the set of slave bosons $\{\phi_{An}\}$. Here $A$ denotes a chosen localized basis state and $n$ relates to the given QP degree of freedom. Two constraints, the first enforcing the normalization of the bosonic content and the second keeping an eye on the match of the bosonic and the fermionic occupation matrix, are established on site-average at saddle-point through the Lagrange-multiplier matrix $\Lambda$. [@lec07] In order to describe inter-atomic correlations adequately, a two-site (cellular-cluster) framework is used. This cluster connects two NN lattice sites between the layers in the Hubbard bilayer and the two impurities in the TIAM. It amounts to a local cluster approach to the electronic self-energy, whereby $\Sigma_{12}(\omega)$ incorporates terms linear in frequency as well as static renormalizations. [@lec07] Therewith the low-energy behavior may be adequately expressed and inter-site correlation functions as well as multiplet weights on the cluster can be retrieved. Importantly, the formalism allows for full spin and orbital rotational invariance, needed to account for the competition between isotropic and anisotropic interactions. In this respect the slave bosons may become true complex numbers and $\Lambda$ can be expanded via Pauli matrices in each orbital sector (with allowed off-diagonal terms between these sectors). Albeit the calculations are formally performed at temperature $T$=0, a small gaussian smearing for the k-point integration introduces a minor $T$ scale. For this reason the energetics are discussed in terms of the free energy $F$. Note that in the numerical solution of the TIAM, a three-orbital model is effectively treated within RISB, whereby the bath enters through its band dispersion. Thus the bath degrees of freedom are not integrated out, but are handled explicitly. In principle, a correlated-bath scenario may also be studied, however we here always keep $U_{\rm bath}$=0. Nevertheless, correlation effects are introduced within the bath due to the coupling to the correlated impurities. The investigated half-filled scenario of the model is either achieved by setting $\varepsilon_d$=$-U/2$ or through an additional Lagrange multiplier fixing the electron occupation on the bath according to the total filling $N$=3.
Hubbard bilayer model\[sec:hubi\]
==================================
The original Hubbard bilayer without DM interaction has already been addressed in several works, [@zie96; @moe99; @fuh06; @kan07; @lec07; @bou08; @haf09; @yos09] most often concerning the electronic phase diagram when varying the ratio $t_\perp/t$. Here however the main interest lies on the ratio $U/|{\bf D}|$ for the coupled square-lattice layers with bandwidth $W$=$8t$. In the following, we restrict the discussion to cases $t_\perp/t$$<$1 with all the energies given in units of the half bandwidth $4t$.
Concerning the electronic phases studied within the current mean-field approach, we restrict the discussion to local cluster orderings, i.e., neglect long-range order parameters suitable for e.g. spin spirals. Such more intricate instabilities are planned to be addressed in more concrete materials-connected future modelings. Here the focus is first on the interplay of the fundamental short-range correlation processes in the strongly correlated metallic regime. Note however that in the present context the cluster description does not account for intra-layer inter-site self-energies.
Half-filled case
----------------
At half filling, each layer accomodates one electron and the whole system is therefore susceptible to a Mott transition. We study two cases, namely the one of weakly coupled layers $(t_\perp$=$0.025)$ and the other with stronger inter-layer hopping $(t_\perp$=$0.1)$.
Figure \[fig:hf1\] shows the phase competition within the half-filled model with increasing the Hubbard $U$. The computations allow for the stabilization of two metallic phases, namely the paramagnetic (PM) and antiferromagnetic-between-layers (AFM) ones.
![(Color online) Energetics, QP weight and multiplet weights with increasing $U$ for the bilayer model at half filling for two values of $t_\perp$ and $D$, respectively. Free energies are normalized to the one of the PM phase. The thick lines in the multiplet-weights plot correspond to the states in the two-particle sector. \[fig:hf1\]](Combination_plot_n2.0.eps){width="8.5cm"}
![(Color online) Evolution of the Fock-state contributions to the local state with the highest slave-boson amplitude (compare with Fig. \[fig:hf1\]) for $U$=3. Note that for $D$$\neq$0 this state “1” is no longer an eigenstate of the $S^2$ and the $S_z$ operator. The inset exhibits the $D$ dependence of the state.\[fig:hfmulti\]](Fockstates.eps){width="8.5cm"}
![(Color online) Same as Fig. \[fig:hf2\], here for the spin-spin correlation functions between the layers. Solid lines: PM phase, dashed lines: AFM phase.\[fig:hf3\]](Spin_data_n2.0.eps){width="8.5cm"}
![(Color online) Same as Fig. \[fig:hf2\], here for the spin-spin correlation functions between the layers. Solid lines: PM phase, dashed lines: AFM phase.\[fig:hf3\]](Spcf_data_n2.0.eps){width="8.5cm"}
From the inspection of the free-energy differences it is nonsurprisingly seen that in general the AFM phase wins over the PM phase at larger $U$. Thereby a smaller $t_\perp$, and hence a smaller bonding/anti-bonding splitting, supports the building-up of the AFM phase, in line with DMFT calculations employing quantum Monte-Carlo solvers for the impurity problem. [@kan07; @haf09] A further gain in AFM free energy is observed at fixed $U$ when introducing the DM interaction, but with only marginal shifts of the phase onset towards smaller $U$. The difference between the two critical $U$=$U_c$ for the two different $t_\perp$ vanishes with $D$, while in the case of $D$=0 the $U_c$ for $t_\perp$=0.1 is clearly smaller. However the general evolution of the QP weight $Z$=$(1-\partial\Sigma/\partial\omega)^{-1}|_{\omega=0}$ with $U$ does not display strong changes with the introduced anisotropic interaction.
In addition, Fig. \[fig:hfmulti\] also displays the slave-boson weight of the local multiplet state that dominates at half filling in the relevant two-particle sector. For $D$=0 the bilayer Hamiltonian (\[eq:biham\]) commutes with $\{S^2,S_z\}$ and thus singlet and triplet states form the local two-particle eigenbasis. As expected, the singlet remains strongest up to the Mott transition, followed by the triplet states (whith their degeneracy lifted when entering the AFM phase). In the case of finite $D$, the picture formally looks very similar, but Hamiltonian and spin operators are no longer commuting operators and the respective two-particle states thus are not true spin eigenfunctions. That is easily understood from the DM interaction favoring a perpendicular alignment of the local spins, contrary to the originally preferred collinear states. It is nicely illustrated in Fig. \[fig:hfmulti\], where the inter-mixing of the singlet and triplet Fock-state building blocks with growing $D$ is exhibited.
Figure \[fig:hf2\] shows the evolution of the spin moments in the two layers with increasing $U$. For $D$=0 only $\langle S_z\rangle$ adopts a nonzero value in the AFM phase, with a steeper increase for larger $t_\perp$. However with finite $D$ also a sizeable $x$ component of $\langle S\rangle$ shows up and grows until $U_c$ is reached. For the smaller $t_\perp$=0.025 the value for $\langle S_x\rangle$ even equals the corresponding $\langle S_z\rangle$ magnitude. A lower $t_\perp$ apparently also effectively increases the relative tendency towards the non-collinear spin alignment driven by the DM coupling. Note that the DM interaction not only modifies the AFM phase, but has an impact in the PM state as well. There $\langle S_{1,y}S_{2,y}\rangle$ exhibits more AFM-like character and the corresponding $(x,z)$ correlation functions show minor weakened AFM-like tendencies, both compared to the $D$=0 case. Close to the Mott transition the larger $t_\perp$ results in a stronger (coherent) spin response for $D$=0, as retrieved from the inter-layer spin correlation functions plotted in Fig. \[fig:hf3\]. For nonzero $D$ the correlation between the $x$ components, i.e. $\langle S_{1,x}S_{2,x}\rangle$, appears to behave especially more disconnected from the $z$ component for the smaller $t_\perp$.
![(Color online) Influence of the DM coupling on the QP weights, spin moments and spin-spin correlation functions for the AFM half-filled bilayer model. The inset shows the evolution of the angle between the spin moments in the two layers.\[fig:hf4\]](Dscan_n2.0_v0.1_inset.eps){width="8.5cm"}
In order to gain further insight into the impact of the DM term, Fig. \[fig:hf4\] depicts explicitly the $D$ dependence for fixed $U$. The Mott transition itself may be tuned over a rather wide range of the anisotropic interaction. Whereas the spin moment in the $x$ direction shows a strong variation with $D$, the spin-spin correlations are only weakly dependent thereon. Albeit no resulting $\langle S_y\rangle$ value exists, the correlations along $y$ still appear to gain strongest in magnitude from a larger $D$. It is also visualized that the angle $\gamma$ between the local spins on the adjacent layers indeed increasingly deviates from the AFM-ideal 180$^{\circ}$ with growing DM interaction. Close to the Mott transition, the value $\gamma$$\sim$100$^{\circ}$ is nearby the DM-ideal value of 90$^{\circ}$.
Hole-doped case
---------------
We now turn to the effects of doping the bilayer model away from half filling. For investigating the simultaneous effects of doping, on-site Coulomb interaction and inter-site DM interaction we set $t_\perp$=0.1 and first fix the Hubbard interaction to $U$=3. As can be seen from Fig.\[fig:hf1\] the value of $U$ puts the system just below the Mott transition at half filling, i.e. strong correlations with the quasiparticle weight $Z$$\sim$0.2 exist.
The results of hole doping $\delta$=$2$$-$$n$ for the system in the filling range $n\in [1.6,2.0]$ are summarized in Fig. \[fig:hd1\]. Let us first discuss the $D$=0 case. Starting from half filling, the system is in the AFM phase for the chosen $U$ value. With increasing $\delta$ the order parameter $\langle S_z\rangle$ decreases, until it vanishes close to $n$=1.74 and the PM phase sets in (at reduced spin-spin correlations and larger QP weight). When including a DM interaction with $D$=0.03 in the model, the situation becomes more intriguing. Again the AFM (C-AFM) phase, now canted in $x$ direction, weakens upon doping from half-filling, however at $n$$\sim$1.76 the Hubbard bilayer system shows a first-order phase transition to a metallic spin-flop (SF) phase.
![(Color online) Influence of hole doping on the bilayer model, with and without DM interaction for $t_\perp$=0.1 and $U$=3. Insets at the top show the free-energy curve, with the region where AFM order is (meta)stable marked in red. C-AFM marks the canted antiferromagnetic and SF the spin-flop phase. The inset in the right-middle depicts the evolution of the angle $\gamma$ between spins. \[fig:hd1\]](Nscan_data_U3.0_inset.eps){width="8.5cm"}
![(Color online) Illustration of the stable local spin configurations according to the local spin-spin angle $\gamma$ on the inter-layer cluster. (a) PM without ordered local moments, (b) AFM with $\gamma$=$180^\circ$, (c) C-AFM with $\gamma$ between $180^\circ$ (pure AFM ordering) and $90^\circ$ (pure DM ordering), (d) SF with $\gamma$$<$$90^\circ$, i.e., weak ferromagnetism with strong canting. \[fig:hd2\]](spins_on_dimer.eps){width="6.5cm"}
The latter one is characterized by the discontinuous jump to a local configuration with an $\langle S_x\rangle$ expectation value [*larger*]{} than $\langle S_z\rangle$. This corresponds to an angle $\gamma$ between the local spins in both layers being lower than $90^\circ$, whereas in the C-AFM phase $\gamma\in [90^\circ,180^\circ]$ holds (see Fig \[fig:hd2\]). The strong decrease of $\gamma$ at the transition point may be observed in the inset of Fig. \[fig:hd1\]. Hence the SF phase displays weak ferromagnetism due to strong canting. Note that neither the spin correlation functions nor the diagonal $Z$ values show a strong signature therein. The SF phase transforms into the usual PM phase at $n$$\sim$1.62.
![(Color online) Phase diagram with $U$ for the doped Hubbard bilayer at filling $n$=1.7 (compare with Fig. \[fig:hd1\]). Insets show again the free-energy curves and the interaction-dependent spin-spin angle $\gamma$.\[fig:hd7\]](Uscan_n1.7_v0.1_inset.eps){width="8.5cm"}
In addition to the doping scan, Fig. \[fig:hd7\] displays the various phases emerging with increasing Hubbard interaction $U$ for fixed hole doping $\delta$=0.3, i.e., $n$=1.7. Without the DM interaction, the standard picture of a stable PM phase at small $U$ and a stable AFM phase at larger $U$ ($U$$>$3.58) remains vital. Note that the $U$ values for AFM stabilization are well above the Mott critical $U$ at half filling. Introducing $D$ stabilizes the metallic SF phase for 1.9$<$$U$$<$3.85, accompanied with the jump in the angle $\gamma$ towards lower values. Therewith the onset of AFM order takes place at slightly larger $U$ than for $D$=0. Hence the finite $D$ enables specific magnetic ordering in a Coulomb interacting regime that is originally not susceptible to such order. Only the $z$ component of the spin correlation function shows a discontinuous behavior at the SF/C-AFM phase boundary.
Two-impurity Anderson model\[sec:tiam\]
========================================
The TIAM [@jay82; @fye87; @jonkot89; @fye89; @schi96; @nis06; @ferr09] belongs to the set of canonical models in the physics of strong electronic correlations, believed to be relevant for the understanding of heavy-fermion systems [@don77]. Via the coupling of the impurities to a bath it contains the single-impurity Kondo physics and as competitor also the RKKY mechanism acting between the impurities. The latter originates from the effective exchange introduced through the impurity-coupling to the same bath. In some works [@moe99; @ferr09] this type of exchange interaction between sites is discussed in the context of two impurities coupled to different baths (similar to the bilayer architecture). But here we try to separate the exchange in an indirect (“RKKY”) one, stemming from effective exchange via the bath, and a direct term, resulting e.g. from an explicit hopping amplitude $t_{12}$ between the impurities (see Eq. (\[eq:impham\])).
Figure \[fig:taim1\] shows the RISB results for the spin correlation functions of the fundamental model with $t_{12}$=$D$=0. Thus the two impurities are only coupled via the bath and exchange can only be mediated therewith.
![(Color online) Interaction-dependent spin correlation functions for the TIAM with $t_{12}$=0. Top: $\langle S^2\rangle$ for $D$=0, bottom: spin-spin correlations for $D$=0 (circles) and $D$=0.05 (squares). The dashed dark line without circles is just the $\langle S_{\rm imp}S_{\rm bath}\rangle$ correlation function mirrored at the zero line. \[fig:taim1\]](Spcf_dimer_t0.eps){width="8.5cm"}
The expectation value $\langle S^2\rangle$=$S(S+1)$ on the impurity quickly rises with $U$ due to the formation of the local moment. It approaches the value 3/4, corresponding to the full $S$=1/2 limit, at large interaction strength. With increasing $U$ a local Fermi liquid is established with a small quasiparticle weight $Z_{\rm imp}$ (see inset Fig. \[fig:taim1\]). The competition between the Kondo screening and the RKKY interaction may be observed from inspection of the spin-spin correlations. From Fig. \[fig:taim1\] it is obvious that $\langle S_{\rm imp}S_{\rm bath}\rangle$, i.e. the correlation between a single impurity and the bath, is always of AFM character with a maximum close to $U_{\rm K}$$\sim$1.6. On the other hand the inter-impurity correlation $\langle S_1S_2\rangle$ is exclusively of FM kind and shows monotonic increase with $U$. The former is associated with the singlet-forming tendencies due to Kondo screening, whereas the latter signals triplet-forming tendencies because of the FM RKKY exchange within the local limit. Close to $U_{\rm K}$ the absolute value of the local RKKY correlation exceeds the singlet-forming amplitude between impurity and bath. The system at larger $U$ is then dominated by the RKKY interaction. [@fye87; @nis06; @ferr09] Within a conventional Schrieffer-Wolff mapping [@schr66] for the Kondo coupling via $J_{\rm K}$=$8V^2/U$, a similar crossover regime would follow also from simple estimates through the associated exchange interactions. For if we understand the RKKY interaction as second-order process, i.e. $J_{\rm RKKY}$$\sim$$J_{\rm K}^2$, then here the two exchange integrals become equivalent for $U$=2, which is the order of magnitude from the numerics. With increasing impurity-bath coupling $V$ the crossover shifts to larger $U$, since $J_{\rm K}^2$ stronger profits therefrom. However note that with our bath bandwidth $W$=6 the present TIAM is surely not in the Kondo-Hamiltonian limit ($U$$\gg$$W$) for the studied interaction range. [@fye89]
![(Color online) Interaction-dependent spin-spin correlations for the TIAM with $t_{12}$=0.2, with and without DM interaction. \[fig:taim2\]](Spcf_t0.2.eps){width="8.5cm"}
Turning on a finite DM term of size $D$=0.05 has nearly no effect at small $U$. However for larger Hubbard interaction rather strong modifications occur especially for the inter-dimer function $\langle S_{1,y}S_{2,y}\rangle$. Remember that the $D$ vector also points in the $y$ direction. Thus an intricate spin-spin coupling scenario arises at large $U$, with still FM alignment in the $(x,z)$ axes and near AFM alignment in the $y$ axes. For $U$$>$5 our mean-field approach yields net local moments in presence of a finite $D$, i.e., a paramagnetic solution is no longer stabilizable. It would thus be very interesting to study the large-$U$ regime of this model beyond mean field (e.g. with the numerical renormalization group approach utilized in Ref. ).
In addition to the basic model with vanishing inter-impurity hopping, Fig. \[fig:taim2\] exhibits the resulting spin correlation functions for the TIAM with $t_{12}$=0.2. Now both $\langle S_{\rm imp}S_{\rm bath}\rangle$ and $\langle S_1S_2\rangle$ display AFM correlations in the weakly interacting limit. This is understood from the direct exchange integral $J_{\rm dir}$=$4t_{12}^2/U$ originating from the introduced dimer coupling. With increasing $U$ the correlation functions develop rather similar as for $t_{12}$=0, yet the overall magnitude is somewhat reduced a small interaction strength. Hence there the direct exchange weakens both, impurity Kondo-screening (due to the stronger inter-impurity link) as well as FM RKKY interaction (since the direct exchange favors AFM behavior). But the crossover point of domination for these processes does not seem to change much with the introduced $t_{12}$. Of course, a very large $t_{12}$ should rank the direct exchange above the other mechanisms, however here we do not investigate this model limit. Finally, when introducing the DM term to the model, effectively four different exchange mechanisms compete with each other: impurity Kondo, RKKY, direct and DM. The latter has indeed again significant effect on the spin correlation between the impurities. For already moderate values of $U$ the dominance of the FM RKKY is lost, turning the system into AFM-like inter-impurity correlations for $U$$>$2. Thus also here the DM interaction severely influences the magnetic correlations for isolated impurities within an itinerant background. It appears to strengthen the singlet-forming tendencies (with stronger response in the ${\bf D}$ direction) in an otherwise triplet-favoring RKKY system at short-range distance.
Summary
=======
A theoretical investigation of effects stemming from the Dzyaloshinskii-Moriya interaction in itinerant systems with strong electronic correlations was presented in this work. In order to study the principle physics on the lattice as well as in the local limit, we elaborated on two prominent model systems, namely the Hubbard bilayer and the one defined by the two-impurity Anderson Hamiltonian. In both cases substantial influence of the DM interaction is found, especially at large coupling where the strong renormalization enhances the impact. The half-filled Hubbard bilayer displays large out-of-axis spin components close to the Mott transition that may severely influence the magnetic response in applied field. Intriguing phenomena in this respect are e.g. observed in quasi-two-dimensional organic compounds. [@kag08] At finite hole doping and larger $U$, the bilayer system with DM interaction exhibits the emergence of a metallic spin-flop phase inbetween the AFM phase at half filling and the PM phase at stronger doping. This finding is of vital importance for many doped Mott systems with anisotropies. For instance, it is well-known that the hole-doped layered cuprates display puzzling phases inbetween the AFM and the superconducting dome and that the DM interaction is not completely negligible at low energy. [@thi88; @cof91; @jur06] Thus it would be very interesting to investigate in some detail whether there is a closer connection between our model results and those observations.
The TIAM is very relevant not only in the context of heavy fermions, but e.g. also for isolated atoms on metallic surfaces. In either case, anisotropic spin terms such as the DM interaction exist in many realistic representants in nature. It results from our studies that the DM term becomes an important player in the hierachy of relevant exchange processes in these contexts. In the local limit it works against the FM tendencies of the RKKY interaction and promotes the singlet formation between the impurities at large local Coloumb interactions. Further research along these lines, e.g. by going beyond mean-field, including the complete $k$ dependence of the impurity-bath coupling or tailoring the modeling towards concrete materials systems is of vital interest to account for generic exchange processes in the strongly correlated itinerant regime.
We thank M. Potthoff and D. Grieger for helpful discussions. Financial support from the Free and Hanseatic City of Hamburg in the context of the Landesexzellenzinitiative Hamburg as well as the DFG-SPP 1386 is gratefully acknowledged. Computations were performed at the local computing center of the University of Hamburg as well as the North-German Supercomputing Alliance (HLRN) under the grant hhp00026.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Hankel tensors arise from signal processing and some other applications. SOS (sum-of-squares) tensors are positive semi-definite symmetric tensors, but not vice versa. The problem for determining an even order symmetric tensor is an SOS tensor or not is equivalent to solving a semi-infinite linear programming problem, which can be done in polynomial time. On the other hand, the problem for determining an even order symmetric tensor is positive semi-definite or not is NP-hard. In this paper, we study SOS-Hankel tensors. Currently, there are two known positive semi-definite Hankel tensor classes: even order complete Hankel tensors and even order strong Hankel tensors. We show complete Hankel tensors are strong Hankel tensors, and even order strong Hankel tensors are SOS-Hankel tensors. We give several examples of positive semi-definite Hankel tensors, which are not strong Hankel tensors. However, all of them are still SOS-Hankel tensors. Does there exist a positive semi-definite non-SOS-Hankel tensor? The answer to this question remains open. If the answer to this question is no, then the problem for determining an even order Hankel tensor is positive semi-definite or not is solvable in polynomial-time. An application of SOS-Hankel tensors to the positive semi-definite tensor completion problem is discussed. We present an ADMM algorithm for solving this problem. Some preliminary numerical results on this algorithm are reported.
[**Key words:**]{} Hankel tensors, generating vectors, sum of squares, positive semi-definiteness, generating functions, tensor completion.
[**AMS subject classifications (2010):**]{} 15A18; 15A69
author:
- 'Guoyin Li [^1] Liqun Qi[^2] Yi Xu [^3]'
title: 'SOS-Hankel Tensors: Theory and Application'
---
Introduction
============
In general, most tensor problems are NP-hard [@HL]. However, most tensor problems in applications have special structures, and they are not NP-hard. In the last few years, there are a number of papers on the Perron-Frobenius theory for nonnegative tensors, and algorithms for computing the largest eigenvalue of a nonnegative tensor [@CPZ; @CQZ; @NQZ]. In particular, in [@HLQS], it was shown that the problem for computing the largest H-eigenvalue of an essentially nonnegative tensor, which includes the problem for computing the largest eigenvalue of a nonnegative tensor as a special problem, can be solved by solving a semi-definite linear programming problem. Thus, this problem is polynomial-time solvable, not NP-hard. This method can be used to find the smallest H-eigenvalue of a Z tensor, thus to be used to determine a given Z tensor is an M tensor or not, while an even order symmetric M tensor is positive semi-definite [@ZQZ].
The problem for determining a given even order symmetric tensor is positive semi-definite or not has important applications in engineering and science [@Hi; @Qi; @QYW; @QYX; @Re; @Sh]. In general, this problem is NP-hard. However, for special tensor classes, such as even order symmetric Z tensors, as pointed above, this problem is polynomial time solvable. In 2014, more classes of structured tensors have been identified, either such tensors are easily to be identified, and they are positive semi-definite or positive definite in the even order symmetric case, or there are easily checkable conditions to identify such tensors are positive semi-definite or not. The former includes Hilbert tensors [@SQ], diagonally dominated tensors [@QS], B tensors [@QS], double B tensors [@LL], quasi-double B tensors [@LL] and H$^+$ tensors [@LWZZL]. The latter includes Cauchy tensors [@CQ].
In [@HLQ], a new class of positive semi-definite tensors, called SOS (sum-of-squares) tensors, was introduced. SOS tensors are positive semi-definite tensors, but not vice versa. SOS tensors are connected with SOS polynomials, which are significant in polynomial theory [@Hi; @Re; @Sh] and polynomial optimization [@Las; @Lau]. In particular, as stated above, the problem to identify a given general even order symmetric tensor is positive semi-definite or not is NP-hard, while the problem to identify a given general even order symmetric tensor is SOS or not is equivalent to solving a semi-definite linear programming problem [@Las; @Lau], thus not NP-hard, but polynomial time solvable. However, no special structured tensor class was identified to be SOS tensors in [@HLQ].
Hankel tensors arise from signal processing and some other applications [@BB; @DQW; @PDV; @Qi15]. Hankel tensors are symmetric tensors. In [@Qi15], positive semi-definite Hankel tensors were studied. Each Hankel tensor is associated with an Hankel matrix. If that Hankel matrix is positive semi-definite, then the Hankel tensor is called a strong Hankel tensor. It was proved that an even order strong Hankel tensor is positive semi-definite. A symmetric tensor is a Hankel tensor if and only if it has a Vandermonde decomposition. If the coefficients of that Vandermonde decomposition are nonnegative, then the Hankel tensor is called a complete Hankel tensor. It was proved that an even order complete Hankel tensor is also positive semi-definite. The relation between strong and complete Hankel tensors was not known in [@Qi15]. An example of a positive semi-definite Hankel tensor, which is neither strong nor complete Hankel tensor was also given in [@Qi15].
In this paper, we study positive semi-definite and SOS Hankel tensors. We introduce completely decomposable tensors. Even order completely decomposable tensors are SOS tensors, but not vice versa. We show that complete Hankel tensors are strong Hankel tensors, while strong Hankel tensors are completely decomposable. Thus, both even order complete Hankel tensors and even order strong Hankel tensors are SOS Hankel tensors. From now on, we call SOS Hankel tensors as SOS-Hankel tensors. Then we show that for any even order $m=2k\ge 4$, there are SOS-Hankel tensors, which are not completely decomposable. This includes the example given in [@Qi15]. We also found some other examples of SOS-Hankel tensors which are not strong Hankel tensors. Does there exist a positive semi-definite non-SOS Hankel tensor? The answer to this question remains open. If the answer to this question is no, then the problem for determining an even order Hankel tensor is positive semi-definite or not is solvable in polynomial-time.
The remainder of this paper is organized as follows. Some preliminary knowledge is given in the next section. In Section 3, we introduce completely decomposable tensors and discuss their properties. We prove completely Hankel tensors are strong Hankel tensors, and strong Hankel tensors are completely decomposable in Section 4. Some SOS-Hankel tensors which are not strong Hankel tensors are given in Section 5. Then we raise the question on positive semi-definite non-SOS Hankel tensors and have some discussion. Finally, in Section 6, we discuss an application of SOS-Hankel tensors to the positive semi-definite tensor completion problem. We present an ADMM algorithm for solving this problem. Some preliminary numerical results on this algorithm are reported.
Preliminaries
=============
Throughout this paper, we use small letters $a, b, \alpha, \beta, \lambda, \cdots$, for scalars, bold small letters ${{\bf x}}, {{\bf y}}, {{\bf u}}, {{\bf v}}, \cdots$, for vectors, capital letters $A, B, C, I, \cdots$, for matrices, calligraphic letters ${{\cal A}}, {{\cal B}}, {{\cal C}}, {{\cal H}}, \cdots$, for tensors. We use ${{\bf 0}}$ to denote the zero vector in $\Re^n$, and ${{\bf e}}_i$ to denote the $i$th unit vector in $\Re^n$, for $i \in [n]$. We use $x_i$ to denote the $i$th component of ${{\bf x}}$.
Denote that $[n] := \{ 1, \cdots, n \}$. Throughout this paper, $m, n$ and $k$ are integers, and $m, n \ge 2$. A tensor ${{\cal A}}= (a_{i_1\cdots i_m})$ of order $m$ and dimension $n$ has entries $a_{i_1\cdots i_m}$ with $i_j \in [n]$ for $j \in [m]$. Tensor ${{\cal A}}$ is said to be a symmetric tensor if its entries $a_{i_1\cdots i_m}$ is invariant under any index permutation. Denote the set of all the real symmetric tensors of order $m$ and dimension $n$ by ${\rm S}_{m, n}$. Then ${\rm S}_{m, n}$ is a linear space. Throughout this paper, we only discuss real symmetric tensors. We use $\| {{\cal A}}\|$ to denote the Frobenius norm of tensor ${{\cal A}}= (a_{i_1\cdots i_m})$, i.e., $\| {{\cal A}}\| = \sum_{i_1, \cdots, i_m \in [n]} a_{i_1 \cdots i_m}^2$.
Let ${{\bf x}}\in \Re^n$. Then ${{\bf x}}^{\otimes m}$ is a rank-one symmetric tensor with entries $x_{i_1}\cdots x_{i_m}$. For ${{\cal A}}\in {\rm S}_{m, n}$ and ${{\bf x}}\in \Re^n$, we have a homogeneous polynomial $f({{\bf x}})$ of $n$ variables and degree $m$, $$\label{e2.0}
f({{\bf x}}) = {{\cal A}}{{\bf x}}^{\otimes m} \equiv \sum_{i_1, \cdots, i_m \in [n]}a_{i_1\cdots i_m}x_{i_1}\cdots x_{i_m}.$$ Note that there is a one to one relation between homogeneous polynomials and symmetric tensors.
If $f({{\bf x}}) \ge 0$ for all ${{\bf x}}\in \Re^n$, then homogeneous polynomial $f({{\bf x}})$ and symmetric tensor ${{\cal A}}$ are called [**positive semi-definite**]{} (PSD). If $f({{\bf x}}) > 0$ for all ${{\bf x}}\in \Re^n, {{\bf x}}\not = {{\bf 0}}$, then $f({{\bf x}})$ and ${{\cal A}}$ are called [**positive definite**]{}. Clearly, if $m$ is odd, there is no positive definite symmetric tensor and there is only one positive semi-definite tensor ${{\cal O}}$. Thus, we assume that $m = 2k$ when [ we discuss positive definite and semi-definite tensors (polynomials). Positive semi-definite polynomials, or called nonnegative polynomials,]{} have important significance in mathematics, and are connected with the 17th Hilbert problem [@Hi; @Re; @Sh]. The concepts of positive semi-definite and positive definite symmetric tensors were introduced in [@Qi]. They have wide applications in science and engineering [@Qi; @QY; @QYW; @QYX]. Denote the set of all positive semi-definite symmetric tensors in ${\rm S}_{m, n}$ by ${\rm PSD}_{m, n}$. Then it is a closed convex cone and has been studied in [@HLQ; @QY].
Let $m=2k$. If $f({{\bf x}})$ can be decomposed to the sum of squares of polynomials of degree $k$, then $f({{\bf x}})$ is called a [**sum-of-squares (SOS) polynomial**]{}, and the corresponding symmetric tensor ${{\cal A}}$ is called an [**SOS tensor**]{} [@HLQ]. SOS polynomials play a central role in the modern theory of polynomial optimization [@Las; @Lau]. SOS tensors were introduced in [@HLQ]. Clearly, an SOS polynomial (tensor) is a PSD polynomial, but not vice versa. Actually, this was shown by young Hilbert [@CL; @Hi; @Mo; @Re; @Sh] that for homogeneous polynomial, only in the following three cases, a PSD polynomial definitely is an SOS polynomial: 1) $n = 2$; 2) $m = 2$; 3) $m=4$ and $n=3$. For tensors, the second case corresponds to the symmetric matrices, i.e., a PSD symmetric matrix is always an SOS matrix. Hilbert proved that in all the other possible combinations of $m=2k$ and $n$, there are non-SOS PSD homogeneous polynomials [@Sh]. The most well-known non-SOS PSD homogeneous polynomial is the Motzkin polynomial [@Mo] $$f_M({{\bf x}}) = x_3^6 + x_1^2x_2^4 + x_1^4x_2^2 - 3 x_1^2x_2^2x_3^2.$$ By the Arithmetic-Geometric inequality, we see that it is a PSD polynomial. But it is not an SOS polynomial [@Re; @Sh]. The other two non-SOS PSD homogeneous polynomials with small $m$ and $n$ are given by Choi and Lam [@CL] $$f_{CL1}({{\bf x}}) = x_4^4 + x_1^2x_2^2 + x_1^2x_3^2 + x_2^2x_3^2 - 4x_1x_2x_3x_4$$ and $$f_{CL2}({{\bf x}}) = x_1^4x_2^2 + x_2^4x_3^2 + x_3^4x_1^2 - 3x_1^2x_2^2x_3^2.$$ Denote the set of all SOS tensors in ${\rm S}_{m, n}$ by ${\rm SOS}_{m, n}$. Then it is also a closed convex cone [@HLQ]. Thus, ${\rm SOS}_{m, 2} = {\rm PSD}_{m, 2}$, ${\rm SOS}_{2, n} = {\rm PSD}_{2, n}$ and ${\rm SOS}_{4, 3} = {\rm PSD}_{4, 3}$. But for other $m =2k \ge 4, n \ge 3$, we have ${\rm SOS}_{m, n} \subsetneq {\rm PSD}_{m, n}$.
[ Let ${{\bf v}}= (v_0, \cdots, v_{(n-1)m})^\top$. Define ${{\cal A}}= (a_{i_1\cdots i_m}) \in {\rm S}_{m, n}$ by $$\label{e2.1}
a_{i_1\cdots i_m} = v_{i_1+\cdots + i_m-m},$$ for $i_1, \cdots, i_m \in [n]$.]{} Then ${{\cal A}}$ is a [**Hankel tensor**]{} and ${{\bf v}}$ is called the [**generating vector**]{} of ${{\cal A}}$. In (\[e2.0\]), if ${{\cal A}}$ is a Hankel tensor, [ then]{} $f({{\bf x}})$ is called a [**Hankel polynomial**]{}. We see that a sufficient and necessary condition for ${{\cal A}}= (a_{i_1\cdots i_m}) \in S_{m, n}$ to be a Hankel tensor is that whenever $i_1+\cdots + i_m = j_1+ \cdots + j_m$, $$\label{e2.2}
a_{i_1\cdots i_m} = a_{j_1\cdots j_m}.$$ Hankel tensors arise from signal processing and other applications [@BB; @DQW; @PDV; @Qi15].
By (\[e2.2\]), the three non-SOS PSD polynomials $f_M({{\bf x}})$, $f_{CL1}({{\bf x}})$ and $f_{CL2}({{\bf x}})$ are not Hankel polynomials. These three polynomials are still non-SOS PSD polynomials if we switch the indices of their variables. However, if we switch the indices of their variables and add some terms to them to make them become Hankel polynomials, then the resulted Hankel polynomials are not positive semi-definite. Hence, no non-SOS PSD Hankel polynomials are resulted. There are other examples of non-SOS PSD polynomials [@Re]. None of them are Hankel polynomials.
Note that for $n = 2$, all symmetric tensors are Hankel tensors.
Completely Decomposable Tensors
===============================
Let ${{\cal A}}\in {\rm S}_{m, n}$. If there are vectors ${{\bf x}}_j \in \Re^n$ for $j \in [r]$ such that $${{\cal A}}= \sum_{j \in [r]} {{\bf x}}_j^{\otimes m},$$ then we say that ${{\cal A}}$ is a [**completely $r$-decomposable tensor**]{}, or a [**completely decomposable tensor**]{}. If ${{\bf x}}_j \in \Re^n_+$ for all $j \in [r]$, then ${{\cal A}}$ is called a [**completely positive tensor**]{} [@QXX]. If ${{\bf x}}_j \in \Re^n_+$ for all $j \in [r]$, then ${{\cal A}}$ is called a [**completely positive tensor**]{}. [ We note that ]{} any odd order symmetric tensor is completely decomposable. Hence, it does not make sense to discuss odd order completely decomposable tensor. But it is still meaningful to show an odd order symmetric tensor is completely $r$-decomposable, as this means that its symmetric rank [@CGLM] is not greater than $r$.
Denote the set of all completely $r$-decomposable tensors in ${\rm S}_{m, n}$ by ${\rm CD}^r_{m, n}$, and the set of all completely decomposable tensors in ${\rm S}_{m, n}$ by ${\rm CD}_{m, n}$. The following theorem will be useful in the next section.
[**(Closedness and convexity of the tensor cones)**]{} \[t3.1\] The following statements hold:
- If the order $m$ is an even number, then ${\rm SOS}_{m,n}$ and ${\rm PSD}_{m,n}$ are closed and convex cones.
- If the order $m$ is an even number, ${\rm CD}_{m,n}^r$ is a closed cone for each integer $r$.
- If the order $m$ is an odd number, then
- ${\rm CD}_{m,n}^{r}$ is either a nonconvex cone such that ${\rm CD}_{m,n}^{r}=-{\rm CD}_{m,n}^{r}$ or the whole symmetric tensor space ${\rm S}_{m,n}$. In particular, for any $r<n$, ${\rm CD}_{m,n}^{r}$ is a nonconvex cone such that ${\rm CD}_{m,n}^{r}=-{\rm CD}_{m,n}^{r}$;
- ${\rm CD}_{m,n}^{r}$ can be non-closed for some integer $r$.
- If the order $m$ is an even number, then ${\rm CD}_{m,n}$ is a closed and convex cone. If $m$ is an odd number, then ${\rm CD}_{m,n}={\rm S}_{m,n}$.
[**Proof.**]{}
\[Proof of [(i)]{}\] The closed and convexity of positive semidefinite cone ${\rm PSD}_{m,n}$ can be directly verified from the definition. The convexity of the SOS tensor cone ${\rm
SOS}_{m,n}$ also directly follows from the definition. To see the closedness of the SOS tensor cone ${\rm SOS}_{m,n}$, let ${{\cal A}}_k \in {\rm SOS}_{m,n}$ with $\mathcal{A}_k \rightarrow
{{\cal A}}\in {\rm S}_{m,n}$. So, $f_k({{\bf x}})={{\cal A}}_k{{\bf x}}^{\otimes m}$ is an SOS polynomial and $f_k(x) \rightarrow f(x)$ with $f({{\bf x}})={{\cal A}}{{\bf x}}^{\otimes m}$. Note from [@Lau Corollary 3.50] that the set of all sum-of-squares polynomials on $\Re^n$ (possibly nonhomogeneous) with degree at most $m$ is a closed cone. So, $f$ is also an SOS polynomial. Therefore, ${{\cal A}}$ is an SOS tensor.
\[Proof of [(ii)]{}\] To see the closedness of ${\rm CD}^r_{m,n}$, we let ${{\cal A}}_k \in {\rm CD}^r_{m,n}$ with ${{\cal A}}_k
\rightarrow {{\cal A}}$. Then, for each integer $k$, there exist ${{\bf x}}^{k,j} \in \Re^n$, $j \in [r]$, such that ${{\cal A}}_k = \sum_{j\in [r]} \left({{\bf x}}^{k,j}\right)^{\otimes m}$. As ${{\cal A}}_k \rightarrow {{\cal A}}$, $\left\{\|{{\cal A}}_k\|\right\}_{k
\in \mathbb{N}}$ is a bounded sequence. Note that $\|{{\cal A}}_k\|^2 \ge \sum_{j \in [r]}\left(\sum_{i=1}^n
\left(x^{k,j}_i\right)^m\right)^2$. So, $\{{{\bf x}}^{k,j}\}_{k \in \mathbb{N}}$, $j=1,\ldots,r$, are bounded sequences. By passing to subsequences, we can assume that ${{\bf x}}^{k,j} \ \rightarrow {{\bf x}}^j$, $j=1,\ldots,r$. Passing to the limit, we have $${{\cal A}}= \sum_{j\in [r]} \left({{\bf x}}^j\right)^{\otimes m} \in {\rm CD}^r_{m,n}.$$ Thus, the conclusion follows.
\[Proof of [(iii) (a)]{}\] First of all, it is easy to see that ${\rm CD}_{m,n}^r$ is a cone. From the definition of ${\rm CD}_{m,n}^r$, if ${{\cal A}}\in {\rm CD}_{m,n}^r$, then [ there exist ${{\bf u}}_1,\ldots,{{\bf u}}_r \in \Re^n$]{} such that ${{\cal A}}=\sum_{i=1}^r u_i^{\otimes m}$. As $m$ is odd, $-\mathcal{A}=\sum_{i=1}^r (-u_i)^{\otimes m}$, and so, $-\mathcal{A} \in {\rm CD}_{m,n}$. So, ${\rm CD}_{m,n}^r = -{\rm CD}_{m,n}^r$. To finish the proof, it suffices to show that if $CD_{m,n}^r$ is convex, then it must be the whole space ${\rm S}_{m,n}$. To see this, we suppose that ${\rm CD}_{m,n}^r$ is convex. Then, ${\rm CD}_{m,n}^r$ is a convex cone with ${\rm CD}_{m,n}^r = -{\rm CD}_{m,n}^r$. This implies that ${\rm CD}_{m,n}^r$ is a subspace. Now, take ${{\cal A}}\in ({\rm CD}_{m,n}^r)^{\bot}$ where $L^{\bot}$ denotes the orthogonal complement in ${\rm S}_{m,n}$ of the subspace $L$. Then, $\langle {{\cal A}}, {{\bf x}}^{\otimes m} \rangle={{\cal A}}{{\bf x}}^{\otimes m}=0$ for all ${{\bf x}}\in \Re^n$. This shows that ${{\cal A}}={{\cal O}}$. So, $({\rm CD}_{m,n}^r)^{\bot}=\{{{\cal O}}\}$, and hence ${\rm CD}_{m,n}^r={\rm S}_{m,n}$.
Now, let us consider the case when $r<n$. To see that assertion that ${\rm CD}_{m,n}^{r}$ is a nonconvex cone such that ${\rm CD}_{m,n}^{r}=-{\rm CD}_{m,n}^{r}$, we only need to show ${\rm CD}_{m,n}^r \neq {\rm S}_{m,n}$. To see this, let ${{\bf e}}_i$ denotes the vector whose $i$th coordinate is one and the others are zero, $i \in [n]$, and define $\overline{{{\cal A}}}:=\sum_{i\in [n]} {{\bf e}}_i^{\otimes m}$. In particular, we see that $\overline{{{\cal A}}} {{\bf e}}_j^{\otimes (m-1)}={{\bf e}}_j$ for all $j\in [n]$. We now show that $\overline{{{\cal A}}} \notin {\rm CD}_{m,n}^r$. Otherwise, [ there exist ${{\bf u}}_1,\ldots,{{\bf u}}_r \in \Re^n$]{} such that $\overline{{{\cal A}}}=\sum_{i\in [r]} ({{\bf u}}_i)^{\otimes m}$. This implies that, for each $j \in [n]$, $${{\bf e}}_j=\overline{{{\cal A}}} {{\bf e}}_j^{\otimes (m-1)}=\sum_{i\in [r]} \alpha_{ij} {{\bf u}}_i \mbox{ with } \alpha_{ij}=\left(({{\bf u}}_i)^\top {{\bf e}}_j\right)^{m-1} \in \Re.$$ This implies that ${{\bf e}}_j \in {\rm span}\{{{\bf u}}_1,\ldots,{{\bf u}}_r\}$ for all $j \in [n]$, and so, ${\rm dim}{\rm span}\{{{\bf u}}_1,\ldots,{{\bf u}}_r\} \ge n$. This is impossible as we assume that $r<n.$
\[Proof of [(iii) (b)]{}\] We borrow an example in [@CGLM] to illustrate the possible non-closedness of the cone ${\rm
CD_{m,n}^r}$ in the case where $m$ is odd [^4]. Let $m=3$ and $n=2$. For each $\epsilon>0$, let $$f_{\epsilon}(x_1,x_2)=\epsilon^2(x_1+\epsilon^{-1}x_2)^3+ \epsilon^2(x_1-\epsilon^{-1}x_2)^3$$ and let ${{\cal A}}_{\epsilon}$ be the associated symmetric tensor, that is, ${{\cal A}}_{\epsilon}{{\bf x}}^{\otimes m}=f_{\epsilon}({{\bf x}})$ for all ${{\bf x}}\in \Re^2$. Clearly ${{\cal A}}_{\epsilon} \in {\rm CD}_{3,2}^2$. Then, $f_{\epsilon}\rightarrow f$ as $\epsilon \rightarrow 0$ where $f(x_1,x_2)=6x_1x_2^2$. Let ${{\cal A}}$ be the associated symmetric tensor of $f$. Then, we have ${{\cal A}}_{\epsilon} \rightarrow {{\cal A}}$. We now see that ${{\cal A}}\notin {\rm CD}_{3,2}^2$. To see this we proceed by contradiction. Then there exist $a_1,b_1,a_2,b_2 \in \Re$ such that $$6x_1x_2^2=(a_1x_1+a_2x_2)^3+(b_1x_1+b_2x_2)^3.$$ By comparing the coefficients, we have $a_1=-b_1$, $a_2=-b_2$ and $a_1a_1^2+b_1b_2^2=2$, which is impossible.
\[Proof of [(iv)]{}\] From the definition, ${\rm CD}_{m,n}$ is a convex cone. As [ ${\rm CD}_{m,n} \subseteq {\rm S}_{m,n}$, the dimension of ${\rm CD}_{m,n}$ is at most $I(m,n)$ where $I(m,n)=\left(\begin{array}{cc}
n+m-1 \\
n-1
\end{array}
\right)$]{}. By Carathéodory’s theorem and the fact that ${\rm CD}_{m,n}$ is a convex cone, ${\rm CD}_{m,n}={\rm CD}_{m,n}^r$ with $r=I(m,n)+1$. Thus, [(ii)]{} implies that ${\rm CD}_{m,n}$ is a closed convex cone if the order $m$ is an even number. We now consider the case when $m$ is an odd number. Similar to the argument in [(iii)]{}, we see that ${\rm CD}_{m,n}=-{\rm CD}_{m,n}$. So, ${\rm CD}_{m,n}$ is a subspace when $m$ is odd. Now, take ${{\cal A}}\in ({\rm CD}_{m,n})^{\bot}$. Then, ${{\cal A}}{{\bf x}}^{\otimes m}=0$ for all ${{\bf x}}\in \Re^n$. This shows that ${{\cal A}}={{\cal O}}$. So, $({\rm CD}_{m,n})^{\bot}=\{{{\cal O}}\}$, and hence ${\rm CD}_{m,n}={\rm S}_{m,n}$.
The intricate non-closedness of the set which consists of all rank-$r$ symmetric tensors over the complex field was discovered and examined in detailed in [@CGLM]. This fact shows the significant difference between a matrix and a tensor with order greater than $2$. Our results here shows that the same feature can happen for the complete $r$-decomposable symmetric tensors where the order is an odd number.
As we have seen in Theorem \[t3.1\] [(iii)(a)]{}, the cone ${\rm CD}_{m,n}^r$ is not convex when $m$ is an odd number and $r<n$. We now illustrate this by a simple example.
[**(Illustrating the possible nonconvexity of ${\rm CD}_{m,n}^r$)**]{} Consider $m=3$, $n=2$ and $r=1$. Then, ${\rm CD}_{3,2}^1=\{{{\cal A}}\in {\rm S}_{3,2}: {{\cal A}}{{\bf x}}^{\otimes 3}=\left({{\bf u}}^\top {{\bf x}}\right)^3 \mbox{ for some } {{\bf u}}\in \Re^2\}$. Let ${{\cal A}}_i \in {\rm S}_{3,2}$ be such that ${{\cal A}}_i {{\bf x}}^{\otimes 3}= x_i^3$, $i=1,2$. We now claim that $\frac{1}{2}{{\cal A}}_1+\frac{1}{2}{{\cal A}}_2 \notin {\rm CD}_{3,2}^1$. Otherwise, there exists ${{\bf u}}=(u_1,u_2)^\top \in \Re^2$ such that $$\frac{1}{2}x_1^3+\frac{1}{2}x_2^3=(u_1x_1+u_2x_2)^3.$$ Comparing the coefficients, we see that $u_1=u_2=\sqrt[3]{\frac{1}{2}}$ and $3u_1u_2^3=3u_1^3u_2=0$. This is impossible. So, we have $\frac{1}{2}{{\cal A}}_1+\frac{1}{2}{{\cal A}}_2 \notin {\rm CD}_{3,2}^1$, and hence ${\rm CD}_{3,2}^1$ is not convex.
Strong Hankel Tensors and Complete Hankel Tensors
=================================================
Suppose that ${{\cal A}}$ is a Hankel tensor defined by (\[e2.1\]). Let $A = (a_{ij})$ be an $\lceil
{(n-1)m+2 \over 2} \rceil \times \lceil {(n-1)m+2 \over 2} \rceil$ matrix with $a_{ij} \equiv v_{i+j-2}$, where $v_{2\lceil {(n-1)m
\over 2} \rceil}$ is an additional number when $(n-1)m$ is odd. Then $A$ is a Hankel matrix, associated with the Hankel tensor ${{\cal A}}$. Clearly, when $m$ is even, such an associated Hankel matrix is unique. Recall from [@Qi15] that ${{\cal A}}$ is called a [**strong Hankel tensor**]{} if [ there exists an associated Hankel matrix $A$ is positive semi-definite]{}. Thus, whether a tensor is a strong Hankel tensor or not can be verified by using tools from matrix analysis. It has also been shown in [@Qi15] that ${{\cal A}}$ is a strong Hankel tensor if and only if it is a Hankel tensor and there exists an absolutely integrable real valued function $h:(-\infty, +\infty) \rightarrow [0,+\infty)$ such that its generating vector ${{\bf v}}= (v_0, v_1, \ldots, v_{(n-1)m})^\top$ satisfies $$\label{strong_Hankel}
v_k=\int_{-\infty}^{\infty}t^k h(t) dt, \ \ \ k=0,1,\ldots,(n-1)m.$$ Such a real valued function $h$ is called the [**generating function**]{} of the strong Hankel tensor ${{\cal A}}$.
A vector ${{\bf u}}= (1, \gamma, \gamma^2, \ldots, \gamma^{n-1})^\top$ for some $\gamma \in \Re$ is called a Vandermonde vector [@Qi15]. If tensor ${{\cal A}}$ has the form $${{\cal A}}=\sum_{i\in [r]} \alpha_i({{\bf u}}_i)^{\otimes m},$$ where ${{\bf u}}_i$ for $i = 1, \ldots, r$, are all Vandermonde vectors, then we say that $\mathcal{A}$ has a Vandermonde decomposition. It was shown in [@Qi15] that a symmetric tensor is a Hankel tensor if and only if it has a Vandermonde decomposition. If the coefficients $\alpha_i$ for $i = 1, \ldots, r$, are all nonnegative, then $\mathcal{A}$ is called a [**complete Hankel tensor**]{} [@Qi15]. Clearly, a complete Hankel tensor is a completely decomposable tensor. Unlike strong Hankel tensors, there is no clear method to check whether a Hankel tensor is a complete Hankel tensor or not, as the Vandermonde decompositions of a Hankel tensor are not unique. It was proved in [@Qi15] that even order strong or complete Hankel tensors are positive semi-definite, but the relationship between strong Hankel tensors and complete Hankel tensors was unknown there.
In the following, we discuss the relation between strong Hankel tensors and complete Hankel tensors.
\[p4.1\] Let $m,n \in \mathbb{N}$. All the strong Hankel tensors of order $m$ and dimension $n$ form a closed convex cone. All the complete Hankel tensors of order $m$ and dimension $n$ form a convex cone. A complete Hankel tensor is a strong Hankel tensor. On the other hand, whenever $m$ is a positive even number and $n \ge 2$, there is a strong Hankel tensor which is not a complete Hankel tensor.
[**Proof.**]{} By definition, it is easy to see that all the complete Hankel tensors of order $m$ and dimension $n$ form a convex cone. Since each strong Hankel tensor of order $m$ and dimension $n$ is associated with a positive semi-definite Hankel matrix and all such positive semi-definite Hankel matrices form a closed convex cone, all the strong Hankel tensors of order $m$ and dimension $n$ form a closed convex cone.
Consider a rank-one complete Hankel tensor ${{\cal A}}={{\bf u}}^{\otimes
m}$, where ${{\bf u}}= (1, \gamma, \gamma^2, \ldots, \gamma^{n-1})^\top$. Then we see the generating vector of ${{\cal A}}$ is [ ${{\bf v}}= (1, \gamma,
\gamma^2, \ldots, \gamma^{(n-1)m})^\top$]{}. When $m$ is even, let $A$ be a Hankel matrix generated by ${{\bf v}}$. When $m$ is odd, let $A$ be a Hankel matrix generated by $(1, \gamma, \gamma^2, \ldots,
\gamma^{(n-1)m})^\top$. Then we see that $A$ is positive semi-definite. Thus, $\mathcal{A}$ is a strong Hankel tensor. From the definition of complete Hankel tensors, we see that a complete Hankel tensor is a linear combination of strong Hankel tensors with nonnegative coefficients. Since all the strong Hankel tensors of order $m$ and dimension $n$ form a closed convex cone, this shows that a complete Hankel tensor is a strong Hankel tensor.
Finally, assume that $m$ is a positive even number and $n \ge 2$. Let ${{\cal A}}={{\bf e}}_n^{\otimes m}$ where ${{\bf e}}_n=(0,\ldots,0,1)\in \Re^n$. It is easy to verify that ${{\cal A}}$ is a strong Hankel tensor. We now see that ${{\cal A}}$ is not a complete Hankel tensor. Assume on the contrary. Then, there exist $r \in \mathbb{N}$, $\alpha_i \ge 0$ and ${{\bf v}}_i \in \Re^n$, $i=1,\ldots,r$ such that $${{\cal A}}=\sum_{i\in [r]} \alpha_i ({{\bf v}}_i)^{\otimes m} \mbox{ with }
{{\bf v}}_i=\left(1,\gamma^i,\ldots,(\gamma^i)^{n-1}\right).$$ By comparing with the $(1,\ldots,1)$-entry, we see that $\alpha_i=0$, $i=1,\ldots,r$. This gives ${{\cal A}}={{\cal O}}$ which makes a contradiction.
In the following theorem, we will show that when the order is even, a strong Hankel tensor is indeed a limiting point of complete Hankel tensors.
[(**Completely decomposability of strong Hankel tensors)**]{}\[th:2\] Let $m,n \in \mathbb{N}$. Let ${{\cal A}}$ be an $m$th-order $n$-dimensional strong Hankel tensor. If the order $m$ is an even number, then $\mathcal{A}$ is a completely decomposable tensor and a limiting point of complete Hankel tensors. If the order $m$ is an odd number, then $\mathcal{A}$ is a completely $r$-decomposable tensor with $r=(n-1)m+1$.
[**Proof.**]{} Let $h$ be the generating function of the strong Hankel tensor $\mathcal{A}$. Then, for any ${{\bf x}}\in \Re^n$, $$\begin{aligned}
\label{eq:1}
f({{\bf x}}):={{\cal A}}{{\bf x}}^{\otimes m} & = & \sum_{i_1,i_2,\ldots,i_m=1}^n v_{i_1+i_2+\ldots+i_m-m} x_{i_1}x_{i_2} \ldots x_{i_m} \nonumber \\
& = & \sum_{i_1,i_2,\ldots,i_m=1}^n \left(\int_{-\infty}^{+\infty} t^{i_1+i_2+\ldots+i_m-m} h(t) dt \right)\, x_{i_1}x_{i_2} \ldots x_{i_m} \nonumber \\
&=&\int_{-\infty}^{+\infty} \left(\sum_{i_1,i_2,\ldots,i_m=1}^n t^{i_1+i_2+\ldots+i_m-m} x_{i_1}x_{i_2} \ldots x_{i_m} \right)h(t)dt \, \nonumber \\
&=& \int_{-\infty}^{+\infty} \left(\sum_{i=1}^n t^{i-1} x_{i}\right)^m h(t)dt = \lim_{l \rightarrow +\infty} f_l(x),\end{aligned}$$ where $$f_l({{\bf x}})=\int_{-l}^{l} \left(\sum_{i=1}^n t^{i-1} x_{i}\right)^m h(t)dt.$$ By the definition of Riemann integral, for each $l \ge 0$, we have $f_l({{\bf x}}) = \lim_{k\rightarrow \infty} f_l^k({{\bf x}})$, where $f_l^k({{\bf x}})$ is a polynomial defined by $$f_l^k({{\bf x}}):=\sum_{j=0}^{2kl} \frac{\left(\sum_{i=1}^n
(\frac{j}{k}-l)^{i-1} x_i \right)^m h(\frac{j}{k}-l)}{k}.$$ Fix any $l \ge 0$ and $k \in \mathbb{N}$. Note that $$\begin{aligned}
f_l^k({{\bf x}}):=\sum_{j=0}^{2kl} \frac{\left(\sum_{i=1}^n
(\frac{j}{k}-l)^{i-1} x_i \right)^m h(\frac{j}{k}-l)}{k}&= & \sum_{j=0}^{2kl}\left(\sum_{i=1}^n\frac{
(\frac{j}{k}-l)^{i-1} h(\frac{j}{k}-l)^{\frac{1}{m} }}{k^{\frac{1}{m}}} x_i\right)^m \\
& = & \sum_{j=0}^{2kl} (\langle {{\bf u}}_j, {{\bf x}}\rangle)^m,\end{aligned}$$ where ${{\bf u}}_j=\frac{
h\left(\frac{j}{k}-l\right)^{\frac{1}{m} }}{k^{\frac{1}{m}}}\left(1,\frac{j}{k}-l,\ldots,(\frac{j}{k}-l)^{n-1}\right)$. Here ${{\bf u}}_j$ are always well-defined as $h$ takes nonnegative values. Define ${{\cal A}}_l^k$ be a symmetric tensor such that $f_l^k({{\bf x}})={{\cal A}}_l^k {{\bf x}}^{\otimes m}$. Then, it is easy to see that each ${{\cal A}}_k^l$ is a complete Hankel tensor and thus a completely decomposable tensor. [ Note from Theorem \[t3.1\] (iv) that]{} the completely decomposable tensor cone $CD_{m,n}$ is a closed convex cone when $m$ is even. It then follows that ${{\cal A}}=\lim_{k \rightarrow \infty}\lim_{l \rightarrow
\infty}{{\cal A}}_k^l$ is a [ completely decomposable tensor and a limiting point of complete Hankel tensors.]{}
To see the assertion in the odd order case, we use a similar argument as in [@Qi15]. Pick real numbers $\gamma_1,\ldots,\gamma_{r}$ with $r=(n-1)m+1$ and $\gamma_i \neq
\gamma_j$ for $i \neq j$. Consider the following linear equation in $\alpha=(\alpha_1,\ldots,\alpha_r)$ with $$v_k=\sum_{i=1}^r \alpha_i \gamma_i^{k}, \ k=0,\ldots,(n-1)m.$$ Note that this linear equation always has a solution say $\bar \alpha= (\bar \alpha_1,\ldots,\bar \alpha_r)$ because the matrix in the above linear equation is a nonsingular Vandermonde matrix. Then, we see that $${{\cal A}}_{i_1,\ldots,i_m} =v_{i_1+\ldots+i_m-m}=\sum_{i=1}^r \bar
\alpha_i \gamma_i^{i_1+\ldots+i_m-m}=\sum_{i=1}^r \bar \alpha_i
\left(({{\bf u}}_i)^{\otimes m}\right)_{i_1,\ldots,i_m},$$ where ${{\bf u}}_i \in \Re^n$ is given by ${{\bf u}}_i=(1,\gamma_i,\ldots,\gamma_i^{n-1})^T$. This shows that ${{\cal A}}=\sum_{i\in [r]} \bar \alpha_i ({{\bf u}}_i)^{\otimes m}$. Now, as $m$ is an odd number, we have $${{\cal A}}=\sum_{i=1}^r \left({\bar \alpha_i}^{\frac{1}{m}}
{{\bf u}}_i\right)^{\otimes m} .$$ Therefore, $\mathcal{A}$ is a completely decomposable tensor and the last conclusion follows.
From the preceding theorem and Proposition \[p4.1\], we have the following corollary.
[**(Non-closedness of the even order complete Hankel tensor cone)**]{} \[example:02\] When $m$ is even and $n \ge 2$, the cone which consists of all the complete Hankel tensors of order $m$ and dimension $n$ is not closed. Its closure is the cone which consists of all the strong Hankel tensors of order $m$ and dimension $n$.
\[cor:2\] Let $n \in \mathbb{N}$ and let $m$ be an even number. Let $\mathcal{A}$ be an $m$th-order $n$-dimensional strong Hankel tensor. Then ${{\cal A}}$ is an SOS tensor and a PSD tensor.
[**Proof.**]{} This is a direct consequence of the above theorem and the fact that ${\rm CD}_{m,n} \subseteq {\rm SOS}_{m,n} \subseteq {\rm PSD}_{m,n}$ for even order $m$.
[**Question 1**]{} For Corollary \[example:02\], what is the situation if the order is odd?
We have seen that the strong Hankel tensor can be regarded as a checkable sufficient condition for positive semi-definite Hankel tensor. We now provide a simple necessary condition for positive semi-definite Hankel tensors. This condition can be verified by solving a feasibility problem of a semi-definite programming problem. For a set $C$ with finite elements, we use $\sharp\, C$ to denote the number of elements in the set $C$.
[**(SDP-type necessary condition for PSD Hankel tensors)**]{} Let $n \in \mathbb{N}$ and let $m$ be a positive even number. Let ${{\cal A}}$ be an $m$th-order $n$-dimensional positive semi-definite Hankel tensor defined by (\[e2.1\]) with generating vector ${{\bf v}}= (v_0, \cdots, v_{(n-1)m})^\top$. Denote $\alpha(m,k)=\sharp\{
(i_1,\ldots,i_m): i_1+\ldots+i_m=m+k\}$, $k=0,\ldots,(n-1)m$. Then, there exists a symmetric $\left(\frac{(n-1)m}{2}+1\right) \times
\left(\frac{(n-1)m}{2}+1\right)$ positive semi-definite matrix $Q$ such that $$\label{eq:SDP}
\alpha(m,k) \, v_k = \sum_{\alpha+\beta=k} Q_{\alpha,\beta}, \ k=0,1,\ldots,(n-1)m.$$
[**Proof.**]{} As ${{\cal A}}$ is an $m$th-order $n$-dimensional positive semi-definite Hankel tensor and ${{\bf v}}$ is its generating vector, $${{\cal A}}{{\bf x}}^{\otimes m}= \sum_{i_1,i_2,\ldots,i_m=1}^n v_{i_1+i_2+\ldots+i_m-m} x_{i_1}x_{i_2} \ldots x_{i_m} \ge 0 \mbox{ for all } {{\bf x}}\in \Re^n.$$ Consider ${{\bf x}}(t)=(x_1(t),\ldots, x_n(t))^\top$ with $x_i(t)=t^{i-1}$, $i=1,\ldots,n$ and $t \in \Re$. Then, for all $t \in \Re$, we have $$\phi(t)={{\cal A}}{{\bf x}}(t)^{\otimes m}=\sum_{i_1,i_2,\ldots,i_m=1}^n v_{i_1+i_2+\ldots+i_m-m} \, t^{i_1+i_2+\ldots+i_m-m}=\sum_{k=0}^{(n-1)m} \sum_{i_1+i_2+\ldots+i_m-m=k}v_k\, t^k \ge 0.$$ As $\phi$ is a one-dimensional polynomial which always takes nonnegative values, $\phi$ is a sums-of-squares [ polynomial [@Lau; @Las]. Define $w_t=\left(1,t,\ldots,t^{\frac{(n-1)m}{2}}\right)^T$. So,]{} there exists a symmetric $\left(\frac{(n-1)m}{2}+1\right) \times
\left(\frac{(n-1)m}{2}+1\right)$ positive semi-definite matrix $Q$ such that $$\sum_{k=0}^{(n-1)m} \alpha(m,k) v_k\, t^k=\sum_{k=0}^{(n-1)m} \sum_{i_1+i_2+\ldots+i_m-m=k}v_k\, t^k =\phi(t)= w_t^TQw_t,$$ which is further equivalent to (by comparing the entries) the relation (\[eq:SDP\]).
[**(Positive definiteness of strong Hankel tensors)**]{} Let $n \in \mathbb{N}$ and let $m$ be an even number. Let ${{\cal A}}$ be an $m$th-order $n$-dimensional strong Hankel tensor. Suppose that the generating function $h$ takes positive value almost everywhere. Then $\mathcal{A}$ is a positive definite tensor.
[**Proof.**]{} From (\[eq:1\]), for any ${{\bf x}}\in \Re^n$, $$\begin{aligned}
f({{\bf x}}):={{\cal A}}{{\bf x}}^{\otimes m} & = & \int_{-\infty}^{+\infty} \left(\sum_{i=1}^n t^{i-1} x_{i}\right)^m h(t)dt \, .\end{aligned}$$ [ From Corollary \[cor:2\], ${{\cal A}}{{\bf x}}^{\otimes m} \ge 0$ for all ${{\bf x}}\in \Re^n$.]{} Suppose that there exists $\bar {{\bf x}}\neq {{\bf 0}}$ such that ${{\cal A}}\bar {{\bf x}}^{\otimes m}=0$. Then, $$\int_{-\infty}^{+\infty} \left(\sum_{i=1}^n t^{i-1} \bar x_{i}\right)^m h(t)dt =0.$$ From our assumption that $h$ takes positive value almost everywhere, for each $l \ge 0$, we have $$\left(\sum_{i=1}^n t^{i-1} \bar x_{i}\right)^m = 0 \mbox{ for almost every } t \in [-l,l].$$ By the continuity, this shows that $\sum_{i=1}^n t^{i-1} \bar x_{i} \equiv 0 \mbox{ for all } t \in [-l,l]$. So, for each $l \ge 0$, we have $\bar x_1+ \sum_{i=2}^n t^{i-1} \bar x_i \equiv 0 \mbox{ for all } t \in [-l,l]$. Letting $t=0$, we have $\bar x_1=0$. Then, $$\sum_{i=2}^n t^{i-1} \bar x_i \equiv 0 \mbox{ for all } t \in [-l,l].$$ Repeating this process, we have $\bar x_2=\ldots=\bar x_n=0$. So, $\bar {{\bf x}}={{\bf 0}}$. Therefore, we see that ${{\cal A}}$ is positive definite.
[**(Complete positivity of strong Hankel tensors)**]{} Let $m,n \in \mathbb{N}$. Let ${{\cal A}}$ be an $m$th-order $n$-dimensional strong Hankel tensor with a generating function $h$ on $\Re$. Suppose that $\{t \in \mathbb{R}: h(t) \neq 0\}
\subseteq \mathbb{R}_+$. Then ${{\cal A}}$ is a completely positive tensor.
[**Proof.**]{} As $\{t \in \Re : h(t) \neq 0\} \subseteq \Re_+$, we have for any $x\in \Re^n$ $$\begin{aligned}
\label{eq:11}
f({{\bf x}}):={{\cal A}}{{\bf x}}^{\otimes m}
&=& \int_{-\infty}^{+\infty} \left(\sum_{i=1}^n t^{i-1} x_{i}\right)^m h(t)dt = \int_{0}^{+\infty} \left(\sum_{i=1}^n t^{i-1} x_{i}\right)^m h(t)dt.\end{aligned}$$ Then, using similar line of argument as in the preceding theorem, we see that $$f({{\bf x}})=\lim_{l \rightarrow +\infty}\lim_{k \rightarrow \infty} f_l^k({{\bf x}}),$$ where $f_l^k({{\bf x}})$ is given by $$\begin{aligned}
f_l^k({{\bf x}}):=\sum_{j=0}^{kl} \frac{\left(\sum_{i=1}^n
(\frac{j}{k})^{i-1} x_i \right)^m h(\frac{j}{k})}{k}&= & \sum_{j=0}^{kl}\left(\sum_{i=1}^n\frac{
(\frac{j}{k})^{i-1} h(\frac{j}{k})^{\frac{1}{m} }}{k^{\frac{1}{m}}} x_i\right)^m
= \sum_{j=0}^{kl} (\langle {{\bf u}}_j, {{\bf x}}\rangle)^m,\end{aligned}$$ where ${{\bf u}}_j=\left(\frac{
h(\frac{j}{k})^{\frac{1}{m} }}{k^{\frac{1}{m}}},\frac{
(\frac{j}{k}) h(\frac{j}{k}-l)^{\frac{1}{m}
}}{k^{\frac{1}{m}}},\ldots,\frac{ (\frac{j}{k})^{n-1}
h(\frac{j}{k})^{\frac{1}{m} }}{k^{\frac{1}{m}}}\right)^\top \in
\Re^n_+$ (as $h$ takes nonnegative values). Define ${{\cal A}}_l^k$ be a symmetric tensor such that $f_l^k({{\bf x}})={{\cal A}}_l^k {{\bf x}}^{\otimes m}$. So, each ${{\cal A}}_k^l$ is a completely positive tensor. Note that the completely positive cone $CP_{m,n}$ is a closed convex cone [@QXX]. It then follows that ${{\cal A}}=\lim_{k \rightarrow \infty}\lim_{l \rightarrow \infty}{{\cal A}}_k^l$ is also a completely positive tensor.
Other SOS-Hankel Tensors
========================
There are other positive semi-definite Hankel tensors, which are not strong Hankel tensors. In [@Qi15], there is such an example. That example is for order $m=4$. We now extend it to $m = 2k$ for any integer $k \ge 2$. We will show that such tensors are not completely decomposable [ (and so, is also not a strong Hankel tensor by Theorem \[th:2\])]{}, but they are still SOS-Hankel tensors.
Let $m = 2k$, $n$ = 2, $k$ is an integer and $k \ge 2$. [ Let $v_0 = v_m = 1$, $v_{2l} = v_{m-2l} = -{1 \over \left({m \atop 2l}\right)}$, $l=1,\ldots,k-1$, and $v_j = 0$ for other $j$. Let ${{\cal A}}= (a_{i_1\cdots i_m})$ be defined by $a_{i_1\cdots i_m} = v_{i_1+\cdots +i_m-m},$ for $i_1, \cdots, i_m = 1, 2$.]{} Then ${{\cal A}}$ is an even order Hankel tensor. [ For any ${{\bf x}}\in \Re^2$, we have $${{\cal A}}{{\bf x}}^{\otimes m} = x_1^m - \sum_{j=1}^{k-1} x_1^{m-2j} x_2^{2j} + x_2^m = \sum_{j=0}^{k-2}\left(x_1^{k-j}x_2^j - x_1^{k-j-2}x_2^{j+2}\right)^2.$$ Thus,]{} ${{\cal A}}$ is an SOS-Hankel tensor, hence a positive semi-definite Hankel tensor. On the other hand, ${{\cal A}}$ is not a completely decomposable tensor. Assume that ${{\cal A}}$ is a completely decomposable tensor. Then there are vectors ${{\bf u}}_j = (a_j, b_j)^\top$ for $j \in [r]$ such that ${{\cal A}}= \sum_{j=1}^r {{\bf u}}_j^m.$ Then for any ${{\bf x}}\in \Re^2$, $${{\cal A}}{{\bf x}}^{\otimes m} = \sum_{p=1}^r (a_px_1+b_px_2)^m = \sum_{j=0}^m \sum_{p=1}^r \left({ m \atop j}\right)a_p^{m-j}b_p^jx_1^{m-j}x_2^j.$$ [ On the other hand, $${{\cal A}}{{\bf x}}^{\otimes m} = x_1^m - \sum_{j=1}^{k-1} x_1^{m-2j} x_2^{2j} + x_2^m.$$]{} Comparing the coefficients of $x_1^{m-2}x_2^2$ in the above two expressions of ${{\cal A}}{{\bf x}}^m$, we have $$\sum_{p=1}^r \left({ m \atop 2}\right)a_p^{m-2}b_p^2 = -1.$$ This is impossible. Thus, ${{\cal A}}$ is not completely decomposable.
[**Question 2**]{} Is there an even order completely decomposable Hankel tensor, which is not a strong Hankel tensor?
We may also construct an example for $m=6$ and $n=3$. Let ${{\cal A}}\in {\rm S}_{4, 3}$ be a Hankel tensor generated by ${{\bf v}}= (v_0 = \alpha, 0, \cdots, 0, v_6=1, 0, \cdots, 0, v_{12}=\alpha)^\top$. Then $$f({{\bf x}})\equiv {{\cal A}}{{\bf x}}^{\otimes 6} =\alpha x_1^6 + x_2^6+30x_1x_2^4x_3+90x_1^2x_2^2x_3^2+ 20x_1^3x_3^3 +\alpha x_3^6.$$ We now show that $f$ is PSD if $\alpha \ge 480\sqrt{15}+10$. Indeed, $$\label{e1}
f({{\bf x}}) = 10\left(x_1^3 +x_3^3\right)^2 + {x_2^2 \over 2} \left(x_2^2 + 30x_1x_3\right)^2 + \left[(\alpha-10)(x_1^6 +x_3^6) + {1 \over 2}x_2^6 - 360x_1^2x_2^2x_3^2 \right] \ge 0,$$ where $\left[(\alpha-10)(x_1^6 +x_3^6) + {1 \over 2}x_2^6 - 360x_1^2x_2^2x_3^2 \right] \ge 0$ because of the arithmetic-geometric inequality. Note that $\left[(\alpha-10)(x_1^6 +x_3^6) + {1 \over 2}x_2^6 - 360x_1^2x_2^2x_3^2 \right] \ge 0$ is a diagonal minus tail form and all positive semi-definite diagonal minus tail forms are SOS [@FK Theorem 2.3]. Thus, if $\alpha \ge 480\sqrt{15}+10$, $f({{\bf x}})$ is also SOS.
On the other hand, we may see that the Hankel tensor ${{\cal A}}$ is not a strong Hankel tensor. Let $A = (a_{ij})$ be generated by ${{\bf v}}$. Then $A$ is a $7 \times 7$ Hankel matrix, with $a_{11} = a_{77} = \alpha$, $a_{44} = a_{35} = a_{53} = a_{26} = a_{62} = a_{17} = a_{71} =1$ and $a_{ij}=0$ for other $(i, j)$. Let ${{\bf y}}= (0, 0, 1, 0, -1, 0, 0, 0)^\top$. Then ${{\bf y}}^\top A {{\bf y}}= -2 < 0$. Hence $A$ is not PSD and ${{\cal A}}$ is not a strong Hankel tensor.
Naturally, we have the following question:
[**Question 3**]{} Are there PSD non-SOS-Hankel tensors?
In a certain sense, this question looks like the Hilbert problem under the Hankel constraint.
The following question is connected with the above question.
[**Question 4**]{} Is the problem for determining an even order Hankel tensor is positive semi-definite or not solvable in polynomial-time?
As discussed before, if there are no PSD non-SOS-Hankel tensors, then the problem for determining an even order Hankel tensor is positive semi-definite or not is solvable in polynomial-time.
Before answering Question 3, we may try to answer an easier question.
[**Question 5**]{} Are there PSD non-SOS-Hankel tensors of order $6$ and dimension $3$?
If there are no PSD non-SOS-Hankel tensors of order $6$ and dimension $3$, then it indicates that the Hilbert problem for PSD non-SOS polynomials has different answer if such polynomials are restricted to be Hankel polynomials.
An Application
==============
Positive Semidefinite Hankel Tensor Completion Problem
------------------------------------------------------
An interesting problem is to fit the tensor data with prescribed Hankel structure of low rank. That is, given a tensor ${{\cal X}}\in {\rm S}_{m,n}$, one try to solve the following optimization problem: $$\begin{aligned}
& \min_{\} \in {\rm S}_{m,n}}& \frac{1}{2}\|{{\cal A}}-{{\cal X}}\|^2 \\
&\mbox{ subject to } & {{\cal A}}\mbox{ is a positive semi-definite Hankel tensor} \\
& & \mathcal{A} \mbox{ is of low rank}.\end{aligned}$$ As a PSD Hankel tensor and the rank for tensor are hard to determine. One could consider the following alternative $$\begin{aligned}
& \min_{{{\cal A}}\in {\rm S}_{m,n}}& \frac{1}{2}\|{{\cal A}}-{{\cal X}}\|^2 \\
&\mbox{ subject to } & {{\cal A}}\mbox{ is a strong Hankel tensor} \\
& & \mbox{the associated Hankel matrix of ${{\cal A}}$ is of low rank}.\end{aligned}$$ Define $l={(n-1)m+2 \over 2}$. This problem can be rewritten as $$\begin{aligned}
& \min_{v \in \mathbb{R}^{(n-1)m+1}}& \frac{1}{2}\sum_{i_1,\ldots,i_m=1}^n\|v_{i_1+\ldots+i_m-m}-{{\cal X}}_{i_1,\ldots,i_m}\|^2 \\
&\mbox{ subject to } & A_{\alpha,\beta}=v_{\alpha+\beta-2}, \ \alpha,\beta=1,\ldots,l, \\
& & A \mbox{ is a positive semi-definite matrix} \mbox{ and } A
\mbox{ is of low rank}.\end{aligned}$$ The low rank constraint is often nonconvex and nonsmooth, and so, the problem is still a hard problem to solve. As the trace norm promotes a low rank solution, one popular approach is to solve its nuclear norm heuristic (see, for example [@JMZ]) the following form: $$\begin{aligned}
& \min_{v \in \mathbb{R}^{(n-1)m+1}}& \frac{1}{2}\sum_{i_1,\ldots,i_m=1}^n\|v_{i_1+\ldots+i_m-m}-\mathcal{X}_{i_1,\ldots,i_m}\|^2 + \mu \|A\|_{\rm tr} \\
&\mbox{ subject to } & A_{\alpha,\beta}=v_{\alpha+\beta-2}, \ \alpha,\beta=1,\ldots,l, \mbox{ and } A \succeq 0,\end{aligned}$$ where $\|A\|_{\rm tr}$ denotes the trace norm of $A$ and is defined as the sum of all eigenvalues of $A$. Define two linear maps $M:\Re^{(n-1)m+1} \rightarrow {\rm S}^l:={\rm S}_{2,l}$ and ${{\cal P}}:\Re^{(n-1)m+1} \rightarrow {\rm S}_{m,n}$ by $$M \, v=\big(v_{\alpha+\beta-2}\big)_{1\le \alpha,\beta \le r} \mbox{ and } ({{\cal P}}({{\bf v}})_{i_1,\ldots,i_m}= v_{i_1+\ldots+i_m-m}.$$ Then, the trace norm problem can be further simplified as $$\begin{aligned}
(TCP) & \displaystyle \min_{{{\bf v}}\in \Re^{(n-1)m+1}, A \in S^l }& \frac{1}{2} \|{{\cal P}}({{\bf v}})-{{\cal X}}\|^2 + \mu \|A\|_{\rm tr} \\
&\mbox{ subject to } & A-M\, {{\bf v}}=0, \mbox{ and } A \succeq 0.\end{aligned}$$ The associated augmented Lagrangian for $(TCP)$ can be formulated as $$L_{\rho}({{\bf v}},A,Z)=\frac{1}{2} \|{{\cal P}}({{\bf v}})-{{\cal X}}\|^2 + \mu \|A\|_{\rm tr}+ {\rm Tr}[Z (A-M\, {{\bf v}})]+\frac{\rho}{2}\|A-M {{\bf v}}\|^2.$$ We now propose an alternating direction method of multiplier (ADMM) to solve $(TCP)$.
[**ADMM for solving $(TCP)$**]{}
Step 0. Given $\mathcal{X} \in S_{m,n}$. Choose an initial point $(v^0,A^0,Z^0) \in \mathbb{R}^{(n-1)m+1} \times S^l \times S^l$ and $\rho>0$. Set $k=1$.
Step 1. Find ${{\bf v}}^k={\rm argmin}_{{\bf v}}\{L_{\rho}({{\bf v}},A^{k-1},Z^{k-1})\}$.
Step 2. Find $A^k={\rm argmin}_A\{L_{\rho}(v^k,A,Z^{k-1}): A \succeq 0\}$.
Step 3. Let $Z^k=Z^{k-1}+\rho (A^k -M \, v^k)$.
Step 4. Let $k=k+1$ and go back to Step 1.
The computational cost of (ADMM) is not heavy which makes it suitable for solving large size problem. In fact, we note that, in Step 1, ${{\bf v}}^k$ can be found by solving a linear equation in $v$: $$({{\cal P}}^*{{\cal P}}+\rho M^*M) v = {{\cal P}}^*{{\cal X}}+M^*(\rho A^{k-1}+Z^{k-1}),$$ where ${{\cal P}}^*$ and $M^*$ are the adjoint mappings of ${{\cal P}}$ and $M$ respectively. Moreover, in Step 2, $A^k$ indeed has a closed form solution as $$\begin{aligned}
A^k &= & {\rm argmin}_A\{L_{\rho}({{\bf v}}^k,A,Z^{k-1}): A \succeq 0\} \\
&= & {\rm argmin}_A\left\{\mu \|A\|_{\rm tr}+ {\rm Tr}[Z^{k-1} (A-M\, {{\bf v}})]+\frac{\rho}{2}\|A-M {{\bf v}}^k\|^2: A \succeq 0\right\} \\
& = & {\rm argmin}_A\left\{\mu {\rm Tr}[I_l A] + {\rm Tr}[Z^{k-1} A]+\frac{\rho}{2}\|A-M {{\bf v}}^k\|^2: A \succeq 0\right\} \\
& = & {\rm argmin}_A\left\{\frac{\rho}{2}\|A+ \frac{1}{\rho}\left(\mu I_l+Z^{k-1}\right)-M {{\bf v}}^k\|^2: A \succeq 0 \right\} \\
& = & \sum_{i=1}^l \max\{\sigma_i,0\} {{\bf u}}_i{{\bf u}}_i^\top\end{aligned}$$ where $\sigma_i$ and ${{\bf u}}_i$ are obtained from the SVD decomposition of $M \,v^k-\frac{1}{\rho}(\mu I_l+Z^{k-1})$, that is, $$M \, {{\bf v}}^k-\frac{1}{\rho}\left(\mu I_l+Z^{k-1}\right)= \sum_{i=1}^l \sigma_i {{\bf u}}_i{{\bf u}}_i^\top.$$ The convergence of the ADMM method has been well-studied by a lot of researchers. For simplicity purpose, we omit the details and refer the interested reader to [@HY].
Numerical Tests
---------------
To illustrate the algorithm of ADMM for solving (TCP), we first generate two random instances of symmetric tensors. Then, we solve the corresponding positive semidefinite Hankel tensor completion problem via the proposed ADMM algorithm. In our numerical test, we set $\mu$ and $\rho$ in the ADMM algorithm as $0.1$ and $10$ respectively. Our proposed algorithm works very well by using these parameters.
Consider a $4$th-order $3$-dimensional symmetric tensor $\mathcal{X}$ given by
$\mathcal{X}(:,:,1,1) =\left(
\begin{array}{ccc}
-0.2972 & 0.4307 & 0.4444\\
0.4307 & -0.4029 & -0.0274\\
0.4444 &-0.0274 & 0.0647\\
\end{array}
\right),$ $\mathcal{X}(:,:,2,1) =\left(
\begin{array}{ccc}
0.4307& -0.4029 & -0.0274\\
-0.4029 & 0.1085 & 0.1760\\
-0.0274 & 0.1760 & -0.2574\\
\end{array}
\right),$
$\mathcal{X}(:,:,3,1) =\left(
\begin{array}{ccc}
0.4444 & -0.0274 & 0.0647\\
-0.0274 & 0.1760 & -0.2574\\
0.0647 & -0.2574 & -0.3208\\
\end{array}
\right),$ $\mathcal{X}(:,:,1,2) =\left(
\begin{array}{ccc}
0.4307 & -0.4029 & -0.0274\\
-0.4029 & 0.1085 & 0.1760\\
-0.0274 & 0.1760 & -0.2574\\
\end{array}
\right),$
$\mathcal{X}(:,:,2,2) =\left(
\begin{array}{ccc}
-0.4029 & 0.1085 & 0.1760\\
0.1085 & 0.9152 & -0.0821\\
0.1760 & -0.0821 & -0.2815\\
\end{array}
\right),$ $\mathcal{X}(:,:,3,2) =\left(
\begin{array}{ccc}
-0.0274 & 0.1760 & -0.2574\\
0.1760 & -0.0821 & -0.2815\\
-0.2574 & -0.2815 & 0.2773\\
\end{array}
\right),$
$\mathcal{X}(:,:,1,3) =\left(
\begin{array}{ccc}
0.4444 & -0.0274 & 0.0647\\
-0.0274 & 0.1760 & -0.2574\\
0.0647 & -0.2574 & -0.3208\\
\end{array}
\right),$ $\mathcal{X}(:,:,2,3) =\left(
\begin{array}{ccc}
-0.0274 & 0.1760 & -0.2574\\
0.1760 & -0.0821 & -0.2815\\
-0.2574 & -0.2815 & 0.2773\\
\end{array}
\right),$
$\mathcal{X}(:,:,3,3) =\left(
\begin{array}{ccc}
0.0647 & -0.2574 & -0.3208\\
-0.2574 & -0.2815 & 0.2773\\
-0.3208 & 0.2773 & -0.5347\\
\end{array}
\right).$
Solving the strong Hankle tensor completion problem for $\mathcal{X}$ via the proposed ADMM method, we obtain $$v=(0.0086, 0.0056, 0.0036, 0.0022, 0.0014, 0.0009, 0.0006, 0.0004, 0.0002)^T$$ and the associated Hankel matrix $A= (A_{ij})_{1 \le i \le j \le 5}$ with $A_{ij}=v_{i+j-2}$ is of rank $1$.
Consider a given $4$th-order $3$-dimensional symmetric tensor $\mathcal{X}$ given by
$\mathcal{X}(:,:,1,1) =\left(
\begin{array}{ccc}
-0.7384 & 0.2309 & 0.3538\\
0.2309 & -0.4025 & 0.2401\\
0.3538 & 0.2401 & -0.2167\\
\end{array}
\right),$ $\mathcal{X}(:,:,2,1) =\left(
\begin{array}{ccc}
0.2309 & -0.4025 & 0.2401\\
-0.4025 & 0.1324 & -0.1888\\
0.2401 & -0.1888 & -0.1495\\
\end{array}
\right),$
$\mathcal{X}(:,:,3,1) =\left(
\begin{array}{ccc}
0.3538 & 0.2401 & -0.2167\\
0.2401 & -0.1888 & -0.1495\\
-0.2167 & -0.1495 & 0.3234\\
\end{array}
\right),$ $\mathcal{X}(:,:,1,2) =\left(
\begin{array}{ccc}
0.2309 & -0.4025 & 0.2401\\
-0.4025 & 0.1324 & -0.1888\\
0.2401 & -0.1888 & -0.1495\\
\end{array}
\right),$
$\mathcal{X}(:,:,2,2) =\left(
\begin{array}{ccc}
-0.4025 & 0.1324 & -0.1888\\
0.1324 & -0.3712 & 0.0019\\
-0.1888 & 0.0019 & -0.1546\\
\end{array}
\right),$ $\mathcal{X}(:,:,3,2) =\left(
\begin{array}{ccc}
0.2401 & -0.1888 & -0.1495\\
-0.1888 & 0.0019 & -0.1546\\
-0.1495 & -0.1546 & -0.0395\\
\end{array}
\right),$
$\mathcal{X}(:,:,1,3) =\left(
\begin{array}{ccc}
0.3538 & 0.2401 & -0.2167\\
0.2401 & -0.1888 & -0.1495\\
-0.2167 & -0.1495 & 0.3234\\
\end{array}
\right),$ $\mathcal{X}(:,:,2,3) =\left(
\begin{array}{ccc}
0.2401 & -0.1888 & -0.1495\\
-0.1888 & 0.0019 & -0.1546\\
-0.1495 & -0.1546 & -0.0395\\
\end{array}
\right),$
$\mathcal{X}(:,:,3,3) =\left(
\begin{array}{ccc}
-0.2167 & -0.1495 & 0.3234\\
-0.1495 & -0.1546 &-0.0395\\
0.3234 & -0.0395 & 0.9162\\
\end{array}
\right).$
Solving the strong Hankle tensor completion problem for $\mathcal{X}$ via the proposed ADMM method, we obtain $$v=( 0, 0, -0.0001, 0.0003, 0.0001, -0.0024, 0.0120, -0.0390, 0.7741)^T$$ and and the associated Hankel matrix $A= (A_{ij})_{1 \le i \le j \le 5}$ with $A_{ij}=v_{i+j-2}$ is of rank $2$.
These preliminary numerical results show that the algorithm of ADMM for solving (TCP) is efficient.
[abc99xyz]{}
R. Badeau and R. Boyer, “Fast multilinear singular value decomposition for structured tensors”, [*SIAM Journal on Matrix Analysis and Applications **30***]{} (2008) 1008-1021.
K.C. Chang, K. Pearson and T. Zhang, “Perron Frobenius Theorem for nonnegative tensors”, [*Communications in Mathematical Sciences **6***]{} (2008) 507-520.
K.C. Chang, L. Qi and T. Zhang, “A survey on the spectral theory of nonnegative tensors”, [*Numerical Linear Algebra with Applications **20***]{} (2013) 891-912.
H. Chen and L. Qi, “Positive definiteness and semi-definiteness of even order symmetric Cauchy tensors”, to appear in: [*J. Indust. Management Optim.*]{}, arXiv:1405.6363.
M.D. Choi and T.Y. Lam, “Extremal positive semidefinite forms”, [*Mathematische Annalen **231***]{} (1977) 1-18.
P. Comon, G. Golub, L.H. Lim and B. Mourrain, “Symmetric tensors and symmetric tensor rank”, [*SIAM Journal on Matrix Analysis and Applications **30***]{} (2008) 1254-1279.
W. Ding, L. Qi and Y. Wei, “Fast Hankel tensor-vector products and application to exponential data fitting”, January 2014. arXiv:1401.6238
C. Fidalgo and A. Kovacec, “Positive semidefinite diagonal minus tail forms are sums of squares”, [*Mathematische Zeitschrift **269***]{} (2011) 629-645.
D.R. Han and X.M. Yuan, “Local linear convergence of the alternating direction method of multipliers for quadratic programs”, [*SIAM Journal on Numerical Analysis **51***]{} (2013) 3446-3457.
D. Hilbert, “Über die Darstellung definiter Formen als Summe von Formenquadraten”, [*Mathematical Annals, **32***]{} (1888) 342-350.
C.J. Hillar and L.H. Lim, “Most tensor problems are NP hard”, [*Journal of the ACM **60***]{} (2013), Article No. 45.
S. Hu, G. Li and L. Qi, “A tensor analogy of Yuan’s alternative theorem and polynomial optimization with sign structure”, [*Journal of Optimization Theory and Applications*]{} (2014) DOI 10.1007/s10957-014-0652-1.
S. Hu, G. Li, L. Qi and Y. Song, “Finding the maximum eigenvalue of essentially nonnegative symmetric tensors via sum of squares programming”, [*Journal of Optimization Theory and Applications **158***]{} (2013) 717-738.
B. Jiang, S.Q. Ma and S.Z Zhang, “Tensor principal component analysis via convex optimization”, [*Mathematical Programming*]{} (2014) DOI 10.1007/s10107-014-0774-0.
J.B. Lasserre, “Global optimization with polynomials and the problem of moments”, [*SIAM Journal on Optimization **11***]{} (2001) 796-817.
M. Laurent, “Sums of squares, moment matrices and optimization over polynomials”, [*Emerging Applications of Algebraic Geometry*]{}, Vol. 149 of IMA Volumes in Mathematics and its Applications, M. Putinar and S. Sullivant eds., Springer, (2009) pp. 157-270.
C. Li and Y. Li, “Double B-tensor and quasi-double B-tensor”, [*Linear Algebra and Its Applications **466***]{} (2015) 343-356.
C. Li, F. Wang, J. Zhao, Y. Zhu and Y. Li, “Criterions for the positive definiteness of real supersymmetric tensors”, [*Journal of Computational and Applied Mathematics **255***]{} (2014) 1-14.
T.S. Motzkin, “The arithmetic-geometric inequality”, In: [*Inequalities*]{}, O. Shisha ed., Academic Press, New York, (1967) pp. 205-224.
M. Ng, L. Qi and G. Zhou, “Finding the largest eigenvalue of a non-negative tensor”, [*SIAM Journal on Matrix Analysis and Applications **31***]{} (2009) 1090-1099.
J.M. Papy, L. De Lauauwer and S. Van Huffel, “Exponential data fitting using multilinear algebra: The single-channel and multi-channel case”, [*Numerical Linear Algebra with Applications **12***]{} (2005) 809-826.
L. Qi, “Eigenvalues of a real supersymmetric tensor”, [*Journal of Symbolic Computation **40***]{} (2005) 1302-1324.
L. Qi, “Hankel tensors: Associated Hankel matrices and Vandermonde decomposition”, [*Communications in Mathematical Sciences **13***]{} (2015) 113-125.
L. Qi and Y. Song, “An even order symmetric B tensor is positive definite”, [*Linear Algebra and Its Applications **457***]{} (2014) 303-312.
L. Qi, C. Xu and Y. Xu, “Nonnegative tensor factorization, completely positive tensors and a hierarchically elimination algorithm”, [*SIAM Journal on Matrix Analysis and Applications **35***]{} (2014) 1227-1241.
L. Qi and Y. Ye, “Space tensor conic programming”, [*Computational Optimization and Applications **59***]{} (2014) 307-319.
L. Qi, G. Yu and E.X. Wu, “Higher order positive semi-definite diffusion tensor imaging”, [*SIAM Journal on Imaging Sciences **3***]{} (2010) 416-433.
L. Qi, G. Yu and Y. Xu, “Nonnegative diffusion orientation distribution function”, [*Journal of Mathematical Imaging and Vision **45***]{} (2013) 103-113.
B. Reznick, “Some concrete aspects of Hilbert’s 17th problem”, [*Contemporary Mathematics **253***]{} (2000) 251-272.
N. Shor, [*Nondifferentiable Optimization and Polynomial Problems*]{}, Kluwer Academic Publishers, Boston, 1998.
Y. Song and L. Qi, “Infinite and finite dimensional Hilbert tensors”, [*Linear Algebra and Its Applications **451***]{} (2014) 1-14.
L. Zhang, L. Qi and G. Zhou, “M-tensors and some applications”, [*SIAM Journal on Matrix Analysis and Applications **35***]{} (2014) 437-452
[^1]: Department of Applied Mathematics, University of New South Wales, Sydney 2052, Australia. E-mail: [email protected] (G. Li). This author’s work was partially supported by Australian Research Council.
[^2]: Department of Applied Mathematics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong. E-mail: [email protected] (L. Qi). This author’s work was partially supported by the Hong Kong Research Grant Council (Grant No. PolyU 502510, 502111, 501212 and 501913).
[^3]: Department of Mathematics, Southeast University, Nanjing, 210096 China. Email: [email protected] (Y. Xu).
[^4]: This example was used to show the set consisting of all tensors with symmetric rank over the complex field less or equal to a fixed number $l$ can be non-closed. Here we show that this example can also be used to show the non-closedness of complete $r$-decomposable tensor cones with odd order.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Dileptons provide direct observables of the electromagnetic current-current correlator in the hot and/or dense medium formed in heavy-ion collisions. In this article an overview is given about the status of the theoretical understanding of the dilepton phenomenology in heavy-ion collisions from studies of the in-medium properties of hadrons and partons within many-body theory and about connections to fundamental questions concerning the chiral phase transition.'
address: 'Cyclotron Institute and Physics Department, Texas A&M University, College Station, Texas 77843-3366, USA'
author:
- Hendrik van Hees
title: 'The dilepton probe in heavy-ion collisions'
---
Introduction
============
In heavy-ion collisions electromagnetic probes, i.e., photons and lepton pairs (“virtual photons”) provide one of the most valuable possibilities to study the interior of the hot and dense medium created in the interaction over its whole history since their spectra are nearly unaffected by final-state interactions [@Rapp:1999ej].
This paper will be restricted to invariant-mass ($M$) and transverse-momentum ($q_T$) spectra of dileptons whose rate is given by [@Shuryak:1980tp; @MT84; @Gale:1990pn] $$\label{MT}
\frac{{\mathrm{d}}N_{ll}}{{\mathrm{d}}^4 x {\mathrm{d}}^4 q} = -\frac{\alpha^2}{3 \pi^3}
\frac{L(M)}{M^2} \operatorname{Im}\Pi_{\mathrm{em}\mu}^{\mu}(M,q;T,\mu_B) f_B(q_0,T),$$ where $\alpha\simeq 1/137$ denotes the fine-structure constant, $M=q_0^2-q^2$ the invariant mass of the lepton pair of energy, $q_0$, and three-momentum, $q$, $T$ the temperature, $\mu_B$ the baryon chemical potential, $f_B$ the Bose distribution, and $L(M)$ the lepton-phase space factor. As is known from $e^+ e^- \rightarrow \mathrm{hadrons}$, in the vacuum the retarded hadronic electromagnetic (em.) current correlator, $\Pi_{\mathrm{em}}$, at low invariant masses, $M \lesssim
M_{\mathrm{dual}}$ is well described by the vector-meson dominance (VMD) model for the light vector mesons $\rho$, $\omega$, and $\phi$ and by the perturbative QCD (pQCD) continuum at higher masses ($M \gtrsim
M_{\mathrm{dual}}$), where $M_{\mathrm{dual}} \simeq 1.5\,\mathrm{GeV}$ denotes a “duality scale”. The hadronic “resonance part” is dominated by the isovector channel ($\rho$ meson).
For a theoretical description of dilepton production in relativistic heavy-ion collisions (HICs) thus the first goal must be an understanding of the in-medium spectral properties of the light vector mesons in the hadronic and the QGP in the partonic phase of the fireball evolution. This review of recent progress in this field is organized as follows: In Sect. \[sect.emcc\] first the constraints on the em. current correlator in strongly interacting matter from QCD as the underlying fundamental theory of strong interactions will be discussed, followed by a brief summary on effective hadronic models. In Sect. \[sect.phen\] these models are confronted with data from ultrarelativistic HICs. Sect. \[sect.concl\] contains brief conclusions and an outlook.
The electromagnetic current correlator in strongly interacting matter {#sect.emcc}
=====================================================================
**Ab-initio constraints.** In the vacuum and at low temperatures and densities the light-quark sector of QCD is governed by (approximate) *chiral symmetry* which is spontaneously broken by the formation of a quarks condensate, ${\left \langle \bar{\psi} \psi \right \rangle} \neq 0$, in the QCD vacuum which manifests itself in the mass splitting of chiral partners in the hadron spectrum, as can be seen, e.g., in the experimental determination of the isovector-vector and -axialvector current correlator through $\tau \rightarrow \nu+n \pi$ decays [@aleph98; @opal99]. From (lattice) QCD at finite temperature one expects a decrease of the quark condensate at high temperatures and/or densities and restoration of chiral symmetry. Thus the mass spectra of hadrons are expected to soften and to degenerate with their pertinent chiral partners above a critical temperature, $T_c$. One indication of the interrelation of chiral symmetry and confinement is that in lattice-QCD (lQCD) calculations the critical temperature, $T_c \simeq 160$-$190 \, \mathrm{MeV}$, for the chiral and the deconfinement (cross-over) transitions coincide [@Karsch:1994hm].
From hadronic modeling two microscopic mechanisms for chiral symmetry restoration (CSR) have emerged: On the one hand it has been conjectured that the hadron masses drop to zero at the critical point due to the melting of the quark condensate) [@Brown:1991kk]. On the other hand within phenomenological hadronic many-body models the hadron spectra show a significant broadening with little mass shifts (“melting-resonance scenario”) [@Gale:1990pn; @Gale:1993zj; @Rapp:1997fs; @Rapp:1999us]. A direct relation between in-medium spectral properties of hadrons, in our context particularly vector mesons, to QCD is provided by QCD sum rules which relate moments of the (in-medium) spectral functions for various currents in different isospin channels in the space-like region to the pertinent quark and four-quark condensates. Detailed studies [@Asakawa:1993pq; @Leupold:1997dg; @Klingl:1997kf; @Ruppert:2005id] show that (in cold nuclear matter) both the “dropping-mass” and the “melting-resonance” scenarios for CSR are compatible with QCD sum rules. One objective for the investigation of dileptons in high-energy heavy-ion collisions is thus to gain insight in the in-medium spectral properties of the light vector mesons through the em. current correlator to constrain the mechanism leading to the softening of the spectral functions. Due to the experimental problems to assess the spectral properties of the axial-vector channel, an indirect theoretical approach may be to connect chiral hadronic models, describing successfully the measured dilepton observables, with CSR via finite-temperature Weinberg-sum rules [@Weinberg:1967kj; @Kapusta:1993hq] which relate moments of the *difference of vector- and axial-vector-current correlators* to quark and four-quark condensates, i.e., order parameters of chiral symmetry, providing constraints on these models from lQCD.
**Hadronic many-body theory.** A (model-independent) approach to assess medium modifications of vector (and axial-vector) mesons is based on the chiral reduction formalism, providing a low-density expansion of the in-medium vector- and axial-vector current correlators in terms of the corresponding vacuum quantities which are taken from experimental data such as $\tau \rightarrow n \pi \nu$. The most intriguing feature of these kind of models is the “in-medium mixing” of the vector- with the axial-vector current correlator due to pions in the medium [@Dey:1990ba] which may provide a mechanism for the onset of CSR. However, the applicability of the low-density approximation is restricted to very low temperatures and densities.
Another ansatz is to use chiral models in various realizations of chiral symmetry. One possibility is to describe vector- (and axial-vector) mesons as gauge bosons within the *(generalized) hidden-local symmetry models* [@bandokugo84; @Harada:2003jx]. Here a particular realization of chiral symmetry, the *vector manifestation*, becomes possible, where the chiral partner of the longitudinal $\rho$ meson is the pion, thus providing a definite chiral model for the “dropping-mass scenario”. A detailed renormalization-group analysis shows that such a model together with Wilson-matching of the effective model to QCD (leading to an “intrinsic” temperature/density dependence of the effective-model parameters), inevitably leads to a vanishing $\rho$ mass at the critical point and a violation of VMD [@HS05].
In the hadronic many-body theory (HMBT) approach, starting from a phenomenological Lagrangian to describe the vacuum properties of the vector mesons, the in-medium modifications of their spectral properties are evaluated within finite-temperature/density quantum-field theory, involving non-perturbative techniques such as the dressing of, e.g., the pion propagator to account for the modification of the $\rho$-meson’s pion cloud and implementation of interactions of the $\rho$-meson with mesons and baryons in the medium (for a review, see [@Rapp:1999ej]). It is characteristic for such models that the various excitations result in a *substantial broadening of the vector mesons and small mass shifts*, i.e., a realization of the “melting-resonance scenario” of CSR. An intriguing property in connection with the model proposed in [@Rapp:1999us] is that the resulting dilepton-emission rates, cf. Eq. (\[MT\]), match that of the hard-thermal-loop improved pQCD rate, when both are extrapolated to the expected chiral-phase transition temperature $T_c \simeq 160$-$190\,\mathrm{MeV}$, i.e., a kind of “quark-hadron duality” [@Rapp:1999us]. This behavior is consistent with the smoothness of quark-number susceptibilities in the corresponding isovector channel across the phase transition in recent lQCD calculations [@Allton:2005gk].
Another approach to assess in-medium properties of vector mesons is the use of empirical scattering amplitudes and dispersion-integral techniques to assess the in-medium $\rho$-meson propagator via the $T\varrho$ approximation [@Eletsky:2001bb].
Dilepton phenomenology in heavy-ion collisions {#sect.phen}
==============================================
In this Section we compare theoretical models of the in-medium em. current correlator to experimental results from ultrarelativistic heavy-ion collisions. For such a comparison, not only detailed models for the in-medium behavior of the correlation function itself in both partonic (QGP) and hadronic states of the medium are required, but also a description of its entire “thermal evolution” over which the rate (\[MT\]) has to be integrated to compare to the experimental observables.
The bulk of hot and dense matter created in ultrarelativistic heavy-ion collisions at the CERN SPS and RHIC can be successfully described by ideal hydrodynamics [@Kolb:2003dz], which implies local thermal equilibrium. Thus the medium is characterized by a temperature- and collective-flow field. In [@vanHees:2007th; @vanHees:2006ng] a simple thermal model has been used, which after a “plasma-formation time” describes the medium by an ideal-gas equation of state of quarks and gluons which according to recent lQCD calculations [@Fodor:2004nz; @Karsch:2007vw] undergoes a phase transition at $T_c \simeq 160$-$190 \,\mathrm{GeV}$ to a hadron-resonance gas. Thermal models [@Andronic:2005yp; @Becattini:2005xt] for the yields of various hadron species indicate that at a temperature close to the phase transition of $T_{\mathrm{ch}} \simeq 160$-$175 \, \mathrm{GeV}$, inelastic reactions within the medium cease, and the corresponding particle ratios are fixed (*chemical freeze-out*), before the particles decouple and freely stream to the detector. This *thermal freeze-out* occurs at temperatures around $T_{\mathrm{fo}} \simeq
90$-$130 \, \mathrm{GeV}$ (depending on the system size). This evolution of the medium is implemented through a cylindrical homogeneous fireball model, including radial flow and longitudinal expansion [@vanHees:2007th]. The temperature is inferred from the equation of state (massless gluons and $N_f=2.3$ effective quark flavors in the QGP and a hadron-resonance gas model in the hadronic phase) and the assumption of an isentropic expansion in accordance with ideal-fluid dynamics. Between the pure QGP and hadronic phases a standard volume partition for a mixed phase is employed. The hadronic phase is characterized by the build-up of hadron-chemical potentials to keep the particle-number ratios fixed at the observed values. The largest uncertainty is the total fireball lifetime which has been adjusted to the total experimental yield.
![Left panel: dimuon excess spectrum [@vanHees:2007th] with an equation of state with $T_c=T_{\mathrm{ch}}=175\,\mathrm{MeV}$ in semicentral $158\,A\mathrm{GeV}$ In-In collisions compared to data by the NA60 collaboration [@Arnaldi:2006jq]; middle panel: dielectron excess spectrum from the same model for central $158\,A\mathrm{GeV}$ Pb-Au collisions compared to data by the CERES/NA45 [@Adamova:2006nu] collaboration. The dash-dotted line shows the result with a $\rho$-meson spectral function including only medium modifications in a meson gas, underlining the importance of baryon effects; right panel: the dilepton excess spectrum based on the implementation of different equations of state in the fireball evolution (EoS-A: $T_c=T_{\mathrm{ch}}=175\,\mathrm{MeV}$, EoS-B: $T_c=T_{\mathrm{ch}}=160\,\mathrm{MeV}$, EoS-C: $T_c=160\,\mathrm{MeV}$, $T_{\mathrm{ch}}=160\,\mathrm{MeV}$).[]{data-label="fig.1"}](dndm-semi2.eps){width="\textwidth"}
![Left panel: dimuon excess spectrum [@vanHees:2007th] with an equation of state with $T_c=T_{\mathrm{ch}}=175\,\mathrm{MeV}$ in semicentral $158\,A\mathrm{GeV}$ In-In collisions compared to data by the NA60 collaboration [@Arnaldi:2006jq]; middle panel: dielectron excess spectrum from the same model for central $158\,A\mathrm{GeV}$ Pb-Au collisions compared to data by the CERES/NA45 [@Adamova:2006nu] collaboration. The dash-dotted line shows the result with a $\rho$-meson spectral function including only medium modifications in a meson gas, underlining the importance of baryon effects; right panel: the dilepton excess spectrum based on the implementation of different equations of state in the fireball evolution (EoS-A: $T_c=T_{\mathrm{ch}}=175\,\mathrm{MeV}$, EoS-B: $T_c=T_{\mathrm{ch}}=160\,\mathrm{MeV}$, EoS-C: $T_c=160\,\mathrm{MeV}$, $T_{\mathrm{ch}}=160\,\mathrm{MeV}$).[]{data-label="fig.1"}](ceres-cent-excess.eps){width="\textwidth"}
![Left panel: dimuon excess spectrum [@vanHees:2007th] with an equation of state with $T_c=T_{\mathrm{ch}}=175\,\mathrm{MeV}$ in semicentral $158\,A\mathrm{GeV}$ In-In collisions compared to data by the NA60 collaboration [@Arnaldi:2006jq]; middle panel: dielectron excess spectrum from the same model for central $158\,A\mathrm{GeV}$ Pb-Au collisions compared to data by the CERES/NA45 [@Adamova:2006nu] collaboration. The dash-dotted line shows the result with a $\rho$-meson spectral function including only medium modifications in a meson gas, underlining the importance of baryon effects; right panel: the dilepton excess spectrum based on the implementation of different equations of state in the fireball evolution (EoS-A: $T_c=T_{\mathrm{ch}}=175\,\mathrm{MeV}$, EoS-B: $T_c=T_{\mathrm{ch}}=160\,\mathrm{MeV}$, EoS-C: $T_c=160\,\mathrm{MeV}$, $T_{\mathrm{ch}}=160\,\mathrm{MeV}$).[]{data-label="fig.1"}](tot-EoS-compar.eps){width="\textwidth"}
**Invariant-mass spectra.** While earlier dilepton measurements at the SPS have shown an enhancement of the dilepton yield at invariant masses in the low-mass region (LMR), $2m_l \leq M \leq 1\,\mathrm{GeV}$, a definite conclusion concerning the nature of the expected CSR could not be reached since models for in-medium modifications of the $\rho$ meson based on either the “dropping-mass” or the “melting-resonance” scenario could describe the data within the experimental mass resolution and errors. Only recently with the precision reached in the measurement of dimuon-invariant-mass ($M$) spectra by the NA60 collaboration [@Arnaldi:2006jq] in $158\,\mathrm{GeV}$ In-In collisions, it could be shown that models predicting a broadening of the vector mesons with small mass shifts (cf. left panel of Fig. \[fig.1\]) seem to be favored compared to those implementing the “dropping-mass scenario”. As can be seen in the middle panel of Fig. \[fig.1\] the same model is also consistent with a recent analysis of $158\,A\mathrm{GeV}$ Pb-Au data on the dielectron-$M$ spectrum by the CERES/NA45 [@Adamova:2006nu] collaboration. As shown by the comparison with the model only implementing mesonic medium effects, baryonic processes are the prevalent effect leading to the massive broadening of the $\rho$ meson necessary to explain the observed dilepton enhancement in the LMR (including the related enhancement below the two-pion threshold).
While in the LMR the observed access yield over the standard hadronic cocktail is mostly due to the emission from the medium-modified light vector mesons, in the intermediate-mass region (IMR), $1\,\mathrm{GeV}$$\,\leq$$\,M$$\,\leq$$\,1.5\,\mathrm{GeV}$, it is either dominated by hadronic “multi-pion processes” (estimated using chiral-mixing formulas) or $q\bar{q}$-annihilation in the QGP phase (given by hard-thermal loop resummed $\bar{q}q$ annihilation) [@vanHees:2006ng; @vanHees:2007th], depending on the equation of state ($T_c$) as will be described below. As shown in the right panel of Fig. \[fig.1\], despite small deviations in the overall yield (which can be adjusted by slight variations of the fireball lifetime), the spectra are robust against details of the equation of state within the boundaries of $T_c$ from lQCD calculations and $T_{\mathrm{ch}}$ from thermal-model analyses. The insensitivity of the dilepton spectra with respect to the equation of state reflects the “quark-hadron duality” of the dilepton rates in the relevant temperature range close to $T_c$ (see Sect. \[sect.emcc\]).
![Left panel: Comparison of dielectron-$M$ spectra based on a hadronic-many-body calculation [@Rapp:2002mm] with recent data from $200\,A\mathrm{GeV}$-Au+Au collisions at RHIC from the PHENIX collaboration [@phenix:2007xw]; middle panel: $m_T$-dimuon spectrum in the mass range $0.6\,\mathrm{GeV} \leq M \leq
0.9\,\mathrm{GeV}$ compared with the NA60 data in $158A\mathrm{GeV}$ In-In collisions [@Damjanovic:2007qm; @Arnaldi:2007ru]; right panel: effective slopes from $q_T$ spectra compared to NA60 data for different equations of state and transverse acceleration of the fireball.[]{data-label="fig.2"}](phenix-dileptons-rapp-theo.eps){width="\textwidth"}
![Left panel: Comparison of dielectron-$M$ spectra based on a hadronic-many-body calculation [@Rapp:2002mm] with recent data from $200\,A\mathrm{GeV}$-Au+Au collisions at RHIC from the PHENIX collaboration [@phenix:2007xw]; middle panel: $m_T$-dimuon spectrum in the mass range $0.6\,\mathrm{GeV} \leq M \leq
0.9\,\mathrm{GeV}$ compared with the NA60 data in $158A\mathrm{GeV}$ In-In collisions [@Damjanovic:2007qm; @Arnaldi:2007ru]; right panel: effective slopes from $q_T$ spectra compared to NA60 data for different equations of state and transverse acceleration of the fireball.[]{data-label="fig.2"}](dndmt0609-semi2.eps){width="\textwidth"}
![Left panel: Comparison of dielectron-$M$ spectra based on a hadronic-many-body calculation [@Rapp:2002mm] with recent data from $200\,A\mathrm{GeV}$-Au+Au collisions at RHIC from the PHENIX collaboration [@phenix:2007xw]; middle panel: $m_T$-dimuon spectrum in the mass range $0.6\,\mathrm{GeV} \leq M \leq
0.9\,\mathrm{GeV}$ compared with the NA60 data in $158A\mathrm{GeV}$ In-In collisions [@Damjanovic:2007qm; @Arnaldi:2007ru]; right panel: effective slopes from $q_T$ spectra compared to NA60 data for different equations of state and transverse acceleration of the fireball.[]{data-label="fig.2"}](slopes-summary.eps){width="\textwidth"}
Using the spectral functions from the above described chiral-reduction approach [@Steele:1997tv] within a hydrodynamic description of the fireball evolution [@Dusling:2006yv; @Dusling:2007kh] leads to similar results for the mass region below and above the $\rho$ region but less broadening in the resonance region which is to be expected from the low-density (virial) expansion treatment of the medium effects. The spectral-functions, based on empirical $\rho$-scattering data [@Eletsky:2001bb] have been implemented within another fireball-evolution model (using a cross-over QGP-hadron phase transition) [@Ruppert:2007cr], showing results for the $M$ spectra comparable to those in [@vanHees:2006ng; @vanHees:2007th] but with less enhancement in the mass region below the $\rho$ (particularly below the two-pion threshold), which may be traced back to the use of the $T\varrho$ approximation to the medium modifications. In the IMR the model in [@Ruppert:2007cr] shows a large fraction of dilepton emission from the partonic phase. We close our brief review on invariant-mass spectra with the remark, that the enhancement of the dilepton yield in the LMR, observed by the PHENIX collaboration in $200A\,\mathrm{GeV}$ Au-Au collisions at RHIC [@phenix:2007xw], cannot be described with the present models [@Rapp:2002mm; @Dusling:2007su].
**Transverse-momentum spectra.** The dimuon transverse-momentum ($q_T$) spectra and pertinent effective-slope fits by the NA60 collaboration [@Damjanovic:2007qm; @Arnaldi:2007ru] provide information which is sensitive to the temperature and collective flow of the medium due to the *blue shift* of the dileptons radiated from a moving thermal source. While the model in [@vanHees:2006ng] describes the $q_T$ spectra for $q_T\lesssim 1\,\mathrm{GeV}$ reasonably well (which is consistent with the agreement in the inclusive $M$ spectra) they were underpredicted at $q_T \gtrsim 1 \, \mathrm{GeV}$ although the fireball parameterization of the temperature and flow agrees well with results from a hydrodynamic calculation [@Dusling:2007kh]. Thus sources for dileptons at high $q_T$ have been investigated, including an improved description of dileptons from $\rho$ decays after thermal freeze-out which benefit from the maximal blue shift due to the fully developed transverse flow [@Rapp:2006cj; @vanHees:2007th]. In this connection it is important to note that the $q_T$ spectrum for emission from a thermal source cf. (\[MT\]) is softer by a Lorentz factor $M/E=1/\gamma$ compared to that from a freely streaming $\rho$ meson due to the dilation of its lifetime. In former calculations the standard description for dileptons from freeze-out $\rho$ decays has been to prolong the fire-ball lifetime for this contribution by $1/\Gamma_{\rho}
\sim 1\,\mathrm{fm}/c$ [@Rapp:1999us; @vanHees:2006ng]. In addition decays of hard “primordial” $\rho$ mesons, produced in the initial hard $NN$ collisions which are subject to jet quenching through the medium but leaving the fireball without equilibrating, have been taken into account. Another source of hard dileptons is Drell-Yan annihilation in primordial $NN$ collisions which has been extrapolated to small invariant masses by imposing constraints from the real-photon point. Finally, $t$-channel-meson (e.g., $\omega$) exchange contributions to the yield of thermal dileptons have been studied. Although the latter show the hardest $q_T$ spectra among all thermal sources, their absolute magnitude is insufficient to resolve the discrepancies in comparison to the data at high $q_T$, which however has improved through the above described more detailed implementation of the hard (non-thermal) dilepton sources (cf. Fig. \[fig.1\] middle panel). The model of [@Ruppert:2007cr] shows larger slopes (also compared to the hydrodynamic fireball simulation [@Dusling:2007kh]).
Finally a study of different parameters for the equation of state has been conducted. The “standard scenario” in [@Rapp:1999us; @vanHees:2006ng] uses $T_c=T_{\mathrm{ch}}=175\,\mathrm{GeV}$ (EoS-A). To investigate the sensitivity of the dilepton spectra with respect to uncertainties in the equation of state, $T_c$ has been varied within the boundaries of $160$-$190\,\mathrm{GeV}$ given by different lQCD calculations [@Fodor:2004nz; @Karsch:2007vw] (EoS-B: $T_c=160\,\mathrm{MeV}$, EoS-C: $T_c=190\,\mathrm{MeV}$), using a chemical freeze-out temperature of $T_{\mathrm{ch}}=160\,\mathrm{GeV}$ to cover the range of thermal-model fits to hadron-number ratios in heavy-ion collisions. For EoS-C a chemically equilibrated hadronic phase between $T_c=190\,\mathrm{MeV}$ and $T_{\mathrm{ch}}=160\,\mathrm{MeV}$ has been assumed. The agreement of the model with the measured $M$ spectra is robust. Variations in the absolute yield can be adjusted by slight changes in the fireball lifetime. It is important to note that in the IMR the partition of the dilepton yield in hadronic and partonic contributions depends sensitively on the equation of state: A scenario like EoS-B with a low critical temperature results in a long QGP+mixed phase, leading to a parton-dominated regime in the IMR, while EoS-C with a high critical temperature describes a hadron-dominated source since the QGP and mixed phase become shorter. Thus, contrary to suggestions in the literature [@Arnaldi:2007ru; @Specht:2007ez], a definite conclusion whether the dilepton yield in the IMR is originating from partonic or hadronic sources can not be drawn at present. In the right panel of Fig. \[fig.2\] effective slopes extracted from the $q_T$ spectra by a fit to ${\mathrm{d}}N_{ll}/(m_T {\mathrm{d}}m_T)=C
\exp(-m_T/T_{\mathrm{eff}})$ are shown. The slopes from the calculations with EoS-B and EoS-C benefit from the larger freeze-out temperature of $T_{\mathrm{fo}}=136\,\mathrm{MeV}$ compared to $T_{\mathrm{fo}}=120\,\mathrm{MeV}$ for EoS-A. To reach the measured effective slopes however, an enhancement of the transverse acceleration of the fireball (from $a_{\perp}=0.085 c^2/\mathrm{fm}$ to $a_{\perp}=0.1 c^2/\mathrm{fm}$), leading to larger blue shifts in the spectra, is necessary. The slopes in the IMR are not so sensitive to the radial flow since the emission in this region is dominated by radiation from (either partonic or hadronic, depending on $T_c$) sources at earlier times where the flow is smaller.
Conclusions and outlook {#sect.concl}
=======================
In conclusion, the confrontation of phenomenological models for the em. current correlator of strongly interacting matter with precise data on dilepton emission in high-energy heavy-ion collisions provides a unique opportunity for a better understanding of the nature of chiral symmetry restoration. Models based on the application of many-body theory to phenomenological hadronic models to assess the medium modifications of the em. current correlator, predicting a strong broadening of the light-vector-meson spectrum with small mass shifts, are favored by the data compared to models implementing a dropping-mass scenario (as, e.g., implied by the intrinsic temperature dependencies of the model parameters of the generalized hidden-local symmetry model due to “Wilsonian matching” with QCD close to $T_c$ [@Harada:2006hu] although a final confrontation of this particular realization of chiral symmetry with dilepton data in HICs has to be completed by an implementation of baryonic interactions). However, the origin of the large dilepton enhancement in the LMR observed by the PHENIX collaboration in $200 \,A\mathrm{GeV}$ Au-Au collisions at RHIC remains unexplained so far.
Future investigations will have to find even closer connections between em. observables in heavy-ion collisions and the chiral phase transition. One possibility is the extension of the hadronic-model calculations with a detailed study of the in-medium properties of both the vector and the axial-vector correlator within a chiral framework, constrained by lQCD calculations of chiral order parameters via Weinberg-sum rules.
**Acknowledgment.** I thank the conference organizers for the invitation to an interesting meeting, R. Rapp for the fruitful collaboration and S. Damjanovic and H. Specht for stimulating discussions. This work was supported in part by the U.S. National Science Foundation under grant no. PHY-0449489.
[10]{} url \#1[`#1`]{}urlprefix\[2\]\[\][[\#2](#2)]{}
Rapp R and Wambach J 2000 *Adv. Nucl. Phys.* **25** 1
Shuryak E V 1980 *Phys. Rept.* **61** 71
McLerran L D and Toimela T 1985 *Phys. Rev. D* **31** 545
Gale C and Kapusta J I 1991 *Nucl. Phys. B* **357** 65
Barate R *et al.* (ALEPH Collaboration) 1998 *Eur. Phys. J. C* **4** 409
Ackerstaff K *et al.* (OPAL Collaboration) 1999 *Eur. Phys. J. C* **7** 571
Karsch F and Laermann E 1994 *Phys. Rev. D* **50** 6954
Brown G E and Rho M 1991 *Phys. Rev. Lett.* **66** 2720
Gale C and Lichard P 1994 *Phys. Rev. D* **49** 3338
Rapp R, Chanfray G and Wambach J 1997 *Nucl. Phys. A* **617** 472
Rapp R and Wambach J 1999 *Eur. Phys. J. A* **6** 415
Asakawa M and Ko C M 1993 *Phys. Rev. C* **48** 526
Leupold S, Peters W and Mosel U 1998 *Nucl. Phys. A* **628** 311
Klingl F, Kaiser N and Weise W 1997 *Nucl. Phys. A* **624** 527
Ruppert J, Renk T and M[ü]{}ller B 2006 *Phys. Rev. C* **73** 034907
Weinberg S 1967 *Phys. Rev. Lett.* **18** 507
Kapusta J I and Shuryak E V 1994 *Phys. Rev. D* **49** 4694
Dey M, Eletsky V L and Ioffe B L 1990 *Phys. Lett. B* **252** 620
Bando M and Kugo T 1984 *Phys. Rev. Lett.* **54** 1215
Harada M and Yamawaki K 2003 *Phys. Rept.* **381** 1
Harada M and Sasaki C 2006 *Phys. Rev. D* **73** 036001
Allton C R *et al.* 2005 *Phys. Rev. D* **71** 054508
Eletsky V L, Belkacem M, Ellis P J *et al.* 2001 *Phys. Rev. C* **64** 035202
Kolb P F and Heinz U W 2003
H and Rapp R 2008 *Nucl. Phys. A* **806** 339
H and Rapp R 2006 *Phys. Rev. Lett.* **97** 102301
Fodor Z and Katz S D 2004 *JHEP* **04** 050
Karsch F 2007 *J. Phys. G* **34** S627
Andronic A, Braun-Munzinger P and Stachel J 2006 *Nucl. Phys. A* **772** 167
Becattini F, Manninen J and Gazdzicki M 2006 *Phys. Rev. C* **73** 044905
Arnaldi R *et al.* (NA60 Collaboration) 2006 *Phys. Rev. Lett.* **96** 162302
Adamova D *et al.* (CERES Collaboration) 2006
Rapp R 2002
Afanasiev S *et al.* (PHENIX Collaboration) 2007
Damjanovic S *et al.* (NA60 Collaboration) 2007 *Nucl. Phys. A* **783** 327
Arnaldi R *et al.* (NA60 Collaboration) 2007
Steele J V, Yamagishi H and Zahed I 1997 *Phys. Rev. D* **56** 5605
Dusling K, Teaney D and Zahed I 2007 *Phys. Rev. C* **75** 024908
Dusling K and Zahed I 2007
Ruppert J, Gale C, Renk T *et al.* 2007
Dusling K and Zahed I 2007
Rapp R, [van Hees]{} H and Strong T 2007 *Braz. J. Phys.* **37** 779
Specht H 2007
Harada M and Sasaki C 2006 *Phys. Rev. D* **74** 114006
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[We find the general charged rotating black hole solutions of the maximal supergravities in dimensions $4\le D\le 9$ arising from toroidally compactified Type II string or M-theories. In each dimension, these are obtained by acting on a generating solution with classical duality symmetries. In $D=4$, $D=5$ and $6\le D \le 9$ the generating solution is specified by the ADM mass, $[{{D-1}\over 2}]$-angular momentum components and five, three and two charges, respectively. We discuss the BPS-saturated (static) black holes and derive the $U$-duality invariant form of the area of the horizon. We also comment on the $U$-duality invariant form of the BPS mass formulae. ]{}'
address: |
$^1$ School of Natural Science, Institute for Advanced Study,\
Olden Lane, Princeton, NJ 08540, U.S.A.,\
$\ $\
$^1$ Department of Applied Mathematics and Theoretical Physics,\
University of Cambridge, Cambridge CB3 9EW, U.K.\
and\
$^2$ Physics Department, Queen Mary and Westfield College,\
Mile End Road, London E1 4NS, U.K.
author:
- 'Mirjam Cvetič$^1$ [^1] and Christopher M. Hull$^2$'
date: June 1996
title: 'Black Holes and U-Duality'
---
Introduction
============
Black holes play an important role in string theory and recent developments (for a review, see [@Horowitz]) have shown that string theory makes it possible to address their microscopic properties, in particular the statistical origin of the black hole entropy and possibly issues of information loss. The starting point in such investigations is the classical black hole solution and the aim of this paper is to find such solutions in toroidally compactified string theories in dimensions $4\le D \le 9$.
The general solutions are found from a particular black hole “generating solution”, which is specified by a canonical choice of the asymptotic values of the scalar fields, the ADM mass, $[{{D-1}\over
2}]$-components of angular momentum and a (minimal) number of charge parameters. The most general black hole, compatible with the “no-hair theorem”, is then obtained by acting on the generating solution with classical duality transformations. These are symmetries of the supergravity equations of motion, and so generate new solutions from old. They do not change the $D$-dimensional Einstein-frame metric but do change the charges and scalar fields. We first consider transformations, belonging to the maximal compact subgroup of duality transformations, which preserve the canonical asymptotic values of the scalar fields and show that all charges are generated in this way. Another duality transformation can be used to change the asymptotic values of the scalar fields.
For the toroidally compactified heterotic string such a program is now close to completion. Particular examples of solutions had been obtained in a number of papers (for a recent review and references, see [@Horowitz]). In dimensions $D=4$, $D=5$ and $6\le
D\le 9$ the generating solution has five, three and two charge parameters, respectively. The charge parameters of the generating solution are associated with the $U(1)$ gauge fields arising from Kaluza-Klein (momentum modes) and two-form (winding modes) sectors for at most two toroidally compactified directions. The general black hole solution is then obtained by applying to the generating solution a subset of transformations, belonging to the maximal compact subgroup of the $T$- and $S$-duality transformations [@SEN]. The explicit expression for the generating solution has been obtained in $D=5$ [@CY5r] and $D\ge 6$[@CYNear; @Llatas], however, in $D=4$ only the five charge static generating solution [@CY4s] (see also [@JMP]) and the four charge rotating solutions [@CY4r] were obtained.
The BPS-saturated solutions of the toroidally compactified heterotic string have non-singular horizons only for $D=4$ static black holes [@CY; @CTII] and $D=5$ black holes with one non-zero angular momentum component [@TMpl]. In $6\le D\le 9$ the BPS-saturated black holes have singular horizons with zero area. The explicit $T$- and $S$-duality invariant formulae for the area of the horizon and the ADM mass for the general BPS-saturated black holes were given for $D=4$ in [@CY; @CTII] and for $D=5$ in [@TMpl]. In particular, the area of the horizon of the BPS-saturated black holes [*does not*]{} depend on the asymptotic values of the scalar fields [@LWI; @CTII; @S; @FK], suggesting a microscopic interpretation.
The purpose of this paper is to study properties of the classical black hole solutions of [*toroidally compactified Type II string theory or ${\rm M}$-theory*]{}, in dimensions $4\le D\le 9$, thus completing the program for the toroidally compactified superstring vacua. We identify the minimum number of charge parameters for the generating solutions, which fully specifies the space-time metric of the [*general*]{} black hole in $D$-dimensions. The “toroidally” compactified sector of the heterotic string and the Neveu-Schwarz-Neveu-Schwarz (NS-NS) sector of the toroidally compactified Type II string have the [*same*]{} effective action and so have the same classical solutions. In this paper we shall show that the generating solutions for black holes in the toroidally compactified Type II string theory (or M-theory) are the [*same*]{} as the ones of toroidally compactified heterotic string. Note that it could have been the case that a more general generating solution with one or more RR charges was needed. Applying $U$-duality transformations to the generating solution generates all black holes of toroidally compactified Type II string theory (or M-theory).
We further address the BPS-saturated solutions, identify the $U$-duality invariant expression for the area of the horizon, i.e. Bekenstein-Hawking (BH) entropy for the general BPS-saturated black holes and outline a procedure to obtain the manifestly $U$-duality invariant mass formulae.
The paper is organized as follows. In Section II we summarize the symmetries of the effective action of toroidally compactified Type II string and obtain the general solution by applying a compact subgroup of duality transformations to the generating solution. In Section III we concentrate on general static BPS-saturated black holes in $4\le D\le 9$ and derive the $U$-duality invariant expression for the area of the horizon. In Appendix A the effective action of the NS-NS sector of toroidally compactified Type II string is given. The explicit form of some of the generating solutions is given in in Appendix B.
Toroidally Compactified Type II String Theory
===============================================
Symmetries
----------
The low-energy effective action for the Type II string or M-theory, toroidally compactified to $D$-dimensions, is the maximal supergravity theory, which has a continuous duality symmetry $U$ of its equations of motion [@CJ] (see Table I, first column). This has a maximal compact subgroup $C_U$ (second column in Table I). In the quantum theory the continuous classical symmetry $U$ is broken to a discrete subgroup ${ Q}_{ U}$ [@HT] (third column in Table I) which is the $U$-duality symmetry of the string theory. However, we will sometimes refer to the group $U$ as the classical $U$-duality.
Solution Generating Technique
-----------------------------
The general black hole solution is obtained by acting on generating solutions with $U$-duality transformations.
The scalar fields take values in the coset $U/C_U$ and can be parameterised by a $U$-valued matrix ${\cal V}(x)$ which transforms under rigid $U$-transformations from the right and local $C_U$ transformations from the left [@CJ]. The Kaluza-Klein and and antisymmetric tensor $U(1)$ gauge fields also transform under $U$. It is convenient to define ${\cal M}={\cal V}^t{\cal
V}$ which is inert under $C_U$ and transforms under $U$ as ${\cal M}{} \to { \Omega} {\cal M}{}{
\Omega}^T $ (${\Omega} \in { U}$).
The asymptotic value ${\cal
M}_{\infty} $ of ${\cal M}$ can be brought to the canonical value ${\cal M}_{0\, \infty}= {\bf
1}$ by a suitable $U$-duality transformation $\Omega_0$. The canonical value ${\cal M}_{0\, \infty}$ is preserved by $C_U$ and the most general solution with the asymptotic behaviour ${\cal
M}_{\infty} ={\cal
M}_{0\, \infty}$ is obtained by acting on the generating solution with a subset of $C_U$ transformations, i.e. the $C_U$ orbits which are of the form $C_U/C_0$ where $C_0$ is the subgroup preserving the generating solution. In particular, with this procedure the complete set of charges is obtained. Indeed, the generating solution is labelled by $n_0$ charges ($n_0=5,3,2$ for $D=4,5,\ge 6$, respectively) and if the dimension of the $C_U$ orbits is $n_1$, then $n_0+n_1$ is the correct dimension of the vector space of charges for the general solution, as we shall check in the following Section. Black holes with arbitrary asymptotic values of scalar fields ${\cal M}_{\infty}$ can then be obtained from these by acting with $\Omega_0$.
We shall seek the general charged rotating black hole solutions. In $D=4$, such solutions are specified by electric [*and*]{} magnetic charges, while in $D>4$ they carry electric charges only (once all $(D-3)$-form gauge fields have been dualised to vector fields).
Black Holes in Various Dimensions
=================================
We will first propose generating solutions for Type II string (or M-theory) black holes in dimensions $4\le D\le
9$ and then go on to show that the action of duality transformations generates all solutions. Remarkably, the generating solutions are the same as those used for the heterotic string. We will discuss only the charge assignments here, and give the explicit solutions in Appendix B.
Charge Assignments for the Generating Solution {#cgs}
----------------------------------------------
### D=4
The generating solution is specified in terms of [*five*]{} charge parameters. It is convenient to choose these to arise in the NS-NS sector of the compactified Type II string as follows. We choose two of the toroidal dimensions labelled by $i=1,2$ and let $A_{\mu\,
i}^{(1)}$ be the two graviphotons (corresponding to $G_{\mu i}$) and $A_{\mu\, i}^{(2)}$ the two $U(1)$ gauge fields coming from the antisymmetric tensor (corresponding to $B_{\mu i}$) (see Appendix A). Corresponding to these four $U(1)$ gauge fields there are four electric charges $Q^{(1),(2)}_i$ and four magnetic ones $P^{(1),(2)}_i$. The generating solution, however carries the following five charges: $ Q_1\equiv Q_1^{(1)},\ Q_2\equiv Q_1^{(2)}, \ P_1\equiv P_2^{(1)},
P_2\equiv P_2^{(2)}$ and $q\equiv Q_2^{(1)}=-Q_2^{(2)}$. It will be useful to define the left-moving and right-moving charges $Q_{i\, L,R}\equiv Q_i^{(1)}\mp Q_i^{(2)}$ and $P_{i\, L,R}\equiv P_i^{(1)}\mp P_i^{(2)}$ ($i=1,2$). The generating solution then carries five charges associated with the first two compactified toroidal directions of the NS-NS sector, where the dyonic charges are subject to the constraint ${\vec P}_R{\vec Q}_R=0$. We choose the convention that all the five charge parameters are positive.
### D=5
In $D=5$ the generating solution is parameterised by three (electric) charge parameters: $Q_1\equiv Q_1^{(1)}, \ Q_2\equiv Q_1^{(2)},$ and $ {\tilde Q}$. Here the electric charges $Q_i^{(1),(2)}$ arise respectively from the graviphoton $A_{\mu\, i}^{(1)}$ and antisymmetric tensor $A_{\mu\, i}^{(2)}$ $U(1)$ gauge fields of the $i$-$th$ toroidally compactified direction of the NS-NS sector, and $\tilde Q$ is the electric charge of the gauge field, whose field strength is related to the field strength of the two-form field $B_{\mu\nu}$ by duality transformation (see Appendix A). Again we choose the convention that all three charges are positive.
### $6\le D\le 9$
In $6\le D\le 9$ the generating solution is parameterised by two electric charges: $Q_1\equiv Q_1^{(1)}, \ Q_2\equiv Q_1^{(2)}$. Again, the electric charges $Q_i^{(1),(2)}$ arise respectively from the graviphoton $A_{\mu\, i}^{(1)}$ and antisymmetric tensor $A_{\mu\, i}^{(2)}$ $U(1)$ gauge fields of the $i$-$th$ toroidally compactified direction and we use the convention that both charges are positive.
Note that the explicit form of the generating solutions with the above charge assignments is the [*same*]{} as the one of the toroidally compactified heterotic string, since the corresponding NS-NS sector of the toroidally compactified string and the “toroidal” sector of the heterotic string are the same.
Action of Duality Transformations on Generating Solution
--------------------------------------------------------
### D=4
The $N=8$ supergravity has 28 abelian gauge fields and so the general black hole solution carries 56 charges (28 electric and 28 magnetic). The ${ U}$-duality group is $E_{7(7)}$, the maximal compact subgroup ${ C_U}$ is $SU(8)$ and the $T$-duality subgroup is $SO(6,6)$ . We use the formulation with rigid $E_7$ symmetry and local $SU(8)$ symmetry [@CJ]. The 56 charges fit into a vector ${\cal Z}$ transforming as a [**56**]{} of $E_7$. In the quantum theory, ${\cal Z}$ is constrained to lie in a lattice by charge quantisation [@HT]. This “bare” charge vector can be “dressed” with the asymptotic value ${\cal V}_\infty$ of the scalar field matrix ${\cal V}$ to form $$\bar {\cal Z}={\cal V}_\infty{\cal Z}=
\left ( \matrix{
q^{ab}\cr
p_{ab}\cr}
\right ) \ ,
\label{zv}$$ which is invariant under $E_7$ but transforms under local $SU(8)$. The 28 electric and 28 magnetic dressed charges are $q_{ab}$ and $p_{ab}$ ($a,b=1, \cdots , 8$ and $q_{ab}=-q_{ba}$,$p_{ab}=-p_{ba}$). They can be combined to form the $Z_{4\,AB}$ matrix ($A,B=1, \cdots , 8$ are $SU(8)$ indices) transforming as the complex antisymmetric representation of $SU(8)$, by defining $
Z_{4\,AB}= (q^{ab}+ip_{ab}) (\Gamma^{ab})^B{}_A$ where $(\Gamma^{ab})^B{}_A$ are the generators of $SO(8)$ in the spinor representation [@CJ]. The matrix $Z_{4\,AB}$ appears on the right hand side of the anticommutator of chiral two-component supercharges $$[Q _{A\alpha}, Q_{B\beta}]= C_{\alpha\beta} Z_{4\,AB} \ ,$$ and thus corresponds to the matrix of 28 complex central charges. An $SU(8)$ transformation $ Z_4\rightarrow Z^0_4=({\cal U} Z_4{\cal U}^T)$ brings this charge matrix to the skew-diagonal form: $$Z^0_4= \left ( \matrix{
0& \lambda_1&0&0&0&0&0&0 \cr
-\lambda_1&0&0&0&0&0&0&0 \cr
0&0&0& \lambda_2&0&0&0&0 \cr
0&0&-\lambda_2&0&0&0&0&0\cr
0&0&0&0&0& \lambda_3&0&0\cr
0&0&0&0&-\lambda_3&0&0&0\cr
0&0&0&0&0&0&0& \lambda_4\cr
0&0&0&0&0&0&-\lambda_4&0\cr}
\right )\ ,
\label{diag4}$$ where the complex $\lambda_{i}$ ($i=1,2,3,4$) are the skew eigenvalues.
For the generating solution with the five charge parameters $Q_{1,2}$, $P_{1,2}$ and $q$ (see Subsection \[cgs\]) the eigenvalues are [^2]: $$\begin{aligned}
\lambda_1&=&Q_{1R} + P_{2R},\cr
\lambda_2&=&Q_{1R} -P_{2R},\cr
\lambda_3&=&Q_{1L} + P_{2L}+2iq,\cr
\lambda_4&=&Q_{1L} - P_{2L}-2iq
\label{fivepar}\end{aligned}$$ (recall $Q_{1\, L,R}\equiv Q_1\mp Q_2$ and $P_{2\, L,R}\equiv P_1\mp P_2$).
We now consider the action of duality transformations on the generating solution and show that all $D=4$ black hole solutions are indeed generated. The $U$-duality group $E_7$ has a maximal subgroup $SO(6,6)\times SL(2,{\bf
R})$ where $SO(6,6)$ is the $T$-duality group and $ SL(2,{\bf R})$ is the $S$-duality group. (Strictly speaking, the duality groups are discrete subgroups of these.) The [**56**]{} representation of $E_7$ decomposes as $${\bf 56} \to (12,2) \oplus (32,1)$$ under $SO(6,6)\times SL(2,{\bf R})$ and thus the 56 charges ${\cal Z}$ decompose into $12$ electric and $12$ magnetic charges in the NS-NS sector, and 32 charges in the Ramond-Ramond (RR) sector. We choose for now the asymptotic value of the scalars to be the canonical one, i.e. ${\cal V}_\infty={\cal V}_{0\infty}\equiv {\bf
1}$.
The maximal compact symmetry of the $T$-duality group is $SO(6)_L\times SO(6)_R\sim SU(4)_L\times SU(4)_R$, and under $ SU(4)_L\times
SU(4)_R\subset SU(8)$ the complex representation ${\bf 28}$ decomposes into complex representations $(12,1)+(1,12)+(4,4)$. This decomposition corresponds to splitting the $8\times 8$ matrix of charges $Z_4$ into $4\times 4$ blocks. The two $4\times 4$ diagonal blocks $Z_R$ and $Z_L$, transform respectively as the antisymmetric complex representations of $SU(4)_{R,L}\sim SO(6)_{R,L}$ and represent the $12+12$ charges of the NS-NS sector. The off-diagonal blocks correspond to the 16 complex RR charges.
The maximal compact subgroup of $SO(6,6)\times SL(2,{\bf R})$ is $SO(6)_L\times
SO(6)_R\times
SO(2)$ and it preserves ${\cal V}_{0\infty}$. The subgroup that preserves the charges of the generating solution is $SO(4)_L\times SO(4)_R$. Thus acting on the generating solution with $SO(6)_L\times SO(6)_R\times
SO(2)$, gives orbits corresponding to the 19-dimensional space $${SO(6)_L\times SO(6)_R \over SO(4)_L\times SO(4)_R} \times SO(2)\ .
\label{osp}$$ As the generating solution has five charges, acting on the generating solution with $SO(6)_L\times
SO(6)_R\times
SO(2)$ gives the required $5+19=24$ NS-NS charges, i.e. the 24 NS-NS charges are parameterised in terms of the five charges of the generating solution and the 19 coordinates of the orbit space (\[osp\]).
The above procedure is closely related to that for $D=4$ toroidally compactified heterotic string vacua [@CY; @CTII], where the general black hole with the $5+51=56$ charges is obtained from the same five-parameter generating solution, and the 51 coordinates of the orbit $${SO(22)_L\times SO(6)_R\over SO(20)_L\times SO(4)_R} \times SO(2) \ .$$
We can now generalise this procedure to include the RR charges. The group $C_U=SU(8)$ preserves the canonical asymptotic values of the scalar fields and only the subgroup $SO(4)_{L}\times SO(4)_{R}$ leaves the generating solution invariant. Then acting with $SU(8)$ gives orbits $$SU(8)/[SO(4)_L\times SO(4)_R]
\label{4du}$$ of dimension $63-6-6=51$. The $SU(8)$ action then induces $51$ new charge parameters, which along with the original five parameters provide charge parameters for the general solution with 56 charges. Finally, the general black hole with arbitrary asymptotic values of the scalars is obtained from these 56-parameter solutions by acting with a $E_7$ transformation. This transformation leaves the central charge matrix $Z_4$ and its eigenvalues $\lambda_i$ invariant, but changes the asymptotic values of scalars and “dresses” the physical charges. The orbits under $E_7$ are the 70-dimensional coset $E_7/SU(8)$, as expected.
The fact that the same five-parameter generating solution that was used for the $D=4$ toroidally compactified heterotic string should be sufficient to generate all black holes with NS-NS charges of $D=4$ toroidally compactified Type II is unsurprising, given the equivalence between the “toroidal” sector of the heterotic string and the NS-NS sector of the Type II string. However, it is interesting that the procedure outlined above is also sufficient to generate [*all*]{} RR charges of the general black hole solution, as it could have been the case that a more general generating solution carrying one or more RR charges was needed.
### D=5
The ${ U}$-duality group is $E_{6(6)}$, the maximal compact subgroup ${ C_U}$ is $USp(8)$ and the $T$-duality group is $SO(5,5)$ with its maximal compact subgroup $SO(5)_L\times SO(5)_R$. In this case there are 27 abelian gauge fields and the 27 electric charges (dressed with asymptotic values of the scalar fields) transform as a [**27**]{} of USp(8) and can be assembled into an $8\times 8$ matrix $Z_{5\,AB}$ ($A,B=1,\dots, 8$) with the properties [@Cremmer]: $$Z_5^{AB\, *}=\Omega^{AC}\Omega^{BD} Z_{5\,CD}, \ \ \Omega^{AB}Z_{5\,AB}=0 ,
\label{Z5}$$ where $\Omega$ is the $USp(8)$ symplectic invariant, which we take to be $$\begin{aligned}
\Omega= \left ( \matrix{
0&1&0&0&0&0&0&0 \cr
-1&0&0&0&0&0&0&0 \cr
0&0&0&1&0&0&0&0 \cr
0&0&-1&0&0&0&0&0\cr
0&0&0&0&0& 1&0&0\cr
0&0&0&0&-1&0&0&0\cr
0&0&0&0&0&0&0& 1\cr
0&0&0&0&0&0&-1&0\cr}
\right ) .
\label{explo}\end{aligned}$$ With $\Omega$ given by (\[explo\]), the $Z_5$ charge matrix can be written in the following form: $$\begin{aligned}
Z_5= \left ( \matrix{
0& z_{12}& z_{13}&z_{14}& \cdots \cr
-z_{12}&0&-z_{14}^*&z_{13}^*&\cdots \cr
-z_{13}^*&z_{14}&0&z_{34}&\cdots\cr
-z_{14}^*&-z_{13}&-z_{34}&0&\cdots\cr
\cdots&\cdots&\cdots&\cdots&\cdots}
\right ).
\label{explZ0}\end{aligned}$$ Here $z_{12},\ z_{34}, \ z_{56}$ are real and satisfy $ z_{12}+ z_{34}+
z_{56}=0$.
The matrix $Z_5$ occurs in the superalgebra and represents the 27 (real) central charges. It can be brought into a skew-diagonal form of the type (\[diag4\]) by an $USp(8)$ transformation $ Z_5\rightarrow Z^0_5=({\cal U} Z_5 {\cal U}^T)$. The four real eigenvalues $\lambda_i$ are subject to the constraint: $\sum_{i=1}^4\lambda_i=0$.
The generating solution is parameterised by three charges $Q_1\equiv Q_1^{(1)},\ Q_2\equiv Q_1^{(2)}$ and $\tilde Q$ (see Subsection \[cgs\]). The four (constrained, real) eigenvalues $\lambda_i$ are then $$\begin{aligned}
\lambda_1&=& \tilde Q+Q_{1R},\cr
\lambda_2&=&\tilde Q-Q_{1R},\cr
\lambda_3&=&-\tilde Q+Q_{1L},\cr
\lambda_4&=&-\tilde Q-Q_{1L},
\label{threepar}\end{aligned}$$ where $Q_{1\, L,R}\equiv Q_1\mp Q_2$. These indeed satisfy the constraint $\sum_{i=1}^4\lambda_i=0$.
The three parameter solution is indeed the generating solution for black holes in $D=5$. The $USp(8)$ duality transformations preserve the canonical asymptotic values of the scalar fields and the subgroup $SO(4)_L\times
SO(4)_R[\subset
SO(5)_L\times SO(5)_R]
\subset USp(8)$ preserves the generating solution. Acting with $USp(8)$ on the generating solution gives orbits $$USp(8)/[SO(4)_L\times SO(4)_R]\ ,
\label{5du}$$ and thus induces $36 -4\times 3=24$ new charge parameters, which along with the original three charge parameters provide 27 electric charges for the general solution in $D=5$.
### D=6
The ${ U}$-duality group is $SO(5,5)$, the maximal compact subgroup ${ C_U}$ is $SO(5)\times
SO(5)$ and the $T$-duality group is $SO(4,4)$ which has maximal compact subgroup $SO(4)_L\times SO(4)_R\sim [SU(2)\times SU(2)]_L\times [SU(2)\times
SU(2)]_R
$. There are 16 abelian vector fields and the bare charges ${\cal Z}$ transform as a [**16**]{} (spinor) of $SO(5,5)$. The dressed charges transform as the $(4,4)$ representation of $SO(5)\times
SO(5)$ and can be arranged into a $4 \times 4 $ charge matrix $Z_6$.
Under $ [SU(2)\times SU(2)]_L\times [SU(2)\times SU(2)]_R\subset SO(5)\times
SO(5)$ the $(4,4)$ decomposes into $(2,2,1,1) +(1,1,2,2)+(1,2,2,1)+(2,1,1,2)$. This decomposition corresponds to splitting the $4\times 4$ matrix of charges $Z_6$ into $2\times 2$ blocks. The two $2\times 2$ diagonal blocks $Z_R$ and $Z_L$, transform respectively as $(2,2,1,1)$ and $(1,1,2,2)$ representations of $ [SU(2)\times SU(2)]_L\times [SU(2)\times
SU(2)]_R$ representing the $4+4$ charges of the NS-NS sector. The off-diagonal blocks correspond to $(1,2,2,1)+(2,1,1,2)$ representations of $ [SU(2)\times SU(2)]_L\times [SU(2)\times SU(2)]_R $ and represent 8 RR charges.
The matrix $Z_6$ occurs in the superalgebra and represents the 16 (real) central charges. It can be brought into a skew-diagonal form of the type (\[diag4\]) by an $SO(5)\times SO(5)$ transformation $ Z_6\rightarrow Z^0_6=({\cal U} Z_6 {\cal U}^T)$ with the two eigenvalues $\lambda_i$. The generating solution is parameterised by two charges $Q_1\equiv Q_1^{(1)}, \ Q_2=Q_1^{(2)}$ (see Subsection \[cgs\]). The two eigenvalues $\lambda_i$ are then $$\begin{aligned}
\lambda_1&=& Q_{1R},\cr
\lambda_2&=&Q_{1L},
\label{twopar}\end{aligned}$$ where again $Q_{1\, L,R}\equiv Q_1\mp Q_2$.
The generating solution is preserved by $SO(3)_L\times SO(3)_R [\subset
SO(4)_L\times SO(4)_R]\subset SO(5)\times SO(5)$ so acting with $SO(5)\times SO(5)$ gives $$[SO(5)\times
SO(5)]/[SO(3)_L\times SO(3)_R]
\label{6dtr}$$ orbits, and thus introduces $2 (10-3)=14$ charge parameters, which along with the two charges ($Q_{1,2}$) of the generating solution provide the 16 charge parameters of the general solution in $D=6$.
### D=7
The ${ U}$-duality group is $SL(5,{\bf R})$, the maximal compact subgroup ${ C_U}$ is $SO(5)$ and the $T$-duality group is $SO(3,3)$ with its maximal compact subgroup $SO(3)_L\times SO(3)_R$. There are ten abelian vector fields and the ten bare electric charges transform as the $\bf 10$ representation of $SL(5, {\bf R})$. Dressing of these with asymptotic values of scalars gives the ten central charges which are inert under $SL(5, {\bf R})$ but transform as a $\bf 10$ under $SO(5)$. The dressed charges can be assembled into a real antisymmetric $5\times 5$ charge matrix $Z_{7ij}$, which appears in the superalgebra as the $4 \times 4$ central charge matrix $Z_{7AB}={1 \over 2}Z_{7ij}{\gamma ^{ij}}_{AB}$ where $ {\gamma ^{ij}}_{AB}$ are the generators of $SO(5)$ in the spinor ([**4**]{}) representation. The matrix $Z_{7ij}$ has two real skew eigenvalues, $\lambda_1, \lambda_2$, which for the generating solution correspond to the two charges $Q_{1\, L,R}$.
The subgroup $SO(2)_L\times SO(2)_R\subset SO(3)_L\times SO(3)_R\subset
O(5)$ preserves the generating solution, so that the action of $SO(5)$ gives orbits $$SO(5)/[SO(2)_L\times SO(2)_R]\ ,
\label{6du}$$ thus introducing $10-2=8$ charge parameters, which together with the two charges ($Q_{1,2}$) of the generating solution (see Subsection \[cgs\]) provide the ten charges of the general solution in $D=7$.
### D=8
The ${ U}$-duality group is $SL(3,{\bf R})\times SL(2, {\bf R})$, the maximal compact subgroup ${C_U}$ is $SO(3)\times U(1)$ and the $T$-duality group is $SO(2,2)$ with maximal compact subgroup $SO(2)_L\times SO(2)_R$. There are six abelian gauge fields and the six bare electric charges transform as $({\bf 3},{\bf 2})$ under $SL(3, {\bf R})\times SL(2,{\bf R})$.
No $C_U$ transformations preserve the generating solution, so that the orbits are $$C_U=SO(3)\times U(1)\ ,
\label{7du}$$ and the $C_U$ transformations introduce $(3+1)$-charge parameters, which along with the two charges of the generating solution provide the six charges of the general solution in $D=8$.
### D=9
The ${U}$-duality group is $SL(2,{\bf R})\times {\bf R}^+$ and the maximal compact subgroup ${ C_U}$ is $U(1)$. There are three abelian gauge fields and the three bare electric charges transform as $({\bf 3}, {\bf 1})$ under $SL(2,{\bf R})\times {\bf R}^+$. The action of $$C_U=U(1),
\label{9du}$$ introduces one-charge parameter, which along with the two charges of the generating solution provides the three charges of the general solution in $D=9$.
Entropy and Mass of BPS-Saturated Static Black Holes
====================================================
We now study the properties of static BPS-saturated solutions. In the preceding Section we identified the charge assignments for the generating solutions, which fully specify the space-time of the general black hole solution in $D$-dimensions for toroidally compactified Type II string (or M-theory) vacuum. The explicit form for these solutions has been given in the literature (with the exception of the rotating five-charge solution in $D=4$). The static generating solutions are given in Appendix B. Also, the area of the horizon for the BPS-saturated (as well as for non-extreme solutions) was calculated explicitly. In addition to static BPS-saturated solutions, we shall also consider near-BPS-saturated solutions in $4\le D\le 9$.
The Bogomol’nyi Bound
---------------------
Consider first the $D=4$ case. Standard arguments [@GH] imply that the ADM mass $M$ is bounded below by the moduli of the eigenvalues $\lambda _i$ (\[diag4\]) of the central charge matrix $Z_4$, i.e. $M \ge |\lambda _i|$, $i=1,...,4$. Without loss of generality the eigenvalues can be ordered in such a way that $|\lambda_i|\ge |\lambda_j|$ for $j\ge i$ .
If $M$ is equal to $|\lambda _1|=\cdots =|\lambda_p|$, the solution preserves ${p\over 8}$ of $N=8$ supersymmetry[^3]. For example, if $M=|\lambda
_1|>|\lambda
_{2,3,4}|$ then ${\textstyle{1\over 8}}$ of the supersymmetry is preserved, while for $M=|\lambda _1|=|\lambda _2|=|\lambda _3|=|\lambda _4|$, ${\textstyle{1\over
2}}$ of the supersymmetry is preserved.
The eigenvalues $\lambda _i$ are each invariant under $E_7$ and $SU(8)$. The physical quantities such as the Bekenstein-Hawking (BH) entropy and the ADM mass of BPS-saturated black holes, can then be written in terms of these quantities, which depend on both the bare charges $\cal Z$ and the asymptotic values of the scalar fields parameterised by ${\cal V}_{\infty}$. However, there are special combinations of these invariants for which the dependence on the asymptotic values of scalar fields drops out. In particular, such combinations play a special role in the BH entropy for the BPS-saturated black hole solutions as discussed in the following Subsection. Similar comments apply in $D>4$.
Bekenstein-Hawking Entropy
--------------------------
The BH entropy is defined as $S_{BH}= {\textstyle{1\over {4G_D}}} A_{h}$ where $G_D$ is the $D$-dimensional Newton’s constant and $A_{h}$ is the area of the horizon. Since the Einstein metric is duality invariant, geometrical quantities such as $A_{h}$ should be too. Thus it should be possible to write $A_{h}$ in terms of duality invariant quantities such as the eigenvalues $\lambda _i$, the ADM mass (or the “non-extremality” parameter $m$ defined in Appendix B) and the angular momentum components. However, in the case of BPS-saturated black holes the BH entropy is, in addition, [*independent* ]{} of the asymptotic values of scalar fields. This property was pointed out in [@LWI], and further exhibited for the general BPS-saturated solutions of toroidally compactified heterotic string vacua in $D=4$ [@CTII] and $D=5$ [@CY5r] as well as for certain BPS-saturated black holes of $D=4$ $N=2$ superstring vacua in [@FKS; @S]. This property was explained in terms of “supersymmetric attractors” in [@FK].
When applying the above arguments to BPS-saturated black holes of toroidally compactified Type II string (or M-theory) their BH entropy can be written in terms of a $U$-duality invariant combination of [*bare*]{} charges alone, thus implying that only a very special combination of bare charges can appear in the BH entropy formula. One is then led to the remarkable result that the entropy must be given in terms of the quartic invariant of $E_7$ in $D=4$ and the cubic invariant of $E_6$ in $D=5$, as these are the only possible $U$-invariants of bare charges.[^4] This fact was first pointed out in [@KK] and [@FK], respectively (and independently in [@CHU]). It has been checked explicitly for certain classes of $D=4$ BPS-saturated black holes [@KK]. Below we will extend the analysis to general BPS-saturated black holes in $D=4,5$.
For $6\le D\le 9$ there are [*no*]{} non-trivial $U$-invariant quantities that can be constructed from the bare charges alone, in agreement with the result that there are no BPS-saturated black holes with non-singular horizons and finite BH entropy in $D\ge 6$, as has been shown explicitly in [@KT; @CYNear].
### D=4
The five-parameter static generating solution has the following BH entropy [@CTII]: $$S_{BH}= 2\pi\sqrt{Q_1Q_2P_1P_2-{\textstyle{1\over 4}}q^2(P_1+P_2)^2}.
\label{4ent}$$ We shall now show that (\[4ent\]) can be rewritten in terms of the $E_7$ quartic invariant (of bare charges).
The quartic $E_{7(7)}$ invariant $J_4$, constructed from the charge matrix $Z_{4\,AB}$, is [@CJ]: $$\begin{aligned}
{J_4}&= &{\rm Tr}({ Z_4}^\dagger{ Z_4})^2-{\textstyle{1\over 4}}
({\rm Tr}Z_4^\dagger{
Z_4})^2+\cr
& &{\textstyle {1\over {96}}}(\epsilon_{ABCDEFGH}{ Z_4}^{AB\, *}{
Z_4}^{CD\, *}{
Z_4}^{EF\, *}
{ Z_4}^{GH\, *}+\epsilon^{ABCDEFGH}Z_{4\, AB}Z_{4\, CD} Z_{4\, EF} Z_{4\, GH})\
,
\label{quartic}\end{aligned}$$ which can be written in terms of the skew-eigenvalues $\lambda _i$ by substituting the skew-diagonalised matrix $Z^0_{4}$ (\[diag4\]) in (\[quartic\]) to give (as in [@KK]): $$\begin{aligned}
{J_4}&=&\sum_{i=1}^4 |\lambda_i|^4 -2\sum_{j>i=1}^4
|\lambda_i|^2|\lambda_j|^2\cr
&+&4({ \lambda}_1^*{\lambda}_2^*{\lambda}_3^*{ \lambda}_4^*+
\lambda_1\lambda_2\lambda_3\lambda_4)\ .
\label{diagfour}\end{aligned}$$ For the five parameter generating solution, the $\lambda _i$ are given by (\[fivepar\]), so that (\[diagfour\]) becomes: $$\begin{aligned}
{J_4}&=&16[(Q_{1R}^2-Q_{1L}^2)(P_{2R}^2-P_{2L}^2)-4P_{2R}^2q^2]\cr
&=&16^2[Q_1Q_2P_1P_2-{\textstyle{1\over
4}}q^2(P_1+P_2)^2] \ .
\label{fourfivepar}\end{aligned}$$ Comparing with (\[4ent\]), we learn that for the five-parameter generating solution the BH entropy is given by $$S_{BH}={\pi\over 8}\sqrt{J_4}.
\label{4Jent}$$ This result generalises the one in [@KK], where the result for the four-parameter solution with $q=0$ was established.
Acting on the generating solution with $SU(8)$ transformations to generate the general charged black hole, and then with a $E_7$ transformation to generate the solution with general asymptotic values of scalar fields, leaves the BH entropy (\[4Jent\]) invariant, since $J_4$ is invariant. As the dressing of the charges is by an $E_7$ transformation, i.e. ${\bar {\cal Z}}
={{\cal V}_\infty {\cal Z}}$, the dependence on the asymptotic values of scalar fields ${\cal V}_\infty$ drops out of the BH entropy, which thus can be written in terms of the bare charges alone, as expected.
### D=5
The BH entropy of the three-parameter static BPS-saturated generating solution is [@VS; @TMpl; @CY5r]: $$S_{BH}= 2\pi\sqrt{Q_1Q_2\tilde Q} \ .
\label{5ent}$$ The cubic $E_{6(6)}$ invariant $J_3$ constructed from the charge matrix $Z_{5\,AB}$ is [@Cremmer]: $${J_3}=-\sum_{A,B,C,D,E,F=1}^{8}\Omega^{AB} Z_{5\,BC} \Omega^{CD} Z_{5\,DE}
\Omega^{EF} Z_{5\,FA} \ .
\label{cubic}$$ By transforming $Z_5$ to a skew-diagonal matrix $Z^0_{5}$, given in terms of the four constrained real eigenvalues $\lambda_i$ (\[threepar\]), and using (\[explo\]) for $\Omega_{AB}$, the cubic form $J_3$ can be written as $${J_3}=2\sum_{i=1}^4 \lambda_i^3.
\label{diagfive}$$ For the (three charge parameter) generating solution the eigenvalues are (\[threepar\]) so that (\[diagfive\]) is: $$\begin{aligned}
{J_3}&=&12(Q_{1R}^2-Q_{1L}^2){\tilde Q}\cr
&=&48Q_1Q_2\tilde Q,
\label{Jfivegen}
\end{aligned}$$ which together with (\[5ent\]) implies $$S_{BH}=\pi\sqrt{{\textstyle{1\over {12}}}J_3}.
\label{5Jent}$$ This result gives the entropy in terms of the cubic invariant for the generating solution, and so, by $U$-invariance, for all charged static BPS-saturated $D=5$ black holes, as conjectured in [@FK; @Pol; @CHU].
### $6 \le D \le 9$
A BPS-saturated black hole in $D\ge 6$ dimensions should have a horizon area that is an invariant under $U$-duality constructed from the bare charges alone, involving no scalars. This would involve, for example, constructing an $SO(5,5)$ singlet from tensor products of charges transforming as a chiral spinor [**16**]{} of $SO(5,5)$ in $D=6$ and constructing singlets of $SL(5,{\bf R})$ from tensor products of charges transforming as a [**10**]{}. There are no such non-trivial singlets in $D\ge 6$, so that the only invariant result for the area is zero, which is precisely what is found (see Appendix B). Indeed, the generating solution for BPS-saturated solution in $D\ge 6$ has [*zero*]{} horizon area [@CYNear] and so zero BH entropy.
### Entropy of Non-Extreme Black Holes
We now comment on the BH entropy of non-extreme black holes, in particular, static black holes in $6\le D\le 9$. (For another approach to address the BH entropy of non-extreme black holes, where additional auxiliary charges are introduced, see [@HMS; @KR].) The non-extreme generating solutions are specified in terms of two electric charges $Q_{1,2}$ and a parameter $m$ which measures the deviation from extremality, i.e. the BPS-saturated limit is reached when $m=0$ while the charges $Q_{1,2}$ are kept constant (see Appendix B). The BH entropy is given by (\[DBHent\]) in Appendix B, which in the near-BPS-saturated limit ($Q_{1,2}\gg m$) reduces to the form (\[DBPSBHent\]): $$S_{BH}=4\pi\sqrt{{\textstyle{1\over{(D-3)^2}}}Q_1Q_2{(2m)}^{2\over{D-3}}}.
\label{Dent}$$
The BH entropy of the non-extreme black holes can also be rewritten in a manifestly duality-invariant manner. We demonstrate this for static near-extreme black holes in $D=7$; examples of such black holes in other dimensions are similar. The $5 \times 5$ matrix of dressed charges $Z_{7ij}$ transforms as an antisymmetric tensor under the local $SO(5)$ symmetry but is invariant under the rigid $SL(5,{\bf R})$ duality symmetry. (The central charges are given by the $4
\times 4$ matrix $Z_{7ij} \gamma^{ij}$ where $ \gamma^{ij}$ are the generators of $SO(5)$ in the 4-dimensional spinor representation.) The two skew eigenvalues $\lambda_i$ of $Z_{7ij} $ are given by $\lambda_i=Q_{R,L} $ and these are invariant under $SO(5) \times SL(5,{\bf R})$. The BH entropy in the near-extreme case is then $$S_{BH}= {{\pi}\over 2}\sqrt{ (|\lambda_1|^2- |\lambda_2|^2){(2m)}^{1\over{2}}}\
{},
\label{Denta}$$ which can be rewritten in a manifestly $U$-duality invariant form as $$S_{BH}= {{\pi}\over 2} \left(
[2tr(Y_7^2)-{\textstyle {1\over 2}}(tr Y_7)^2]m \right )^{1\over 4} \ .
\label{Dentb}$$ Here $Y_7\equiv Z_7^tZ_7$ and we have used the relationship: $tr(Y_7^m)=2
(-|\lambda_1|^2)^{m}+(-|\lambda_2|^2)^{m}$. Note that now the entropy does depend on the asymptotic values of the scalar fields.
ADM Masses and Supersymmetry Breaking
-------------------------------------
We now comment on a $U$-duality invariant form of the black hole mass formula for BPS-saturated black holes with different numbers of preserved supersymmetries. We shall derive such expressions in $D=4$. Examples in other dimensions are similar.
As discussed in the beginning of this Section, in $D=4$ the BPS-saturated black holes will preserve ${p\over 8}$ of the $N=8$ supersymmetry if the BPS-saturated ADM mass $M$ is equal to $|\lambda _1|=\cdots =|\lambda_p|$ where $\lambda _i$ ($i=1,\cdots, 4$) are eigenvalues (\[diag4\]) of central charge matrix $Z_4$. Without loss of generality one can order the eigenvalues in such a way that $|\lambda_i|\ge
|\lambda_j|$ for $j\ge i$ . Note also that from (\[diag4\]) $$tr(Y_4^m)= 2\sum _{i=1}^4(-|\lambda_i|^2)^{m}\ ,$$ where $Y_4\equiv Z_4^\dagger Z_4$ and $m=1,\cdots, 5-p$.
### p=4
These solutions preserve ${1\over 2}$ of $N=8$ supersymmetry and $$M=|\lambda_1|=|\lambda_2|=|\lambda _3|=|\lambda_4|.$$ Examples of such solutions are obtained from the generating solution with only one non-zero charge, e.g., those with only $Q_1\ne 0$. The mass can be written in a $U$-invariant form as $$M=-{\textstyle{1\over 8}}trY_4\ .$$
### p=3
These solutions preserve ${3\over 8}$ of $N=8$ supersymmetry thus $$M=|\lambda_1|=|\lambda_2|=|\lambda _3|>|\lambda_4|\ .$$ An example of such a generating solution corresponds to the case with $(Q_1,
Q_2, P_1=P_2)\ne 0$, while non-zero $q$ is determined in terms of other nonzero charges as $q={\textstyle{1\over
2}}[(Q_{1R}+P_{2R})^2-Q_{1L}^2]^{1/2}$. The $U$-duality invariant form of the BPS-saturated mass can now be written in terms of two invariants: $$tr Y_4=-6|\lambda _1|^2-2|\lambda _4|^2, \qquad
tr
(Y_4^2)=
6|\lambda _1|^4+2|\lambda _4|^4$$ as the (larger) root of the quadratic equation: $$48 M^4 +12 tr Y_4 M^2 + (tr Y_4)^2-2 tr(Y_4^2)=0 \ ,$$ $$M^2=-{\textstyle{1\over 8}} tr Y_4 + \sqrt{ {\textstyle{1\over {24}} tr (Y_4^2)
- {\textstyle{1\over {192}}}(tr Y_4)^2}}\ .$$
### p=2
These solutions preserve ${1\over 4}$ of $N=8$ supersymmetry and have the mass $$M=|\lambda_1|=|\lambda_2|>|\lambda _3|\ge|\lambda_4| \ .$$ An example of such a generating solution is the case with only $(Q_1,Q_2)\ne 0$. The general mass can be written in terms of the three invariants $tr (Y_4^m)$ for $m=1,2,3$. The $U$-duality invariant expression for the BPS-saturated mass formula is then given by the (largest) root of a cubic equation; we do not give it explicitly here. Some simplification is obtained if $|\lambda _3|=|\lambda_4|$, in which case only two invariants and a quadratic equation are needed.
### p=1
These solutions preserve ${1\over 8}$ of the $N=8$ supersymmetry and the mass is $$M=|\lambda_1|> |\lambda_2|\ge|\lambda _3|\ge|\lambda_4|\ .$$ Examples of such generating solutions are the case with only $(Q_1,Q_2,P_1)\ne 0$ and the generating solution with all the five charges non-zero is also in this class. The $U$-invariant mass can be written in terms of the four invariants $tr
(Y_4^m)$ for $m=1,2,3,4$ and involves the (largest) root of a quartic equation so we do not give it explicitly here.
We would like to thank K. Chan, S. Gubser, F. Larsen, A. Sen, A. Tseytlin and D. Youm for useful discussions. The work is supported by the Institute for Advanced Study funds and J. Seward Johnson foundation (M.C.), U.S. DOE Grant No. DOE-EY-76-02-3071 (M.C.), the NATO collaborative research grant CGR No. 940870 (M.C.) and the National Science Foundation Career Advancement Award No. PHY95-12732 (M.C.). The authors acknowledge the hospitality of the Institute for Theoretical Physics at the University of California, Santa Barbara, where the work was initiated, the hospitality of the Department of Applied Mathematics and Theoretical Physics of Cambridge (M.C.) and of the CERN Theory Division (M.C.).
2.mm
Appendix A: Effective Action of the NS-NS Sector of Type II String on Tori
==========================================================================
For the sake of completeness we briefly summarize the form of the effective action of the NS-NS sector for the toroidally compactified Type II string in $D$-dimensions (see, e.g., [@MS].). The notation used is that of [@CY4r].
The compactification of the $(10-D)$-spatial coordinates on a $(10-D)$-torus is achieved by choosing the following abelian Kaluza-Klein Ansatz for the ten-dimensional metric $$\hat{G}_{MN}=\left(\matrix{e^{a\varphi}g_{{\mu}{\nu}}+
G_{{m}{n}}A^{(1)\,m}_{{\mu}}A^{(1)\,n}_{{\nu}} & A^{(1)\,m}_{{\mu}}
G_{{m}{n}} \cr A^{(1)\,n}_{{\nu}}G_{{m}{n}} & G_{{m}{n}}}\right),
\label{4dkk}$$ where $A^{(1)\,m}_{\mu}$ ($\mu = 0,1,...,D-1$; $m=1,...,10-D$) are $D$-dimensional Kaluza-Klein $U(1)$ gauge fields, $\varphi \equiv \hat{\Phi} - {1\over 2}{\rm ln}\,{\rm det}\, G_{mn}$ is the $D$-dimensional dilaton field, and $a\equiv {2\over{D-2}}$. Then, the affective action is specified by the following massless bosonic fields: the (Einstein-frame) graviton $g_{\mu\nu}$, the dilaton $e^{\varphi}$, $(20-2D)$ $U(1)$ gauge fields ${\cal A}^i_{\mu}
\equiv (A^{(1)\,m}_{\mu},A^{(2)}_{\mu\,m})$ defined as $A^{(2)}_{\mu\,m} \equiv \hat{B}_{\mu m}+\hat{B}_{mn}
A^{(1)\,n}_{\mu}$, and the following symmetric $O(10-D,10-D)$ matrix of the scalar fields (moduli): $$M=\left ( \matrix{G^{-1} & -G^{-1}C \cr
-C^T G^{-1} & G + C^T G^{-1}C}
\right ),
\label{modulthree}$$ where $G \equiv [\hat{G}_{mn}]$, $C \equiv [\hat{B}_{mn}]$ and are defined in terms of the internal parts of ten-dimensional fields. Then the NS-NS sector of the $D$-dimensional effective action takes the form: $$\begin{aligned}
{\cal L}&=&{1\over{16\pi G_D}}\sqrt{-g}[{\cal R}_g-{1\over (D-2)}
g^{\mu\nu}\partial_{\mu}\varphi\partial_{\nu}\varphi+{1\over 8}
g^{\mu\nu}{\rm Tr}(\partial_{\mu}ML\partial_{\nu}ML)\cr&-&{1\over{12}}
e^{-2a\varphi}g^{\mu\mu^{\prime}}g^{\nu\nu^{\prime}}
g^{\rho\rho^{\prime}}H_{\mu\nu\rho}H_{\mu^{\prime}\nu^{\prime}
\rho^{\prime}}
-{1\over 4}e^{-a\varphi}g^{\mu\mu^{\prime}}g^{\nu\nu^{\prime}}
{\cal F}^{i}_{\mu\nu}(LML)_{ij}
{\cal F}^{j}_{\mu^{\prime}\nu^{\prime}}],
\label{effaction}\end{aligned}$$ where $g\equiv {\rm det}\,g_{\mu\nu}$, ${\cal R}_g$ is the Ricci scalar of $g_{\mu\nu}$, and ${\cal F}^i_{\mu\nu} = \partial_{\mu}
{\cal A}^i_{\nu}-\partial_{\nu} {\cal A}^i_{\mu}$ are the $U(1)^{20-2D}$ gauge field strengths and $H_{\mu\nu\rho} \equiv
(\partial_{\mu}B_{\nu\rho}-{1\over 2}{\cal A}^i_{\mu}L_{ij}
{\cal F}^j_{\nu\rho}) + {\rm cyc.\ perms.\ of}\ \mu , \nu , \rho$ is the field strength of the two-form field $B_{\mu\nu}$.
The $D$-dimensional effective action (\[effaction\]) is invariant under the $O(10-D,10-D)$ transformations ($T$-duality): $$M \to \Omega M \Omega^T ,\ \ \ {\cal A}^i_{\mu} \to \Omega_{ij}
{\cal A}^j_{\mu}, \ \ \ g_{\mu\nu} \to g_{\mu\nu}, \ \ \
\varphi \to \varphi, \ \ \ B_{\mu\nu} \to B_{\mu\nu},
\label{tdual}$$ where $\Omega$ is an $O(10-D,10-D)$ invariant matrix, [*i.e.*]{}, with the following property: $$\Omega^T L \Omega = L ,\ \ \ L =\left ( \matrix{0 & I_{10-D}\cr
I_{10-D} & 0 } \right ),
\label{4dL}$$ where $I_n$ denotes the $n\times n$ identity matrix.
In $D=4$ the field strength of the abelian gauge field is self-dual, i.e. $\tilde{\cal F}^{i\,\mu\nu} = {1\over 2\sqrt{-g}}
\varepsilon^{\mu\nu\rho\sigma}{\cal F}^i_{\rho\sigma}$, and thus the charged solutions are specified by the electric and magnetic charges.
In $D=5$ the effective action is specified by the graviton, 26 scalar fields (25 moduli fields in the matrix $M$ and the dilaton $\varphi$), 10 $U(1)$ gauge fields, and the field strength $H_{\mu\nu\rho}$ of the two-form field $B_{\mu\nu}$. By the duality transformation $
H^{\mu\nu\rho}=-{e^{4\varphi/3}\over{2!\sqrt{-g}}}
\varepsilon^{\mu\nu\rho\lambda\sigma}{\tilde F}_{\lambda\sigma},$ $H_{\mu\nu\rho}$ can be related to the field strength $\tilde F_{\mu\nu}$ of the gauge field $ \tilde A_\mu$, which specifies an additional electric charge $\tilde Q$.
In $D\ge 6$ there the allowed charges are only electric charges associated with the $(20-2D)$ NS-NS sector abelian gauge fields.
Appendix B: Static Generating Solutions
=======================================
For the sake of simplicity we present explicitly only the explicit solution for the non-extreme [*static*]{} generating solution in $D=4$, $D=5$, $6\le D \le 9$ and with four[@CYI; @HLMS], three [@CY5r; @HMS] and two charge[@Peet] parameters of the NS-NS sector, respectively. Note that the full generating solution in $D=4$ is parameterised by [*five*]{} charge parameters. For the explicit form of the rotating generating solution in $D=5$ see [@CY5r] and in $6\le D\le 9$ see [@CYNear; @Llatas], while in $D=4$ the four charge parameter rotating solution is given in [@CY4r] and the five charge static solution is given in [@CY5r]. The parameterisation there is given in terms of the “toroidal” sector of toroidally compactified heterotic string.
We choose to parameterise the generating solutions in terms of the mass $m$ of the $D$-dimensional Schwarzschild solution, and the boost parameters $\delta_i$, specifying the charges of the solution. The notation used is similar to that in [@CTIII].
$D=4$-Four Charge Static Solution
---------------------------------
The expression for the non-extreme dyonic rotating black hole solution in terms of the (non-trivial) four-dimensional bosonic fields is of the following form [^5]: $$\begin{aligned}
ds^2_{E}&=& -\lambda fdt^2+\lambda^{-1}[f^{-1}dr^2+r^2d\Omega_2^2
],\cr
G_{11}&=&{{T_1}\over{T_2}}, \ \
G_{22}={{F_2}\over {F_1}}, \ \
e^{2\varphi}={{F_1F_2}\over{T_1T_2}},
\label{4dsol}\end{aligned}$$ where $ds^2_{E}$ specifies the Einstein-frame ($D$-dimensional) space-time line element, $G_{ij}$ correspond to the internal toroidal metric coefficients and $\varphi$ is the $D$-dimensional dilaton field (see Appendix A). Other scalar fields are constant and assume canonical values (one or zero). Here $$f=1-{{2m}\over r}, \ \ \lambda=(T_1T_2F_1F_2)^{-{1\over 2}}
\label{4fl}$$ and $$\begin{aligned}
T_1&=&1+{{2m{\rm sinh}^2 \delta_{e1}}\over r},\ \
T_2=1+{{2m{\rm sinh}^2 \delta_{e2}}\over r}, \cr
F_1&=&1+{{2m{\rm sinh}^2 \delta_{m1}}\over r}, \ \
F_2=1+{{2m{\rm sinh}^2 \delta_{m2}}\over r},
\label{4dpar}\end{aligned}$$ The ADM mass, and four $U(1)$ charges $Q_1,Q_2,P_1,
P_2$, associated with the respective gauge fields $A^{(1)}_{1\mu},A^{(2)}_{1\mu},A^{(1)}_{2\mu},A^{(2)}_{2\mu}$, can be expressed in terms of $m$, and four boosts $\delta_{e1,e2,m1,m2}$ in the following way: $$\begin{aligned}
M&=&4m({\rm cosh}^2 \delta_{e1}+{\rm cosh}^2 \delta_{e2}+
{\rm cosh}^2 \delta_{m1}+{\rm cosh}^2 \delta_{m2})-8m, \cr
Q_1&=&4m{\rm cosh}\delta_{e1}{\rm sinh}\delta_{e1},\ \ \ \ \
Q_2= 4m{\rm cosh}\delta_{e2}{\rm sinh}\delta_{e2}, \cr
P_1&=&4m{\rm cosh}\delta_{m1}{\rm sinh}\delta_{m1},\ \ \ \ \
P_2 =4m{\rm cosh}\delta_{m2}{\rm sinh}\delta_{m2}.
\label{4dphys}\end{aligned}$$ The BH entropy is of the form[@CYI; @HLMS]: $$S_{BH}=2\pi
(4m)^2\cosh\delta_{e1}\cosh\delta_{e2}\cosh\delta_{m1}\cosh\delta_{m2},
\label{4BHent}$$ which in the BPS-saturated limit ($m\to 0$, $\delta_{e1,e2,m1,m2}\to
\infty$ while keeping $Q_{1,2}$, $P_{1,2}$ finite) reduces to the form [@CY]: $$S_{BH}=2\pi\sqrt{Q_1Q_2P_1P_2}.
\label{4BPSBHent}$$ In the case of the fifth charge parameter $q$ added, the BH entropy of the BPS-saturated black holes becomes [@CTII]: $$S_{BH}=2\pi\sqrt{Q_1Q_2P_1P_2-{\textstyle{1\over 4}}q^2(P_1+P_2)^2}.
\label{4BPSBH5p}$$
$D=5$-Three Charge Static Solution
----------------------------------
The expression for the non-extreme dyonic rotating black hole solution in terms of the (non-trivial) five-dimensional bosonic fields is of the following form [@CY5r; @HMS][^6]: $$\begin{aligned}
ds^2_{E}&=& -\lambda^2 fdt^2+\lambda^{-1}[f^{-1}dr^2+r^2d\Omega_3^2
],\cr
G_{11}&=&{{T_1}\over{T_2}}, \ \
e^{2\varphi}={\tilde T^2\over{T_1T_2}},
\label{5dsol}\end{aligned}$$ with other scalars assuming constant canonical values. Here $$f=1-{{2m}\over r^2}, \ \ \lambda=(T_1T_2\tilde T)^{-{1\over 3}},
\label{5fl}$$ and $$\begin{aligned}
T_1&=&1+{{2m{\rm sinh}^2 \delta_{e1}}\over r^2}, \ \
T_2=1+{{2m{\rm sinh}^2 \delta_{e2}}\over r^2}, \ \
\tilde T=1+{{2m{\rm sinh}^2 \delta_{\tilde e}}\over r^2} \ .
\label{5dpar}\end{aligned}$$ The ADM mass, three charges $Q_{1,2},\tilde Q$ associated with respective gauge fields $A_{1\mu}^{(1)}, \ A_{1\mu}^{(2)}$ and ${\tilde A}_\mu$ (the gauge field related to the two from field $B_{\mu\nu}$ by a duality transformation), are expressed in terms of $m$, and three boosts $\delta_{e1,e2,\tilde e}$ in the following way: $$\begin{aligned}
M&=&2m({\rm cosh}^2 \delta_{e1}+{\rm cosh}^2 \delta_{e2}+
{\rm cosh}^2 \delta_{\tilde e})-3m, \cr
Q_1&=&2m{\rm cosh}\delta_{e1}{\rm sinh}\delta_{e1},\ \
Q_2= 2m{\rm cosh}\delta_{e2}{\rm sinh}\delta_{e2},\ \
{\tilde Q}=2m{\rm cosh}\delta_{\tilde e}{\rm sinh}\delta_{\tilde e}.
\label{5dphys}\end{aligned}$$ The BH entropy is of the form [@HMS]: $$S_{BH}=2\pi (2m)^{3\over 2}\cosh\delta_{e1}\cosh\delta_{e2}\cosh\delta_{\tilde
e}
\label{5BHent},$$ which in the BPS-saturated limit ($m\to 0$, $\delta_{e1,e2,\tilde e}\to
\infty$ with $Q_{1,2}, \ \tilde Q$ finite) reduces to the form [@VS; @TMpl]: $$S_{BH}=2\pi\sqrt{Q_1Q_2\tilde Q}.
\label{5BPSBHent}$$
$6\le D\le 9$-Two Charge Static Solution
----------------------------------------
The expression for the non-extreme dyonic rotating black hole solution in terms of the (non-trivial) five-dimensional bosonic fields is of the following form [@Peet; @HSen; @CYNear; @Llatas][^7]: $$\begin{aligned}
ds^2_{E}&=& -\lambda^{D-3} fdt^2+\lambda^{-1}[f^{-1}dr^2+r^2d\Omega_{D-2}^2
],\cr
G_{11}&=&{{T_1}\over{T_2}}, \ \
e^{2\varphi}={1\over{T_1T_2}},
\label{Ddsol}\end{aligned}$$ while other scalar fields are constant and assume canonical values. Here $$f=1-{{2m}\over r^{D-3}}, \ \ \lambda=(T_1T_2)^{1\over{D-2}},
\label{Dfl}$$ and $$T_1=1+{{2m{\rm sinh}^2 \delta_{e1}}\over r^{D-3}}, \ \
T_2=1+{{2m{\rm sinh}^2 \delta_{e2}}\over r^{D-3}}.
\label{Ddpar}$$ The ADM mass, $U(1)$ charges $Q_1,Q_2$ associated with respective $A_{1\mu}^{(1)}, \ A_{2\mu}^{(2)}$ gauge fields, are expressed in terms of $m$, and two boosts $\delta_{e1,e2}$ in the following way: $$\begin{aligned}
M&=&{{\omega_{D-2}m}\over{8\pi G_D}}[ (D-3)({\rm cosh}^2( \delta_{e1}+{\rm
cosh}^2 \delta_{e2})-(D-4)], \cr
Q_1&=&{{\omega_{D-2}m}\over{8\pi G_D}}(D-3){\rm cosh}\delta_{e1}{\rm
sinh}\delta_{e1},\
Q_2={{\omega_{D-2}m}\over{8\pi G_D}}(D-3){\rm cosh}\delta_{e2}{\rm
sinh}\delta_{e2},\ \
\label{Ddphys}\end{aligned}$$ where $\omega_{D-2}={{2\pi^{{D-1}\over 2}}/ {\Gamma({{D-1}\over 2})}}$.
The BH entropy is of the form: $$S_{BH}={{\omega_{D-2}}\over {2G_D}}
m^{{D-2}\over{D-3}}\cosh\delta_{e1}\cosh\delta_{e2},
\label{DBHent}$$ which in the near-BPS-saturated limit ($Q_{1,2}\gg m$) reduces to the form: $$S_{BH}=4\pi\sqrt{{\textstyle{1\over{(D-3)^2}}}Q_1Q_2{(2m)}^{2\over{D-3}}}.
\label{DBPSBHent}$$
--------------------------------------------------------------------------------------------------------
D Classical Duality -${ U}$ Maximal Compact Subgroup-${ C_U}$ Quantum Duality-${Q_U}$
--- -------------------------------------- ----------------------------------- -------------------------
4 $E_{{7(7)}}$ $SU(8)$ $E_{{7(7)}}({\bf Z})$
5 $E_{{6(6)}}$ $USp(8)$ $E_{{6(6)}}({\bf Z})$
6 $SO(5,5)$ $SO(5)\times SO(5)$ $SO(5,5;{\bf Z})$
7 $SL(5,{\bf R})$ $SO(5)$ $SL(5, {\bf Z})$
8 $SL(3,{\bf R})\times SL(2, {\bf R})$ $SO(3)\times U(1)$ $SL(3, {\bf
Z})\times
SL(2,{\bf Z})$
9 $SL(2,{\bf R})\times {\bf R}^+$ $ U(1)$ $SL(2, {\bf Z}) $
--------------------------------------------------------------------------------------------------------
: The classical and quantum duality symmetries [@HT] for toroidally compactified Type II string in $ 4\le D\le 9$.
[^1]: On sabbatic leave from the University of Pennsylvania.
[^2]: For the four charges $Q_{1,2}$ and $P_{1,2}$ the eigenvalues were given in [@KK].
[^3]: For a related discussion of the number of preserved supersymmetries see [@Kal; @KK].
[^4]: In $D=3$ there is a unique quintic $E_{8(8)}$-invariant which should play a similar role for $D=3$ black hole solutions.
[^5]: The four-dimensional Newton’s constant is taken to be $G_N^{D=4}={1\over 8}$ and we follow the convention of [@MP], for the definitions of the ADM mass, charges, dipole moments and angular momenta.
[^6]: The five-dimensional Newton’s constant is taken to be $G_N^{D=5}={{2\pi}\over 8}$.
[^7]: The $D$-dimensional Newton’s constant is taken to be $G_N^{D}={{(2\pi)^{D-4}}\over 8}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present finite-element solutions of the Laplace equation for the silicon-based trench-isolated double quantum-dot and the capacitively-coupled single-electron transistor device architecture. This system is a candidate for charge and spin-based quantum computation in the solid state, as demonstrated by recent coherent-charge oscillation experiments. Our key findings demonstrate control of the electric potential and electric field in the vicinity of the double quantum-dot by the electric potential applied to the in-plane gates. This constitutes a useful theoretical analysis of the silicon-based architecture for quantum information processing applications.'
author:
- 'S. Rahman$^1$'
- 'J. Gorman$^1$'
- 'C. H. W. Barnes$^1$'
- 'D. A. Williams$^2$'
- 'H. P. Langtangen$^3$'
title: A finite element analysis of a silicon based double quantum dot structure
---
Recent experiments conducted on trench-isolated double quantum-dot (IDQD) structures have successfully demonstrated detection of single-electron polarization,[@Emiroglu:03] and coherent-charge oscillation.[@Gorman:05] This highlights the possibility of constructing charge-based quantum computer circuits in Si, with coherence times of the order $100$ ns.[@NielsonBook; @Loss:98] The architecture for a single qubit device is a complex, three-dimensional structure consisting of a single-electron transistor (SET), an IDQD, and gate electrodes. This makes it difficult to determine theoretically the system evolution by means of a complete and self-consistent Schrödinger-Poisson analysis. This is particularly the case if the analysis were to fully take into account the device geometry and all interactions while performing quantum manipulations within the coherence time of the qubit.[@ft1]
In this paper, we present a significant contribution to such an analysis by a finite-element solution of the Laplace equation with the three-dimensional device geometry and material composition taken into account. The aim of this work is to demonstrate the electrostatic effect on the IDQD structure when voltages are applied to the in-plane control gates of the device.
Figures \[fig:schematic\](a) and \[fig:schematic\](b) show device schematics in the $x$-$y$ and $x$-$z$-planes, respectively. The trench isolation, illustrated in Fig. \[fig:schematic\](b), is formed by high-resolution electron-beam lithography and reactive-ion etching. Each trench is approximately $150$ nm deep and runs into the buried-oxide (BOX) layer of the silicon-on-insulator (SOI) wafer. The active regions of the device elements are P doped Si, which are electrically isolated from other device elements, as seen in Fig. \[fig:schematic\]. In this work, we use rectangular approximations to the device elements and the etched profiles to simplify the analysis.[@ft2]
The small dimensions of the quantum dots and the $20$ nm constriction between them, which is fully depleted and acts as a tunable inter-dot tunnel barrier, result in a significant double-well type confinement potential for the electrons that occupy discrete quantum states on each quantum dot.[@Fujisawa:98; @vanDerWiel:02] The tunable inter-dot coupling causes the wave functions in the two dots to overlap and hybridize so that they may be thought of as pseudo-molecular states of an artificial two atom molecule. The IDQD is an electrically-isolated component that is coupled only capacitively to the rest of the circuit, including the SET for read-out, and the in-plane control gates (G1 to G3) for manipulation. Voltages applied to the gates G1 to G3 are used to tune the electric field in the vicinity of the IDQD, and thus, the confinement potential asymmetry and inter-dot tunnelling. Hence, an electron initially localized on one quantum dot may be allowed to tunnel to the opposite quantum dot by such manipulation.
An electric field is induced on the SET as a result of this polarization process. This modulates the chemical potential of the SET, and, therefore, the conductance through the source and drain leads under a finite SET bias condition. To maximize the change in conductance, the SET is initially tuned to the charge-sensitive regime by an appropriate voltage bias at G4, and under a small source-drain bias to ensure operation in the linear transport regime. Such manipulation of the device has only been shown through experiment so far. Therefore, a thorough theoretical analysis is necessary to complement the recent experimental findings, and build a more comprehensive understanding of the physical mechanisms involved.
Analytic methods exist for calculating the electrostatic potential in two-dimensional electron gases generated by patterned surface gates on GaAs/AlGaAs heterostructures.[@Davies:94; @Davies:95] While these analytic methods yield useful results for such devices, they are unsuitable for trench-isolated Si structures, where the geometry is much more sophisticated. Therefore, numerical methods have to be implemented. The finite-element method is a well-suited means for simulation of geometrically-complicated domains,[@Cook; @Zienkiewicz] and is commonly used to solve Poisson-type equations.[@Iserles; @DpBook2]
In order to determine the electric field throughout the modelled device regions, the numerical solution to the Laplace equation in three-dimensions is performed:
$$\begin{aligned}
\label{eqn:poisson}
\nabla \cdot [ \epsilon (\mathbf{x}) \nabla \phi (\mathbf{x}) ] &=& 0, \qquad \mathbf{x} \in \Omega \in \mathbb{R}^3, \label{eqn:lap1} \\
\phi(\mathbf{x}) &=& D_i, \qquad \mathbf{x} \in \partial \Omega_{D_i}, \label{eqn:lap2}\\
- \epsilon (\mathbf{x}) \frac{\partial \phi }{\partial n} &=&
q,\qquad \mathbf{x} \in \partial \Omega_{N}, \label{eqn:l3}\end{aligned}$$
where $\phi (\mathbf{x})$ is the electrostatic potential and $\epsilon (\mathbf{x})$ is the material dielectric parameter. The dielectric parameter varies discontinuously on moving through the different materials - air ($\epsilon_0 = 8.85 \cdot 10^{-12}
~$Fm$^{-1}$), Si ($11.0 \epsilon_0$) and SiO$_2$ ($4.5 \epsilon_0
$). We apply Dirichlet boundary conditions to the surfaces of the metal gates and to the grounded base of the device, and the Neumann boundary condition to the exposed surfaces. (We apply Neumann boundary conditions with $q=0$, but for generality, we include in our discussion the possibility of non-zero $q$).
The finite element solution of the Laplace equation is well covered in the literature.[@Iserles; @DpBook2; @Mohan] The basic idea of the finite element method is to approximate the unknown fields, for example $\phi$ in the Laplace equation above, by $\widetilde{\phi}$ which is a combination of linearly independent basis functions $N_j$
$$\label{eqn:fem4}
\widetilde{\phi}({{{\mbox{\boldmath $x$}}}}) = \sum_{j=1}^M \phi_j N_j({{{\mbox{\boldmath $x$}}}}) ,$$
where $M$ is the number of basis functions and $\phi_j$ are the expansion coefficients to be determined. In the finite element method, the computational domain $\Omega$ is divided into a number of elements, and the $N_j$ are chosen to be piecewise polynomials such that they are non-zero only in a ‘few’ adjacent elements.[@ft3; @Iserles] The method then requires the substitution of $\widetilde{\phi}$ into the Laplace equation, and the residual $R = \nabla \cdot [\epsilon \nabla \widetilde{\phi}
]$ to be orthogonal to the space spanned by a linearly independent set $\{ W_1,\ldots,W_n\}$. In our calculations, we implement the orthogonality through the weighted residual statement:
$$\begin{aligned}
\int_{\Omega} R W_i d\Omega = 0 , \qquad i=1,\hdots,M ,\end{aligned}$$
and Galerkin’s method i.e. set $W_i$ equal to the basis functions $N_i$ to obtain
$$\begin{aligned}
\label{eqn:fem1}
\int_{\Omega} N_i [
\frac{\partial}{\partial x} \epsilon \frac{\partial \widetilde{\phi}}{\partial
x}
+
\frac{\partial}{\partial y} \epsilon \frac{\partial \widetilde{\phi}}{\partial
y}
+
\frac{\partial}{\partial z} \epsilon \frac{\partial \widetilde{\phi}}{\partial
z}]
d\Omega = 0 {\thinspace . }\end{aligned}$$
Using integration by parts, we reduce the order of derivatives in Eq. (\[eqn:fem1\]);
$$\begin{aligned}
\label{fem2}
- \int_{\Omega}
\epsilon \frac{\partial \widetilde{\phi}}{\partial x} \frac{\partial N_i}{\partial x}
+
\epsilon \frac{\partial \widetilde{\phi}}{\partial y} \frac{\partial N_i}{\partial y}
+
\epsilon \frac{\partial \widetilde{\phi}}{\partial z} \frac{\partial N_i}{\partial z}
d\Omega
\nonumber \\ \nonumber\\
+\int_{\partial \Omega} N_i \epsilon \frac{\partial \widetilde{\phi}}{ \partial n} d\Gamma
=0 {\thinspace . }\end{aligned}$$
The weighted residual method and integration by parts leads to a natural mechanism for the incorporation of derivative boundary condition given by Eq. (\[eqn:l3\]) for the Laplace operator $\nabla \cdot [\epsilon \nabla \widetilde{\phi}]$. Hence, we obtain after expanding the approximation for $\widetilde{\phi}$
$$\begin{aligned}
\label{eqn:fem5}
\int_{\Omega}
\epsilon \frac{\partial N_i}{\partial x} \frac{\partial N_j}{\partial x} \phi_j
+
\epsilon \frac{\partial N_i}{\partial y} \frac{\partial N_j}{\partial y} \phi_j
+
\epsilon \frac{\partial N_i}{\partial z} \frac{\partial N_j}{\partial z} \phi_j
d\Omega
\nonumber\\ \nonumber\\
+ \int_{\partial \Omega} N_i q d\Gamma
=0 ,\end{aligned}$$
where summation is implied over repeated indices. The problem has now been reduced to one of matrix inversion;
$$\label{eqn:femMain}
K \Phi = F ,$$
where the ‘stiffness’ matrix $K$ is given by
$$\label{eqn:femK}
K_{ij} = \int_{\Omega}
\epsilon \frac{\partial N_i}{\partial x} \frac{\partial N_j}{\partial x}
+
\epsilon \frac{\partial N_i}{\partial y} \frac{\partial N_j}{\partial y}
+
\epsilon \frac{\partial N_i}{\partial z} \frac{\partial N_j}{\partial z}
d\Omega ,$$
and the RHS vector is given by
$$\label{eqn:femF}
F_i=-\int_{\partial \Omega} N_i q d\Gamma ,$$
and $\Phi$ is simply the vector of unknowns $\phi_j$. Dirichlet boundary conditions are implemented by forcing prescribed values of $\phi_j$.
In the formulation and solution of Eq. (\[eqn:femMain\]) we have chosen linear basis functions corresponding to 8-noded brick elements. This method therefore has a convergence rate for error of 2.0.[@ft4]
We have employed Gaussian quadrature in three-dimensions for the volume integrals and two-dimensions for surface integrals, and a conjugate gradient method with Incomplete Lower and Upper (ILU) factorization preconditioning to solve the resulting linear system of equations. Sufficient resolution was obtained by a mesh with $107\times 77\times 9$ nodes in the $x$, $y$ and $z$ directions, respectively. We implemented adaptive mesh refinement along $z$ axis to improve accuracy in the vicinity of the active region.
Figures \[fig:2dsolutions\](a) and \[fig:2dsolutions\](b) show cross-sectional slices along orthogonal planes of the full three-dimensional solution for the simulated electric potential. The simulation was performed with the following parameters: the gate potentials of G1, G2 and G3 are set to +$1$ V, $2$ V and -$1$ V, respectively; G4 is set to $-4.8$ V. The voltage chosen for G4 is approximately equal to that used for this gate in the experimental demonstrations in Ref. where single electron polarization of the IDQD was obtained for such a device. The choices of G1, G2 and G3, are also similar to those in the experiments but for this simulation, the exact values are chosen so that three-dimensional illustrations are as clear as possible.
Figures \[fig:2dsolutions\](a) and \[fig:2dsolutions\](b) clearly demonstrate the effect of the applied gate voltages on the potential landscape of the IDQD and the device as a whole; the result of applying a voltage on the in-plane gates is that a significant fraction of the applied voltage is induced on the IDQD, despite the etched trench gap. The abrupt change in the effective permittivity from the metallic gates to the voids, from the voids to the SiO$_2$, and from Si to SiO$_2$, causes some definition of the gates and the IDQD in the plots. The difference between the relative permittivity of air, Si and SiO$_2$ leads to a potential gradient, such that the absolute value of the potential is prone to vanish more rapidly in air, compared with Si and SiO$_2$. However, Fig. \[fig:2dsolutions\](b) clearly demonstrates that for this particular pillar height, which matches the device used in experiment, the potential at the IDQD is due mainly to the electric field vectors that are on a direct path through the trench isolation, and not the underlying substrate. This is consistent with experimental observations and is the preferred mechanism of device operation, since it is relatively easier in design and theoretical analysis, compared with the case where the majority of the electric field is through the semiconductor base and the field lines arrive at the IDQD from several different paths.
=9.0cm
For a more quantitative analysis, we determine the electric potential along the active region of the IDQD as a function of the applied gate voltages. This is shown in Figs. \[fig:xscurves1\](a) and \[fig:xscurves1\](b), where different gates are used to apply the in-plane electric field. Figure \[fig:xscurves1\](a) shows that the effect of varying the voltage applied to gate G2, from -$5$ V to +$5$ V, is to induce a voltage at the IDQD from -$2$ V to +$0.8$ V, respectively. Note that G4 had lowered the overall potential by approximately $0.55$ V in this case.
The data in Fig. \[fig:xscurves1\](a) shows a maximum change of $0.3$ V of the electrostatic potential at the IDQD, when the G2 gate voltage is raised or lowered by $1$ V. This field coupling factor of $\sim$$30~\%$ is approximately one order of magnitude greater than what was observed in experiment.[@Emiroglu:03] However, the measured quantity in experiments is the SET current, and the IDQD coupling terms are inferred from such measurements. The task of calculating such coefficients exactly as measured in experiment is beyond the scope of a purely electrostatic model, since, with the SET present, the global system that must be treated consists of interacting sub-systems of quantum mechanically bound electrons. Therefore, we project that a self-consistent Schrödinger-Poisson analysis of the system would yield results for the coupling coefficients that are closer to actual values observed in experiment.
Our results are also consistent with the experimental demonstrations of Ref. where the voltage on gate G2 is swept continuously from $-5$ V to $+5$ V, with G4 set to $\sim$$-4.8$ V, in order to demonstrate conductance resonances of the SET currents but also resonances due to the single electron polarization of the IDQD. (The device in Ref. had a slight asymmetry in the alignment of the IDQD relative to G2, hence the ability of G2 to polarize the IDQD.)
The abrupt changes in the potential at $115$ nm and $265$ nm are due to the change of relative permittivity at the air-SiO$_2$ interface. The difference between the potential gradient in the air and in the semiconductor regions is more evident in these figures. Figure \[fig:xscurves1\](b) shows the effect of applying voltages of opposite sign to gates G1 and G3. This results in a potential gradient across the IDQD, which has a maximum value of $\sim$$0.007$ Vnm$^{-1}$ in these simulations. This clearly demonstrates an effective mechanism for externally tuning the internal potential asymmetry of the IDQD electronic states.
In the experimental demonstrations of Ref. , a voltage bias is pulsed across in-plane metallic gates, which are placed perpendicularly to an IDQD as in our case, and this was shown to result in the coherent oscillation of a single electron charge present in an IDQD. This is again consistent with our results which suggests a strong electric field is induced at the IDQD due to the electric field at the in-plane gates.
In summary, we have successfully demonstrated, by means of finite-element solutions to the Laplace equation, that the electric potential and potential gradient across the confining region of the IDQD in trench-isolated Si devices may be manipulated effectively by the voltages applied to capacitively-coupled in-plane gates. Our calculations show good correlation with recent experimental demonstrations, where the IDQD electron states are manipulated by such methods.
We thank S. Pfaendler and R. Schumann for comments and useful discussions. SR, JG, and CB acknowledge the support of the EPSRC through the QIP IRC. SR acknowledges the Cambridge-MIT Institute for financial support.
E. G. Emiroglu, D. G. Hasko and D. A. Williams, Appl. Phys. Lett. [**83**]{}, 3942-3044 (2003).
J. Gorman, D. G. Hasko and D. A. Williams, Phys. Rev. Lett. [**95**]{}, 090502 (2005).
D. Loss and D. P. DiVincenzo, Phys.Rev. A [**57**]{}, 120-126 (1998).
M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, [*Cambridge University Press*]{}, (2003).
The most significant interaction is that between the SET and the IDQD because it directly influences the read-out operation and the degree of back-action.
The device elements are in fact roughly rectangular, so this approximation introduces only a small error to the results presented.
W. G. van der Wiel, S. De Francheshi, J. M. Elzerman, T. Fujisawa, S. Tarucha and L. P. Kouwenhoven, Reviews of Modern Physics, [**75**]{}, 1-22, (2003).
T. Fujisawa, T. H. Oosterkamp, W. G. van der Wiel, B. W. Broer, R. Aguado, S. Tarucha and L. P. Kouwenhoven, Science, [**282**]{}, 932-935, (1998).
J. H. Davies, I. A. Larkin and E. V. Sukhorukov J. Appl. Phys. [**77**]{}, 4504-4512 (1995).
J. H. Davies and I. A. Larkin, Phys. Rev. B [**49(7)**]{}, 4800-4809 (1994).
R. D. Cook, D. S. Malkus and M. E. Plesha, [*Concepts and Applications of Finite Element Analysis*]{}, Wiley, (1989).
O. Zienkiewicz, [*The Finite Element Method*]{}, McGraw-Hill, 4th edition (1994).
A. Iserles, [*A first course in the numerical analysis of differential equations*]{}, Cambridge University Press, (1995).
H. P. Langtangen, [*Computational Partial Differential Equations - Numerical methods and Diffpack Programming*]{}, Springer, 2nd edition (2003).
L. Ramdas Ram-Mohan, [*Finite Element and Boundary Element Applications in Quantum Mechanics*]{}, Oxford University Press, 1st edition, (2002).
The choice of piecewise polynomials results in a speed-up over spectral methods with little increase in error.
Higher-order basis functions were also implemented but had not resulted in any significant improvement in the accuracy of the results presented.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Traffic flow prediction, particularly in areas that experience highly dynamic flows such as motorways, is a major issue faced in traffic management. Due to increasingly large volumes of data sets being generated every minute, deep learning methods have been used extensively in the latest years for both short and long term prediction. However, such models, despite their efficiency, need large amounts of historical information to be provided, and they take a considerable amount of time and computing resources to train, validate and test. This paper presents two new spatial-temporal approaches for building accurate short-term prediction along a popular motorway in Sydney, by making use of the graph structure of the motorway network (including exits and entries). The methods are built on proximity-based approaches, denoted backtracking and interpolation, which uses the most recent and closest traffic flow information for each of the target counting stations along the motorway. The results indicate that for short-term predictions (less than 10 minutes into the future), the proposed graph-based approaches outperform state-of-the-art deep learning models, such as long-term short memory, convolutional neuronal networks or hybrid models.'
author:
- 'Adriana-Simona Mihaita$^{1}$, Zac Papachatgis$^{1}$ and Marian-Andrei Rizoiu$^{1}$ [^1]'
bibliography:
- 'IEEE\_ITSC\_2019.bib'
title: '**[Graph modelling approaches for motorway traffic flow prediction]{}**'
---
motorway flow predicting, graph-based prediction, backtracking, interpolation, deep learning analysis.
ACKNOWLEDGMENT {#acknowledgment .unnumbered}
==============
This work has been done as part of the ARC Linkage Project LP180100114. The authors are highly grateful for the support of Transport for NSW, Australia.
[^1]: $^{1}$Adriana-Simona Mihaita, Zac Papachatgis and Marian-Andrei Rizoiu are with the UTS Data Science Institute, University of Technology in Sydney, 81 Broadway Ultimo, NSW, Australia. Corresponding author’s contact: [[[email protected]]([email protected]).]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The distribution of a time integral of geometric Brownian motion is not well understood. To price an Asian option and to obtain measures of its dependence on the parameters of time, strike price, and underlying market price, it is essential to have the distribution of time integral of geometric Brownian motion and it is also required to have a way to manipulate its distribution. We present integral forms for key quantities in the price of Asian option and its derivatives ([*[delta, gamma,theta, and vega]{}*]{}). For example for any $a>0$ $\mathbb{E} \left[ (A_t -a)^+\right] = t -a + a^{2} \,\mathbb{E} \left[ (a+A_t)^{-1} \exp ( \frac{ 2M_t}{a+ A_t} - \frac{2}{a} ) \right]$, where $A_t = \int^t_0 \exp (B_s -s/2)\, ds$ and $M_t =\exp (B_t -t/2).$ [Keywords]{}: Asian option, Derivatives of option prices, Geometric Brownian Motion, Time integral.[Subject Classification]{}: Primary 91B28, 60J65 Secondary 60G99.'
address:
- |
Mathematics Department\
Florida State University\
Tallahassee, FL 32306
- |
Mathematics Department\
Florida State University\
Tallahassee, FL 32306
author:
- Jungmin Choi
- Kyounghee Kim
title: The derivatives of Asian call option prices
---
Introduction
============
The payoff of an Asian option depends on the (geometric or arithmetic) average of prices of a given risky asset over the pre-specified time interval. Under the Black-Scholes framework, one assumes that the price process $\{S_t, t \ge 0\}$ of the risky asset follows $$dS_t = \mu S_t dt + \sigma S_t d B_t, \qquad S_0 >0$$ where $\mu$ and $\sigma$ are given constants and $\{B_t, t\ge 0\}$ is a standard one dimensional Brownian motion. In this setting, it is easy to understand the geometric average; if $0 \le t_1 < t_2$ then in distribution $$\sqrt{S_{t_1}\cdot S_{t_2}} = S_0 \exp \left[\ ( \sigma \sqrt{t_2+ 3 t_1})\cdot \mathcal{N} + (\mu- \frac{\sigma^2}{2}) (t_1+t_2)\ \right]$$ where $\mathcal{N}$ is a standard normal random variable. On the other hand, the distribution of an arithmetic average process is not well understood. A continuous version of the arithmetic average is a time integral of a price process. Using the Inverse Lapalce Transformation, Yor [@Y] proved many interesting identities related to the distribution of geometric Brownian motion, which gives us deeper understanding of functions of geometric Brownian motion and useful information about their time integrals. More detailed research for the relation between the time integral and an Asian option was considered in [@GY]. Using the joint density of $( \int^t_0 \exp(B_s) dW_s, \exp (B_t) ) $ where $B_t, W_t$ are independent Brownian motions (given in [@B]) the moment generating function of the time integral process was computed in [@K] . The method of changing measures was considered to analyze the properties of the time integral process. (See [@GK] [@MY1] [@MY2] [@MY3].) In [@D] the very useful time reversing property is used to analyze the time integral process. Dufresne also provided a certain form for the density function of the time integral of geometric Brownian motion. However the author pointed out the difficulties to use his formula in practice especially when the time integral is over the short time period due to the slow convergence rate.
The payoff of an European style fixed strike Asian option is given by a function of the time integral of the price of the risky asset $S_t$ $$(\frac{1}{|I|} \int_I \, S_t \, dt- \kappa)^+$$ where $\kappa$ is a fixed strike price and $I$ is the pre-specified time interval with the length $|I|$. Under the risk neutral measure $\mathbb{Q}$, we may set the price process $S_t$ given by an SDE, $dS_t = \sigma S_t dB_t$ where $\sigma$ is a constant depending on the risky asset and $B_t$ is a $\mathbb{Q}$-brownian motion. Without loss of generality, we may also assume that the time interval $I=[0, \tau]$ for some $\tau>0$. It follows that the price of European style Asian option is given by $$e^{-r\tau} E_{\mathbb{Q}}\left[(\frac{1}{\tau} \int_0^\tau S_0 \exp\{\sigma B_t - \sigma^2t/2\} \,dt - \kappa)^+\right ].$$ Since $\sigma B_t = B_{\sigma^2 t}$ in law, we can rewrite the above quantity as follows $$\label{E:asian}\frac{S_0}{\tau \sigma^2}e^{-r\tau} E_{\mathbb{Q}}\left[( \int_0^{\sigma^2\tau} e^{ B_t - t/2}\, dt - \frac{\sigma^2\kappa\tau}{S_0})^+\right ].$$ Note that the initial price of the risky asset is not appeared in the time integral. To obtain the the price of Asian option and the derivatives with respect to the asset price we need to understand the following quantities $$\label{E:quantities} E\left[ (A_t - a)^+\right], \ \ \frac{d}{da} E\left[ (A_t - a)^+\right], \ \ {\rm and\ \ } \frac{d^2}{da^2} E\left[ (A_t - a)^+\right]$$ where $A_t = \int^t_0 \exp (B_s -s/2)\, ds$ and $a >0$. In this paper we show that the quantities in (\[E:quantities\]) can be expressed in terms of the expected values of functions of exponential Brownian motion. We believe that these expressions would provide the alternative approach to simulate the Asian option price and its greeks. The simulation results for Asian option price and its greeks were considered by several authors. (For example see [@A], [@BG], [@CSW], [@J].)
Other types of Asian options are also considered. When the strike price depends on the average price, it is called the floating-strike Asian option. (See for example [@HHSW] [@HW].) In [@HW], Henderson and Wojakowski show the very useful symmetries between fixed-strike and floating strike Asian options. They showed that at the starting point of the averaging period there exists an equivalent relation between the floating-strike Asian option and the fixed strike Asian option. However, once the averaging period has begun, the floating strike Asian option can not be re-expressed as a fixed strike option.
In section 2 we discuss the relation between the time integral and the exponential Brownian motion. We summarize the result in [@GK] and present the key proposition. In section 3 we discuss the price of an Asian option and its derivatives (delta, gamma, theta, and vega).
The time integral of exponential Brownian motion.
=================================================
For $\mu \in \mathbb{R}_{\ge 0}$ let us denote $A_t^{(\mu)}$ a time integral of an exponential Brownian motion with drift $\mu$ $$A_t^{(\mu)} := \int^t_0 \exp [B_s + (\mu-1/2) s ] \, ds.$$ When $\mu =0$ we simply use $A_t$ without a superscript. For $ t \ge 0$ and $y >0$ we set $M_t := \exp (B_t-t/2)$ and $$R_t:= - \frac{M_t}{y^{-1}-1/2 \, A_t}\,=\, 2\frac{d}{dt}\log(1-\frac{y}{2}A_t) .$$ It is not hard to see that the process $R_t$ satisfies the following SDE $$dR_t =R_tdB_t - \frac{1}{2} R_t^2 \,dt,\ \ \ \ R_0=-y$$ up to an explosion time $\tau_\infty := \inf \{ t\ge 0: A_t = 2/y\}$. By considering a stopping time $\tau_n = \inf \{t : R_t \le -n\}$ and a Girsanov density process $R_t 1_{\{t < \tau_n\}}$, we define a new measure $\mathbb{Q}$ under which $$\tilde B_t = B_t - \int^t_0 R_s 1_{\{s < \tau_n\}} \, ds= B_t - 2 \log(1-\frac{y}{2} A_{t \wedge \tau_n})$$ is a standard Brownian motion. Let us define $\tilde M = \exp (\tilde B_t -t/2)$ and $\tilde A_t = \int^t_0 \tilde M_s\, ds$. It follows that $$\label{E:dic1}
\tilde M_t = \frac{M_t}{(1-\frac{y}{2}A_{t \wedge \tau_n})^2},\ \ \ 1+\frac{y}{2} \tilde A_{t \wedge \tau_n} = \frac{1}{1-\frac{y}{2}A_{t \wedge \tau_n}}$$ and $$\label{E:dic2}R_{t\wedge \tau_n}= - \frac{\tilde M_{t \wedge \tau_n}}{y^{-1}+ \frac{1}{2} \tilde A_{t \wedge \tau_n}}.$$ The following proposition is a simple modification of the proposition 3.3 in [@GK].
\[T:keylemma\] If $f(x,z)$ is a Borel measurable function and $y>0$ then $$\label{E:key}
\begin{aligned}
E&\left[ f(M_t, A_t); A_t < \frac{2}{y} \right] \\&= e^{-y} E\left[ f(\frac{M_t}{(1+\frac{y}{2} A_t)^2}, \frac{A_t}{1+\frac{y}{2}A_t}) \exp \left( \frac{M_t}{y^{-1}+\frac{1}{2}A_t}\right) \right].
\end{aligned}$$
The proof of the proposition is essentially same as the proof of the proposition 3.3 in [@GK], so we give a brief sketch. (For the detailed proof we refer [@GK].) For fixed $n$ and $y>0$ it is not hard to see that $$\exp(R_t+y)1_{\{\tau_n>t\}} = \exp ( \int^{t \wedge\tau_n}_0 R_s dB_s - \frac{1}{2} \int^{t \wedge \tau_n}_0R_s^2\, ds)$$ satisfies a Novikov condition. Thus we have $$E\left[f(M_t, A_t); \tau_n >t\right] = E_{\mathbb{Q}}\left[ f(M_t, A_t)\exp(-R_t-y)1_{\tau_n>t}\right].$$ Since the event $\{ \tau_n>t\}$ is the same as $\{ \max_{s \le t} \tilde M_s/(y^{-1}+ 1/2 \, \tilde A_s)<n\}$, by letting $n\to \infty$ the r.h.s. becomes $E_{\mathbb{Q}}[f(M_t,A_t) \exp (-R_t-y)]$. Use (\[E:dic1\]) and (\[E:dic2\]) we can rewrite the limit of the r.h.s. in terms of $\tilde M_t, \tilde A_t$ and $\tilde R_t$. Since we are only interested in the quantity, we remove the tilde and get (\[E:key\]).
Let us discuss certain of interesting choices of $f$ in Proposition \[T:keylemma\]. Simple choices like a constant function or power functions allows us to have various relations between a simple martingale $M_t$ and its time integral. We consider the ones who directly related to the problem of pricing the European style Asian option and estimating the sensitivities. First we consider a constant function $f(x,z)\equiv 1$. Using a constant function, it is easy to see that the l.h.s. of (\[E:key\]) is a probability distribution of the time integral of a geometric brownian motion. By setting $a= 2/y$ it follows that
\[L:distribution1\] For a positive $a>0$, we have $$Pr[A_t < a] = e^{-2/a} E \left[ \exp\left(\frac{M_t}{ a/2 +1/2\,A_t}\right)\right].$$
For $\nu \in \mathbb{R}$ we have $M_t^\nu \exp(\nu t/2-\nu^2t/2) = \exp (\nu B_t -\nu^2t/2)$ and thus $$\label{E:drift} E[M_t^\nu \exp(\nu t/2-\nu^2t/2) ; A_t \le 2/y] = Pr[A_t^{(\nu)}\le 2/y].$$ By setting $f(x,z)=x^\nu$ we get
\[L:distribution2\] For a real number $\nu$ and a positive real $a>0$, we have $$\begin{aligned}
Pr& [A_t^{(\nu)} \le a] \\&= a^{2 \nu} e^{-2/a} \mathbb{E} \left[ (a+ A_t^{(\nu)})^{-2\nu} \exp \left ( \frac{2 \exp ( B_t + (\nu-1/2) t)}{a+ A_t^{(\nu)}} \right) \right].
\end{aligned}$$
=1.8in
Figure 2.1. The density function of $A_t$.
Let $g_t$ denote the probability density function for $A_t$. It is known that the density function for $A_t$ is continuous and positive. Figure 2.1 is the result of the simulation ($10000$ simulation for each $a$) using the identity given in the following Lemma \[L:density\]. The $n$-th moments of the time integral process $A_t$ can be computed with simple computation. However it is known that $A_t$ has a heavy tail probability and thus knowing all integer moments does not give a probability density function. In Figure 2.1 we can see the heavy tail probability.
\[L:density\] For a real number $a>0$, the probability density function $g_t$ for $A_t$ satisfies $$g_t(a) = \frac{2}{a^2} Pr[A_t\le a] - \frac{2}{a^2} Pr[A_t^{(1)} \le a].$$
From Lemma \[L:distribution1\] we have $$\begin{aligned}
g_t(a) &= \frac{d}{da} Pr[A_t <a]\\ & =\frac{2}{a^2} Pr[A_t\le a] - \frac{1}{2}e^{-2/a} E\left[\frac{M_t}{(a/2+1/2\, A_t)^2} \exp\left( \frac{M_t}{a/2+1/2\, A_t}\right)\right]
\end{aligned}$$ Since $M_t = \exp (B_t - t/2)$ we use $M_t$ as a Girsanov density function for the second term. Under the new measure $B_t - t/2$ is a standard Brownian motion and thus the second term is equal to the following quantity: $$\frac{1}{2} e^{-2/a} E\left[ \frac{1}{(a/2+1/2\, A_t^{(1)})^2} \exp \left( \frac{2\exp(B_t + t/2)}{ a+A_t^{(1)}}\right) \right].$$ Comparing with Lemma \[L:distribution2\], we can see that the above quantity is the same as $2/a^2 Pr[A_t^{(1)} \le a]$.
Furthermore we have
\[P:joint\] For any $a>0$, $b>0$ and $t>0$ we have $$Pr[M_t <b, A_t <a] = e^{-2/a} \mathbb{E} \left[ \exp \left( \frac{2 M_t}{a+A_t}\right) \, ;\, M_t \le b (1+ \frac{1}{a} A_t)^2\right].$$
Consider $f(x,z) = 1\{x <b\}$ in equation (\[E:key\]). By setting $a= 2/y$ we get an integration form for the joint probability density function for $M_t$ and $A_t$.
The simulation result of the above proposition \[P:joint\] for $t=a$ is shown in the l.h.s. of the following figure and the joint distribution for $t=1$ is shown in the r.h.s.
=1.8in =1.8in
Figure 2.2. Joint distribution of $M_t$ and $A_t$.
The first quantity of interests to have the derivatives of Asian options is the expected value of the maximum function of the time integral subtracted by a constant. Using Proposition \[T:keylemma\] we have the following result :
\[T:moment\] For any $a>0$ and $\nu\in \mathbf{R}$, we have $$\mathbb{E} \left[ (A_t^{(\nu)} -a)^+\right] = \frac{e^{\nu t} -1}{\nu} -a + e^{\nu t /2 - \nu^2 t/2 } a^{2 \nu +2} \,\mathbb{E} \left[ \frac{M_t^\nu}{(a+A_t)^{2 \nu+1} }\exp ( \frac{ 2M_t}{a+ A_t} - \frac{2}{a} ) \right].$$ In particular if $\nu = 0$ we have $$\label{E:max} \mathbb{E} \left[ (A_t -a)^+\right] = t -a + a^{2} \,\mathbb{E} \left[ (a+A_t)^{-1} \exp ( \frac{ 2M_t}{a+ A_t} - \frac{2}{a} ) \right].$$
Let us set $f(x,z) = x^\nu (2/y -z)$. The l.h.s. of (\[E:key\]) becomes $E[2/y-A_t; A_t < 2/y]=E[2/y-A_t] - E[(A_t-2/y)^+]$. Since $2/y -A_t/(1+y/2\, A_t) = 1/(2/y+A_t)$, the second result comes directly from (\[E:key\]). Using the identity (\[E:drift\]) we can change the drift term of the exponential brownian motion to get the first result.
=1.94in
Figure 2.3. Expected value of max function : $a=0.4,t=0.5$.
The above figure is the simulation result of the second identity (\[E:max\]) in Theorem \[T:moment\]. We set $a=0.4$ and $t=0.5$. The dotted line is obtained using the l.h.s. of the equation (\[E:max\]), that is we simulate $A_t$ using Monte Carlo methods and take the maximum function. The solid line is from the r.h.s. of the equation (\[E:max\]). Only with 500 simulation, we can observe that the convergence is faster for the solid line, which was expected because of the discarded simulations of the l.h.s. (when $A_t$ is smaller than $a$).
Now let us consider the derivatives of $E[(A_t-a)^+]$ with respect to $a$. Since $$E[A_t -a; A_t \le a] = - \int^a_0 Pr[A_t \le u]\, du$$ we have that $$\begin{aligned}
E[(A_t-a)^+] \ &=\ E[A_t -a] - E[A_t -a; ; A_t \le a]\\
&= t-a+\int^a_0 Pr[A_t \le u]\, du.
\end{aligned}$$ It follows that $$\frac{d}{da}E[(A_t-a)^+] = -1+Pr[A_t\le a],\ \ {\rm and\ \ }\ \frac{d^2}{da^2}E[(A_t-a)^+] = g_t(a).$$ The following theorem is the direct application of Lemma \[L:distribution1\] and Lemma \[L:density\].
\[T:derivative\] The first and second derivatives of $E[(A_t-a)^+]$ with respect to $a$ is given by following equations: $$\begin{aligned}
\frac{d}{da}E[(A_t-a)^+] \ =\ & -1+e^{-2/a} \mathbb{E} \left[ \exp \left ( \frac{2 \exp ( B_t -t/2)}{a+ A_t}\right) \right],\\
\frac{d^2}{da^2}E[(A_t-a)^+] \ =\ &\ \frac{2}{a^2} e^{-2/a} \mathbb{E} \left[ \exp \left ( \frac{2 \exp ( B_t -t/2)}{a+ A_t}\right) \right] \\ &- \frac{1}{2}e^{-2/a} E\left[\frac{M_t}{(a/2+1/2\, A_t)^2} \exp\left( \frac{M_t}{a/2+1/2\, A_t}\right)\right].
\end{aligned}$$
Sensitivities.
==============
Since $A_t$ satisfies $A_{t+s} =A_t+M_t\tilde A_s$ for $t,s >0$ where $\tilde A$ is an independent copy of $A$, pricing the fixed strike Asian option at time $0 \le t \le $ the expiration date is essentially identical as the pricing the option at the beginning of the averaging period.
One of important activities in Financial Market is managing the risk. One way to measure the risk in the option is estimating the ‘[*Greek letters*]{}’ such as delta, gamma, theta, etc. In this section we use identities obtained in Section 2 to get the ‘[*Greek letters*]{}’ of an European Style Asian option under the Black-Scholes framework; the price process of the risky asset follows under the risk-neutral measure $$dS_t = \sigma S_t dB_t + r S_t dt,$$ where $B_t$ is a standard brownian motion under the risk-neutral measure, $r$ is a constant interest rate and $\sigma$ is a constant volatility of given asset. The expression (\[E:asian\]) is equal to the price of an European style Asian option with the expiration date $\tau$ and the strike price $\kappa$ : $$\label{E:asiancall}
Call = \frac{S_0}{\tau \sigma^2}e^{-r\tau} E\left[( \int_0^{\sigma^2\tau} e^{ B_t - t/2}\, dt - \frac{\sigma^2\kappa\tau}{S_0})^+\right ].$$
\[T:call\] The price of an Asian call at time $t=0$ is given by $$Call_0 = S_0 e^{-r \tau} \left[ 1- \frac{\kappa}{S_0} + \frac{\sigma^2 \kappa^2 \tau}{S_0^2} E \left[(a+A_{\tau \sigma^2})^{-1} \exp \left( \frac{2 M_{\tau\sigma^2}}{a+A_{\tau\sigma^2}}-\frac{2}{a}\right)\right]\right],$$ where $a= \frac{\sigma^2\kappa\tau}{S_0}$, $\tau$ is the expiration date, $\kappa$ is the strike price.
The result comes directly from the combination of Theorem \[T:moment\] and equation (\[E:asiancall\]).
The delta $\Delta$ is the rate of change of the price of the option with respect to the price of the underline asset. Thus we have $$\label{E:delta}\Delta =\frac{ Call }{S_0} -\left. \frac{\kappa}{S_0} e^{-r \tau} \frac{d}{da} E[(A_{\sigma^2 \tau} -a)^+]\right |_{a= \frac{\sigma^2\kappa\tau}{S_0}}.$$ Using Theorem \[T:moment\] and Theorem \[T:derivative\] we have
\[T:delta\] The delta $\Delta_0$ of an Asian call option at time $t=0$ is given by $$\begin{aligned}
\Delta_0= e^{-r\tau} + \frac{\kappa}{S_0} e^{-r \tau-2/a}& \left\{ a E \left[(a+A_{\tau \sigma^2})^{-1} \exp \left( \frac{2 M_{\tau\sigma^2}}{a+A_{\tau\sigma^2}}\right)\right]\right.\\&\left.\left.- E\left[ \exp \left( \frac{2 M_{\tau\sigma^2}}{a+A_{\tau\sigma^2}}\right)\right]\right\}\right|_{a= \frac{\sigma^2\kappa\tau}{S_0}}.
\end{aligned}$$
The gamma $\Gamma$ is the rate of change of the delta with respect to the price of underlying asset. Thus the direct computation using equation (\[E:delta\]) and the results in Theorem \[T:moment\], Theorem \[T:derivative\] and Theorem \[T:delta\] we have
\[T:gamma\] The gamma $\Gamma_0$ of an Asian call option at time $t=0$ is given by $$\Gamma_0= \frac{\sigma^2 \kappa^2\tau}{S_0^3} e^{-r \tau} g_{\tau\sigma^2} (\frac{\sigma^2\kappa\tau}{S_0}),$$ where $g_t$ is a continuous probability density function of $A_t$ given in Lemma \[L:density\].
=1.4in =1.4in
=1.4in =1.4in
Figure 3.1. Derivatives vs. the time to expiration: $S_0=\kappa=\sigma=1$
In Figure 3.1. we use Monte Carlo methods to get Greek letters $\Delta, \Gamma$, $\Theta$ and $\nu$. We take the risk free rate $r=0$ and both the initial stock price and the strike price are equal to 1. We also set for simplicity $\sigma=1$. Notice that we plot greek letters vs. time to expiration. Since we computed greeks at the beginning of the averaging period, time to expiration is the same as the length of the averaging period. In the upper l.h.s. figure we can observe that as time to expiration increases, the value of $\Delta_0$ increases. The plot of $\Gamma_0$ with respect to time to expiration is given in the upper r.h.s. figure. Since $\Gamma$ is the first derivative of $\Delta$ with respect to the stock prices, it reflects our observation in the plot of $\Delta_0$. The lower l.h.s. figure is the plot of $\Theta_0$ and the lower r.h.s. figure is the plot of $\nu_0$.
Since all parameters are constants, the call price satisfies the differential equation $$\frac{\partial Call}{\partial t} + r S \frac{\partial Call}{\partial S} + \frac{1}{2} \sigma^2 S \frac{\partial^2 Call}{\partial S^2} = r \,Call.$$ It follows that $$\Theta_0 + r S_0 \Delta_0 + \frac{1}{2} \sigma^2 S_0 \Gamma_0 = r \,Call_0$$ where $\Theta_0$ is the rate of change of the price of the option at time $t=0$ with respect to time and $Call_0$ is the price of the call option at time $0$.
\[T:theta\] The theta $\Theta_0$ of an Asian call option at time $t=0$ is given by $$\Theta_0= r Call_0 -r S_0 \Delta_0 - \frac{1}{2} \sigma^2 S_0 \Gamma_0$$ where $Call_0$, $\Delta_0$ and $\Gamma_0$ are given in the previous theorems 3.1–3.
Now let us discuss the vega ${\bf \nu}$. The vega $\mathbf{\nu}$ is the rate of change of the price of the option with respect to the volatility of the underlying asset. The call price of European style Asian option is given in (\[E:asiancall\]). Thus we have
$$\begin{aligned}
\frac{\partial Call }{\partial \sigma} &= -\frac{2}{\sigma} \, Call + \frac{2S_0}{\sigma}e^{-r \tau} E\left[ M_{\sigma^2 \tau}- \frac{\kappa}{S_0} \ ;\ A_{\sigma^2 \tau} > \frac{\sigma^2 \kappa\tau}{S_0} \right]\\
&= -\frac{2}{\sigma} \, Call +\frac{2S_0}{\sigma} e^{-r \tau} E\left[M_{\sigma^2 \tau}- \frac{\kappa}{S_0}\right] \\&\phantom{alsdkj} - \frac{2S_0}{\sigma} e^{-r \tau} E\left[M_{\sigma^2 \tau}- \frac{\kappa}{S_0}\,;\, A_{\sigma^2 \tau} < \frac{\sigma^2 \kappa\tau}{S_0} \right].
\end{aligned}$$
It follows that $$\label{E:vega}
\begin{aligned}
\frac{\partial Call }{\partial \sigma} &= -\frac{2}{\sigma} \, Call+ \frac{2S_0}{\sigma} e^{-r \tau} \left( 1- E\left[M_{\sigma^2 \tau} \,;\, A_{\sigma^2 \tau} < \frac{\sigma^2 \kappa\tau}{S_0} \right]\right)\\&-\frac{2 \kappa}{\sigma} \left(1-P \left[A_{\sigma^2 \tau}< \frac{\sigma^2 \kappa\tau}{S_0} \right]\right).
\end{aligned}$$ By setting $\nu=1$ in equation (\[E:drift\]) in Lemma \[L:distribution1\], we see that $$1- E\left[M_{\sigma^2 \tau}\,;\, A_{\sigma^2 \tau} < \frac{\sigma^2 \kappa\tau}{S_0} \right] = P \left[A_{\sigma^2\tau}^{(1)}> \frac{\sigma^2 \kappa\tau}{S_0} \right].$$ Therefore we have
\[T:vega\] The vega $\nu_0$ of an asian call option at time $t=0$ is given by $$\nu_0 = -\frac{2}{\sigma} \, Call+ \frac{2S_0}{\sigma} e^{-r \tau} P \left[A_{\sigma^2\tau}^{(1)}> \frac{\sigma^2 \kappa\tau}{S_0} \right]-\frac{2 \kappa}{\sigma}P \left[A_{\sigma^2 \tau}> \frac{\sigma^2 \kappa\tau}{S_0} \right].$$
Last two quantities in the above theorem can be obtained using Lemma \[L:distribution2\].
[99]{}
J. Andreasen, The pricing of discretely sampled Asian and lookback options: a change of numeraire approach, J. Comp. Finance 1, 15-36 (1998).
P. Bougerol: *Examples de théprèmes locaux sur les groups résolubles*, Ann. Inst. H. Poincaré Sect. B (N.S.) **19**, 369-391 (1983).
M. Broadie and P. Glasserman: *Estimating Security Price Derivatives Using Simulation*, Management Science **42**, 269-285 (1996).
S.L. Chung, M. Shackleton and R. Wojakowski, Efficient quadratic approximation of floating strike Asian option values, preprint (2003).
D. Dufresne: *The integral of geometric Brownian motion*, Adv. in Appl. Probab. **33**, 223-241 (2001).
H. Geman and M. Yor: *Asian Options, Bessel Processes and Perpetuities*, Math. Finance **2**, 349-375 (1993).
V. Goodman and K. Kim: *Exponential Martingales and Time integrals of Brownian Motion*, preprint (2006) math.PR/0612034.
V. Henderson, D. Hobson, W. Shaw and R. Wojakowski, preprint.
V. Henderson and R. Wojakowski, On the Equivalence of Floating and Fixed-strike Asian Options, J. Appl. Prob. 39, 391-394 (2002).
J.C. Hull: *Options, Futures, and Other Derivatives, 6th edition*, Prentice Hall (2006)
N. Ju, Pricing Asian and basket options via Taylor expansion, J. Comp. Finance
K. Kim: *Moment Generating function of the inverse of integral of geometric Brownian Motion*, Proc. Amer. Math. Soc. **132**, 2753-2759 (2004)
H. Matsumoto and M. Yor: *An analogue of Pitman’s 2M – X theorem for exponential Wiener functionals, Part I: A time-inversion approach.*, Nagoya Math. J. **159**, 125-166 (2000).
H. Matsumoto and M. Yor: *An analogue of Pitman’s 2M — X theorem for exponential Wiener functionals, Part II: The role of the generalized inverse Gaussian laws.*, Nagoya Math. J. **162**, 65-86 (2001).
H. Matsumoto and M. Yor: *A relationship between Brownian motions with opposite drifts via certain enlargments of the Brownian filtration*, Osaka Journal of Mathematics **38**, 383-398 (2001).
M. Yor: *On some exponential functionals of Brownian motion*, Adv. in Appl. Probab. **24**, 509-531 (1992).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents a rank rigidity result for negatively curved spaces. Let $M$ be a compact manifold with negative sectional curvature and suppose that along every geodesic in $M$ there is a parallel vector field making curvature $-a^2$ with the geodesic direction. We prove that $M$ has constant curvature equal to $-a^2$ if $M$ is odd dimensional, or if $M$ is even dimensional and has sectional curvature pinched as follows: $-\Lambda^2 < K < -\lambda^2$ where $\lambda/\Lambda > .93$. When $a$ is extremal, i.e. $-a^2$ is the curvature minimum or maximum for the manifold, this result is analogous to rank rigidity results in various other curvature settings where higher rank implies that the space is locally symmetric. In particular, this result is the first positive result for lower rank (i.e. when $-a^2$ is minimal), and in the upper rank case gives a shorter proof of the hyperbolic rank rigidity theorem of Hamenstädt, subject to the pinching condition in even dimension. We also present a rigidity result using only an assumption on maximal Lyapunov exponents in direct analogy with work done by Connell. Our proof of the main theorem uses the ergodic theory of the frame flow developed by Brin and others - in particular the transitivity group associated to this flow.'
author:
- 'David Constantine[^1] [^2]'
date: 'December 19, 2006'
title: Consequences of Ergodic Frame Flow for Rank Rigidity in Negative Curvature
---
\[section\] \[thm\][Definition]{} \[thm\][Lemma]{} \[thm\][Proposition]{}
Introduction
============
Rank rigidity was first proved in the higher Euclidean rank setting by Ballmann [@Ballman] and, using different methods, by Burns and Spatzier [@Burns-Spatzier]. A manifold is said to have higher Euclidean rank if a parallel normal Jacobi field can be found along every geodesic. Ballmann and Burns-Spatzier proved that if an irreducible, compact, nonpositively curved manifold has higher Euclidean rank, then it is locally symmetric. Ballmann’s proof works for finite volume as well and the most general version of this theorem is due to Eberlein and Heber, who prove it under only a dynamical condition on the isometry group of $M$’s universal cover [@Eberlein-Heber]. Hamenstädt showed that a compact manifold with curvature bounded above by -1 is locally symmetric if along every geodesic there is a Jacobi field making curvature -1 with the geodesic direction [@Hamenstadt]. She called this situation higher hyperbolic rank. Shankar, Spatzier and Wilking extended rank rigidity into positive curvature by defining spherical rank. A manifold with curvature bounded above by 1 is said to have higher spherical rank if every geodesic has a conjugate point at $\pi$, or equivalently, a parallel vector field making curvature 1 with the geodesic direction. They proved that a complete manifold with higher spherical rank is a compact, rank one locally symmetric space [@sphericalrank].
These results settle many rank rigidity questions, but leave questions about other curvature settings open (see [@sphericalrank] for an excellent overview). In this paper we prove the following theorem, which can be applied to various settings in negative curvature.
Let $M$ be a compact, negatively curved manifold. Suppose that along every geodesic in $M$ there exists a parallel vector field making sectional curvature $-a^2$ with the geodesic direction. If $M$ is odd dimensional, or if $M$ is even dimensional and satisfies the sectional curvature pinching condition $-\Lambda^2 < K <
-\lambda^2$ with $\lambda/\Lambda > .93$ then $M$ has constant negative curvature equal to $-a^2$.
Note that, unlike previous rank rigidity results, Theorem 1 allows for situations where the distinguished curvature $-a^2$ is not extremal. However, the cases where $-a^2$ is extremal are of particular importance and in these situations the extremality of the distinguished curvature $-a^2$ allows the hypotheses of our theorem to be weakened, as demonstrated in section \[sec:parallel\] of this paper. The folowing two results are then easy corollaries of Theorem 1:
Let $M$ be a compact manifold with sectional curvature $-1 \leq K < 0$. Suppose that along every geodesic in $M$ there exists a Jacobi field making sectional curvature $-1$ with the geodesic direction. If $M$ is odd dimensional, or if $M$ is even dimensional and satisfies the sectional curvature pinching condition $-1 \leq K < -\lambda^2$ with $\lambda > .93$ then $M$ is hyperbolic.
*(compare with Hamenstädt [@Hamenstadt])* Let $M$ be a compact manifold with sectional curvature bounded above by $-1$. Suppose that along every geodesic in $M$ there exists a Jacobi field making sectional curvature $-1$ with the geodesic direction. If $M$ is odd dimensional, or if $M$ is even dimensional and satisfies the sectional curvature pinching condition $-(1/.93)^2 < K \leq
-1$ then $M$ is hyperbolic.
In Corollary 1, $-1$ is the curvature minimum for $M$ and we obtain a new rank rigidity result analogous to those described above. This is the first positive result for lower rank, i.e. when the distinguished curvature value is the lower curvature bound (see section \[sec:conclusion\] for more discussion). In Corollary 2, $-1$ is the curvature maximum for $M$ and we obtain a shorter proof of Hamenstädt’s result, under an added pinching constraint in even dimension.
In [@Connell], Connell showed that rank rigidity results can be obtained using only a dynamical assumption on the geodesic flow, namely an assumption on the Lyapunov exponents at a full measure set of unit tangent vectors. His paper deals with the upper rank situations treated by Ballmann, Burns-Spatzier and Hamenstädt. He proves that having the minimal Lyapunov exponent allowed by the curvature restrictions attained at a full measure set of unit tangent vectors is sufficient to apply the results of Ballman and Burns-Spatzier or Hamenstädt. In the lower rank setting of this paper, this viewpoint translates into
Let $M$ be a compact manifold with sectional curvature $-a^2 \leq K < 0$, where $a > 0$. Suppose that for a full (Liouville) measure set of unit tangent vectors $v$ on $M$ the maximal Lyapunov exponent at $v$ is $a$, the maximum allowed by the curvature restriction. If $M$ is odd dimensional, or if $M$ is even dimensional and satisfies the sectional curvature pinching condition $-a^2 \leq K < -\lambda^2$ with $\lambda/a > .93$ then $M$ is of constant curvature $-a^2$.
The adaptation of Connell’s arguments for this setting is discussed in section \[sec:Lyap\].
The proof of Theorem 1 relies on dynamical properties of the geodesic and frame flows on negatively curved manifolds. We rely heavily on Brin’s work on frame flows (see [@Brin-survey] for a survey), and results on the ergodicity of these flows due to Brin and Gromov in odd dimension [@Brin-Gromov] and to Brin and Karcher in even dimension [@Brin-Karcher]. These results are summarized in section \[sec:background\]. In particular, we utilize the transitivity group $H_v$, defined for any vector $v$ in the unit tangent bundle of $M$, which acts on $v^{\perp} \subset
T^1M$. Essentially, elements of $H_v$ correspond to parallel translations around ideal polygons in $M$’s universal cover. Brin shows that this group is the structure group for the ergodic components of the frame flow (see e.g [@Brin-survey] or [@Brin-gpext]). In section \[sec:trans\] we show that, subject to suitable recursion properties on these ideal polygons, $H_v$ preserves the parallel fields that make curvature $-a^2$ with the geodesic defined by $v$. We then show that these recursion properties are generic and that the elements of $H_v$ vary continuously with the choice of ideal polygon. Thus, all of $H_v$ respects the distinguished fields. Finally, we apply results of Brin-Gromov and Brin-Karcher on the ergodicity of the 2-frame flow which imply that $H_v$ acts transitively on $v^\perp$ and conclude that the curvature of $M$ is constant.
I would like to thank Chris Connell for discussions helpful with the arguments in section \[sec:parallel\] of this paper, Jeffrey Rauch for the proof of Lemma \[Rauchlemma\], and Ben Schmidt for helpful comments on this paper. In particular, special thanks are due to my advisor, Ralf Spatzier, for suggesting this problem, for help with several pieces of the argument and for helpful comments on this paper.
Notation and background {#sec:background}
=======================
Let us begin by fixing some notation and stating the results we will need. Let $M$ be a compact Riemannian manifold with negative sectional curvature and let $\tilde{M}$ be its universal cover. Denote by $T^1M$ and $T^1\tilde{M}$ the unit tangent bundles to $M$ and $\tilde{M}$, respectively. We will denote by $g_t$ the geodesic flow on either of these spaces, and by $F_t$ the frame flow on the Stiefel manifold $St_kM$, the space of ordered orthonormal $k$-frames on $M$. $St_kM$ is a fiber bundle over $T^1M$ with the group $SO(n-1)$ acting on the right; $St_nM$ is a principal bundle with $SO(n-1)$ as structure group. There are standard measures on these spaces, namely Liouville measure on $T^1M$ and $T^1\tilde{M}$, and on $St_kM$ the product measure of Liouville measure and the measure on the fibers inherited from the Haar measure on $SO(n-1)$. Unless otherwise specified, these will be the measures used in all that follows. Let $\gamma_v (t)$ denote the geodesic in $M$ or $\tilde{M}$ with velocity $v$ at time 0. We will denote by $w_v (t)$ a parallel normal vector field along $\gamma_v (t)$ making the distinguished curvature $-a^2$ with $\dot{\gamma}_v (t)$. Finally, $\langle \cdot, \cdot \rangle$ will denote the Riemannian inner product, $R(\cdot, \cdot) \cdot$ will denote the curvature tensor and $K(\cdot, \cdot)$ will denote sectional curvature.
The geodesic flow gives rise to stable and unstable foliations $W^s_g$ and $W^u_g$ of $T^1M$ or $T^1\tilde{M}$. These foliations are absolutely continuous, and the geodesic flow for such an $M$ is ergodic (proved by Anosov, see Brin’s appendix to [@Brin-ergodicity]). The frame flow also gives rise to stable and unstable foliations $W^s_F$ and $W^u_F$ as shown by Brin [@Brin-toptrans]. These foliations allow Brin to define the *transitivity group* in the following way. If $v$ and $v'$ are on the same leaf of $W^s_g$ then for every $n$-frame $\alpha$ above $v$ there is a unique $n$-frame $\alpha'$ above $v'$ such that $\alpha$ and $\alpha'$ belong to the same leaf of $W^s_F$. In particular, the distance between $F_t(\alpha)$ and $F_t(\alpha')$ approaches 0 as $t \rightarrow \infty$ (this is how one determines that $\alpha'$ belongs to the leaf $W^s_F(\alpha)$). Let $p(v, v')$ be the map from the fiber of $St_nM$ over $v$ to the fiber over $v'$ that takes each $\alpha$ to the corresponding $\alpha'$ over $v'$. Note $p(v, v')$ corresponds to a unique isometry between $v^{\perp}$ and $v'^{\perp}$. Once defined on $n$-frames, $p(v, v')$ acts on all $k$-frames; the action on 2-frames will be what we use in this paper and we will abuse notation by using $p(v,
v')$ to denote this restricted action. One can think of $p(v, v')(\alpha)$ as the result of parallel transporting $\alpha$ along $\gamma_v (t)$ out ‘to the boundary at infinity of $\tilde{M}$’ and then back to $v'$ along $\gamma_{v'} (t)$. If $v'$ and $v$ belong to the same leaf of $W^u_g$ there is similarly an isometry corresponding to parallel translation to the boundary at infinity along $\gamma_{-v}$ and back along $\gamma_{-v'}$. This defines the unstable leaves for the frame flow foliation and we will also denote this isometry by $p(v, v')$. Brin (see [@Brin-survey] Defn. 4.4) then defines the transitivity group at $v$ as follows:
Given any sequence $s=\{v_0, v_1, \ldots , v_k\}$ with $v_0=v_k=v$ such that each pair $\{v_i, v_{i+1}\}$ lies on the same leaf of $W^s_g$ or $W^u_g$ we have an isomorphism of $v^{\perp}$ given by $$I(s) = \prod_{i=0}^{k-1} p(v_i, v_{i+1}).$$ The closure of the set of all such isometries is denoted by $H_v$ and is called the transitivity group.
![‘equilateral’ ideal rectangle[]{data-label="fig:equi"}](figure1ca.eps){height=".45\textheight" width=".6\textwidth"}
The idea of the transitivity group is that it is generated by isometries coming from parallel translation around ideal polygons in $\tilde{M}$ with an even number of sides, such as the one shown in figure \[fig:equi\]. Note that here only ‘equilateral’ polygons are allowed; in figure \[fig:equi\] only rectangles for which $W^u_g(v_1)$ is tangent to $W^s_g(v_3)$ are permitted. We will later find it useful to allow general ideal polygons.
The definition of this group arises in Brin’s analysis of the ergodic components of the frame flow. He shows in [@Brin-gpext] that the ergodic components are subbundles of $St_kM$ with structure group a closed subgroup of $SO(n-1)$ (see also [@Brin-survey] section 5 for an overview). In addition, his proof demonstrates that the structure group for the ergodic component is the transitivity group (see [@Brin-survey] Remark 2 or [@Brin-gpext] Proposition 2). This explicit geometric description of the ergodic components is the central tool used in our proof.
We use two results on the ergodicity of the 2-frame flow in our proof.
\[thm:BG\] *(Brin-Gromov [@Brin-Gromov] Proposition 4.3)* If $M$ has negative sectional curvature and odd dimension then the 2-frame flow is ergodic.
\[thm:BK\] *(Brin-Karcher [@Brin-Karcher])* If $M$ has sectional curvature satisfying $-\Lambda^2 < K < -\lambda^2$ with $\lambda / \Lambda > .93$ then the 2-frame flow is ergodic.
Theorem \[thm:BK\] is not directly stated as above in [@Brin-Karcher], rather it follows from remarks made in section 2 of that paper together with Proposition 2.9 and the extensive estimates carried out in the later sections. Note that since the 2-frame flow preserves the parallel fields making curvature $-a^2$, ergodicity of this flow alone seems to indicate that the manifold has constant curvature. However, since the subset of $St_2M$ given by these distinguished fields may, a priori, have zero measure, the result does not follow directly from ergodicity. Instead, we must use the precise description of the ergodic components given by the transitivity group.
The transitivity group and distinguished vector fields {#sec:trans}
======================================================
As noted in the Introduction, the description of the ergodic components in terms of the transitivity group is crucial. In this section we investigate how the distinguished vector fields $w_v(t)$ along $\gamma_v(t)$ behave under the action of the transitivity group and use the results to prove Theorem 1.
To obtain results we will need to assume some dynamical properties of the ideal polygon that produces a given element of $H_v$. For example, consider the ideal rectangle defined by $v$, $v_1$, $v_2$ and $v_3$ as pictured in figure \[fig:equi\]. Note that the distinguished vector field $w_v(t)$ which makes constant curvature $-a^2$ with $\gamma_v(t)$ corresponds uniquely to a parallel normal vector field $P(t)$ along $\gamma_{v_1}(t)$, such that $v_1$ and $P(0)$ make a 2-frame in the leaf of the stable foliation containing the 2-frame $\{v, w_v(0)\}$, that is, $\{v_1, P(0)\} = p(v, v_1)\{v,
w_v(0)\}$. By continuity of the sectional curvature, $K(P(t), \dot{\gamma}_{v_1}(t))
\rightarrow -a^2$ as $t \rightarrow \infty$.
Suppose $\gamma(t)$ is a recurrent geodesic with a parallel normal field $P(t)$ along it such that $K(P(t), \dot{\gamma}(t)) \rightarrow -a^2$ as $t \rightarrow \infty$. Then $K(P(t), \dot{\gamma}(t)) \equiv -a^2$ for all $t$.
Since $\gamma(t)$ is recurrent we can take an increasing sequence $\{ t_k\} $ tending to infinity such that $\dot{\gamma}(t_k)$ approaches $\dot{\gamma}(0)$. Since the parallel field $P(t)$ has constant norm and the set of vectors in $\dot{\gamma}_v(t)^{\perp}$ with this norm is compact, we can, by passing to a subsequence, assume that $P(t_k)$ has a limit $G(0)$. Extend $G(0)$ to a parallel vector field $G(t)$ along $\gamma(t)$.
By construction, $K(G(0), \dot{\gamma}(0)) = \lim_{k \rightarrow \infty} K(P(t_k),
\dot{\gamma}(t_k)) = -a^2$. In addition, for any real number $T$, the recurrence $\dot{\gamma}(t_k) \rightarrow \dot{\gamma}(0)$ implies recurrence $\dot{\gamma}(t_k+T)
\rightarrow \dot{\gamma}(T)$. By continuity of the frame flow, we get that $P(t_k+T)
\rightarrow G(T)$ for the vector field $G$ defined above. Thus $G(t)$ makes curvature $-a^2$ with $\dot{\gamma}(t)$ for any time $t$.
We can repeat the same argument as above, letting $G(t)$ recur along the same sequence of times to produce $G_1(t)$, and likewise $G_i(t)$ recur to produce $G_{i+1}(t)$, forming a sequence of fields all making curvature identically $-a^2$ with the geodesic direction. Now, observe that $G(0) = P(0)\cdot g$ for some $g\in SO(n-1)$. Note here that $g$ is not well defined by looking at $P$ and $G$ alone, but will be well defined if we consider $n$-frame orbits with second vector $P$ recurring to $n$-frames with second vector $G(0)$; this is the $g$ we utilize. By construction and the fact that the $SO(n-1)$ action commutes with parallel translation, $G_i(0)=P(0)\cdot g^{i+1}$. $SO(n-1)$ is compact, so the $\{g^i\}$ have convergent subsequences. In addition, since the terms of this sequence are all iterates of a single element, we can, by adjusting terms of such a subsequence by suitable negative powers of $g$, have the subsequence converge to the identity. Choose a subsequence $\{i_j\}$ such that $g^{i_j+1} \rightarrow id$ as $j
\rightarrow \infty$. These $G_{i_j}(t)$ approach our original field $P(t)$ showing that $P$ makes constant curvature $-a^2$ with $\dot{\gamma}$ as well.
Consider the situation depicted in figure 1. Lemma 3.1 shows that, when $\gamma_{v_1}$ is recurrent in forward time, the map $p(v, v_1)$ preserves the distinguished vector fields in the sense that it sends a vector from one such field, $w_v(0)$, to a vector from another such field along $\gamma_{v_1}$. Thus, if in figure 1 we have that $\gamma_{v_1}$ and $\gamma_{v_3}$ are recurrent in positive time and $\gamma_v$ and $\gamma_{v_2}$ are recurrent in negative time, then the element of $H_v$ given by parallel translation around this ideal rectangle will map $w_v(0)$ to another element of $v^{\perp}$ which is in a parallel field along $\gamma_v$ making curvature $-a^2$. If these sort of recurrence properties held for all ‘equilateral’ ideal polygons based at $v$ we would have that the transitivity group preserves the distinguished vector fields. We cannot assure that these recurrence properties are always present, but ergodicity of the geodesic flow on $M$ indicates that they will be present almost all the time. We now work out the details of this.
First, the ergodicity of the geodesic flow implies that there is a full measure set of vectors $v$ in $T^1M$ which have dense forward and backward orbits under the geodesic flow. Choosing $v$ from this set implies that $\gamma_{v_1}$ will be recurrent in positive time and $\gamma_v$ will be recurrent in negative time. For ideal rectangles, this leaves only the positive time recurrence of $\gamma_{v_3}$ and the negative time recurrence of $\gamma_{v_2}$ lacking. It is convenient at this time to extend the definition of the transitivity group.
Consider the situation depicted in figure 2. Here the unit tangent vectors $v$, $v_1$, $v_2$ and $v_3$ describe an ideal rectangle in $T^1\tilde M$. Each pair $\{v, v_1\}$, $\{v_1, v_2\}$, $\{v_3, v\}$ lies on a leaf of $W^s_g$ or $W^u_g$ and we let $T\in
\mathbb{R}$ be the time such that $g_T(v_2) \in W^s_g(v_3)$ or $W^u_g(v_3)$. In order to make a true ideal rectangle we require that the leaves connecting these pairs alternate between stable and unstable. Note that in figure 2, the leaf containing the first pair, $\{v, v_1\}$ is a stable leaf, but we can similarly start with an unstable leaf. We now make the following definition:
Let $v$, $v_1$, $v_2$, $v_3$ be vectors in $T^1\tilde M$ describing an ideal rectangle as indicated above. Let $\tilde{F}_T$ be the restriction of the time $T$ frame flow map to the frames based at $v_2$. Note that the choice of $v_1$ and $v_3$ uniquely determines this rectangle and define $$h(v_1, v_3) = p(v_3, v) \circ p(g_T(v_2), v_3) \circ \tilde{F}_T \circ p(v_1, v_2)
\circ p(v, v_1).$$ Let $\hat{H}_v$ be the closure of the group generated by all such $h(v_1, v_3)$.
{height=".45\textheight" width=".6\textwidth"}
Note that $\hat{H}_v$ allows parallel translations along *all* ideal rectangles based at $v$. Furthermore, it is easy to see that a parallel translation around any ‘equilateral’ ideal polygon as in the definition of $H_v$ can be broken up into a series of translations around the general ideal rectangles allowed in the definition of $\hat{H}_v$. Thus, $\hat{H}_v \supseteq H_v$. However, $\hat{H}_v$ preserves ergodic components, as frame flow certainly preserves ergodic components. As the group $H_v$ is completely determined by the ergodic component of $n$-frames and $\hat{H}_v$ produces this same ergodic component, $\hat{H}_v \subseteq H_v$. Therefore, $\hat{H}_v = H_v$.
Despite the equality of these groups, for our purposes there is a benefit to allowing this seemingly looser definition. Each $(v_1, v_3) \in W^s_g(v) \times W^u_g(v)$ defines a rectangle used in $\hat{H}_v$. This is in opposition to the case for $H_v$, where only a set of measure zero define allowed rectangles. The advantage of this is the following. Since $M$ is negatively curved, the geodesic flow is ergodic and the unstable foliation $W^u_g$ is absolutely continuous. Thus, there is a full measure set of $v_1 \in T^1M$ with dense negative time orbit and it must intersect the leaf $W^s_g(v)$ in a set of full conditional measure for almost all $v \in T^1M$ (see Appendix to [@Brin-ergodicity] Lemma 5.4). Picking $v_1$ from this set will ensure the needed negative time recurrence of $\gamma_{v_2}$ since $\gamma_{-v_2} \to \gamma_{-v_1}$. Likewise, for almost every $v
\in T^1M$ a full conditional measure set of $v_3 \in W^u_g(v)$ will have the needed positive time recurrence under the geodesic flow. We conclude that we can find a full measure set of $v$ having dense forward and backward orbits *and* (using Fubini’s theorem) with a full measure set of $W^s_g(v) \times W^u_g(v)$ possessing the desired dynamical properties for $(-v_1, v_3)$. Therefore the desired recurrence properties are generic in the set of rectangles used to generate $\hat{H}_v$ for almost all $v$. The final fact needed to prove that the transitivity group preserves the distinguished vector fields is the following:
The map $(v_1, v_3) \mapsto h(v_1, v_3)$ is continuous.
This Lemma follows from the fact that the frame flow admits a continuous foliation, which was proved by Brin [@Brin-toptrans].
First, note that $p(v, v_1)$ and $p(v_3, v)$ are defined by leaves of the foliations for the frame flow. The continuous dependence of these maps on $(v_1, v_3)$ follows precisely from the continuity of the leaves.
Second, as $(v_1, v_3)$ varies, the leaves $W^u_g(v_1)$ and $W^s_g(v_3)$ and the geodesic connecting $\gamma_{v_1}(-\infty)$ to $\gamma_{v_3}(\infty)$ all vary continuously. Thus $v_2$, $T$ and $g_T(v_2)$ will vary continuously. Along with the argument of the previous paragraph, all this implies that the maps $p(v_1, v_2)$ and $p(g_T(v_2), v_3)$ depend continuously on $(v_1, v_3)$. Also, $\tilde{F}_T$ will depend continuously on $(v_1,
v_3)$ as the frame flow is continuous.
Since $h(v_1, v_3)$ is the composition of these maps, the Lemma is proved.
Now we can prove the following result:
\[prop:preserves\] For almost all $v \in T^1M$, if $w_v(t)$ is a parallel field along $\gamma_v(t)$ making constant curvature $-a^2$ with the geodesic direction, then for every element $h$ in the transitivity group we have $K(h(w_v(0)), v) = -a^2$.
As discussed above, for almost all $v \in T^1M$ a full measure set of the $(v_1, v_3) \in
W^s_g(v) \times W^u_g(v)$ give rectangles with the recurrence properties necessary for $h(v_1, v_3) \in \hat{H}(v)$ to map $w_v(0)$ to another distinguished vector field along $\gamma_v$. In particular, this set of ‘nice’ $(v_1, v_3)$ is dense in $W^s_g(v) \times
W^u_g(v)$. Since $h(v_1, v_3)$ depends continuously on $(v_1, v_3)$ and for a dense set of $(v_1, v_3)$, it preserves the distinguished fields, we have that all $h(v_1, v_3)$ preserve the distinguished fields. Since $\hat{H}_v = H_v$ is generated by these elements, the transitivity group preserves the curvature $-a^2$ as desired.
This result gives us the desired relationship between the transitivity group and the distinguished vector fields. We can now apply the results of Brin-Gromov and Brin-Karcher and prove Theorem 1 easily.
Let $M$ be a compact, negatively curved manifold. Suppose that along every geodesic in $M$ there exists a parallel vector field making sectional curvature $-a^2$ with the geodesic direction. If $M$ is odd dimensional, or if $M$ is even dimensional and satisfies the sectional curvature pinching condition $-\Lambda^2 < K < -\lambda^2$ with $\lambda/\Lambda > .93$ then $M$ has constant negative curvature equal to $-a^2$.
We showed in Proposition \[prop:preserves\] that for almost all $v \in T^1M$ the sectional curvature $K(h(w_v(0)), v) = -a^2$ for all $h$ in the transitivity group. In the setting of the theorem, the results of Brin-Gromov and Brin-Karcher tell us that the 2-frame flow is ergodic. In particluar, since the transitivity group gives the ergodic component for this flow, the transitivity group must act transitively on $v^\perp \subset
T^1M$. Thus, $K(\cdot, v)$ is identically $-a^2$. Since this holds for almost all $v$ it holds for all $v$ by continuity of $K(\cdot, \cdot)$, and the theorem is proved.
Parallel fields and Jacobi fields {#sec:parallel}
=================================
In [@sphericalrank] a distinction is made between ‘weak’ and ‘strong’ rank. The existence of *parallel* fields making extremal curvature is called strong rank; the existence only of *Jacobi* fields making extremal curvature is called weak rank. A parallel field can be scaled (by a solution to the real variable version of the Jacobi equation where the standard derivative replaces the covariant derivative) to produce a Jacobi field. Thus, a proof under the less stringent hypothesis of weak rank implies a proof for strong rank. Hamenstädt’s is the sole result prior to this paper for weak rank. She states her main theorem for parallel fields only, but she shows in Lemma 2.1 that in negative curvature a Jacobi field making maximal curvature is a parallel field scaled by a function [@Hamenstadt]. Essentially, she shows that Jacobi fields making maximal curvature grow at precisely the rate one finds for the constant curvature case. Connell accomplishes the same in [@Connell] Lemma 2.3. This, together with some of the arguments below, shows that these Jacobi fields are in fact parallel fields scaled by an appropriate function. Therefore, Corollary 2 is a weak rank result, needing only the Jacobi field hypothesis.
In this section we show that Jacobi fields making *minimal* curvature with the geodesic direction are also scaled parallel fields. This will justify the phrasing of Corollary 1 as a weak rank result.
First, note that we need only consider non-vanishing Jacobi fields; hence it will be enough to prove that stable and unstable Jacobi fields are scaled parallel fields. Stable Jacobi fields are those which have norm approaching zero as $t \to \infty$; unstable Jacobi fields have the same property in the negative time direction. Suppose $J(t)$ is a stable Jacobi field along the geodesic $\gamma(t)$ making curvature $-a^2$ with the geodesic (take $a>0$ now), where $-a^2$ is the curvature minimum for the manifold (the modifications of what follows for unstable Jacobi fields are straightforward). The Rauch Comparison Theorem (see [@doC] Chapt 10, Theorem 2.3) can be used to show that
$$\label{eq:geq}
|J(t)| \geq |J(0)| e^{-at}.$$
We would like to show that equality is achieved in (\[eq:geq\]). Write $J(t)=j(t)U(t)$ where $j(t)=|J(t)|$ and $U(t)$ is a unit vector field. Then the Jacobi equation for $J$ reads:
$$\label{eq:jacobi}
j''U + 2j'U' + jU'' + jR(\dot{\gamma}, U)\dot{\gamma} = 0$$
where $j'$ denotes the standard derivative and $U'$ denotes covariant derivative. Taking the inner product of (\[eq:jacobi\]) with $U$ and noting that $\langle U'', U\rangle =
- \langle U', U' \rangle$ we obtain
$$\label{eq:simpjacobi}
j'' - j(\langle U', U'\rangle +a^2) = 0.$$
We now know the following about the magnitude of $J$: $j \geq 0$ by definition, $\lim_{t
\to \infty} j(t) = 0$ since $J$ is a stable Jacobi field, and $j'' \geq a^2 j$ by (\[eq:simpjacobi\]). These allow the following conclusion; its proof was shown to the author by Jeffrey Rauch:
\[Rauchlemma\] Let $j$ be a non-negative, real valued function satifsying $j'' \geq a^2 j$ and $\lim_{t
\to \infty} j(t) = 0$. Then $j(t) \leq j(0) e^{-at}$.
We have that $a^2j-j'' \leq 0.$ On the interval $0 \leq t \leq R$ for $R\gg1$ define $g_R$ by $g_R(0)=j(0)$, $g_R(R)=j(R)$ and $a^2g_R-g_R'' = 0$. Note that as $R \to
\infty$, $g_R \to j(0)e^{-at}$. We claim that $j \leq g_R$; the Lemma follows in the limit.
This claim is essentially the maximum principle. First, $j \leq g_R$ holds at $0$ and $R$. Now suppose $j-g_R$ has a positive maximum at $c \in (0, R)$. Then $(j''-g_R'')(c)
\leq 0$. However, we know $a^2(j-g_R) - (j''-g_R'') \leq 0$, so a positive value of $j-g_R$ at $c$ together with a negative value of $j''-g_R''$ yields a contradiction. Therefore $j \leq g_R$ holds on all of $[0, R]$ as desired.
This Lemma, together with equation (\[eq:geq\]), tells us that $|J(t)| =
|J(0)|e^{-at}$. Examining equation (\[eq:simpjacobi\]) we see that having the growth rate $e^{-at}$, as in the constant curvature $-a^2$ case, implies that $U' =0$, that is, $J$ is a scaled parallel field, as desired.
The dynamical perspective {#sec:Lyap}
=========================
In this section we discuss how the results of Connell in [@Connell] can be adapted to prove Theorem 2 as a simple consequence of Corollary 1. The necessary changes are for the most part cosmetic; the discussion here is included for completeness, but the author does not claim to have added anything of substance to Connell’s work. The notation below that has not already been assigned follows Connell’s for ease of reference.
Recall that Lyapunov exponents are a tool for measuring long-term asympotic growth rates in dynamical systems (see Katok and Mendoza’s Supplement to [@H-K] section S.2 for an exposition). In the setting of the geodesic flow they can be defined as follows. Let $v
\in T^1M$ and $u \in v^{\perp}$. Let $J_u(t)$ be the unstable Jacobi field along $\gamma_v$ with initial condition $J_u(0) = u$. Then, the *positive Lyapunov exponent at $v$ in the $u$-direction* is $$\lambda_v^+(u) = \limsup_{t \to \infty} \frac{1}{t} log|J_u(t)|.$$ Define $$\lambda_v^+ = \max_{u \in v^{\perp}} \lambda_v^+(u).$$ This is the maximal Lyapunov exponent at $v$; the curvature bound $-a^2 \leq K$ (again, take $a>0$) implies that $\lambda_v^+ \leq a$. Let $$\Omega = \{v \in T^1M : \lambda_v^+ = a\}.$$ We can now rephrase Theorem 2 more succinctly.
Let $M$ be a compact manifold with sectional curvature $-a^2 \leq K < 0$. Suppose that $\Omega$ has full measure with respect to a geodesic flow-invariant measure $\mu$ with full support. If $M$ is odd dimensional, or if $M$ is even dimensional and satisfies the sectional curvature pinching condition $-a^2 \leq K < -\lambda^2$ with $\lambda/a > .93$ then $M$ is of constant curvature $-a^2$.
Connell shows in the upper rank case that the dynamical assumption implies the geometric one, that is, that the manifold in fact has higher rank, allowing the application of an appropriate rank rigidity theorem. He first shows ([@Connell] Proposition 2.4) that along a closed geodesic $\lambda_v^+ = a$ implies the existence of an unstable Jacobi field making curvature $-a^2$ with the geodesic direction. Essentially, if the Jacobi field giving rise to the Lyapunov exponent does not have this curvature, it will continually see non-extremal curvature a positive fraction of the time as it moves around the closed geodesic. This contradicts the supposed value of the Lyapunov exponent. The lower curvature bound version of the argument is exactly the same as that presented by Connell, with the proper inequalities reversed; also note that the work in section \[sec:parallel\] of this paper gives the results analogous to Connell’s Lemma 2.3 necessary for the argument.
It is clear that if a dense set of geodesics have the distinguished Jacobi fields, then all geodesics will. Since the velocity vectors for closed geodesics are dense in $T^1M$, Connell finishes his proof in section 3 of [@Connell] by showing that these vectors are all in $\Omega$ and using the argument of the previous paragraph. Adapted to the setting of Theorem 2 the argument runs as follows. If $w \in T^1M$ is tangent to a closed geodesic and $\lambda_w^+ < a$ the previous paragraph implies that any unstable Jacobi field along $\gamma_w$ must make curvature strictly greater than $-a^2$ for a positive amount of time. By continuity, this will also be true of any unstable Jacobi field along a geodesic $\gamma_v$ in a sufficiently small neighborhood of $\gamma_w$ (in the Sasaki metric on $T^1M$). The ergodic theorem implies that for a full measure set of $v \in T^1M$, $\gamma_v$ will spend a positive fraction of its life in this small neighborhood of the periodic geodesic $\gamma_w$; the positivity follows from the fact that $\mu$ has full support. The intersection of this full measure set with the full measure set $\Omega$ thus contains vectors $v$ which have $\lambda_v^+ = a$ but spend a positive fraction of their life so close to $\gamma_w$ that no Jacobi fields along them can make the minimal curvature $-a^2$ with the geodesic direction during this fraction of the time. In fact, since $\gamma_w$ is compact, so is the closure of this small neighborhood and therefore the curvature between these Jacobi fields and the geodesics, when in this neighborhood, can be bounded away from $-a^2$, i.e. $K(J_u,
\dot{\gamma}_v)>c>-a^2$ for a fixed $c$. Having this curvature bound a positive fraction of the time contradicts $\lambda_v^+ = a$; therefore all closed geodesics must lie in $\Omega$ and the argument is complete.
Again, this version of the argument, relevant for the lower curvature bound situation, is the same as that presented by Connell with the proper inequalities reversed. Thus, the dynamical assumption implies the geometric assumption of Corollary 1 and Theorem 2 follows. Note that for these arguments the extremality of the distinguished curvature is essential; a result that parallels Theorem 1 in allowing non-extremal distinguished curvature cannot be hoped for.
Conclusion {#sec:conclusion}
==========
We conclude with a few remarks on our results in the context of the other rank rigidity theorems. As noted above, Corollary 2 treats the case dealt with by Hamenstädt’s hyperbolic rank rigidity theorem, strong upper hyperbolic rank. Unlike Hamenstädt’s result, the result presented in this paper is limited by the curvature pinching condition in even dimension. However, this paper’s proof is shorter, and has the advantage of telling us which symmetric space $\tilde{M}$ is. Corollary 1 is strong lower hyperbolic rank rigidity, and this result is the first positive result for lower rank rigidity of any sort. Counterexamples to lower rank rigidity in other curvature settings are known; [@sphericalrank] presents an overview, together with counterexamples to weak upper and lower spherical rank rigidity. Counterexamples to lower Euclidean rank rigidity are given in [@Spatzier-Strake]. When the value $-a^2$ is not extremal we have a result of a different type than previous rank rigidity results. Our results also show that spaces of constant negative curvature can not be deformed while maintaining the distinguished parallel vector fields along all geodesics, or while maintaining extremal Lyapunov exponents at a full measure set of $T^1M$, except by scaling the metric.
Note that in even dimension a result as extensive as our odd dimensional result cannot be hoped for. Since parallel translation preserves the complex structure on a Kähler manifold the 2-frame flow will not be ergodic (see [@Brin-Gromov] for some results on unitary frame bundles). These known counterexamples to ergodic frame flow are excluded by requiring $-1<K<-1/4$, leading Brin to conjecture that strict 1/4-pinching implies that the frame flow is ergodic ([@Brin-survey] Conjecture 2.6). A positive answer to this conjecture, or any extended results for ergodicity of the 2-frame flow in negative curvature would extend the results on rank rigidity presented here correspondingly, using the same proof as presented above. One still hopes that lower hyperbolic rank rigidity (in the sense that higher rank implies the space is locally symmetric) could be true without any curvature pinching in even dimensions, perhaps even without the restriction $K < 0$, but such a result would call for a completely different method of proof from that presented here.
[10]{}
W. Ballmann. Nonpositively curved manifolds of higher rank. , 122:597–609, 1985.
W. Ballmann. . Birkhäuser, 1995.
M. Brin. Topological transitivity of one class of dynamical systems and frame flows on manifolds of negative curvature. , 9:8–16, 1975.
M. Brin. Topology of group extensions of anosov systems. , 18:858–864, 1975.
M. Brin. Ergodic theory of frame flows. In A. Katok, editor, [*Ergodic Theory and Dynamical Systems: proceedings, special year, Maryland 1979-1980*]{}, pages 163–183. Birkhäuser, 1981.
M. Brin and M. Gromov. On the ergodicity of frame flows. , 60:1–7, 1980.
M. Brin and H. Karcher. Frame flows on manifolds with pinched negative curvature. , 52(3):275–297, 1984.
K. Burns and R. Spatzier. Manifolds of nonpositive curvature and their buildings. , (65):35–59, 1987.
C. Connell. Minimal lyapunov exponents, quasiconformal structures and rigidity ofr manifolds of nonpositive curvature. , 23:429–446, 2003.
M. P. a. do Carmo. . Birkhäuser, 1992.
P. Eberlein and J. Heber. A differential geometric characterization of symmetric spaces of higher rank. , 71:33–44, 1990.
U. Hamenstädt. A geometric characterization of negatively curved locally symmetric spaces. , 34:193–221, 1991.
A. Katok and B. Hasselblatt. . Cambridge University Press, 1995.
K. Shankar, R. Spatzier and B. Wilking. Spherical rank rigidity and blaschke manifolds. , 128(1):65–81, 2005.
R. J. Spatzier and M. Strake. Some examples of higher rank manifolds of nonnegative curvature. , 65(2):299–317, 1990.
[^1]: Supported by NSF Graduate Research Fellowship
[^2]: <span style="font-variant:small-caps;">Department of Mathematics, University of Michigan, Ann Arbor, MI 48103 U.S.A.</span> *email:* `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Naturally-occurring granular materials often consist of angular particles whose shape and frictional characteristics may have important implications on macroscopic flow rheology. In this paper, we provide a theoretical account for the peculiar phenomenon of auto-acoustic compaction – non-monotonic variation of shear band volume with shear rate in angular particles – recently observed in experiments. Our approach is based on the notion that the volume of a granular material is determined by an effective-disorder temperature known as the compactivity. Noise sources in a driven granular material couple its various degrees of freedom and the environment, causing the flow of entropy between them. The grain-scale dynamics is described by the shear-transformation-zone (STZ) theory of granular flow, which accounts for irreversible plastic deformation in terms of localized flow defects whose density is governed by the state of configurational disorder. To model the effects of grain shape and frictional characteristics, we propose an Ising-like internal variable to account for nearest-neighbor grain interlocking and geometric frustration, and interpret the effect of friction as an acoustic noise strength. We show quantitative agreement between experimental measurements and theoretical predictions, and propose additional experiments that provide stringent tests on the new theoretical elements.'
author:
- 'Charles K. C. Lieou'
- 'Ahmed E. Elbanna'
- 'J. S. Langer'
- 'J. M. Carlson'
title: 'Shear flow of angular grains: acoustic effects and non-monotonic rate dependence of volume'
---
Introduction
============
The purpose of this paper is to explore the peculiar dynamics of a sheared granular material composed of angular grains which are shape-anisotropic and frictional in character. In doing so, we shall provide an explanation of the phenomenon of auto-acoustic compaction, recently observed in a series of experiments by van der Elst *et al.* [@brodsky_2012], in which the sample volume varies reversibly with the applied shear rate in a non-monotonic fashion. Specifically, their experiments found shear band volume reduction by up to $10\%$ at intermediate shear rates between the slow quasi-static and fast grain-inertial flow regimes for angular sand particles, but not for smooth glass beads, both in the presence and absence of tapping – forced, periodic vibrational excitation. The authors of that paper posit that shearing provides a source of acoustic energy that un-jams a granular material and allows the granular medium to explore packing configurations. At intermediate shear rates, acoustic vibrations result in a denser packing, similar to compaction due to externally-driven vibrations [@johnson_2008; @daniels_2005; @daniels_2006; @nowak_1998; @knight_1995]. Other experiments have also found non-monotonic flow rheology in granular media composed of shape-anisotropic grains [@brodsky_2007; @dijksman_2011]. An understanding of the effect of acoustic phenomena in sheared granular flow is especially important in the context of earthquakes, which generate seismic waves that propagate to gouge-filled faults in the vicinity and may cause dramatic reduction in shear strength [@melosh_1979; @melosh_1996].
Our analysis is based primarily on the idea that nonequilibrium states of a granular material are characterized by its compactivity $$\label{eq:compactivity_def}
X = \left( \dfrac{\partial V}{\partial S_C} \right)_{\Lambda_{\alpha}} ,$$ or equivalently, its effective disorder temperature $T_{\text{eff}} = p X$ [@edwards_1989a; @edwards_1989b; @edwards_1989c; @edwards_1990a; @edwards_1990b]. (Here $V$ is the extensive volume of the system, $S_C$ is the configurational entropy, and the $\Lambda_{\alpha}$’s are internal variables that specify the configurational state of the granular subsystem.) The state variable $X$ is a thermodynamically well-defined quantity; we naturally assume that the observable volume $V$ is a function of $X$. In this spirit, we assume that $X$ is determined by an entropy-flow equation that is consistent with the second law of thermodynamics. The compactivity $X$ is increased by external work done on the system, and decreases when entropy flows from the granular system into its environment. This heat flow is governed by various noise sources within the system: noise generated by the driving forces, noise generated by friction between particles, etc. Thus, the disorder temperature and the volume are determined by the interplay between these dissipative effects. Specifically, under generic assumptions regarding its functional form, the frictional noise can describe the competition between shear-induced dilation and acoustic compaction, thereby explaining the non-monotonic variation of sample volume with shear rate.
In the present investigation, the microscopic model for effective-temperature dynamics, grain interactions, and the driving forces in the system is the shear-transformation-zone (STZ) theory of granular flow, originally developed to study shear flow in amorphous molecular solids [@falk_1998; @langer_2011]. In the context of granular media, the STZ theory has been invoked to account for constitutive friction laws in earthquake physics [@daub_2010], glassy phenomena in sheared hard-sphere systems [@lieou_2012], and formation of a finely comminuted gouge layer in fault materials [@lieou_2014]. Under the STZ theoretical framework, plastic deformation can be explained in terms of localized flow defects, or STZ’s, whose density is characterized by the compactivity $X$.
Prior applications of the STZ theory made no assumptions regarding the shape and characteristics of constituent grains. The van der Elst *et al.* experiments, however, clearly demonstrate the significance of grain shape and frictional characteristics in granular flow rheology. The goal of this paper is to provide a quantitative description of these effects. Our proposition is that the large variation of volume is the result of geometric frustration, or lack thereof, between neighboring grains, modeled in terms of an Ising-like internal variable. When coupled with the interpretation of frictional dissipation between particles as a kind of noise, it is possible to quantitatively account for the observed non-monotonic variation of sample volume with shear rate in a granular medium with angular particles.
The rest of our paper is structured as follows. In Sec. II, we repeat the statistical-thermodynamic analysis largely along the lines of [@lieou_2012; @lieou_2014], but incorporate the effect of tapping and inter-particle frictional dissipation. Specifically, we introduce a “frictional noise” which couples the fast, kinetic, and slow, configurational, degrees freedom, and accounts for how frictional dissipation may cause the steady-state sample volume to vary non-monotonically with shear rate. In Sec. III we introduce our microscopic model that describes how the volume varies with the compactivity. The model is a combination of STZ’s and misalignment defects, the latter of which are described by an extra Ising-like internal variable that characterizes the shape effect in terms of grain orientation, interlocking, and geometric frustration. Then, in Sec. IV we present our theoretical predictions which quantitatively match the experimental measurements of van der Elst *et al*. We conclude the paper in Sec. V with a list of proposed experiments and future directions.
Theoretical formulation for volume variation {#sec:2}
============================================
Statistical thermodynamics
--------------------------
As in prior applications of STZ theory [@lieou_2012; @lieou_2014], it is important to quantify the interaction between different components of the granular system; to this end, we turn to the laws of thermodynamics. The developments in this section largely mirror those of [@lieou_2012; @lieou_2014].
Consider a noncrystalline system of hard grains at temperature $T$, with total energy $U_T$. For simplicity we use a single state variable $T$ to characterize all macroscopic and microscopic kinetic-vibrational degrees of freedom of the grains, assumed to be in contact with a thermal reservoir. A consequence of this simplification is that frictional dissipation, among other forms of inelastic grain interaction, simply amounts to the flow of energy from the macroscopic to the microscopic degrees of freedom [@leff_1992; @leff_1993]. It is thus unnecessary to account for friction explicitly in the overall energy balance. This simplification is especially convenient in view of our characterization of inter-particle friction as a kind of noise below. For practical purposes, $T$ can be interpreted as a measure of material preparation. Because grains interact only via contact forces, there is no configurational potential energy, so $U_T$ equals the total energy of the system.
Suppose that this system is driven by a shear stress $s$ in the presence of a pressure $p$. The first law of thermodynamics for this system is $$\begin{aligned}
\label{eq:first_law}
\nonumber \dot{U}_T &=& V s\, \dot{\gamma}^{\text{pl}} - p\, \dot{V} \\ &=& V s\, \dot{\gamma}^{\text{pl}} - p X \dot{S}_C - p \sum_{\alpha} \left( \dfrac{\partial V}{\partial \Lambda_{\alpha}} \right)_{S_C} \dot{\Lambda}_{\alpha}, ~~~~~\end{aligned}$$ where $\dot{\gamma}^{\text{pl}}$ is the plastic shear rate. As in Eq. (\[eq:compactivity\_def\]), $S_C$ is the granular configurational entropy, and the $\Lambda_{\alpha}$ are internal variables that specify the configurational state of the granular subsystem.
Let $S_T$ denote the entropy of all kinetic-vibrational degrees of freedom. Then $$\dot{U}_T = T \dot{S}_T ,$$ and $$\label{eq:S_C}
p X \dot{S}_C = V s \dot{\gamma}^{\text{pl}} - p \sum_{\alpha} \left( \dfrac{\partial V}{\partial \Lambda_{\alpha}} \right)_{S_C} \dot{\Lambda}_{\alpha} - T \dot{S}_T .$$ The second law of thermodynamics states that the total entropy of the system, being the sum of all kinetic-vibrational and configurational degrees of freedom, must be a non-decreasing function of time: $$\label{eq:second_law}
\dot{S} = \dot{S}_C + \dot{S}_T \geq 0 .$$ Substituting Eq. for $\dot S_C$ into the second law above, and using the fact that each individually variable term in the resulting inequality must be non-negative [@coleman_1963; @langer_2009a; @langer_2009b; @langer_2009c], we arrive at the second-law constraints $$\begin{aligned}
\label{eq:W} {\cal W} = V s\, \dot{\gamma}^{\text{pl}} - p \sum_{\alpha} \left( \dfrac{\partial V}{\partial \Lambda_{\alpha}} \right)_{S_C} \dot{\Lambda}_{\alpha} \geq 0; \\
\label{eq:S_T} (p X - T) \dot{S}_T \geq 0 .\end{aligned}$$ In arriving at these two constraints, we have arranged terms in such a way that terms pertaining to the degrees of freedom that belong to the same subsystem are grouped together. The dissipation rate ${\cal W}$, as defined in [@langer_2009a; @langer_2009b; @langer_2009c], is the difference between the rate at which inelastic work is done on the configurational subsystem and the rate at which energy is stored in the internal degrees of freedom. The second constraint implies that $p X - T$ and $\dot{S}_T$ must carry the same sign if they are nonzero, so that $$\label{eq:Q}
T \dot{S}_T = - {\cal K} (X, T) \left( T - p\, X \right)\equiv \,{\cal Q},$$ where ${\cal K} (X, T)$ is a non-negative thermal transport coefficient. It is already clear from this analysis that $p\,X$ plays the role of a temperature. $p\,X$ approaches $T$ in an equilibrating system; and a heat flux ${\cal Q}$ flows between the granular subsystem and the reservoir when the two subsystems are not in thermodynamic equilibrium with each other.
Steady-state compactivity as a function of strain rate
------------------------------------------------------
Let us introduce the dimensionless strain rate, or the inertial number (see for example [@jop_2006]) $$q \equiv \tau \dot{\gamma}^{\text{pl}} ,$$ where $\tau$ is the inertial time scale, to be discussed in greater detail in Sec. III. (In the steady state, the total, imposed shear rate $\dot{\gamma}$ equals the plastic strain rate $\dot{\gamma}^{\text{pl}}$.) We also define the dimensionless compactivity $$\chi \equiv X / v_Z ,$$ where $v_Z$ is the excess volume associated with STZ’s – rare, non-interacting loose spots where irreversible particle rearrangements occur. (In the solidlike – as opposed to hydrodynamic – regime, nonaffine particle displacements occur everywhere. However, 1STZ’s refer to the subset of those that are irreversible and involve local topological change, so our picture of dilute defects remains valid.) The dimensionless compactivity $\chi$ measures the amount of configurational (i.e. structural) disorder in the granular system. Intuitively, $\chi$ increases monotonically with the extensive volume $V$ of the system. It is obvious that external vibrations and inter-particle dissipative mechanisms such as friction play important roles in controlling the configurational state of the granular medium. However, in the absence of these mechanisms, as in hard-sphere systems with zero vibrational noise strength [@lieou_2012], the steady-state compactivity ought to be a function of the strain rate alone: $\chi = \hat{\chi} (q)$. As seen in our hard-sphere analysis and in simulations [@lieou_2012; @liu_2011; @haxton_2012], $\hat{\chi} (q)$ approaches some constant $\hat{\chi}_0$ in the limit of small $q$. On the other hand, $\hat{\chi} (q)$ becomes a rapidly increasing function of $q$ once the shear rate becomes comparable to the rate of intrinsic structural relaxation, to reflect shear-rate dilation in hard-sphere systems. It is customary to write the inverse relation $q (\hat{\chi})$ in the Vogel-Fulcher-Tamann (VFT) form [@lieou_2012; @haxton_2007; @manning_2007b] $$\label{eq:vft}
\dfrac{1}{q} = \dfrac{1}{q_0} \exp \left[ \dfrac{A}{\hat{\chi}} + \alpha_{\text{eff}} (\hat{\chi}) \right] ,$$ where $$\alpha_{\text{eff}} (\hat{\chi}) = \left( \dfrac{\hat{\chi}_1}{\hat{\chi} - \hat{\chi}_0} \right) \exp \left( - 3 \dfrac{\hat{\chi} - \hat{\chi}_0}{\hat{\chi}_A - \hat{\chi}_0} \right) .$$
The quantity $\chi$ evolves according to the first law of thermodynamics. To deduce its equation of motion, return to Eq. for the rate of entropy change $\dot{S}_C$ of the configurational subsystem, let us invoke the quasistationary approximation $\dot{\Lambda}_{\alpha} = 0$ for each internal variable $\Lambda_{\alpha}$, and use Eq. to eliminate $\dot{S}_T$. The result is $$\label{eq:S_C2}
p X \dot{S}_C = V s \dot{\gamma}^{\text{pl}} - {\cal K} (X, T) (p X - T) .$$ To convert this into an equation for the dimensionless compactivity $\chi$, we use the scaling $\chi = X / v_Z$ and $\theta \equiv T / p \, v_Z$. Then, we use the relation $$\begin{aligned}
\label{eq:chiscdot}
\nonumber \chi \dot{S}_C &=& \chi \left( \dfrac{\partial S_C}{\partial \chi} \right)_{\Lambda_{\alpha}} \dot{\chi} + \chi \sum_{\alpha} \left( \dfrac{\partial S_C}{\partial \Lambda_{\alpha}} \right)_{S_C} \dot{\Lambda}_{\alpha} \\ &=& \chi \left( \dfrac{\partial S_C}{\partial \chi} \right)_{\Lambda_{\alpha}} \dot{\chi},\end{aligned}$$ where in the second equality we again used the quasistationary approximation to eliminate the time derivatives of other internal variables. Now comes the crucial step: since the transport coefficient ${\cal K} (\chi, \theta)$ couples the configurational and kinetic-vibrational subsystems, it should consist of additive mechanical, vibrational and frictional contributions. Specifically, $$\label{eq:A}
{\cal K} (\chi, \theta) = \dfrac{V}{\tau} \, {\cal A} \, \left( \Gamma + \xi + \rho \right) .$$ Here, $\Gamma$ is the mechanical noise that pertains to externally applied shear. It will be computed below in Eq. in terms of the rate of entropy generation, and we will show that it is proportional to the work of plastic deformation, i.e., the tensor product $s \dot{\gamma}^{\text{pl}}$ of the shear stress and the plastic strain rate. On the other hand, the dimensionless quantity $\rho$ is a measure of the intensity of externally-imposed acoustic-vibrational motion of the grains [@brodsky_2012], more generally known as tapping. Tapping provides a means to un-jam a granular system so that it can explore packing configurations. In granular experiments, acoustic vibrations have been found to increase the packing fraction [@daniels_2005; @daniels_2006; @nowak_1998; @knight_1995], and trigger stick-slip motion under shear [@johnson_2008]. When $\rho = 0$, the system is fully jammed in the sense that configurational rearrangements can occur only in response to sufficiently large driving forces. In addition, $\xi$ is the system-specific frictional coupling or noise, to be determined based on phenomenology. In contrast to prior STZ analyses in which no assumption whatsoever was made in regard to the dissipative nature of particle interaction [@lieou_2012; @langer_2009c; @langer_2008], the $\rho$ term is replaced by $\xi + \rho$ to reflect the importance of frictional dissipation. ${\cal A}$ is a non-negative quantity to be determined by appealing to the steady-state solution in special cases. We also implicitly subsume all time scales relevant to tapping and friction under the inertial time scale $\tau$ in Eq. .
In anticipation of Pechenik’s hypothesis in Eq. below, we rewrite the first term in Eq. : $s \dot{\gamma}^{\text{pl}} = ( \Gamma / \tau) {\cal B}$, where ${\cal B}$ is a constant. Then, some simple algebra, along with use of Eqs. and , reduce Eq. to $$c^{\text{eff}} \dot{\chi} = \Gamma {\cal B} - {\cal A} \left( \Gamma + \xi + \rho \right) (\chi - \theta),$$ with $c^{\text{eff}}$ being a scalar quantity that describes the capacity of volume dilation. We argued above that in the absence of vibration or inelastic dissipation, $\rho = \xi = 0$, the steady-state compactivity is uniquely determined as $\chi^{\text{ss}} = \hat{\chi} (q)$. This gives ${\cal A} = {\cal B} / (\hat{\chi} (q) - \theta )$ whose non-negativity incidentally implies the constraint that $\hat{\chi} (q) > \theta$, i.e., stirring the system drives the slow, configurational degrees of freedom out of equilibrium with the fast, kinetic-vibrational degrees of freedom. Then, in general, the steady-state compactivity is given by $$\label{eq:X_ss}
\chi^{\text{ss}} = \dfrac{\Gamma \hat{\chi}(q) + (\xi + \rho) \theta}{\Gamma + (\xi + \rho)} .$$
As we alluded to above, the extensive volume $V$ of the system – measured in [@brodsky_2012] by the change in shear band thickness – is an increasing function of the compactivity $\chi$. The explicit functional form of $V(\chi)$ will be discussed later in Sec. III in the context of a microscopic model and internal state variables. Thus Eq. , in effect, describes the variation of system volume with shear rate and noise strength.
With $\hat{\chi}(q)$ being an increasing function of the dimensionless strain rate $q$, how can we understand the non-monotonic variation of shear band thickness – and therefore the compactivity $\chi$ – with shear rate, as observed in the experiments of van der Elst *et al.* [@brodsky_2012], within Eq. ? Specifically, can we account for the decrease in $\chi$ at intermediate strain rates, and shear-rate dilation at large $q$? The answer lies in the frictional noise term $\xi$. Intuitively, $\xi$ should be a scalar function of the plastic work of shearing; thus $\xi = \xi (\Gamma)$. It induces correlations between particle velocities and enhances non-local effects [@kamrin_2014]. Let us first focus on the case when vibrations are absent. In the limit of vanishingly small strain rate $q$, the mechanical noise $\Gamma \rightarrow 0$; Eq. then shows that $\chi^{\text{ss}} \rightarrow \theta$ provided that $\xi \neq 0$. Because $\hat{\chi} (q)$ increases monotonically with $q$ and exceeds $\theta$, $\chi^{\text{ss}}$ must also be an increasing function of $q$, which is contrary to the non-monotonicity of volume variation with shear rate. A resolution to this dilemma is that $\xi (\Gamma) = 0$ at zero shear rate, to reflect the fact that friction does not dissipate energy when no slipping takes place. In fact, if $\xi (\Gamma) \rightarrow 0$ faster than $\Gamma$ at small shear rates – say, if $\xi (\Gamma) \sim \Gamma^2$ for small $\Gamma$, as in Newtonian friction – then $\chi^{\text{ss}} \rightarrow \hat{\chi} (q = 0) = \hat{\chi}_0$ in that limit. (Indeed, if we interpret $\xi$ as some kind of energy, in analogy to $\Gamma$, then because the energy associated with an inelastic collision between grains is proportional to the square of their relative velocity, it is plausible for $\xi (\Gamma) \sim \Gamma^2$.) As $q$ increases so that $\xi (\Gamma)$ becomes large enough, it is possible for the steady-state compactivity $\chi^{\text{ss}}$ to fall below $\hat{\chi} (q)$: $\theta < \chi^{\text{ss}} < \hat{\chi} (q)$, because $\hat{\chi} (q) > \theta$.
In the opposite limit of large shear rate, the experiments indicate that shear-induced dilation must once again dominate, and that inter-particle friction becomes less important. Thus we stipulate that $\xi (\Gamma)$ saturates at large $\Gamma$ so that $\chi^{\text{ss}} \rightarrow \hat{\chi} (q)$ in Eq. . We now check that this assumption is consistent with experimental findings in the presence of tapping. For small $q$, Eq. with $\rho \neq 0$ indicates that $\chi^{\text{ss}} \rightarrow \theta < \hat{\chi}_0$, so that tapping does increase the packing fraction in the slow quasistatic limit. On the other hand, the boundedness of both $\xi$ and $\rho$ shows that $\chi^{\text{ss}} \rightarrow \hat{\chi}(q)$ in the fast inertial regime, independent of friction and tapping, and coinciding with the $\rho = 0$ behavior as seen by the overlap of the two curves in Fig. \[fig:Vq\] below in that limit.
Having argued that non-monotonic variation of volume with shear rate is indeed possible, let us now turn our attention to formulating a microscopic model that accounts for the flow rheology of angular grains, and quantifies volume variation with configurational disorder.
Microscopic model
=================
Internal variables and system volume
------------------------------------
The extensive volume $V$ of the granular packing plays a central role in this paper. As such, we characterize the volume in terms of the configurational state of the system, and derive equations of motion for the corresponding internal variables. We emphasize that the non-monotonic variation of volume with shear rate as a result of inter-particle friction is not unique to the microscopic model to be introduced here. In fact, any model for which the volume $V$ varies monotonically with the compactivity $X$ ought to qualitatively describe this non-monotonicity. However, the value of our model lies in its ability to provide a physical account of the relationship between grain-scale configuration and volume, as well as a good quantitative fit to the experimental data. Readers who are not interested in the microscopic details may skip directly to Sec. III. E, where we summarize the formulas that will be used in the ensuing analysis.
Recall our physical picture that in dense granular flow, irreversible particle rearrangements occur at rare, non-interacting soft spots with excess free volume known as STZ’s. The applied shear stress defines a direction relative to which STZ’s can be classified according to orientation, with total numbers $N_+$ and $N_-$ respectively. Upon application of shear stress in the “plus” direction, STZ’s of the minus type easily deform to become plus-type STZ’s. However, plus-type STZ’s rarely flip and acquire the minus orientation; rather, they are annihilated readily by noise. If the total number of grains equals $N$, we define the intensive variables $$\Lambda = \dfrac{N_+ + N_-}{N}; \quad m = \dfrac{N_+ - N_-}{N_+ + N_-}$$ as the density and orientational bias of STZ’s.
On the other hand, in angular grains, shape anisotropy and geometric frustration allows for the distinct possibility for grains interlocking, which reduces local volume, independent of the presence of localized slip events. Said differently, because of the absence of infinite-fold symmetry, neighboring grains that do not align with one another contribute excess volume. The simplest way to describe this is to assume that there is an extra Ising-like order parameter, $\eta$, that describes grain orientation (this orientation is independent of the direction of shear stress). Specifically, let $N^G_+$ and $N^G_-$ denote the number of grains in each of the two orientations, and define $$\label{eq:eta_def}
\eta = \dfrac{N^G_+ - N^G_-}{N} .$$ Of course, $-1 \leq \eta \leq 1$, as it should. Unlike STZ’s which are rare, isolated defects, each grain is associated with a particular direction; that is, $N^G_+ + N^G_- = N$.
Denote by $v_Z$ and $v_a$ the excess volumes associated with STZ’s and misalignments. The assumption that the system volume does not depend on STZ orientation, but depends on nearest-neighbor interactions in an Ising-like manner similar to the binary clusters recently invoked in a model of glass transition [@langer_2013], allows us to write down the extensive volume of the system as follows: $$\begin{aligned}
\label{eq:V}
\nonumber V &=& V_0 + N \Lambda v_Z - N \eta^2 v_a + V_1 (S_1) \\\nonumber &=& V_0 + N \Lambda v_Z - N \eta^2 v_a \\ & & + V_1 (S_C - S_Z (\Lambda, m) - S_G (\eta) ). ~~~~~\end{aligned}$$ Here, $V_0 = N a^3$ is the total volume of grains. $S_C$ is the configurational entropy, consisting of the entropy $S_Z$ associated with STZ’s, $S_G$ associated with grain orientation, and $S_1$ for all other configurational degrees of freedom. Correspondingly, $V_1$ is the volume associated with those degrees of freedom, to be discussed in greater detail towards the end of this Section. Then, under the assumption the STZ’s and grain alignments are two-state entities, we can compute $S_Z$ and $S_G$ easily by counting the number of possible configurations of distributing $N_+$ and $N_-$ STZ’s of each orientation, and $N^G_+$ and $N^G_-$ orientation states for each grain, among $N$ sites [@langer_2009c]. The result is $$\begin{aligned}
\label{eq:S_Z} S_Z (\Lambda, m) &=& N S_0 (\Lambda) + N \Lambda \psi (m);\\
\label{eq:S_G} S_G (\eta) &=& N \psi (\eta) ,\end{aligned}$$ where $$\begin{aligned}
\label{eq:S_0} S_0 (\Lambda) &=& - \Lambda \ln \Lambda + \Lambda;\\
\label{eq:psi}\nonumber \psi(m) &=& \ln 2 - \dfrac{1}{2} (1 + m) \ln (1 + m) \\ & & - \dfrac{1}{2} (1 - m) \ln (1 - m).\end{aligned}$$
Equations of motion
-------------------
In order to study the dynamics of the system, the preceding analysis needs to be supplemented with equations of motion for each of the internal variables. We first look at STZ dynamics. As usual, the STZ equation of motion for $N_+$ and $N_-$ is given by: $$\label{eq:master}
\tau \dot{N}_{\pm} = {\cal R} (\pm s) N_{\mp} - {\cal R} (\mp s) N_{\pm} + \tilde{\Gamma} \left( \dfrac{1}{2} N^{\text{eq}} - N_{\pm} \right) .$$ The corresponding strain rate is $$\label{eq:strainrate}
\dot{\gamma}^{\text{pl}} = \dfrac{2\,v_0}{\tau V} \left[ {\cal R}(s) N_- - {\cal R}(-s) N_+ \right],$$ where we define the volume of the plastic core of an STZ to be $v_0$, which ought to be proportional to $a^3$. Because we are describing simple rather than pure shear, there is a factor of $2$ up front.
Some comments on the various quantities that appear in Eqs. and are now in order. The quantity $\tau = a \sqrt{\rho_G / p}$, where $\rho_G$ denotes the material density of the grains, is the inertial time scale that characterizes the typical duration of a pressure-driven particle rearrangement event [@lieou_2012; @lieou_2014; @liu_2011; @haxton_2012]. It is proportional to the average time between successive grain-grain collisions. This time scale also applies in a dense granular medium where inter-particle friction is important, as long as the friction is proportional to the normal force at the contact interface. Its product with the shear rate $\dot{\gamma}^{\text{pl}}$ gives the so-called inertial number, the magnitude of which determines the flow regime of dense granular flow [@jop_2006]. ${\cal R}(\pm s)$ represent the rates (in units of $\tau^{-1}$) at which the STZ’s are making forward and backward transitions. The term proportional to $\tilde{\Gamma}$ represents the rates of STZ creation and annihilation; $N^{\text{eq}}$ is the steady-state, total number of STZ’s. Specifically, $\tilde{\Gamma}/\tau$ is an attempt frequency consisting of additive vibrational and mechanical parts: $$\tilde{\Gamma} = \rho +\Gamma.$$ Because friction should play no role in the creation or annihilation of STZ’s, $\xi$ does not appear in the expression for $\tilde{\Gamma} / \tau$.
After making these remarks on the elements of the extensive STZ equations of motion, we can rewrite them exclusively in terms of the intensive state variables $\Lambda$ and $m$ as follows: $$\begin{aligned}
\label{eq:Lambda} \tau\, \dot{\Lambda} &=& \tilde{\Gamma} ( \Lambda^{\text{eq}} - \Lambda ) ; \\
\label{eq:m} \tau\, \dot{m} &=& 2\, {\cal C}(s) ( {\cal T}(s) - m ) - \tilde{\Gamma} m - \tau \dfrac{\dot{\Lambda}}{\Lambda} m ; ~~~~~ \\
\label{eq:D_pl} \tau \,\dot{\gamma}^{\text{pl}} &=&2\, \epsilon_0\,\Lambda\, {\cal C}(s) ({\cal T}(s) - m),\end{aligned}$$ where $\epsilon_0 = N\, v_0 / V_0$ is independent of $a$, and $\Lambda^{\text{eq}} = N^{eq} / N$. In writing Eq. we have implicitly made the approximation $\Lambda \ll 1$ and $V_1 \ll V_0$ in Eq. so that $V \approx V_0$ in Eq. . We also define $$\label{eq:calC}
{\cal C}(s) = \dfrac{1}{2} \left( {\cal R}(s) + {\cal R}(-s) \right) ,$$ and $${\cal T}(s) = \dfrac{{\cal R}(s) - {\cal R}(-s)}{{\cal R}(s) + {\cal R}(-s)} .$$
In analogy to the master equation for STZ transitions, we propose that the simplest possible equation of motion that describes the change in the number of grains is of the form $$\label{eq:NG}
\tau \dot{N}_{\pm}^G = {\cal R}_{\pm}^G N_{\mp}^G - {\cal R}_{\mp}^G N_{\mp}^G .$$ Here, ${\cal R}_{\pm}^G$ is a yet-to-be specified rate factor for the transition between orientations, absorbing all other relevant time scales such as the inverse tapping frequency; we expect that it is proportional to the sum of mechanical and vibrational noise strengths $\tilde{\Gamma}$. A key difference between Eq. and its counterpart, Eq. for STZ transitions, is the absence of an extra term which, in Eq. , describes the creation and annihilation of STZ’s and the approach of STZ density to an equilibrium value. The reason behind this is two-fold. Firstly, as we alluded to before, every grain must belong to either one of the two orientations, but a given grain need not be part of an STZ at a given instant. Secondly, the effect of noise is already accounted for in the rate factor ${\cal R}_{\pm}^G$ which is not directly related to the direction of the shear stress $s$. With this, the STZ equations of motion is supplemented by an extra equation for the temporal evolution of grain orientational bias $\eta$: $$\label{eq:eta}
\tau \dot{\eta} = 2 {\cal C}^G ({\cal T}^G - \eta),$$ where $${\cal C}^G = \dfrac{1}{2} ( {\cal R}^G_+ + {\cal R}^G_- ); \quad {\cal T}^G = \dfrac{{\cal R}^G_+ - {\cal R}^G_-}{{\cal R}^G_+ + {\cal R}^G_-} .$$
Dissipation rate and thermodynamic constraints
----------------------------------------------
At this point in the development, the second law of thermodynamics provides useful constraints on various ingredients of the equations of motion, and on steady-state dynamics of the system. To this end, we now substitute this and Eqs. , , , and into Eq. for the dissipation rate ${\cal W}$ which, according to the second law of thermodynamics, must be non-negative. We also use the approximation $V \approx V_0$ where appropriate. The result is $$\begin{aligned}
\label{eq:W2}
\nonumber \tau\, \dfrac{{\cal W}}{N} &=& - \tilde{\Gamma}\, p\, X\, \Lambda\, m \dfrac{d \psi}{dm} - p\, \tilde{\Gamma}\,(\Lambda^{\text{eq}} - \Lambda) \\\nonumber && \times \left[ v_Z + X \left( \ln \Lambda - \psi (m) + m \,\dfrac{d \psi}{dm} \right) \right] \\\nonumber && + 2\, \Lambda\, {\cal C}(s)\,\Bigl( {\cal T}(s) - m\Bigr) \left( v_0 s + p X \dfrac{d \psi}{dm} \right) \\ && + 2\, {\cal C}^G \, \Bigl( {\cal T}^G - \eta \Bigr) \left( X \dfrac{d \psi}{d \eta} + 2\, \eta \, v_a \right) . ~~~~~\end{aligned}$$ The second-law constraint, ${\cal W} \geq 0$, must be satisfied by all possible motions of the system; this is guaranteed if each of the four terms in Eq. is non-negative [@coleman_1963; @langer_2009a; @langer_2009b; @langer_2009c]. (Indeed, only the third term depends explicitly on the shear stress $s$, while the second term is proportional to $\dot{\Lambda}$, and $\eta$ appears only in the fourth term. The entire expression must be non-negative irrespective of $s$ and $\dot{\Lambda}$, and $\eta$.) The first term automatically satisfies this requirement because, from Eq. , we have $$\label{eq:dpsidm}
\dfrac{d \psi}{dm} = - \dfrac{1}{2} \ln \left( \dfrac{1 + m}{1 - m} \right) = - \tanh^{-1} (m) ,$$ so that the product $- m ( d \psi / dm)$ is automatically non-negative.
The non-negativity constraint on the second term in Eq. can be written in the form $$- \dfrac{\partial F}{\partial \Lambda} (\Lambda^{\text{eq}} - \Lambda) \geq 0 ,$$ where $F$ is a free energy given by $$F (\Lambda, m) = p \left[ v_Z \Lambda - X S_0 (\Lambda) - X \Lambda \left( \psi(m) - m \dfrac{d \psi}{d m} \right) \right].$$ $\Lambda^{\text{eq}}$ must be the value of $\Lambda$ at which $ \partial F / \partial \Lambda$ changes sign, so that $$\label{eq:Lambda_eq}
\Lambda^{\text{eq}} = \exp \left[ - \dfrac{v_Z}{X} + \psi(m) - m \dfrac{d \psi}{dm} \right] \approx 2\, \exp \left( - \dfrac{v_Z}{X} \right). ~~~~~$$ Thus, the STZ density in this non-equilibrium situation is given by a Boltzmann-like expression in which the compactivity plays the role of the temperature.
As for the third term, we have $$\label{eq:Ts}
\Bigl( {\cal T}(s) - m \Bigr) \left( v_0 s + p X \dfrac{d \psi}{dm} \right) \geq 0.$$ The two factors on the left-hand side must be monotonically increasing functions of $s$ that change sign at the same point for arbitrary values of $m$. According to Eq. , this is possible only if $${\cal T}(s) = \tanh \left( \dfrac{v_0 s}{p X} \right).$$
Finally, the extra constraint associated with the new internal variable $\eta$ is $$2 {\cal C}^G ({\cal T}^G - \eta) \left( X \dfrac{d \psi}{d \eta} + 2 \eta v_a \right) \geq 0.$$ In the same spirit as for the third term, this holds if and only if each of the two quantities in the pair of parenthesis change sign at the same value of $\eta$; thus $${\cal T}^G = \tanh \left( \dfrac{2 \eta v_a}{X} \right) .$$ According to Eq. , the steady-state value of $\eta$ is then given by the solution to the equation $$\label{eq:eta_eq}
\eta^{\text{eq}} = \tanh \left( \dfrac{2 \eta^{\text{eq}} v_a}{X} \right).$$ This is reminiscent of the familiar spontaneous symmetry breaking in the magnetization of an Ising ferromagnet. If $X > X_c \equiv 2 v_a$, then the only solution to Eq. is $\eta^{\text{eq}} = 0$; this applies to a “dilute” granular packing, for which interlocking is no longer important. On the other hand, if $X < X_c$, there are two non-zero solution in the steady state, $\eta^{\text{eq}} = \pm \eta_0$, the absolute value of which decreases with increasing $X$ in this regime.
Mechanical noise and steady-state dynamics
------------------------------------------
In quasistationary or steady-state situations such as the experiments by van der Elst *et al.* [@brodsky_2012] that we analyze in this paper, a major simplification comes from setting $\dot{\Lambda} = \dot{m} = \dot{\eta} = 0$, so that $\Lambda = \Lambda^{\text{eq}}$ and $\eta = \eta^{\text{eq}}$. To determine the mechanical noise strength $\Gamma$ and the stationary value of $m$, we invoke Pechenik’s hypothesis [@langer_2009c; @pechenik_2003], which states that the mechanical noise strength $\Gamma$ is proportional to the mechanical work per STZ. The plastic work per unit volume is simply $\dot{\gamma}^{\text{pl}} s$. To convert this rate into a noise strength with dimensions of inverse time, we multiply by the volume per STZ, $V_0 / (N \Lambda^{\text{eq}})$, and divide by an energy conveniently written in the form $\epsilon_0 (V_0 / N)\, s_0$. Here, $s_0$ is a system-specific parameter with the dimensions of stress. The resulting expression for $\Gamma$ is $$\label{eq:Gamma}
\Gamma = \dfrac{\tau \dot{\gamma}^{\text{pl}} s}{\epsilon_0 s_0 \Lambda^{\text{eq}}} = \dfrac{2 s}{s_0} {\cal C}(s) \Bigl( {\cal T}(s) - m \Bigr).$$ With this result, the stationary version of Eq. reads $$2\, {\cal C}(s) \Bigl( {\cal T}(s) - m \Bigr) \left( 1 - \dfrac{m s}{s_0} \right) - m\, \rho = 0 .$$ The stationary value of $m$ is then given by $$\begin{aligned}
\label{eq:m_eq}
\nonumber m^{\text{eq}} (s) &=& \dfrac{s_0}{2s} \left[ 1 + \dfrac{s}{s_0} {\cal T}(s) + \dfrac{\rho}{2 {\cal C}(s)} \right] \\ &&- \dfrac{s_0}{2s} \sqrt{\left[ 1 + \dfrac{s}{s_0} {\cal T}(s) + \dfrac{\rho}{2 {\cal C}(s)} \right]^2 - 4 \dfrac{s}{s_0} {\cal T}(s)} . ~~~~~\end{aligned}$$ In particular, when $\rho = 0$, we have $$m^{\text{eq}} =
\begin{cases}
{\cal T}(s), & \text{if $(s / s_0)\, {\cal T}(s) < 1$} ; \\
s_0/s, & \text{if $(s / s_0)\, {\cal T}(s) \geq 1$}.
\end{cases}$$ An important consequence of this is that the yield stress for a completely jammed system is the solution of the equation $$s_y \, {\cal T}(s_y) = s_y \tanh \left( \dfrac{v_0 s_y}{p X}\right) = s_0.$$ If the temperature-like quantity $p\,X$ is small in comparison with $\epsilon_0 a^3 s_0$, then $s_y \approx s_0$. Thus $s_0$ sets, in effect, the minimum flow stress of the system in the absence of tapping. When $\rho \neq 0$, however, the system is un-jammed, and flows at arbitrarily small shear stress $s$.
Finally, let us return to using dimensionless variables $q = \tau \dot{\gamma}^{\text{pl}}$ and $\chi = X / v_Z$. The steady-state version of Eq. for the strain rate becomes $$\label{eq:q_s}
q \equiv \tau \dot{\gamma}^{\text{pl}} = 4\,\epsilon_0\,e^{-\,1/\chi}\,{\cal C}(s) \left[ \tanh \left( \dfrac{\epsilon_0 s}{\epsilon_Z p \chi} \right) - m^{\text{eq}} (s) \right] ,$$ where $\epsilon_Z \equiv v_Z / a^3$. Now, from Eq. above, the system volume varies monotonically with the compactivity as follows: $$\label{eq:V2}
\dfrac{V}{V_0} = 1 + \Lambda^{\text{eq}} \epsilon_Z - (\eta^{\text{eq}})^2 \epsilon_a + \dfrac{V_1}{V_0} ,$$ where $\epsilon_a = v_a / a^3$, $\Lambda^{\text{eq}} = 2 e^{-1 / \chi^{\text{ss}} }$, and $\eta^{\text{eq}}$ satisfies $\eta^{\text{eq}} = \tanh (2 \eta^{\text{eq}} \epsilon_a / \epsilon_Z \chi^{\text{ss}} )$. We now specify the volume $V_1$ associated with all other configurational degrees of freedom. Because STZ’s are rare density fluctuations whose existence results in denser spots nearby, their contribution to the system volume should be small in comparison to the effects of grain interlocking and the packing fraction itself, the latter being subsumed in $V_1$. The simplest assumption is that this volume varies linearly with the compactivity $\chi$: $\Lambda^{\text{eq}} \epsilon_Z + V_1 / V_0 \approx \epsilon_1 (\chi - \chi_r)$, where the effective volume expansion coefficient $\epsilon_1$ is assumed to be a constant, and $\chi_r$ is an offset that can be conveniently chosen to equal $\hat{\chi}_0$. Thus $$\label{eq:V3}
\dfrac{V}{V_0} = 1 - (\eta^{\text{eq}})^2 \epsilon_a + \epsilon_1 (\chi - \hat{\chi}_0) .$$
Summary: steady-state relations between volume, compactivity and shear rate
---------------------------------------------------------------------------
Summarizing, the steady-state system volume $V$ varies with the compactivity $\chi^{\text{ss}}$ according to Eq. : $$\label{eq:V4}
\dfrac{V}{V_0} = 1 - (\eta^{\text{eq}})^2 \epsilon_a + \epsilon_1 (\chi^{\text{ss}} - \hat{\chi}_0) ,$$ where $\eta^{\text{eq}}$ satisfies $\eta^{\text{eq}} = \tanh (2 \eta^{\text{eq}} \epsilon_a / \epsilon_Z \chi^{\text{ss}} )$. On the other hand, the steady-state compactivity is controlled by the driving forces and the frictional noise according to Eq. : $$\label{eq:X_ss2}
\chi^{\text{ss}} = \dfrac{\Gamma \hat{\chi}(q) + (\xi + \rho) \theta}{\Gamma + (\xi + \rho)} .$$ Here, the dimensionless strain rate is related to the compactivity $\chi^{\text{ss}}$ by $$\label{eq:q_s2}
q = 4\,\epsilon_0\,e^{-\,1/\chi^{\text{ss}} }\,{\cal C}(s) \left[ \tanh \left( \dfrac{\epsilon_0 s}{\epsilon_Z p \chi} \right) - m^{\text{eq}} (s) \right] ,$$ with $m^{\text{eq}}(s)$ given by Eq. . Eqs. , and are the primary relations that describe the variation of system volume, or shear band thickness, with the imposed shear rate $q$.
Analysis of experiments by van der Elst et al.
==============================================
Figure 1 shows the rescaled experimental data of van der Elst *et al.* [@brodsky_2012] along with the corresponding results for the theoretical model presented in this paper. The data points represent averages over repeated experimental measurements, and the error bars represent the corresponding standard deviation. The curves show the corresponding results for our theoretical model. The experiments were performed on angular sand particles, as well as spherical glass beads, sheared in a cylindrical torsional rheometer with parallel plate geometry. They observe pronounced, reversible non-monotonic variation of shear band thickness as a function of shear rate in angular sand, but not in glass beads. Specifically, in the absence of tapping, the angular sand shear band thickness approaches a constant value at very slow shear rates, then dips by a maximum of roughly $10\%$ at intermediate shear rates, before increasing again at fast shear rates. The shear band thickness also varies non-monotonically with the shear rate in the presence of tapping; it is smaller than in the absence of tapping in the slow, quasistatic regime, but coincides with the no-tapping behavior in the fast, inertial regime. This is in addition to the slow compaction of the non-shearing bulk, which is not shown in the figure, and may be interpreted as aging in the presence of gravity.
 Variation of steady-state volume $V$ or shear band thickness, normalized by the volume $V_0 = N a^3$ of grains, as a function of the dimensionless shear rate $q = \tau \dot{\gamma}^{\text{pl}}$. Results are shown for both angular sand particles and spherical glass beads. The data points indicate the average measurement over experimental runs (16 for angular sand without tapping, 30 for angular sand with tapping, 17 for glass beads without tapping, and 6 for glass beads with tapping). The error bars indicate the standard deviation of the measurements. The curves show results of the calculations using our theoretical model; parameter values used to compute these curves are summarized in Table \[tab:parameters\].](fig_Vq2.eps){width=".45\textwidth"}
In computing the theoretical curves in Fig. \[fig:Vq\] we used Eqs. and for the steady-state volume, but drop the $\eta$-dependent term in Eq. for spherical glass beads for which the misalignments in angular grains have no counterpart. We also made a number of simplifications, and appealed to the observations to determine a number of elements in our theory. First, we neglect the diffusion of configurational disorder across the shear band boundary, and neglect aging effects in the non-shearing bulk. This assumption is justified as long as the initial state is one with a small degree of configurational disorder, for which prior STZ analyzes indeed predict the emergence of a disorder-limited shear band that relaxes very slowly if at all, with a sharp shear band boundary [@lieou_2014; @manning_2007a; @manning_2009; @daub_2009], and within which the distribution of configurational disorder is uniform. Thus, we confine the subsequent analysis to within the shear band, and assume that internal state variables carry no spatial dependence.
Then we use the shear band thickness in the fast and slow shear rate limits, and in the absence of tapping ($\rho = 0$), to constrain the angular sand frictional noise strength $\xi$ which first appeared in Eq. . We argued in Sec. II above that $\chi^{\text{ss}} \rightarrow \hat{\chi}(q = 0) = \hat{\chi}_0$ in the limit of vanishingly slow shear rate, and that $\chi^{\text{ss}} \rightarrow \hat{\chi} (q)$ and diverges at large shear rate, but that $\theta < \chi^{\text{ss}} < \hat{\chi}(q)$ between the two limits, implying that it is plausible for $\xi (\Gamma) \sim \Gamma^2$ at small $\Gamma$, and saturates at large $\Gamma$. One way to interpolate between the slow, quasistatic and fast, inertial behaviors is to assume that $\xi$ takes the form $$\label{eq:xi}
\xi (\Gamma) = \xi_0 \tanh ( \beta \Gamma^2 ) .$$ In the ensuing analysis, we use $\xi_0 = 1.2$ and $\beta = 20$ for angular sand. On the other hand, we set $\xi = 0$ for spherical glass beads. In fact, the data points in Fig. \[fig:Vq\] for spherical glass beads sheared in the absence of tapping indicate a small degree of compaction at $q \sim 10^{-2}$, suggestive of a small, nonzero $\xi$. Thus, setting $\xi = 0$ for glass beads results in a small misfit between theory and data, which is not surprising because there is nonzero interparticle friction even between spherical glass beads. However, we do not attempt to model that behavior.
Van der Elst *et al.* measured the shear rate in terms of the rotation angular velocity $\omega$ of the shear cell. (The geometry is equivalent to that of a rectangular shear cell with periodic boundary conditions in the shearing direction, at least locally.) To convert this to our shear rate, we use the estimate $\dot{\gamma}^{\text{pl}} = r \omega / h$, where $r$ is the shear cell radius and $h$ is the shear band height. In their experiments, $r = 9.5$ mm and $h \simeq 1$ mm. Next, the angular sand particles have a typical diameter of $a = 350$ $\mu$m, with mass density $\rho_G = 1600$ kg m$^{-3}$, and the experiments were performed at a confining pressure of $p \simeq 7$ kPa. This gives the inertial time scale $\tau = a \sqrt{\rho_G / p} = 1.67 \times 10^{-4}$ s, so that the conversion formula between the rotation velocity $\omega$ and our dimensionless shear rate $q$ is $$q = (1.59 \times 10^{-3} \text{s}) \omega .$$ We note in passing that the tapping frequency in the experiments is 40.2 kHz, the inverse of which is only an order of magnitude smaller than the inertial time scale. Recall our argument in Eq. that misalignments are created and annihilated by noise; thus the fact that the inertial and tapping time scales are comparable justifies our assumption that both the STZ density $\Lambda$ and the misalignment bias $\eta$ are functions of the same compactivity $\chi$. This need not be the case if the two time scales are several orders of magnitude apart; we will comment on its implication in Sec. V.
The stress measurements were too noisy for us to be able to constrain parameters associated with STZ dynamics. However, based on other shearing experiments on angular grains [@mair_1999; @mair_2002], the yield stress should be about 0.4 times the confining pressure, so we have chosen $s_0 = 0.4 p$. We have also chosen ${\cal C}(s) \simeq {\cal C}_0 = 1$ to be a constant, because the STZ transition rate should not be very sensitive to the shear stress to pressure ratio provided that $s / p < 1$. The parameter fitting procedure reveals that the steady-state volume variation is insensitive to these STZ dynamics parameters in comparison with those involved in the choice of $\hat{\chi} (q)$. We choose $\theta = 0.2$, $\rho = 0$ and $\theta = 0.18$, $\rho = 5 \times 10^{-4}$ in the absence and presence of tapping, respectively. With $\theta$ subsuming all kinetic degrees of freedom, its different values in the two cases reflect the expectation that tapping can increase the packing fraction of a generic granular assembly [@daniels_2005; @daniels_2006; @nowak_1998; @knight_1995]. (Following [@edwards_1998], it might be possible to determine $\theta$ using the fluctuation-dissipation theorem and the Langevin equation, but that gives us another adjustable parameter interpreted as a drag coefficient, so we regard $\theta$ as an adjustable parameter itself.) The other parameters in our calculation are summarized in Table \[tab:parameters\]. While our model produces qualitative agreement with the experiment over a wide range of parameter values, the parameter values used in the present analysis have been chosen to provide quantitative fit with the experimental measurements and, based on past experience, are physically reasonable. It may be possible to further constrain the parameters of microscopic origin with additional simple experiments; for example, stress parameters may be constrained using slide-hold-slide experiments.
[ >m[.08]{} >m[.50]{} >m[.11]{} >m[.27]{}]{} Parameter & Description & Value & Reference or remark\
$p$ & Confining pressure & 7 kPa & Constrained by experiment [@brodsky_2012]\
$s_0$ & Minimum flow stress & 2.8 kPa & Determined empirically [@mair_1999; @mair_2002]\
$\tau$ & Inertial time scale & $ 1.67 \times 10^{-4}$ s & Constrained by experiment [@brodsky_2012]\
${\cal C}_0$ & Characteristic STZ transition rate & 1 & Microscopic [@lieou_2012]\
$\xi_0$ & Maximum frictional noise strength & 1.2 & Adjustable parameter\
$\beta$ & Parameter in frictional noise strength, Eq. & 20 & Adjustable parameter\
$\rho$ & Tapping intensity & 0, $5 \times 10^{-4}$ & Adjustable parameter\
$\theta$ & Kinetic temperature & 0.2, 0.18 & Adjustable parameter\
$\hat{\chi}_0$ & Steady-state dimensionless compactivity in $q \rightarrow 0$ limit & 0.3 & Adjustable parameter\
$\hat{\chi}_1$ & Parameter in VFT expression, Eq. & 0.02 & [@lieou_2012]\
$\hat{\chi}_A$ & Parameter in VFT expression, Eq. & 0.33 & [@lieou_2012]\
$A$ & Parameter in VFT expression, Eq. & 2 & [@lieou_2012]\
$q_0$ & Critical strain rate & 2 & [@lieou_2012]\
$\epsilon_0$ & Plastic core volume per STZ in units of grain volume & 1.5 & Microscopic [@lieou_2012]\
$\epsilon_Z$ & Excess volume per STZ in units of grain volume & 0.5 & Determined Microscopic [@lieou_2012]\
$\epsilon_1$ & Effective volume expansion coefficient & 0.3 & Microscopic\
$\epsilon_a$ & Misalignment defect volume in units of grain volume & 0.1 & Microscopic\
 Theoretical results for the variation of steady-state compactivity $\chi^{\text{ss}}$ with the dimensionless shear rate $q = \tau \dot{\gamma}^{\text{pl}}$, for both angular sand particles and spherical glass beads. Parameter values for each of these curves are summarized in Table \[tab:parameters\].](fig_Xq2.eps){width=".45\textwidth"}
In Fig. \[fig:Xq\], we plot the variation of the dimensionless steady-state compactivity $\chi^{\text{ss}}$ with the shear rate $q$. $\chi^{\text{ss}}$ varies in the same qualitative manner as the volume $V$, as it should, for the two quantities are related to each other in a monotonic fashion according to Eq. . Finally, in Fig. \[fig:sq\], we show our model prediction for the variation of the shear stress to pressure ratio $s / p$ with the shear rate $q$. Obviously, in the absence of vibrations, the granular medium does not flow until the shear stress exceeds the threshold $s_0$. On the other hand, tapping un-jams the system and causes it to flow at arbitrarily small shear stress $s$. This is a hallmark of glassy behavior, as seen in other amorphous solids [@lieou_2012; @liu_2011; @haxton_2012; @langer_2008]. The pronounced weakening is also hypothesized as a consequence of acoustic fluidization in earthquake faults [@melosh_1979; @melosh_1996]. In addition, the shear stress $s$ increases faster with the shear rate $q$ when $q > 10^{-3}$ in angular sand than in glass beads, conforming with the intuition that more work is necessary to cause angular, frictional particles to flow under shear.
 Theoretical results for the variation of shear stress to pressure ratio $s/p$ with the dimensionless shear rate $q = \tau \dot{\gamma}^{\text{pl}}$, for both angular sand particles and spherical glass beads. Parameter values for each of these curves are summarized in Table \[tab:parameters\].](fig_sq2.eps){width=".45\textwidth"}
Concluding remarks
==================
In this paper, we are proposing a fundamentally new, thermodynamic interpretation of an unexpected experimental observation – the non-monotonic variation of steady-state shear band thickness, or sample volume, as a function of shear rate, in a granular medium composed of angular, frictional particles. In our theory, this volume is determined by the effective disorder temperature (or “compactivity”) of the grains. As a macroscopic state variable, the compactivity is controlled by a variety of microscopic mechanisms including STZ transitions, intergranular friction, and the strength of noisy fluctuations generated by collisions and external tapping. For example, energy dissipated by friction effectively “cools” the system and reduces its volume. The various microscopic mechanisms are described by physical parameters whose orders of magnitude can be estimated empirically and from experimental data. We have argued that if the frictional “noise strength” $\xi$ varies quadratically with the shear rate at the quasistatic shear limit – in conformity with the energy dissipation associated with inelastic collisions between particles – and saturates in the inertial, fast-shear limit, then it is possible for the steady-state compactivity and therefore the volume to show non-monotonic variation with shear rate. In other words, energy dissipated by friction effectively “cools” the system and reduces its volume.
In addition, we have introduced a microscopic model that quantifies the volume of the granular assembly $V$ as a function of the compactivity $\chi$. The most salient feature of this model is the combination of STZ’s, soft spots with excess free volume that facilitate grain rearrangement, and misalignment defects, arising from grain interlocking and geometric frustration ubiquitous to angular particles. With a judicious choice of parameters, we have shown excellent quantitative agreement between our theory and the experimental measurements of van der Elst *et al.* [@brodsky_2012] on sheared angular sand particles and spherical glass beads. In our opinion, the fact that this picture fits together as well as it does is strong evidence in favor of its validity. This is not pure phenomenology. It is a systematic attempt to develop a first-principles theory of a complex and important class of physical phenomena.
We emphasize again that qualitatively, the non-monotonic variation of sample thickness with shear rate is a consequence of frictional noise alone, and not uniquely described by our microscopic model of STZ’s and misalignments. One example of an alternative, simple model is the linear model $V = C \chi + D$ without the misalignment term, where the effective volume expansion coefficient $C$ is a constant. However, we have not been able to fit quantitatively the experimental data with this model nearly as closely as with the model of misalignments, which amplifies the amount of compaction at intermediate strain rates. Therefore, we have chosen to adopt the present model of STZ’s and misalignments, the latter of which is necessary to account for the rather substantial amount of compaction observed in the transitional regime, between the slow, quasistatic and fast, inertial limits.
A key feature of the Ising-like model of misalignments is the existence of a “ferromagnetic” transition (see Eq. above); it happens above a critical compactivity $\chi_c$ at which the volume variation as a function of shear rate ought to display a small cusp, at a shear rate apparently not probed by the van der Elst *et al.* experiments. Our binary, Ising-like model of misalignments might also be useful in formulating a description of other glassy phenomena in granular materials, in a manner similar to binary clusters introduced in [@langer_2013].
A central assumption in the paper is that both misalignments and STZ’s are governed by the same “configurational temperature” or compactivity $\chi$, the justification of which is that the inertial time scale $\tau = a \sqrt{\rho_G / p}$ and the inverse tapping frequency differ by less than an order of magnitude. Had this not been the case, it might be necessary to characterize the system with as many as three temperature-like quantities: the kinetic temperature $\theta$, the “noise” compactivity $\chi_K$ that pertains to the vibrational subsystem and governs the misalignment bias $\eta$, and the configurational compactivity $\chi_C$ that pertains to the shearing, configurational subsystem and governs the STZ density $\Lambda$, all falling out of equilibrium with one another. When that happens, the variation of volume $V$ as a function of shear rate $q = \tau \dot{\gamma}^{\text{pl}}$ in the presence of tapping need not coincide with that in the absence of tapping, in the fast shear rate limit. There are multiple ways to separate the vibrational and configurational time scales, and verify or dismiss our speculation. For example, one could conduct the shear experiment at substantially higher confining pressure $p$ to shorten the inertial time scale $\tau$ of the configurational subsystem, or tap the system at a higher frequency so that the inertial and vibrational time scales are at least several orders of magnitude apart. In the former case, we speculate that the vibrational subsystem would fall out of equilibrium with the configurational subsystem, with $\chi_K > \chi_C$; in the latter case, $\chi_C > \chi_K$. Either way, the coupling between the different subsystems would differ from that in the present paper (cf. Eq. ), and qualitatively distinct behaviors might emerge. Such experiments might provide the most stringent tests yet of non-equilibrium thermodynamics.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Nicholas van der Elst and Emily Brodsky for sharing their data with us, and for illuminating discussions. Additionally, we thank Kenneth Kamrin, Bulbul Chakraborty, Karen Daniels, Karin Dahmen, and Michael Cates for sharing their insights. This work was supported by an Office of Naval Research MURI Grant No. N000140810747, NSF Grant No. DMR0606092, and the NSF/USGS Southern California Earthquake Center, funded by NSF Cooperative Agreement EAR-0529922 and USGS Cooperative Agreement 07HQAG0008, and the David and Lucile Packard Foundation. Additionally, JSL was supported in part by the U.S. Department of Energy, Office of Basic Energy Sciences, Materials Science and Engineering Division, DE-AC05-00OR-22725, through a subcontract from Oak Ridge National Laboratory.
[5]{}
N. J. van der Elst, E. E. Brodsky, P.-Y. Le Bas, and P. A. Johnson, J. Geophys. Res. **117**, B09314 (2012).
P. A. Johnson, H. Savage, M. Knuth, J. Gomberg, and C. Marone, Nature **451**, 57 (2008).
K. E. Daniels and R. P. Behringer, Phys. Rev. Lett. **94**, 168001 (2005).
K. E. Daniels and R. P. Behringer, J. Stat. Mech. **2006**, P07018 (2006).
E. R. Nowak, J. B. Knight, E. Ben-Naim, H. M. Jaeger, and S. R. Nagel, Phys. Rev. E **57**, 1971 (1998).
J. B. Knight, C. G. Fandrich, C. N. Lau, H. M. Jaeger, and S. R. Nagel, Phys. Rev. E **51**, 3957 (1995).
K. Lu, E. E. Brodsky, and H. P. Kavehpour, J. Fluid Mech. **587**, 347 (2007).
J. A. Dijksman, G. H. Wortel, L. T. H. van Dellen, O. Dauchot, and M. van Hecke, Phys. Rev. Lett. **107**, 108303 (2011).
H. J. Melosh, J. Geophys. Res. **84** (B13), 7513 (1979).
H. J. Melosh, Nature **379**, 601 (1996).
S. F. Edwards and R. B. S. Oakeshott, Physica A **157**, 1080 (1989).
A. Mehta and S. F. Edwards, Physica A **157**, 1091 (1989).
S. F. Edwards and R. B. S. Oakeshott, Physica D **38**, 88 (1989).
S. F. Edwards, Rheol. Acta **29**, 493 (1990).
S. F. Edwards, J. Phys. Condensed Matter **2**, SA63 (1990).
M. L. Falk and J. S. Langer, Phys. Rev. E **57**, 7192 (1998).
M. L. Falk and J. S. Langer, Ann. Rev. Cond. Matt. Phys. **2**, 353 (2011).
E. G. Daub and J. M. Carlson, Ann. Rev. Cond. Matt. Phys. **1**, 397 (2010).
C. K. C. Lieou and J. S. Langer, Phys. Rev. E **85**, 061308 (2012).
C. K. C. Lieou, A. E. Elbanna, and J. M. Carlson, Phys. Rev. E **89**, 022203 (2014).
A. J. Mallinckrodt and H. S. Leff, Am. J. Phys. **60**, 356 (1992).
H. S. Leff and A. J. Mallinckrodt, Am. J. Phys. **61**, 121 (1993).
B. D. Coleman and W. Noll, Arch. Ration. Mech. Anal. **13**, 167 (1963).
E. Bouchbinder and J. S. Langer, Phys. Rev. E **80**, 031131 (2009).
E. Bouchbinder and J. S. Langer, Phys. Rev. E **80**, 031132 (2009).
E. Bouchbinder and J. S. Langer, Phys. Rev. E **80**, 031133 (2009).
P. Jop, Y. Forterre, and O. Pouliquen, Nature (London) **441**, 727 (2006).
T. K. Haxton, M. Schmiedeberg and A. J. Liu, Phys. Rev. E **83**, 031503 (2011).
T. K. Haxton, Phys. Rev. E **85**, 011503 (2012).
T. K. Haxton and A. J. Liu, Phys. Rev. Lett. **99**, 195701 (2007).
J. S. Langer and M. L. Manning, Phys. Rev. E **76**, 056107 (2007).
J. S. Langer, Phys. Rev. E **77**, 021502 (2008).
K. Kamrin and G. Koval, Comput. Particle Mech., in press (2014).
J. S. Langer, Phys. Rev. E **88**, 012122 (2013).
J. S. Langer and L. Pechenik, Phys. Rev. E **68**, 061507 (2003).
M. L. Manning, J. S. Langer and J. M. Carlson, Phys. Rev. E **76**, 056106 (2007).
M. L. Manning, E. G. Daub, J. S. Langer, and J. M. Carlson, Phys. Rev. E **79**, 016110 (2009).
E. G. Daub and J. M. Carlson, Phys. Rev. E **80**, 066113 (2009).
K. Mair and C. Marone, J. Geophys. Res. **104**, 28899 (1999).
K. Mair, K. M. Frye, and C. Marone, J. Geophys. Res. **107** (B10), 2219 (2002).
S. F. Edwards and D. V. Grinev, Phys. Rev. E **58**, 4758 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For first passage percolation on $\Z^2$ with i.i.d. bounded edge weights, we consider the upper tail large deviation event; i.e., the rare situation where the first passage time between two points at distance $n$, is macroscopically larger than typical. It was shown by Kesten [@Kes86] that the probability of this event decays as $\exp (-\Theta(n^2))$. However the question of existence of the rate function i.e., whether the log-probability normalized by $n^2$ tends to a limit, had remained open. We show that under some additional mild regularity assumption on the passage time distribution, the rate function for upper tail large deviation indeed exists. Our proof can be generalized to work in higher dimensions and for the corresponding problem in last passage percolation as well. The key intuition behind the proof is that a limiting metric structure which is atypical causes the upper tail large deviation event. The formal argument then relies on an approximate version of the above which allows us to dilate the large deviation environment to compare the upper tail probabilities for various values of $n.$'
address:
- 'Riddhipratim Basu, International Centre for Theoretical Sciences, Tata Institute of Fundamental Research, Bangalore, India'
- 'Shirshendu Ganguly, Department of Statistics, UC Berkeley, Berkeley, CA, USA'
- 'Allan Sly, Department of Mathematics, Princeton University, Princeton, NJ, USA'
author:
- Riddhipratim Basu
- Shirshendu Ganguly
- Allan Sly
bibliography:
- 'delocalization.bib'
title: Upper Tail Large Deviations in First Passage Percolation
---
Introduction and main result
============================
First passage percolation is a popular model of fluid flow through inhomogeneous random media, where one puts random weights on the edges of a graph and considers the first passage time between two vertices, which is obtained by minimizing the total weight among all paths between the two vertices. First passage percolation on Euclidean lattices was introduced by Hammersley and Welsh [@HW65] in 1965 and has been studied extensively both in statistical physics and probability literature ever since. This model served as one of the motivations of developing the theory of subadditive stochastic processes and the early progresses using subadditivity was made by Hammersley-Richardson-Kingman [@HW65; @K73; @R73] and culminated in the proof of the celebrated Cox-Durrett shape theorem [@CD81] establishing the first order law of large number behaviour for passage times between far away points. Further progress was made into the 80s and 90s through efforts of Kesten [@Kes86; @Kes87; @Kes93] and Talagrand [@Tal94] establishing concentration inequalities for passage times; Newman and others [@New95] on more geometric aspects of the model. Much progress has been made since [@BKS04; @H08] including a flurry of results in the last five years [@Cha11; @AD14; @DH14; @DH17]. Despite this impressive progress, most of the fundamental questions still remain major mathematical challenges, see the survey [@ADH15] for a comprehensive history as well as an extensive list of the major open problems in this field.
One other reason planar first passage percolation came into prominence is that this model is believed to be in the KPZ universality class that was introduced by Kardar, Parisi and Zhang [@KPZ86] in 1986. Using non rigorous renormalization group techniques, KPZ predicted universal scaling exponents for many (1+1)-dimensional growth models including first and last passage percolation under very general conditions on the passage time distribution (precise definitions later). An explosion of rigorous results in the last 18 years starting with the seminal work of Baik, Deift and Johansson [@BDJ99] has now verified the KPZ prediction for a handful of models including last passage percolation with Exponential, Geometric or Bernoulli passage times. However, this progress has been mostly restricted to the so-called exactly solvable (or, integrable) models where exact formulae are available using deep connections to algebraic combinatorics, representation theory and random matrix theory; and extremely detailed information has been obtained about such models by analyzing those formulae. Although the same results are qualitatively expected to hold for a much larger class of models, these methods rely very crucially on the exact formulae, and moving beyond the exactly solvable models remains a major challenge.
Our focus in this paper is such a problem in the non-integrable setting of first passage percolation in the large deviation regime. The question first arose in the work of Kesten [@Kes86] who considered the probability of large deviation events in first passage percolation. Postponing the precise definitions momentarily, let us first describe informally the set-up. Consider the passage time $\T_n$ from $(0,0)$ to $(n,0)$. The shape theorem dictates that under some regularity conditions $\frac{{{\mathbf{T}}}_n}{n}\to \mu$ almost surely for some $\mu \in (0,\infty)$. The study of large deviations is concerned with the unlikely events $\{{{\mathbf{T}}}_n \geq (\mu+{\varepsilon})n\}$ (upper tail) and $\{{{\mathbf{T}}}_n \leq (\mu-{\varepsilon})n\}$ (lower tail). In the classical theory of large deviations, log of such probabilities suitably scaled (by the so-called speed of large deviations) converges to a function of ${\varepsilon}$, known as the rate function. For first passage percolation, Kesten [@Kes86] showed the large deviation speed of $n$ and existence of the rate function for the lower tail using a subadditive argument. For the upper tail Kesten showed a large deviation speed on $n^2$ for bounded edge weight distribution, however the existence of rate function remained open (see Open Question 18, in [@ADH15]). Our main result in this paper (see Theorem \[t:ldp\] below) answers this question establishing the existence of rate function for the upper tail, thereby establishing first such result beyond the exactly solvable models.
Model definitions and statement of result {#mods}
-----------------------------------------
We start with formal definitions of standard first passage percolation on $\Z^d$, $d\geq 2$. Let $E(\Z^d)$ denote the set of all nearest neighbour edges in $\Z^d$. Let $\nu$ be a probability measure supported on the non-negative real line. Let $\Pi=\{X_e: e \in E(\Z^d)\}$ denote a field of i.i.d. random variables where each $X_e$ (called the passage time of the edge $e$) has distribution $\nu$. For a sequence $\gamma=e_1e_2\cdots e_{k}$ of neighbouring edges (called a path), the passage time of the path, denoted by $\ell(\gamma)$,[^1] is defined as $$\ell(\gamma)=\sum_{i=1}^{k} X_{e_i}.$$ For any two vertices $u$ and $v$, the first passage time between $u$ and $v$, denoted ${{\mathbf{PT}}}(u,v)$ is defined as the infimum of $\ell(\gamma)$ where $\gamma$ varies over all paths starting at $u$ and ending at $v$. Let $\mathbf{0}$ denote the origin. Under very mild conditions on $\nu$, it is a fundamental fact that for all $v\in \Z^d$, there exists $\mu(d,\nu,v)\geq 0$ such that $$\lim_{n\to \infty}\frac{{{\mathbf{PT}}}(\mathbf{0},nv)}{n}=\mu(d,\nu,v)$$ almost surely. For the special case when $v=(1,0,\ldots, 0)$ denote the unit vector along the first co-ordinate, we denote the limiting constant by just $\mu$, also known as the time constant in the literature. For the rest of this paper we shall focus on the planar case ($d=2$) of the above model. Although our main result extends to higher dimensions with little to no change, we choose to work in two dimensions to avoid additional notational overhead. From now on, we shall be in the setting of standard first passage percolation on $\Z^2$ unless otherwise mentioned. Let $\mathbf{n}:=(n,0)$ and let us denote the passage time ${{\mathbf{PT}}}(\mathbf{0},\mathbf{n})$ by ${{\mathbf{T}}}_n$. As mentioned above we are concerned with the probability of the upper tail large deviation event: $$\label{utdef}
\sU_{\zeta}(n):=\{{{\mathbf{T}}}_{n}\ge (\mu+\zeta) n\}.$$ for some $\zeta>0$. Throughout the paper we will assume the rather general condition that the passage time distribution has a continuous density on a compact interval $[0,b]$. Even though we believe our proof methods can be used to extend our result beyond this assumption, the former will help make some of the proofs cleaner. For future reference we record this assumption below.
\[d:pt\] For $b>0$, let $\cP(b)$ denote the set of all probability measures with support $[0,b]$ and a continuous density.
It is well known that if $\nu\in \cP(b)$ for any $b>0$, then we have $0<\mu<b$ (e.g. see [@HW65]). Also observe that for $\nu\in \cP(b)$, we have deterministically that ${{\mathbf{T}}}_{n}\leq bn$. So while considering the large deviation event $\sU_{\zeta}$ in the above scenario it suffices to consider $\zeta \in (0,b-\mu)$. Our main theorem shows that the large deviation rate function exists in the above setting.
\[t:ldp\] Consider standard first passage percolation on $\Z^2$ with passage time distribution $\nu\in \cP(b)$ for some $b>0$. Then for $\zeta\in (0,b-\mu)$ there exists $r=r(\nu,\zeta)\in (0,\infty)$ such that $$\lim_{n\to \infty} -\frac{\log \P (\sU_{\zeta}(n))}{n^2}=r.$$
A couple of remarks are in order. First, there is nothing special about the direction $(1,0)$; the same result holds for any unit vector $v$ with different rate function $r$, with minor adjustments in the proof. Also, a variant of this result holds in higher dimensions as well where the speed of the large deviation is $n^{d}$ rather than $n^2$ (See e.g., ). The same argument proving Theorem \[t:ldp\] can be used to prove the higher dimensional analogue. However, in this paper we shall only concentrate on proving Theorem \[t:ldp\].
Observe that the condition in Theorem \[t:ldp\] is not optimal and we have not made an attempt to make it the weakest possible. It is however important to observe that some condition is needed to ensure even the $n^2$ speed of the large deviation. Together with the standard assumptions that the mass at $0$ is less than the critical bond percolation probability on $\Z^2$ and that the edge distribution is not degenerate at a single point, Kesten assumed boundedness. It is easy to see that the boundedness assumption cannot be completely removed. For example, if the passage times are exponentially distributed, just increasing all the passage times around the origin by $(\mu+ \zeta) n$, would force the large deviation event, while its probability being only exponentially small in $n$. One can however prove Kesten’s result for passage times with sufficiently fast decaying tails, and one believes that the rate function will exist in such a case too possibly under some additional assumptions. However, in this paper we have not pursued those directions, and instead focussed on proving the result in the simplest possible case that is still sufficiently general to be of interest.
Background and Related Works
----------------------------
First passage percolation can be thought of as putting a random metric on $\Z^d$, where the distance between two vertices is given by the first passage time between them. As alluded to above, the most fundamental result about first passage percolation says that under suitable rescaling these metrics converge almost surely to a deterministic metric on $\R^d$ in a pointed Gromov-Hausdorff sense. More precisely we have the following. Suppose $\nu\in \cP(b)$ for some $b\in (0,\infty)$ (actually the result is valid more generally, one only needs some moment condition and that the mass of any atom at $0$ is sufficiently small), and let $\tilde{B}(t)$ denote the set of all vertices that are within distance $t$ of $\mathbf{0}$ in the FPP metric, and let $B(t)=\tilde{B}(t)+[-\frac{1}{2}, \frac{1}{2}]^{d}$. Then there exists a non-random compact convex set $\cB=\cB_{\nu}$ with obvious symmetries such that for each ${\varepsilon}>0$ $$\label{e:shape}
\P\left((1-{\varepsilon})\cB_\nu \subset \frac{B(t)}{t} \subset (1+{\varepsilon})\cB_{\nu} \text{ for all large }t \right)=1.$$ The set $\cB$ is called the limit shape for this model. Recall the limiting constant $\mu(d,\nu,v)$ in direction $v$. It is not hard to see that $\mu(d,\nu,\cdot)$ can be extended to a norm in $\R^d$ and $\cB$ is the unit ball corresponding to this norm. The shape theorem implies that at large scales, the distance function in the FPP metric in a fixed direction grows approximately linearly, and the convexity of the limit shape is then just a consequence of triangle inequality.
The shape theorem is a law of large number result, and the natural next question of obtaining fluctuations has been extensively investigated. The moderate deviation estimates are interesting, as in $d=2$, KPZ scaling predicts a fluctuation exponent of $1/3$, however the best known fluctuation and concentration bounds (for ${{\mathbf{T}}}_n$) have so far been proved at $n^{1/2+o(1)}$ scale [@Kes93; @Tal94; @BKS04]. In this paper, we are looking at the large deviation regime, i.e., where we consider a linear deviation of ${{\mathbf{T}}}_n$ from its long term value. Although we recall standard results only for ${{\mathbf{T}}}_n$; qualitatively same results hold in all directions. Also we are assuming throughout that the passage time distribution is in $\cP(b)$ for some $b$ although many of these results hold under weaker assumptions.
Kesten [@Kes86] considered both upper and lower tail large deviations for first passage percolation. Let $\sL_{\zeta}(n):=\{{{\mathbf{T}}}_{n}\le (\mu-\zeta) n\}$ (throughout this section for brevity we will use ${{\mathbf{T}}}_n$ to denote the passage time between $(0,0,\ldots,0)$ and $(n,0,\ldots,0)$ in $\Z^d$ although it was initially defined only for $\Z^2$) denote the lower tail large deviation event. Using a subadditive argument, Kesten showed that for $\zeta\in (0,\mu)$, $$\label{e:lt}
\lim_{n\to \infty} -\frac{\log \P(\sL_{\zeta}(n))}{n}=r_{\ell}(\zeta) \in (0,\infty).$$
For the upper tail large deviations, Kesten showed that $$\label{e:ut}
0 < \liminf_{n\to \infty} - \frac{\log \P(\sU_{\zeta}(n))}{n^d} \leq \limsup_{n\to \infty} - \frac{\log \P(\sU_{\zeta}(n))}{n^d}<\infty.$$ The existence of the limit was left open and this open question was re-iterated in [@ADH15] (See Question 18), which we answer in our Theorem \[t:ldp\].
Observe that the speed of large deviations is different in upper and lower tails. This is not unexpected and can be intuitively explained as follows. For ${{\mathbf{T}}}_n$ to be much smaller than $\mu n$, one needs only one path that is atypically small; however it is much more unlikely for ${{\mathbf{T}}}_n$ to be atypically large, since typically one can find $n^{d-1}$ many ‘parallel’ short paths between the origin and $(n,0,0,\ldots,0)$ which are disjoint except at the beginning and the end. Thus to attain the upper tail event all such paths need to be large, each of which costs $e^{-\Theta(n)}$ and hence the total cost is at least $(e^{-\Theta(n)})^{n^{d-1}}$. Indeed this feature is quite common in many growth models, e.g. last passage percolation, parabolic Anderson model and deviation of the spectrum of GUE (see [@cranston] and the references therein).
As a matter of fact, among the only cases of growth models where the existence of rate function is known for both tails are the so-called exactly solvable models of last passage percolation. As an illustration, we only describe the result for the case of exponential directed last passage percolation in $\Z^2$ [@Jo99]; however the same qualitative result is known in the case of Poissonian directed last passage percolation in $\R^2$ [@DZ1] and last passage percolation on $\Z^2$ with geometric edge weights [@Jo99]. Consider the following last passage percolation model on $\Z^2$ where each vertex is equipped with an i.i.d. sample of $\mbox{Exp}(1)$ random variable. As before, the weight of any path is the sum of weights on it. The difference from the first passage percolation model is that we only consider up/right directed paths and the last passage time between two vertices is calculated by maximizing the weight over all such paths between the two vertices. This is one of the first exactly solvable models rigorously shown to be in the KPZ universality class by Johansson [@Jo99] using exact determinantal formulae. Let $L_n$ denote the last passage time from $(0,0)$ and $(n,n)$. It is well known [@Ro81] that $\frac{L_n}{n}\to 4$ almost surely as $n\to \infty$. Johansson proved large and moderate deviation estimates for $L_n$. In particular he proved that $$\lim_{n\to \infty} \frac{\log \P(L_{n}\geq (4+\zeta)n)}{n}=-I_{u}(\zeta);\quad \zeta>0~\text{and}$$ $$\lim_{n\to \infty} \frac{\log \P(L_{n}\leq (4-\zeta)n)}{n^2}= -I_{\ell}(\zeta);\quad \zeta \in (0,4).$$ The functions $I_{\ell}$ and $I_{u}$ could in principle be explicitly evaluated there. Observe that for last passage percolation, as expected, the role of upper tail and lower tail is reversed but qualitatively there is no other difference from the FPP case. We list below a few other results worth mentioning: a similar result as above in the context of Poissonian LPP by Dueschel and Zeitouni in [@DZ1]. Still within KPZ universality class, but in the framework of particles systems, functional large deviation principle for Totally Asymmetric Simple Exclusion Process (TASEP), which is closely connected to Exponential LPP, was obtained, for the $n$-speed tail by Varadhan and Jensen [@varadhan1; @Jensen00] and for the $n^2$-speed tail recently by Olla and Tsai in [@OT17].
However the above results concerning LDP at speed $n^2$, use some form of integrability and the proofs rely heavily on the nature of the passage time distributions which are intimately connected to the integrable features in these models. Although the large deviation behaviour is expected to be universal, the existence of the rate function was not even known for any other non-integrable model of last passage percolation. It is left to the reader to check that all our arguments will remain valid, in fact become simpler, for a general last passage percolation model (with bounded edge weights, say).
Although as far as we are aware, our result is the first one proving the existence of a large deviation rate function for the $n^2$-speed tail for point to point passage times in a non-integrable setting, one variant of such a result was proved by Chow-Zhang [@CZ] in the case of line-to-line first passage time in standard first passage percolation where the open problem addressed by Theorem \[t:ldp\] was also mentioned. Formally Chow-Zhang considers the minimum passage time over all paths with one endpoint in $A=\{(0,i):i\in \{0,1,\ldots, n\}\}$ and the other endpoint in $B=\{(n,i):i\in \{0,1,\ldots, n\}\}$ and moreover they consider the geodesic restricted to lie in the square $[0,n]^2.$ Let us denote the passage time by ${{\mathbf{T}}}^*_n$. It is a standard result [@Kes86] that $\frac{{{\mathbf{T}}}^*_n}{n}\to \mu$ almost surely as $n\to \infty$. In [@CZ], Chow and Zhang showed that for $\zeta>0$ $$\lim_{n\to \infty} -\frac{\log \P(\sU_{\zeta}(n))}{n^2}$$ exists and is nontrivial. The appropriate variant of their result holds in all dimensions. Even though the specific geometric setting considered in [@CZ] causes significant simplification, and in particular rules out backtracks of the geodesic and does not create a necessity for the metric space dilation approach in this paper, it is worth mentioning that the argument in [@CZ] is a multi sub-additive argument, which bears resemblance with our approach at least at a high level (see Section \[outline\] for more details).
Finally we end this section with a brief discussion about a related line of work concerning geometric consequences of large deviation events in first/last passage percolation. Formally one considers the measure obtained by conditioning on the large deviation events, and investigates how does the geometry of the random field of weights change? These questions were considered in the setting of exactly solvable Poissonian last passage percolation for the upper tail (i.e., the tail with large deviation speed $n$) by Deuschel and Zeitouni who, in [@DZ1], showed that under the upper tail large deviation event, the maximizing paths between two far away points is with high probability localized around the straight line segment joining the two endpoints. For the harder lower tail case, in a recent paper [@BGS17A] we showed that forcing the large deviation event makes the path delocalized with high probability. Although we choose to work in the setting of last passage percolation in the latter, our argument goes beyond the integrable setting under certain distributional assumptions. (see remarks in [@BGS17A] for more details.)
A brief outline of the paper {#outline}
----------------------------
The argument of proving Theorem \[t:ldp\] is quite involved and has many pieces going into the proof. The purpose of this section is to provide a broad overview of the steps of the argument. At a very high level, our argument intuitively is predicated on the existence of a limiting metric structure as in even in the upper tail large deviation regime, which roughly implies that conditional on the large deviation event, the distances in a fixed direction grow linearly at large scales, and as the direction is varied the gradient changes in a reasonably regular way. The reason to expect this is intimately tied to the reason behind the $n^2$ speed of large deviation, which causes the edge distributions of $\Theta(n^2)$ many edges to change.
Although we believe the above statement to be true, for the purposes of the proof it suffices to have sub-sequential limits. In fact the exact statement that we prove in much less refined. (see Proposition \[t:stable\]).
For the remainder of the paper let $b>0$ and $\nu\in \cP(b)$ be fixed. Recall that $\mu$ denotes the time constant in the $x$-direction for the standard first passage percolation on $\Z^2$ with $\nu$-distributed edge weights. Let $\zeta\in (0, b-\mu)$ be fixed. For $n\in \N$, let $a_{n}=a_n(\zeta)$ be defined by $$a_{n}=\log \P(\sU_{\zeta}(n)).$$ Theorem \[t:ldp\] will follow easily from the following multi-subadditive result.
\[p:subadditive\] For each ${\varepsilon}>0$, there exists $N_0>0$ such that the following holds. For all $n\in \N$ with $n>N_0$ there exists $M_0=M_0(n)$ such that for all $m>M_0$ we have $$\frac{a_{m}}{m^2} \geq \frac{a_{n}}{n^2} -{\varepsilon}.$$
Most of this paper is devoted to proving Proposition \[p:subadditive\], but before we outline its proof let us quickly finish the proof of Theorem \[t:ldp\] assuming the above.
Let $$\begin{aligned}
\label{limsupinf}
\kappa &= \limsup_{n\to \infty} \frac{a_n}{n^2};~\text{and}\\
\kappa' &= \liminf_{n\to \infty} \frac{a_n}{n^2}.\end{aligned}$$ By Kesten’s result we know that $-\infty < \kappa' \leq \kappa <0$ and hence it suffices to prove that for all ${\varepsilon}>0,$ we have $\kappa'\geq \kappa -2{\varepsilon}$. Fix ${\varepsilon}>0$ and let $N_0$ be such that the conclusion of Proposition \[p:subadditive\] holds. Pick $N_1>N_0$ such that $\frac{a_{N_1}}{N_1^2}\geq \kappa -{\varepsilon}/2$, and pick $N_2> M_0(N_1)$ as in Proposition \[p:subadditive\] such that $\frac{a_{N_2}}{N_2^2}\leq \kappa' +{\varepsilon}/2$. Proposition \[p:subadditive\] now implies that $\kappa'\geq \kappa-2{\varepsilon}$, as required. This completes the proof of the theorem.
The rest of this paper proves Proposition \[p:subadditive\]. Observe that to prove the proposition, we need to obtain a lower bound to $\P(\sU_{\zeta}(m))$ in terms of $\P(\sU_{\zeta}(n))$ for $m\gg n\gg 1$. First (and the most important) step is to construct an event with probability at least $\P(\sU_{\zeta}(n))^{m^2/n^2}$ (upto an error of $e^{-o(m^2)}$) on which we shall have $\{{{\mathbf{T}}}_{m} \geq (\mu+\zeta') m\}$ for $\zeta'$ smaller but arbitrarily close to $\zeta$.
**Throughout the article, for notational brevity we will be omitting the floor signs and ignoring any rounding issue since they will not have any effect on the nature of the arguments.**
Formally we have the following proposition.
\[p:pathconstruct\] For each ${\varepsilon}'\in (0,\zeta)$ and ${\varepsilon}>0$, there exists $N_0$ and $H_0$ such that for all $n>N_0$ and $m>nH_0$ we have $$\log \P(\sU_{\zeta-{\varepsilon}'}(m)) \geq \frac{m^2}{n^2}\log \P(\sU_{\zeta}(n))-{\varepsilon}m^2.$$
Once we have Proposition \[p:pathconstruct\] at our disposal, all we need to prove Proposition \[p:subadditive\] is a way to compare $\P(\sU_{\zeta}(n))$ and $\P(\sU_{\zeta'}(n))$ when $\zeta$ and $\zeta'$ are close. To this end we have the following proposition which essentially says that if the rate function exists it must be continuous in $\zeta$.
\[p:cont\] For each ${\varepsilon}>0$, there exists ${\varepsilon}'>0$ such that for all $n$ sufficiently large we have $$\frac{\log \P(\sU_{\zeta-{\varepsilon}'}(n))}{n^2} \leq \frac{\log \P(\sU_{\zeta}(n))}{n^2}+{\varepsilon}.$$
Our assumption of the edge distribution possessing a continuous density (see Definition \[d:pt\]) is essentially only used in the proof of the above. Although this result can be proven much more generally we have not made such an attempt in this paper. It is easy to complete the proof of Proposition \[p:subadditive\] using Propositions \[p:pathconstruct\] and \[p:cont\].
The proof follows immediately by noticing that $$\frac{a_{m}}{m^2} \geq \frac{\log \P(\sU_{\zeta-{\varepsilon}'}(m))}{m^2}-{\varepsilon}\ge \frac{a_{n}}{n^2} -2{\varepsilon},$$ where the first inequality is the content of Proposition \[p:cont\] and the second inequality is the content of Proposition \[p:pathconstruct\].
The rest of this paper deals with proving Propositions \[p:pathconstruct\] and \[p:cont\]. Proof of Proposition \[p:cont\] is easier. Essentially one shows that to change the passage time ${{\mathbf{T}}}_n$ by ${\varepsilon}' n$ it suffices to increase the passage times of all the edges inside a box of size $O(n)$ by $O({\varepsilon}')$. The cost of such a change can be made as small as possible in the exponential scale by choosing ${\varepsilon}'$ small enough and using the continuity of the density of $\nu$. The only subtle point is that since the variables are supported on $[0,b]$, one cannot increase the values of the edges that already have values close to $b.$ However by choosing the parameters carefully we ensure that there are not too many edges of the latter kind and that the geodesic necessarily passes through many edges whose values are away from $b$ for which the perturbation strategy works. The formal proof appears in Section \[s:cont\]. The remainder of this section presents an outline of the proof of Proposition \[p:pathconstruct\], which is really the heart of this paper.
![As outlined below, the main idea behind the proof of Theorem \[t:ldp\] involves dilating the limiting metric space structure. The figure illustrates a situation when the dilation factor is $5$.[]{data-label="fig1"}](dilate.pdf)
For the purpose of facilitating illustration, we shall only outline the proof in the special case $m=2n$. Also we shall pretend, for the time being that the event $\{{{\mathbf{T}}}_n \geq (\mu+\zeta)n\}$ only depends on the edges weights in the box $B=\llbracket 0, n\rrbracket \times \llbracket-\frac{n}{2}, \frac{n}{2}\rrbracket$ where $\llbracket a,b \rrbracket:= [a,b]\cap \Z$. Observe that this is not deterministically true because the paths are allowed to backtrack. However, we pretend this for the moment for the sake of exposition. In fact the above is true with high probability if one replaces $B$ by a box of side length being a large ($\nu$ dependent) constant times $n$ and centered at the origin. This is what we will do throughout the rest of the paper.
Let ${\varepsilon}$ be an arbitrary small positive number. Suppose that $\P({{\mathbf{T}}}_{n}\geq (\mu+\zeta)n)=p$. So our task is to create an environment on $B_1=\llbracket 0, 2n\rrbracket \times \llbracket-n, n\rrbracket$ with probability at least $p^4$ (upto an error $e^{-o(n^2)}$ on which we shall have $\{{{\mathbf{T}}}_{2n}\geq (\mu+\zeta-{\varepsilon})2n\}$. The basic idea of such a construction is as follows. We condition on the large deviation event $\{{{\mathbf{T}}}_{n}\geq (\mu+\zeta)n\}$ and look at an environment $\omega$ in $B$. We show that with high probability $\omega$ is such that $B$ can be tiled by sub-boxes of size $k\times k$ which we will call ‘tiles’ (see Figure \[fig2\]), most of the tiles are **stable**. We describe below roughly the notion of stability which makes precise the notion of a limiting metric space structure as alluded to at the beginning of Section \[outline\].
- Consider a tile and for each $\bz$ in the tile and any $\theta \in \bS^1$, let $\bz_1$ and $\bz_2$ be points such that $\bz,\bz_1,\bz_2$ lie in a straight line making angle $\theta$ with the $x-$axis and $\|\bz-\bz_1\|_2=\|\bz_1-\bz_2\|_2=k$ where $\|\cdot\|_2$ denotes the Euclidean norm. Then the box is said to stable if $$\label{stab908}
{{\mathbf{PT}}}(\bz,\bz_1)=(1+o(1)){{\mathbf{PT}}}(\bz_1,\bz_2)=(1+o(1))\frac{{{\mathbf{PT}}}(\bz,\bz_2)}{2},$$
(see Section \[gradstabsec\] for formal definitions). Thus the above says that for any $\bz$ in the tile, the passage time from $\bz$ acts approximately like a linear function in every direction $\theta$ at scale $k.$ Note that this linear function a priori depends on the environment $\omega.$ However it can be shown that with significant probability the environments $\omega$ approximately yield the same linear function. Even though for the actual proof we will need a result stronger in many senses, below we illustrate how to exploit the above property.
![Figure illustrating the proof sketch below where for every path $\gamma'$ in the dilated environment $\omega'$ there exists a path $\gamma$ obtained by scaling down the endpoints of the excursions. However note that a priori $\gamma_{i}$ need not be excursions even though $\gamma'_i$ are by definition. The former are just taken to be the shortest path in the environment $\omega$ between the end points of $\gamma'_i$ divided by $2.$ Here $k=\frac{n}{2}.$[]{data-label="fig2"}](dilaoutline.pdf)
Given a tiling of $B$ in to stable $k\times k$ tiles we construct an environment on $B_1$ by independently sampling environments $\omega_1,\omega_2,\omega_3,\omega_4$ on $B$ with the same law as $\omega$. Using the latter, we now tile $B_1$ using tiles of size $2k\times 2k$ where each such tile is formed from $4$ tiles of size $k \times k$ (one from each $\omega_i$) as illustrated in Figure \[fig2\]. Let us call the constructed environment $\omega'.$ Given such a construction, we would be done once we establish the following two properties:
1. The constructed event has probability comparable to $p^4$ which follows quite easily since we picked four independent copies of environments in $\sU_{\zeta}(n)$ to obtain the environment on $B_1$.
2. To show that any path $\gamma'$ in $\omega'$ between $\bo{0}$ and $\bo{2n}$ has length at least $(\mu+\zeta-{\varepsilon})2n$ for some small ${\varepsilon}.$
To show the latter we decompose $\gamma'$ into excursions $\gamma'_1,\gamma'_2,\ldots$ where each $\gamma'_i$ resides in a tile of size $2k\times 2k$ and $\gamma'_i$ and $\gamma'_{i+1}$ reside in separate tiles. Thus $|\gamma'|=\sum_{i}|\gamma'_i|.$ Now the key is to observe that for such a $\gamma'$ one can create a path $\gamma$ in the environment $\omega$ between $\bo{0}$ and $\bo{n}$ such that $\gamma$ is a concatenation of paths $\gamma_1,\gamma_2,\ldots$. This is done by just taking $\gamma_i$ to be the shortest path between points which are the endpoints of $\gamma'_i$ scaled down by a factor $2.$ (see Figure \[fig2\]). However note that $\gamma_{i}$ need not be excursions even though $\gamma'_i$’s are by definition. The former are just taken to be the shortest path in the environment $\omega$ between the points obtained by dividing the end points of $\gamma'_i$ by $2.$
The stability of the tiles now imply that $|\gamma_{i}|=(1+o(1))\frac{|\gamma'_i|}{2}$. Thus it follows that $$\label{dila453}
|\gamma'|=\sum_{i}|\gamma'_i|=2(1+o(1))\sum_i|\gamma_i|=2(1+o(1))|\gamma|\ge 2(1+o(1))(\mu+\zeta)n$$ where the last inequality follows by definition as $\gamma$ is a path between $\bo{0}$ and $\bo{n}$ in the environment $\omega$ which is in $\sU_{\zeta}(n).$
There are a few obstacles in making this outline rigorous and some work is needed to circumvent them as we briefly outline below.
\(1) The most important step is to prove that $B$ can be divided into such stable tiles. In fact we prove that there exists a tiling of $B$ where most tiles are stable, i.e., the total number of points in unstable tiles is $o(n^2)$. This essentially is a property of a general metric structure on $B$ which is bi-Lipschitz with respect to the Euclidean metric (the fact that the FPP metric has this property is a consequence of the shape theorem in . We record this observation in Lemma \[lb90\]). The formal stability result is Proposition \[t:stable\] in this paper and the proof is provided in Section \[s:sproof\] where a detailed outline of the proof and an elaborate explanation of the key ideas can be found.
Intuitively the result says that any sub-sequential limiting metric structure due to its bi-Lipschitz nature should have a reasonably smooth gradient function. Thus the size of the tiles capture the scale at which an approximate smoothness is witnessed. However formally we show (see Proposition \[t:stable\]) that all but at most a small fraction of tiles are stable and the unstable tiles can be handled by replacing all the edge values in those by values close to $b$ (recall that $\nu$ is supported on $[0,b]$). This operation only can increase the passage time and hence makes the upper tail event more likely and on the other hand it only costs $e^{-o(n^2)}$ in probability and hence does not change any of the conclusions.
\(2) Finally we describe briefly another point which we have swept under the carpet so far. All the discussion above describes how to construct a $2n\times 2n$ environment out of an $n\times n$ environment preserving (upto an error) the upper tail large deviation event. However observe that in order to prove Proposition \[p:pathconstruct\], we need to be able to dilate the original environment by factor $h=\frac{m}{n}$ which could be arbitrarily large. To ensure that the error term $(1+o(1))$ in does not blow up we will in fact modify the notion of stable tiles which allows dilation by an arbitrary factor $h$. To ensure this we prove that stable tiles have a couple of additional properties:
- First of all we need to ensure stability at most locations at many consecutive length scales rather than just two as in .
- More importantly, we show that as the direction vector is varied at a given location, the gradient field has approximate convexity properties. This result should be thought of as a weak analogue of the convexity of the limiting shape in in the upper tail large deviation regime and this will enable us to compare the distance function between the $k\times k$ box and the $kh\times kh$ box. The formal convexity statement is stated as Proposition \[conc1\] and the proof is presented in Section \[pconc1\].
Organization
------------
We finish off this introduction by describing the organization of the remainder of this paper. In Sections \[s:grad\] and \[prelim\] we set up the notation and make a precise statement of the stabilization result Proposition \[t:stable\]. We also make precise definition and statement of the regularity results of the gradient field. The proofs of these results are postponed until later. In Section \[s:construct\] we use these results to prove Proposition \[p:pathconstruct\]. In Sections \[s:cont\] and \[pconc1\] we provide the proofs of the continuity of rate function (Proposition \[p:cont\]) and approximate convexity of the distance function (Proposition \[conc1\]) respectively. Finally in Section \[s:sproof\] we prove the stability result Proposition \[t:stable\] to complete the argument. For easy reference, below we summarize some of the notations and the parameters (already defined or to be defined later), that will be used frequently throughout the article.
[**Notation**]{} [**Defined in**]{} [**Short Informal Description**]{}
--------------------------------------------------------------------- --------------------------- ---------------------------------------------------------------
$\sU_{\zeta}(n)\, (\sL_{\zeta}(n))$ See Upper-tail (lower tail) events
at scale $n$.
$\Pi=(X_e: e\in E(\Z^2))$ Section \[s:grad\] Typical noise space.
$ \Pi^{\sU}=(X^{\sU}_{e}: e\in E(\Z^2))$ Section \[s:grad\] Noise space conditioned on
$\sU_{\zeta}(n)$.
$\sC$ Lemma \[apriori23\] We will restrict ourselves to a box
size $\sC n$ outside which the geodesic
to $\bo{n}$ does not escape w.h.p..
$\sU^*_{\zeta}(n)\, $ See Upper-tail event restricted
inside box of size $4\sC n$.
${{\mathbf{Box}}}(r)$ (${{\mathbf{L-Box}}}(r)$) Section \[s:grad\] The continuous box (lattice box)
$[-r,r]^2$ ($\llbracket -r,r\rrbracket^2$).
$\bS^1(\eta)$ See Discretized unit circle:
$[0,\eta,2\eta,\ldots, 2\pi]$.
$\sE$ Lemma \[lb90\] Lower bound on passage times.
$\sS(\bz,\theta,\ell,k)$ See Discrete segment.
${{\mathbf{PT}}}(\cdot,\cdot)$ Section \[mods\] Passage time.
$\bz \text{ is }(\delta,\theta,\ell,k)-{{\mathbf{Stable}}}$ See Distance function grows linearly.
in direction $\theta$.
${{\mathbf{Tile}}}_n(j,v)$ Definition \[deftile543\] Tile of size $\frac{n}{2^j}$ of ${{\mathbf{Box}}}(n)$ in
corresponding to $v\in{{\llbracket{1,2^j}\rrbracket^2}}$.
$\grad(\bz,\theta,\ell)$ See Gradient function.
${{\mathbf{Grid}}}_n(j)$ Section \[prelim\] Points with spacing $\frac{n}{2^j}$ in ${{\mathbf{Box}}}(n).$
${{\mathbf{Grid}}}_n(\ell;j)$ Section \[prelim\] Points with spacing $\ell$ on
the edges of ${{\mathbf{Grid}}}_n(j).$
$ \overset{\eta_1,\ell_1,j_1}{{{\mathbf{Proj}}}}_n(\bo{z},\bo{w})$ See Projected distance for pairs
of points in a grid at scale $\eta_1.$
: Table of glossaries
Acknowledgements {#acknowledgements .unnumbered}
----------------
Research of RB is partially supported by a Simons Junior Faculty Fellowship and a Ramanujan Fellowship from Govt. of India. SG is supported by a Miller Research Fellowship.
Formal definitions and notations {#s:grad}
================================
Throughout the remainder of this paper we shall fix a passage time distribution $\nu$ that satisfies the hypothesis of Theorem \[t:ldp\], i.e., it is supported on $[0,b]$ with a continuous density function. This in particular implies that passage times are not-concentrated on one point and there is no mass at $0$, which in turn implies that the shape theorem holds. For this passage time distribution and a direction vector $\bo{v}\in \bS^1$, we shall denote by $\mu_{\bo{v}}$ the time constant in direction $\bo{v}$ (as introduced in the previous section, for $\bo{v}=(1,0)$ we shall drop the subscript). Under these condition one can prove the following basic concentration estimate (see e.g. [@Kes86]) for each ${\varepsilon}>0$, $\bo{v}\in \bS^1$, some $c>0$ and all $n$ sufficiently large we have ($\lfloor n\bo{v} \rfloor$ is the vertex in $\Z^2$ obtained by taking co-ordinate wise integer parts of $n\bo{v}$): $$\label{e:concentration1}
\P(|{{\mathbf{PT}}}({\mathbf{0}, \lfloor n\bo{v} \rfloor})-\mu_{\bo{v}} n|\geq {\varepsilon}n)\leq e^{-cn}.$$ We shall use many times and often implicitly without referring to it. Notice that we are concerned with the large deviation regime whereas is for typical environments. To use it in the large deviation regime we need a tool to compare the environment in the large deviation regime with the typical environment. This is provided by the FKG inequality. Let $\Pi=(X_e: e\in E(\Z^2))$ and $\Pi^{\sU}=(X^{\sU}_{e}: e\in E(\Z^2))$ be the typical and conditional (on $\sU_{\zeta}(n)$) edge weight environments respectively. Let ${{\mathbf{Geo}}}_{\sU}(\cdot,\cdot)$ denote the geodesics in the environment $\Pi^{\sU}$. The following lemma is a well known consequence of the FKG inequality (see for e.g. Strassen’s Theorem).
\[FKG\] There exists a coupling $(\Pi,\Pi^{\sU})$ such that almost surely, for each edge $e$ we have $X_{e}\le X^{\sU}_e.$
There are two main consequences of Lemma \[FKG\] that will be useful for us. First, this will provide lower bounds on the FPP metric conditional on $\sU_{\zeta}$, and second, it will enable us to restrict our attention to finite boxes. Before we proceed with the relevant statements, we extend the function ${{\mathbf{PT}}}$ from $\Z^2\times \Z^2$ to $\R^2\times \R^2$; this will reduce notational complexities significantly. There is not one canonical way to this, we choose the following extension for concreteness. For every $x,y\in \R^2$ define ${{\mathbf{PT}}}(x,y):={{\mathbf{PT}}}(\hat x, \hat y)$ where $\hat x$ and $\hat y$ are the nearest lattice points to $x,y$ respectively, (in case of a tie, we choose the one which is smallest in the usual lexicographic order on $\Z^2.$). We introduce some more useful notations. Throughout we will use ${{\mathbf{Box}}}(r)$ (resp. ${{\mathbf{L-Box}}}(r)$) to denote the box $[-r,r]^2\subseteq \R^2$ (resp. the box $\llbracket -r,r \rrbracket^2 \subseteq \Z^2$).
The next lemma shows that, geodesics do not wander too much even in the large deviation regime. Let $\mu_{\min}=\min_{v\in \bS^1} \mu_v$. It a consequence of that $\mu_{\min}>0$. Let us fix $\sC= \frac{4b}{\mu_{\min}}$. This $\sC$ will be important for us and will be fixed throughout the paper.
\[apriori23\] For all $\zeta \in (0, b-\mu)$There exists $c>0$ such that for all $n$ sufficiently large we have, $$\P({{\mathbf{Geo}}}_{\sU}(\bo{0}, \bo{n})\subset {{\mathbf{Box}}}(\sC n))\geq 1-e^{-cn}.$$
Observe that together with an union bound over the lattice points on the boundary of ${{\mathbf{Box}}}(\sC n)$ implies that ${{\mathbf{PT}}}(\mathbf{0}, \Z^2\setminus {{\mathbf{Box}}}(\sC n))\ge 2 bn$ with exponentially small failure probability in the typical environment. Lemma \[FKG\] implies that the same is true for the environment $\Pi^{\sU}$.
The next lemma shows that the in the environment $\Pi^{\sU}$, with high probability the FPP metric within ${{\mathbf{Box}}}(\sC n)$ is lower bounded by a constant multiple of the Euclidean metric.
\[lb90\] There exists $\alpha, c>0$ such that for all sufficiently large $n$, with conditional ( on $\sU_{\zeta}$) probability at least $1-e^{-c\sqrt{n}}$ the following holds: for any two points $\bz, {{\mathbf{w}}}\in {{\mathbf{Box}}}(4\sC n)$ with $|\bz-{{\mathbf{w}}}|\ge \sqrt{n},$ the passage time between $\bz$ and ${{\mathbf{w}}}$ restricted inside ${{\mathbf{Box}}}(4\sC n)$ is at least $\alpha |z-w|.$
This high probability event described above will be useful for us, and we denoted by $\sE$ for future reference.
The lemma follows by taking a union bound over all pairs of points in ${{\mathbf{L-Box}}}(4 \sC n)$ with mutual distance at least $\sqrt{n}-3$, and using Lemma \[FKG\] together with (take $\alpha= \frac{1}{2}\mu_{\min}$ for example).
Thus from the above lemmas we can restrict ourselves to ${{\mathbf{Box}}}(4\sC n)$ by defining the event $\sU^*_{\zeta}=\sU_{\zeta}\cap \sE$ where we take $\alpha=\frac{1}{2}\mu_{\min}.$ Notice that $\sU^*_{\zeta}$ is just a function of the edges in ${{\mathbf{Box}}}(4\sC n).$ Now by Lemma \[lb90\] $$\label{e:comparison}
(1-e^{-c\sqrt n})\P(\sU^*_{\zeta})\leq \P(\sU_{\zeta})$$ This allows us to work with $\sU^*_{\zeta}$ instead of $\sU_\zeta$ and this is what we will do throughout the article.
Gradients and stability {#gradstabsec}
-----------------------
To precisely state the stabilization that we have alluded to, we need to develop some more notation. For our purposes, we shall be comparing distance functions for fixed directions, so we introduce the following notation. For $\bz \in \R^2,$ and $\theta \in \mathbb{S}^1$ (the unit circle), let $\L_{\theta, \bz}=\{\bz+\lambda\theta: \lambda>0\}$, i.e., in the standard parametrization of $\mathbb{S}^1$, $\L_{\theta, \bz}$ denote the the ray starting from $\bz$ in the direction $\theta$. We shall consider a sequence of equally spaced points along $\L_{\theta, \bz}$ defined as follows. For $\bz\in \R^2, \theta\in \mathbb{S}^1, k\in \N$ and $\ell>0$, let us define the discrete segment $$\label{disseg}
\sS(\bz,\theta,\ell,k)=[\bz_0,\bz_1,\ldots,\bz_k]$$ where $\bz_0=\bz$ and $\bz_{i+1}=\bz_i+\ell \theta$, see Figure \[fig3\].
![$k$ points spaced at distance $\ell$ along a line making angle $\theta$ with the $x-$axis forming $\sS(\bz,\theta,\ell,k)$.[]{data-label="fig3"}](disseg.pdf)
We define the passage time for the segment $\sS$ by $$\label{pt12}
{{\mathbf{PT}}}(\bz,\theta,\ell,k):=\sum_{i=0}^{k-1} {{\mathbf{PT}}}(\bz_i,\bz_{i+1}).$$
Now note that the starting point and ending points of $\sS({\bz,\theta, \frac{\ell}{2}, 2k})$ and $\sS({\bz,\theta, \ell, k})$ are the same and the former is obtained from the latter by subdividing subintervals of the latter in to equal halves.
As an easy consequence of the triangle inequality we have the following straightforward lemma.
\[monotone\] ${{\mathbf{PT}}}(\bz,\theta,\frac{\ell}{2}, 2k)\ge {{\mathbf{PT}}}(\bz,\theta,\ell,k).$
The main arguments in this paper rely on a notion of stability of the passage time at a point $\bz.$ Fix a tolerance parameter $\delta>0.$ For $k\in \N$, $\ell>0$ and $\theta\in \mathbb{S}^1$, we say that $\bz\in \R^2$ is $(\delta, \theta, \ell, k)-{{\mathbf{Stable}}}$ (with respect to any edge weight configuration $\Pi$) if for $1\le k'\le k,$ $$\begin{aligned}
\label{stabnot32}
\frac{k'{{\mathbf{PT}}}(\bz,\theta,\ell,1)}{(1+\delta)}\le {{\mathbf{PT}}}(\bz,\theta,\ell k', 1)\le(1+\delta)k' {{\mathbf{PT}}}(\bz,\theta,\ell,1).\end{aligned}$$
In words, $\bz\in \R^2$ is $(\delta, \theta, \ell, k)-{{\mathbf{Stable}}}$ if the passage time from $z$ to $z+(\theta, \ell k')$ can be approximated up to a $(1+\delta)$ multiplicative error by $k'$ times the passage time from $z$ to $z+(\theta, \ell)$ for all $1\le k'\le k$. This captures the linear growth of the distance function.
In the following for convenience we would work with a discretized version of $\bS^1$. For any $\eta>0,$ let $$\label{discirc}
\bS^1(\eta)=\{0,\eta,2\eta,\ldots 2\pi-\eta\}.$$ ($\eta$ is assumed to have the required properties to avoid rounding issues. Also throughout the article, we will use $\theta$ interchangeably to denote an angle or a unit vector making the corresponding angle with the $x-$axis. The usage will be clear from context.) In the sequel we will say that $\bz$ is $(\delta, \bS^1(\eta),\ell,k)-{{\mathbf{Stable}}}$ if $\bz$ is $(\delta,\theta,\ell,k)-{{\mathbf{Stable}}}$ for each $\theta \in \bS^1(\eta)$ and similarly we will say that $\bz$ is $(\delta,\ell,k)-{{\mathbf{Stable}}}$ if $\bz$ is $(\delta,\theta,\ell,k)-{{\mathbf{Stable}}}$ for each $\theta \in \bS^1$.
With this preparation, we can now state an initial version of our stabilization result.
\[t:stable\] Fix $\delta,{\varepsilon},\eta >0,$ and $k\in \N$ and $J_1\in \N$. There exists $J_2\in \N$ such that for all large enough $n$, conditioned on $\sU^*_{\zeta}(n)$ the following holds: there exists $J_1\leq j\leq J_2$ (random depending on $\Pi \in \sU^*_{\zeta}(n)$ ) such that $$\#\{\bz\in {{\mathbf{L-Box}}}(\sC n): \bz~\text{is not}~(\delta, \bS^1(\eta),\frac{\sC n}{2^j},k)-{{\mathbf{Stable}}}\} \leq {\varepsilon}n^2 .$$
Proof of Proposition \[t:stable\] is rather technical and is postponed until Section \[s:sproof\]. This is one of the three main ingredients of our proof, and we state this result in terms of ${{\mathbf{L-Box}}}(\sC n)$ so that it can directly be fed into many of the later arguments. However, for the next few definitions and results it will be notationally convenient to work with boxes of size $n$.
We next define the gradient function for ${{\mathbf{Stable}}}$ points naturally in the following way: For $\theta \in \bS^1,$ and $\ell \in \N,$ let $$\label{grad87}
\grad(\bz,\theta,\ell)=\frac{{{\mathbf{PT}}}(\bz, \bz+ (\theta,\ell))}{\ell}.$$ An easy consequence of the notion of stability is that the gradient function stays almost constant over a range of values of $\ell.$
\[eascons\] Fix $j\in \N$. On the event $\sE$ (see Lemma \[lb90\]), for all sufficiently large $n$, for any $\ell\ge \frac{n}{2^j}$, and for any $(\delta,\bS^1(\eta), \ell,k)-{{\mathbf{Stable}}}$ point $\bz,$ for any $\frac{k}{4}\ell \le \ell', \ell''\le k\ell,$ and for any $\theta \in \bS^1$ $$\grad(\bz,\theta,\ell')=\left(1+O\bigl(\eta+\delta +\frac{1}{k}\bigr)\right)\grad(\bz,\theta,\ell'').$$
The above lemma without the $O(\eta)$ term in the multiplicative factor follows immediately from definition of stability for all $\theta$ in $\bS^1(\eta)$. However we need to extend this to all $\theta\in \bS^{1}$, and a further approximation is necessary. For any $\theta\in \bS^1$ let $\hat \theta$ be the closest point in $\bS^1(\eta).$ Then by triangle inequality for any $\ell$, it follows that $$\begin{aligned}
\label{tria1}
{{\mathbf{PT}}}(\bz,\bz+(\theta,\ell))&\le {{\mathbf{PT}}}(\bz,\bz+(\hat \theta,\ell))+b\eta \ell,\\
\nonumber
{{\mathbf{PT}}}(\bz,\bz+(\hat \theta,\ell))&\le {{\mathbf{PT}}}(\bz,\bz+( \theta,\ell))+b\eta \ell;\end{aligned}$$ since the edge variables are bounded by $b$. This completes the proof of the lemma with the addition of the $O(\eta)$ term in the multiplicative error.
Note that Proposition \[t:stable\] claims that most points in ${{\mathbf{Box}}}(\sC n)$ are stable. We now prove a stable point implies stability in a neighbourhood with slightly worse parameters.
\[stab23\] For $k>C>m>0$, $\ell>0$ and for any $\bz$ which is $(\delta,\theta,\ell,k)-{{\mathbf{Stable}}}$ and any $\bz'$ such that $|\bz-\bz'|\le \ell m$ we have $\bz'$ is $(\delta',\theta,C \ell, \frac{k}{C})-{{\mathbf{Stable}}},$ where $\delta'= \delta +O(\frac{m}{C}).$
The proof follows by another application of triangle inequality where we observe the following, analogous to : For any $\ell',$ $$\begin{aligned}
\label{tria2}
{{\mathbf{PT}}}(\bz,\bz+(\theta,\ell'))&\le {{\mathbf{PT}}}(\bz',\bz'+(\theta,\ell'))+b \ell m,\\
\nonumber
{{\mathbf{PT}}}(\bz',\bz'+(\theta,\ell'))&\le {{\mathbf{PT}}}(\bz,\bz+(\theta,\ell'))+b \ell m.\end{aligned}$$ Hence for any $\ell'\ge C\ell$ it follows that, ${{\mathbf{PT}}}(\bz,\bz+(\theta,\ell'))=(1+O(\frac{m}{C})){{\mathbf{PT}}}(\bz',\bz'+(\theta,\ell')),$ (see Figure \[fig4\] for an illustration).
An immediate but important corollary is the following smoothness of the gradient field which we state without proof.
\[smoothgrad\] Given $\delta$ and $\delta'$ as in Lemma \[stab23\], for all large $k$ and all large $n$ and any $\bz, \bz'$ satisfying the hypothesis of that lemma, and for all $\theta \in \bS^{1},$ $$\frac{1}{1+\delta'}\le \frac{\grad(\bz,\theta,\ell)}{\grad(\bz',\theta, C\ell)}< 1+\delta'.$$
![Stability for the discrete segment formed by the red points implies the stability for the nearby segment formed by the blue points.[]{data-label="fig4"}](stab.pdf)
Stability of Tiles
------------------
In this subsection we introduce the notion of stability of tiles parallel to the notion of stability for points, which will be convenient for the proofs. The section contains a few lemmas which even though quite similar to the ones already stated, have various associated quantifiers which could make it a little hard to read and the reader can choose to skip the straightforward proofs in this section. This will not affect readability of the future sections.
Given a lattice box ${{\mathbf{L-Box}}}(n)$ we will often think of it as made up of boxes of a particular scale $j$, i.e. think of the box as being naturally tiled using boxes of size $n/2^j$. Note that one can define a natural bijection between the set of tiles and the set $\llbracket 1, 2^j\rrbracket^2$. We will use this bijection to denote the tile corresponding to $v\in \llbracket 1, 2^j\rrbracket^2$ by ${{\mathbf{Tile}}}_n(j,v)$ (see Figure \[fig5\]).
\[deftile543\]For any $v\in \llbracket 1, 2^j \rrbracket^2$, a tile ${{\mathbf{Tile}}}_n(j,v)$ is said to be $(\delta, \bS^1(\eta), \ell, k, {\varepsilon})-{{\mathbf{Stable}}}$ if at least $1-{\varepsilon}$ fraction of the lattice points in ${{\mathbf{Tile}}}_n(j,v)$ are $(\delta, \bS^{1}(\eta),\ell,k)-{{\mathbf{Stable}}}.$
**In the sequel we will choose $\ell=\frac{n}{2^{j+m}}$ and $k=2^{2m}$ for some $j,m\ll \log_2(n)$, where the choice of $j$ and $m$ will vary through the paper and will depend on some other parameters relevant for specific applications.**
![The first figure illustrates the tiling an $n\times n$ box in to tiles of size $\frac{n}{4}.$ Thus the set of tiles has a natural bijection with $\llbracket 1, 4\rrbracket^2$.[]{data-label="fig5"}](tile.pdf)
Using Lemma \[stab23\], we now prove that if at least $(1-{\varepsilon})$ fraction of the lattice points in a ${{\mathbf{Tile}}}_{n}(j,v)$ are stable for some values of the parameters, then all the points are stable for a slightly different range of parameters.
\[ssttile\] Let $j,m \in \N$, and $\ell=\frac{n}{2^{j+m}}, k=2^{2m}$. Fix $\delta>0$. There exists $C>0$ sufficiently large such that for all sufficiently small ${\varepsilon}>0$, on $\sE,$ the following holds for all sufficiently large $n$: if ${{\mathbf{Tile}}}_n(j,v)$ is $(\delta,\bS^{1}(\eta),\ell,k, {\varepsilon})-{{\mathbf{Stable}}}$, then ${{\mathbf{Tile}}}_n(j,v)$ is $(2\delta, \bS^{1}(\eta), \ell',k', 0)-{{\mathbf{Stable}}}$ where $\ell'=\max(\frac{n}{2^{j}}C\sqrt{{\varepsilon}},\ell)$ and $k'=k\ell/\ell'$.
Observe that for every ${{\mathbf{Tile}}}_n(j,v)$ that is $(\delta,\bS^{1}(\eta), \ell,k, {\varepsilon})-{{\mathbf{Stable}}}$ and any $\bz \in {{\mathbf{Tile}}}_n(j,v)$ there exists ${{\mathbf{w}}}\in {{\mathbf{Tile}}}_n(j,v)$ with $|\bz-{{\mathbf{w}}}| \le 8\sqrt {\varepsilon}\frac{n}{2^j}$ and ${{\mathbf{w}}}$ is $(\delta,\bS^1(\eta), \ell,k)-{{\mathbf{Stable}}}.$ This is because the existence of a $\bz$ for which there is no such ${{\mathbf{w}}}$ contradicts the hypothesis that ${{\mathbf{Tile}}}_n(j,v)$ is $(\delta, \bS^{1}(\eta), \ell,k, {\varepsilon})-{{\mathbf{Stable}}}$. The proof now follows from Lemma \[stab23\] by for $C$ sufficiently large (and ${\varepsilon}$ sufficiently small).
From now on we will call a $(\delta,\bS^{1}(\eta),\ell',k', 0)-{{\mathbf{Stable}}}$ tile as a $(\delta,\bS^{1}(\eta),\ell',k')-{{\mathbf{Stable}}}$ tile. We now show that the above in fact implies stability for all angles $\theta \in \bS^1.$
\[l:gradtile\] Let $j, m$ be as in the previous lemma. Then on $\sE,$ the following holds for all sufficiently large $n$: for a $(\delta,\bS^1(\eta), \ell,k)-{{\mathbf{Stable}}}$ ${{\mathbf{Tile}}}_n(j,v)$ for $\bz, \bz' \in {{\mathbf{Tile}}}_n(j,v)$ we have for all $\theta \in \bS^1,$ $$\begin{aligned}
\label{stabgrad12}
\frac{1}{1+\delta'}&\le \frac{\grad(\bz,\theta,k_1\ell)}{\grad(\bz',\theta, k_2 \ell)}< 1+\delta',\end{aligned}$$ with $\delta'= O(\delta+\eta+ \frac{1}{2^m})$ and $1\le k_1 ,k_2 \le k.$
The proof is quite similar to that of Lemma \[ssttile\]. Recalling if for any $\theta \in \bS^1$, $\hat \theta$ is the closest point in $\bS^1(\eta),$ then for any $k_1\le k,$ $$\begin{aligned}
|{{\mathbf{PT}}}(\bz,\bz+(\theta, k_1\ell))- {{\mathbf{PT}}}(\bz,\bz+(\hat \theta, k_1\ell))|\le b\eta k_1 \ell,\\\end{aligned}$$ which along with the hypothesis that $\bz$ is $(\delta,\bS^1(\eta), \ell,k)-{{\mathbf{Stable}}}$ implies that $$\begin{aligned}
\frac{1}{1+O(\delta+\eta)}&\le \frac{\grad(\bz,\theta,k_1\ell)}{\grad(\bz,\theta, k_2 \ell)}< 1+O(\delta+\eta)).\end{aligned}$$ Now another application of triangle inequality as in , shows that for any $\bz,\bz' $ as in the statement of the lemma, $$\begin{aligned}
{{\mathbf{PT}}}(\bz,\bz+(\theta, k \ell))&\le {{\mathbf{PT}}}(\bz',\bz'+(\theta,k \ell))+O(b\frac{n}{2^j}).\end{aligned}$$ Hence using the fact that $k\ell=\frac{2^m n}{2^j},$ it follows that, $
\frac{1}{1+O(\delta+\eta+\frac{1}{2^m})}\le \frac{\grad(\bz,\theta,k\ell)}{\grad(\bz',\theta, k \ell)}< 1+O(\delta+\eta+\frac{1}{2^m}).
$
Thus from now on, we shall refer to a tile as $(\delta,\ell,k)-{{\mathbf{Stable}}}$ if is satisfied with $\delta$ in place of $\delta'$. Now for a $(\delta,\ell,k)-{{\mathbf{Stable}}}$ ${{\mathbf{Tile}}}_n(j,v)$ as above, allows us to define a gradient function not for every individual point $\bz$ but for the whole tile itself.
\[tilegrad\] For a $(\delta,\ell,k)-{{\mathbf{Stable}}}$ ${{\mathbf{Tile}}}_n(j,v)$ define for any $\theta \in \bS^1,$ $$\grad_n((j,v),\theta)= \grad_n((j,v),\theta, \ell):=\grad(\bz,\theta, \ell)$$ for the center point $\bz$ of ${{\mathbf{Tile}}}_n(j,v).$
Observe that even though this definition implicitly depends on $\ell$, we shall drop it from our notation as the length scale $\ell$ will always be clear from the context. The reason for calling the the gradient function for ${{\mathbf{Tile}}}_n(j,v)$ is the following: even if we replace the centre of ${{\mathbf{Tile}}}_n(j,v)$ by any arbitrary $\bz\in {{\mathbf{Tile}}}_n(j,v)$, the value of the gradient changes only by a multiplicative factor of $(1+\delta)$; in all our applications, by proper choice of parameters $\delta$ will be made arbitrarily close to zero.
With the above preparation we shall now go back to the setting of Proposition \[t:stable\] and show that there exists a scale $j$ such that, conditional on $\sU^*_{\zeta}(n)$, with probability bounded below most of the scale $j$ tiles in ${{\mathbf{Box}}}(\sC n)$ are stable.
\[goodscale1\] Conditional on $\sU^*_{\zeta}(n),$ (recall that this was an event on ${{\mathbf{L-Box}}}(4\sC n)$). Then given $\eta, m,\delta,{\varepsilon}_1, J_1$ such that $\frac{1}{2^m}\ge \sqrt{{\varepsilon}_1},$ there exists a constant $J_2$ such that for all large enough $n,$ there exists a scale $J_1\le j\le J_2 $ (depending on $n$) such that with probability at least $\frac{1}{J_2}$, for all but ${\varepsilon}_1$ fraction of $v\in \llbracket 1, 2^{j}\rrbracket^2,$ ${{\mathbf{Tile}}}_{\sC n}(j,v)$ is $(\delta,\bS^1(\eta),\ell,k,{\varepsilon}_1)-{{\mathbf{Stable}}}$ (see Definition \[deftile543\]) where $\ell=\frac{\sC n}{2^{j+m}}$ and $k=2^{2m}.$
Note that from the statement of Proposition \[t:stable\] choosing $k=2^{4m}$ and ${\varepsilon}={\varepsilon}_1^2$ it follows that there exists a scale $j$ such that with probability at least $\frac{1}{J_2}$ ($J_2$ appearing in the statement of Proposition \[t:stable\]) the fraction of points $\bz$ in ${{\mathbf{L-Box}}}(\sC n)$ which are not $(\delta,\bS^{1}(\eta),\ell,k)-{{\mathbf{Stable}}}$ is at most ${\varepsilon}n^2$ where $\ell=\frac{\sC n}{2^{j+m}}$ and $k=2^{2m}.$ Thus the total fraction of $v\in \llbracket 1, 2^{j}\rrbracket^2$ such that ${{\mathbf{Tile}}}_{\sC n}(j,v)$ is not $(\delta, \bS^1(\eta), \ell,k,{\varepsilon}_1)-{{\mathbf{Stable}}}$ is at most ${\varepsilon}_1$ since other wise the total fraction of points $\bz\in {{\mathbf{L-Box}}}(\sC n)$ that are not $(\delta, \bS^1(\eta),\ell,k)-{{\mathbf{Stable}}}$ will be more than ${\varepsilon}={\varepsilon}_1^2$ contradicting the conclusion of Proposition \[t:stable\].
The above result along with Lemma \[ssttile\] now implies that most of the tiles are stable with the parameter ${\varepsilon}$ being set to $0$, and other parameters slightly worsened.
\[most34\] Given small enough $\delta_1,{\varepsilon}_1>0$ and a positive integer $m_1,$ such that $\frac{1}{2^{m_1}}\ge {\varepsilon}_1^{1/4},$ and $J_1 \in \N$ there exists $J_2$ such that for all large enough $n,$ conditioned on $\sU^*_{\zeta}(n)$ there exists $J_1\le j_1< J_2$ (depending on $n$) such that with probability at least $\frac{1}{J_2}$ the fraction of $v\in \llbracket 1, 2^{j_1}\rrbracket^2$ such that ${{\mathbf{Tile}}}_{{\sC n}}(j_1,v)$ is not $(\delta_1,\ell_1, k_1)-{{\mathbf{Stable}}}$ is at most ${\varepsilon}_1$ where $\ell_1=\frac{n}{2^{j_1+m_1}}$ and $k_1=2^{2m_1}$.
The proof will follow by first using Lemma \[goodscale1\] with some choice of parameters $\eta, \delta, m, {\varepsilon}_1, J_1$ which implies the existence of $j_1$ such that with probability at least $\frac{1}{J_2}$, for all but ${\varepsilon}_1$ fraction of $v\in \llbracket 1, 2^{j_1}\rrbracket^2,$ ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ are $(\delta,\bS^1(\eta),\ell,k,{\varepsilon}_1)-{{\mathbf{Stable}}}$ (see Definition \[deftile543\]) where $\ell=\frac{\sC n}{2^{j_1+m}}$ and $k=2^{2m}$ for some values of $\delta$ and $m$. We will now apply Lemma \[ssttile\] to conclude from the above that all but ${\varepsilon}_1$ fraction of $v\in \llbracket 1, 2^{j_1}\rrbracket^2,$ ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ are $(2\delta, \bS^{1}(\eta), \ell',k', 0)-{{\mathbf{Stable}}}$ where $\ell'=\max(\frac{n}{2^{j_1}}C\sqrt{{\varepsilon}},\ell)$ and $k'=k\ell/\ell'$ for some $C.$ Now applying Lemma \[l:gradtile\] we conclude that each tile of the latter kind is in fact $(\delta'',\ell',k')-{{\mathbf{Stable}}}$ for some $\delta''>0.$ It can now be verified that our initial choice of parameters can be made such that $\delta'',\ell',k'$ matches the parameters in Lemma \[most34\].
Throughout the article Lemma \[most34\] will govern our choices of parameters.
Technical preliminaries {#prelim}
=======================
As mentioned in our proof strategy, we shall take a configuration from the large deviation regime at some length scale $n$, and replicate/dilate the same configuration to obtain a configuration at a higher length scale. The obvious problem one notices is that for continuous passage time distributions, each configuration has probability $0$. Hence to carry out our proof strategy, we will not be able to work with the edge weight configurations directly. We will project it to a discrete set of $e^{o(n^2)}$ many elements and pick the most likely one among them (still in the large deviation regime). We shall employ the following discretization.
Note that by the upper bound on the support of the edge variables, deterministically, ${{\mathbf{PT}}}(\bx,\by)\le 3b|\bo{x}-\bo{y}|+2b$ for any $\bo{x},\bo{y} \in \R^2$. Now for a discretization parameter $\eta_1,$ we will discretize the normalized distances (passage time divided by Euclidean distance) to be in the set $\{0, \eta_1,2\eta_1,\ldots, 3b\}$ (again assuming that $\eta_1$ is chosen to avoid rounding issues) and project the distance functions ${{\mathbf{PT}}}(\cdot,\cdot)$ onto a discrete space accordingly.
To define things formally, first let the set of all points in ${{\mathbf{Box}}}(n)\cap \frac{n}{2^j}\Z^2$ be called ${{\mathbf{Grid}}}_n(j).$ We will also need the following variant. Let $\ell=\frac{n}{2^{j+m}}$ for some some $m\in \N$. By ${{\mathbf{Grid}}}_n(\ell;j),$ we shall denote the set of all points in ${{\mathbf{Grid}}}_n(j+m)$ which intersect the line segments joining the nearest neighbors in ${{\mathbf{Grid}}}_n(j)$ thought of as elements of $\frac{n}{2^j}\Z^2$ (see Figure \[fig6\]).
![Figure illustrating the various grid points. Intersection of the brown lines denote ${{\mathbf{Grid}}}_n(j)$, intersection of the black lines with the black lines as well as the brown lines denote the points in ${{\mathbf{Grid}}}_n(j+\frac{m}{4})$ which are not in ${{\mathbf{Grid}}}_n(j)$, and similarly points on red lines and blue lines denote points in ${{\mathbf{Grid}}}_n(j+\frac{m}{2})$, and ${{\mathbf{Grid}}}_n(j+m)$ respectively which are not in the previous coarser grid. []{data-label="fig6"}](Grid.pdf)
Now given $\eta_1,\ell_1,j_1$ with $\ell_1=\frac{\sC n}{2^{j_1+m_1}}$ let the projection map $\overset{\eta_1,\ell_1,j_1}{{{\mathbf{Proj}}}}: {{\mathbf{Grid}}}_{\sC n}(j_1+m_1)\times {{\mathbf{Grid}}}_{\sC n}(j_1+m_1) \to \R_{+}$ be defined as follows: for any $\bo{z},\bo{w} \in{{\mathbf{Grid}}}_{\sC n}(j_1+m_1),$ $$\label{projectedfunction}
\overset{\eta_1,\ell_1,j_1}{{{\mathbf{Proj}}}}(\bo{z},\bo{w})=\eta_1\left\lfloor\frac{{{\mathbf{PT}}}(\bo{z},\bo{w})}{\eta_1 |\bo{z}-\bo{w}|}\right\rfloor|\bz-\bo{w}|.$$
Observe that the function $\overset{\eta_1,\ell_1,j_1}{{{\mathbf{Proj}}}}$[^2] is random but we choose to suppress the dependence on the underlying noise for brevity. We will also drop the dependence on $\eta_1,\ell_1,j_1$ in the notation whenever there is no scope of confusion. Observe that a very basic counting argument yields that the cardinality of the image set of $\overset{\eta_1,\ell_1,j_1}{{{\mathbf{Proj}}}}$ denoted $\sP\sV_{\eta_1,\ell_1,j_1}$ satisfies
$$\label{imageset}
|\sP\sV_{\eta_1,\ell_1,j_1}|\leq e^{O(2^{2(j_1+m_1)})\log\frac{1}{\eta_1}}=e^{o(n^2)}$$
where $m_1$, as above, is defined by $\ell_1=\frac{\sC n}{2^{j_1+m_1}}$. Note that ${{{\mathbf{Proj}}}}$ induces a weighted graph with vertex set ${{\mathbf{Grid}}}_{\sC n}(j_1+m_1),$ and the weight on any edge $(\bo{z},\bo{w})$ being ${{{\mathbf{Proj}}}}(\bo{z},\bo{w})$. It will also be useful to extend the definition of ${{\mathbf{Proj}}}$ to a larger set of pairs. For all pairs of points $\bo{z}, \bo{w} \in {{\mathbf{Box}}}(\sC n)$ we will extend the definition, by letting ${{\mathbf{Proj}}}(\bo{z}, \bo{w})={{{\mathbf{Proj}}}}(\hat{\bo{z}},\hat{\bo{w}}),$ where $\hat{\bo{z}},\hat{\bo{w}}$ are the nearest points to $\bo{z}, \bo{w}$ respectively in ${{\mathbf{Grid}}}_{\sC n}(j_1+m_1),$ (as before breaking ties by picking the smallest in the lexicographic order). Note that if $\bz$ and ${{\mathbf{w}}}$ get rounded to the same point, then ${{\mathbf{Proj}}}(\bo{z}, \bo{w})$ is zero which is not a realistic definition. However we will only be interested in pairs $\bz$ and ${{\mathbf{w}}}$ that are reasonably far apart so that the above issue will not arise and hence we will not bother about this aspect of the definition.
The first thing we show now is that the error introduced by using ${{\mathbf{Proj}}}(\cdot, \cdot)$ instead of ${{\mathbf{PT}}}(\cdot, \cdot)$ can be neglected at sufficiently large length scales. For reasons that will become clear momentarily, we shall work with ${{\mathbf{Stable}}}$ tiles, although the approximation is valid independent of that. Fix $\delta_1,{\varepsilon}_1$ and $m_1$ as in Lemma \[most34\], which then guarantees that there exists $j_1$ with probability bounded away from zero, such that for all but ${\varepsilon}_1$ fraction of $v \in \llbracket 1, 2^{j_1}\rrbracket^2$, ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ is $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}}$ where $\ell_1$ and $k_1$ are $\frac{\sC n}{2^{j_1+m_1}}$ and $2^{2m_1}$ respectively. For later reference let us call $v \in \llbracket1, 2^{j_1}\rrbracket^2$ such that ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ is not $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}}$ as $(\delta_1,\ell_1, k_1)-{{\mathbf{Unstable}}}.$
Fixing a value of $\eta_1$ (to be specified later and $\ll \delta_1$) we will now consider the projection map $\overset{\eta_1,\ell_2,j_1}{{{\mathbf{Proj}}}}$ where $\ell_2=\frac{\sC n}{2^{j_1+\frac{m_1}{2}}}$.
\[compare1\] Fix $\delta_1,\eta_1, \ell_1, \ell_2, j_1$ as above and conditioned on $\sU^*_{\zeta}(n),$ consider $v\in \llbracket 1, 2^{j_1}\rrbracket^2$ such that ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ is $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}}.$ Then for any $\bo{z},\bo{w}\in {{\mathbf{Grid}}}_{\sC n}(j_1+m_1/2)$ such that $\bo{z},\bo{w} \in {{\mathbf{Tile}}}_{\sC n}(j_1,v)$ we have the following: $$1\leq \frac{{{\mathbf{PT}}}(\bo{z}, \bo{w})}{{{\mathbf{Proj}}}(\bo{z}, \bo{w})}\leq 1+O(\eta_1).$$
Observe that by definition $${{\mathbf{Proj}}}(\bo{z}, \bo{w})\leq {{\mathbf{PT}}}(\bo{z}, \bo{w}) \leq {{\mathbf{Proj}}}(\bo{z}, \bo{w})+ O(\eta_1) |\bo{z}-\bo{w}|.$$ The proof follows immediately by noticing that since $\bo{z}$ and $\bo{w}$ are at distance at least $\ell_2$ and on $\sU^*_{\zeta}(n)$ by definition ${{\mathbf{PT}}}(\bo{z}, \bo{w})\ge \alpha |\bo{z}- \bo{w}|.$
We now define a gradient function corresponding to the projected distances analogous to . As in the above setting let $\bo{z}, \bo{w}\in {{\mathbf{Grid}}}_{\sC n}(j_1+m_1/2)$ and let $\theta$ and $d>0$ be such that $\bo{w}=\bo{z}+(\theta,d)$. Then let $$\label{grad88}
\grad_{{{\mathbf{Proj}}}}(\bz,\theta, d)=\frac{{{\mathbf{Proj}}}(\bo{z}, \bo{w})}{|\bo{z}-\bo{w}|}.$$ Once we have defined the projected gradients only for pairs of points in ${{\mathbf{Grid}}}_{\sC n}(j_1+m_1/2)$ we define projected gradients in all directions at a slightly coarser scale, i.e. for all points in ${{\mathbf{Grid}}}_{\sC n}(j_1+m_1/4)$. For any $\bo{z}\in {{\mathbf{Tile}}}_{\sC n}(j_1,v)\cap {{\mathbf{Grid}}}_{\sC n}(j_1+m_1/4)$ and for any $\theta\in \bS^1$ and $\frac{n}{2^{j_1}}>d>\frac{n}{2^{j_1+m_1/4}}$ let $$\label{grad89}
\grad_{{{\mathbf{Proj}}}}(\bz,\theta, d)=\frac{{{\mathbf{Proj}}}(\bo{z}, \bo{w})}{d},$$ where $\bo{w}$ is the closest point to $\bo{z}+(\theta,d)$ in ${{\mathbf{Grid}}}_{\sC n}(j_1+m_1/2)$. Note that ${{\mathbf{Grid}}}_{\sC n}(j_1+m_1/4)\subset {{\mathbf{Grid}}}_{\sC n}(j_1+m_1/2)$. Thus is defined via . If ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ is $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}}.$ then as in along with an application of triangle inequality as in , the following result about smoothness of the projected gradient field follows whose proof we omit.
\[smoothgradproj\] For any $\bz, \bz' \in {{\mathbf{Tile}}}_{\sC n}(j_1,v)$ and $\theta_1, \theta_2 \in \bS^1,$ such that $|\theta_1-\theta_2|\le \eta_1$ and $d_1,d_2$ such that $\grad_{{{\mathbf{Proj}}}}(\bz,\theta_1, d_1)$ and $\grad_{{{\mathbf{Proj}}}}(\bz',\theta_2, d_2)$ are defined via then $$\frac{1}{1+O(\delta_1+\eta_1+2^{-m_1/4})}\le \frac{\grad_{{{\mathbf{Proj}}}}(\bz,\theta_1, d_1)}{\grad_{{{\mathbf{Proj}}}}(\bz',\theta_2, d_2)}< 1+O(\delta_1+\eta_1+2^{-m_1/4}).$$
Note that above we choose $|\theta_1-\theta_2|\le \eta_1$ where the latter appeared in the definition of ${{\mathbf{Proj}}}.$ This is done deliberately to avoid introducing new notation since for us any small enough value of $\eta_1$ would serve both the purposes.
This allows us to define a projected gradient for the entire tile as we did in Definition \[tilegrad\].
\[tilegrad2proj\] If ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ is $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}},$ then let $$\grad_{{{\mathbf{Proj}}}}((j_1,v),\theta):=\grad_{{{\mathbf{Proj}}}}(\bz,\theta, d)$$ for some arbitrary $\bz\in {{\mathbf{Tile}}}_n(j_1,v)\cap {{\mathbf{Grid}}}_{\sC n}(j_1+m_1/2),$ and $d$ such that the RHS is defined via . Note that the definition depends on the choice of $\bz$ and $d$ but only up to a multiplicative factor of $(1+O(\delta_1+\frac{1}{2^{m_1/4}})),$ which can be made arbitrarily close to one by choosing the parameters appropriately. Hence for concreteness we choose $\bz$ to be the center point of ${{\mathbf{Tile}}}_n(j,v)$ and $d=\frac{n}{2^{j_1+m_1/8}}.$
Essentially the fact that ${{\mathbf{PT}}}$ satisfies the triangle inequality (by definition) is what leads to the convexity of the limit shape $\cB$ in . One might therefore hope that ${{{\mathbf{Proj}}}}$ satisfies an approximate triangle inequality. To formally state things, it would be convenient to consider the following function on entire $\R^2$ given by the following: for any $\bo{w}=(\theta,r)$, $$\|\bo{w}\|_{(j,v)}=\|(\theta,r)\|_{(j,v)}=r\grad_{{{\mathbf{Proj}}}}((j,v),\theta).$$ Note that as in Definition \[tilegrad2proj\], this definition implicitly depends on the choice of $\bz$ and $d.$ The next lemma shows the approximate convexity of the above defined function which allows us to think of the above as roughly a norm.
\[conc1\]If ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ is $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}},$ then for any set of vectors $\bo{w}_1,\bo{w}_2,\ldots,\bo{w}_t$, if $\bo{w}=\sum_{i=1}^t \bo{w}_i$ then $$\|\bo{w}\|_{(j_1,v)}\le (1+O(\delta_1+\frac{1}{2^{\frac{m_1}{16}}}))\left(\sum_{i=1}^t\|\bo{w}_i\|_{(j_1,v)}\right).$$
The proof even though relies on an approximate triangle inequality is a little technical and is postponed to Section \[pconc1\]. For the next result, given $\delta_1, {\varepsilon}_1,$ and $m_1$ satisfying the hypothesis of Lemma \[most34\], let $j_1$ be the scale obtained from that lemma and recall the definitions of $\ell_1$ and $k_1$ from the statement of the same. Recalling $\eta_1$, ${{\mathbf{Proj}}}=\overset{\eta_1,\ell_2,j_1}{{{\mathbf{Proj}}}}$ where $\ell_2=\frac{\sC n}{2^{j_1+\frac{m_1}{2}}}$ from Lemma \[compare1\]. consider the set of images $\sP\sV_{\eta_1,\ell_2,j_1}$ from .
In the sequel to avoid introducing new notation we will in fact denote $\frac{\log\P(\sU^*_{\zeta}(n))}{n^2}$ by $\kappa$ even though it was used to define the lim sup of $\frac{\log\P(\sU_{\zeta}(n))}{n^2}$ in . We now state the following easy consequence of the pigeon-hole principle.
\[proj23\] Given the parameters as above and ${\varepsilon}_4>0$ there exists $\Im\in \sP\sV_{\eta_1,\ell_2, j_1}$ and $A \subset \llbracket 1, 2^{j_1} \rrbracket^2 $ such that $|A|= {\varepsilon}_1 2^{2j_1}$, such that $$\log\frac{\P(\sU^*_{\zeta}(n)\cap {{\mathbf{Proj}}}^{-1}(\Im)\cap \bigl\{\{v \in \llbracket 1, 2^{j_1}\rrbracket^2 :v \text{ is } (\delta_1, \ell_1, k_1) -{{\mathbf{Unstable}}}\}\subset A\bigr\})}{n^2}\ge \kappa-{\varepsilon}_4,$$ for all large enough $n.$
Recall the trivial bound mentioned in , $$|\sP\sV_{\eta_1,\ell_2, j_1}|=e^{O(2^{2(j_1+m_1)}\log\frac{1}{\eta_1})}=e^{O(1)}.$$ Moreover the possible subsets $A$ of $\llbracket 1,2^{j_1}\rrbracket^2$ of size at most ${\varepsilon}_1 2^{2j_1}$ is at most $e^{O(H({\varepsilon}_1))2^{2j_1}}$ where $H(\cdot)$ is the entropy functional. Thus by pigeon-hole principle the result follows.
Henceforth, for $A$ as in Lemma \[proj23\], we will denote the above event i.e., $$\label{basev23}
\sU^*_{\zeta}(n) \cap {{\mathbf{Proj}}}^{-1}(\Im)\cap \bigl\{\{v \in \llbracket 1,2^{j_1}\rrbracket^2 :v \text{ is } (\delta_1, \ell_1, k_1) -{{\mathbf{Unstable}}}\}\subset A\bigr\},$$ which will be our building block for later constructions as $${{\mathbf{Base-event}}}:={{\mathbf{Base-event}}}(\eta_1,\delta_1,\ell_1,k_1,{\varepsilon}_1).$$ Note that the above definition should also contain $A$ as a parameter which we are suppressing to avoid cluttering.
Constructing a Large Deviation Event at a Higher Scale {#s:construct}
======================================================
In this section we prove Proposition \[p:pathconstruct\]. With the definition and results from the previous section at our disposal, following the strategy outlined in Section \[outline\], for any $n_1\gg n$, we now proceed to creating the favourable event ${{\mathbf{Fav}}}:={{\mathbf{Fav}}}(n_1)$ which will imply $\sU_{\zeta'}(n_1)$ where $\zeta'\ge \zeta-O({\varepsilon}),$ for some small ${\varepsilon},$ and moreover, $$\frac{\log\P({{\mathbf{Fav}}}(n_1))}{n_1^2}\ge \frac{\log\P(\sU^*_{\zeta}(n))}{n^2}-O({\varepsilon}).$$
We start by defining certain key ingredients: Fixing ${\varepsilon}_6>0,$ for brevity we adopt the following abbreviations $$\begin{aligned}
\label{abbre34}
{{\mathfrak{n}}}_0:=\sC n_1(1+2{\varepsilon}_6),\,\, {{\mathfrak{n}}}_1:=\sC n_1(1+{\varepsilon}_6),\,\,{{\mathfrak{n}}}_2:=\sC n_1, {{\mathfrak{n}}}_3:= \sC n(1+{\varepsilon}_6), {{\mathfrak{n}}}_4:= \sC n. \end{aligned}$$ Moreover in the sequel we will denote ${{\mathbf{Box}}}({{\mathfrak{n}}}_i)$ as ${\mathfrak{B}}_i.$ ${{\mathbf{Fav}}}$ will be a function of the edges in $\mathfrak{B}_0,$ with the property that on the event ${{\mathbf{Fav}}},$ $${{\mathbf{PT}}}_{{\mathfrak{B}}_0}(\bo{0},{\mathfrak{B}}_0^c)\ge bn_1, \text{ and }
{{\mathbf{PT}}}_{{\mathfrak{B}}_0}(\bo{0}, \bo{n_1})\ge (\mu+\zeta-O({\varepsilon}))n_1,$$ for some small ${\varepsilon}.$ Clearly this implies that ${{\mathbf{Fav}}}\subset \sU_{\zeta'}(n_1)$ for some $\zeta'=\zeta-O({\varepsilon})$. The basic geometry we shall be working with is the following. Fix $j\in \N$. Tile the box ${\mathfrak{B}}_0$ by ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_0}(j,v)$ for $v\in [2^j]^2$. Now each such tile is a square of size $\frac{{{\mathfrak{n}}}_0}{2^j}$. For $v\in [2^j]^2$, consider the square with the same centre as ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_0}(j,v)$ and side length $\frac{{{\mathfrak{n}}}_1}{2^j}$. Call this square (closed) ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j,v)$; see Figure \[fig8\]. It follows that neighbouring ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j,v)$’s are separated by vertical and horizontal strips of width at most ${\varepsilon}_6 \frac{{{\mathfrak{n}}}_1}{2^j}$. For obvious reasons, the set of all edges in ${\mathfrak{B}}_0$ that does not belong to any ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j,v)$ is called ${{\mathbf{Corridor}}}^{\rm{ext}}(j,{{\mathfrak{n}}}_0)$. ($\rm{ext}$ stands for exterior, we will also consider corridors inside ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j,v)$) **Without loss of generality we shall assume that $\bo{0}=(0,0)$ and $\bo{n_1}=(n_1,0)$ are at the center of some (different) ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j,v)$’s**.
![The figure illustrates the basic structural definitions inside ${\mathfrak{B}}_0$. On the LHS the figure shows the ${{\mathbf{Corridor}}}^{\rm{ext}}(j,{{\mathfrak{n}}}_0)$ (red region) and the tiling of the remaining area by ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j,v).$ The RHS zooms into one particular ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j,v),$ (the south-east one) and shows $\bo{B}_{{{\mathfrak{n}}}_4}(j,v,w)$ and the surrounding $\bo{C}_{{{\mathfrak{n}}}_4}(j,v,w)$ which form a part of ${{\mathbf{Corridor}}}^{\rm{int}}(j,{{\mathfrak{n}}}_0)$.[]{data-label="fig8"}](fav1.pdf)
Our construction of ${{\mathbf{Fav}}}$ will have two steps:
(i) Specifying the environment inside ${{\mathbf{Tile}}}^*_{{{\mathfrak{n}}}_1}(j,v)$ for various $v \in \llbracket1,2^j\rrbracket^2.$
(ii) Specifying the environment in ${{\mathbf{Corridor}}}^{\rm{ext}}(j, {{\mathfrak{n}}}_0)$.
Part (i) involves a large deviation environment in the smaller scale $n$, whereas for the second part we just make all the edge weights close to $b$. We shall formalize part (i) later, but for now let us make part (ii) formal as follows. Let ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,j)$ denote the event that the passage time on each edge in ${{\mathbf{Corridor}}}^{\rm{ext}}(j, {{\mathfrak{n}}}_0)$ is in $[b-{\varepsilon}_7,b]$ for some small but fixed ${\varepsilon}_7$. As the total number of the edges in ${{\mathbf{Corridor}}}^{\rm{ext}}(j, {{\mathfrak{n}}}_0)$ is $O({\varepsilon}_6 {{\mathfrak{n}}}_0^2)$, it follows that $-\log \P ({{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,j))= O({\varepsilon}_6 n_1^2)$ (the constant in the $O(\cdot)$ notation depends on ${\varepsilon}_7,$ and ${\varepsilon}_6$ will be chosen to be much smaller than ${\varepsilon}_7$ depending on the edge distribution $\nu$).
Recalling that the goal is to create an event on which the FPP distance (within the box ${\mathfrak{B}}_0$) between $\bo{0}$ and $\bo{n_1}$ is forced to be large, having constructed ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,j)$ we are left to do two more things:
1. Specifying the environments inside ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j,v)$ using the large deviation ${{\mathbf{Base-event}}}$ defined in Lemma \[proj23\].
2. Using the above showing that any path $\gamma$ between $\mathbf{0}$ and $\bo{n_1}$ contained in ${\mathfrak{B}}_0$ has length $(\mu+\zeta-O({\varepsilon}))n_1$. However to be able to use the properties of ${{\mathbf{Base-event}}}$ (in particular the stability properties) we need some regularity properties of $\gamma.$ Hence the first step, given such an arbitrary $\gamma$ is to preprocess it to obtain another path $\cP(\gamma)$ from $\mathbf{0}$ and $\bo{n_1}$ such that the path $\cP(\gamma)$ has the desired regularity properties, and, on the event ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,j)$, has length within a factor $(1+o(1))$ of the length of $\gamma$.
To accomplish the first part for any $j$ and $v \in {{\llbracket{1,2^j}\rrbracket^2}},$ it will be convenient to think of each ${{\mathbf{Tile}}}^*_{{{\mathfrak{n}}}_1}(j,v)$ as naturally made up of $(\frac{n_1}{n})^2$ copies of ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_3}(j,v).$ We will denote the copy of the tile as $\bo{A}_{{{\mathfrak{n}}}_3}(j,v,w)$ for $w \in \llbracket 1, \frac{n_1}{n} \rrbracket^2,$ and as before each $\bo{A}_{{{\mathfrak{n}}}_3}(j,v,w)$ can be thought of as a copy of ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_4}(j,v)$ to be called $\bo{B}_{{{\mathfrak{n}}}_4}(j,v,w)$ surrounded by an annulus $\bo{C}_{{{\mathfrak{n}}}_4}(j,v,w)$ of width $\frac{{\varepsilon}_6}{2}\frac{{{\mathfrak{n}}}_4}{2^j}$, (see Figure \[fig8\]). As before we denote the union of edges in $\bo{C}_{{{\mathfrak{n}}}_4}(j,v,w)$ union over $v\in {{\llbracket{1,2^j}\rrbracket^2}}$ and $w \in \llbracket 1, \frac{n_1}{n} \rrbracket^2$ as ${{\mathbf{Corridor}}}^{\rm{int}}(j, {{\mathfrak{n}}}_0).$ Now similar to ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,j)$ let ${{\mathbf{Barrier}}}^{\rm{int}}({{\mathfrak{n}}}_0,j)$ denote the event that the passage time on each edge in ${{\mathbf{Corridor}}}^{\rm{int}}(j, {{\mathfrak{n}}}_0)$ is in $[b-{\varepsilon}_7,b]$ and similar considerations as before show that $-\log \P ({{\mathbf{Barrier}}}^{\rm{int}}({{\mathfrak{n}}}_0,j))= O({\varepsilon}_6 n_1^2)$.
We now prescribe the environment inside $\bo{B}_{{{\mathfrak{n}}}_4}(j,v,w)$. Since there are many parameters involved, to avoid repetition throughout this section we will work with the choice of parameters as in Lemma \[proj23\]. Note that this causes us from now to work with a specific scale $j_1$ and not a generic scale $j.$ Recall the set $A$ of size ${\varepsilon}_1 2^{2j_1}$ in the statement of Lemma \[proj23\].
Construction of ${{\mathbf{Fav}}}$:
-----------------------------------
At a high level the event ${{\mathbf{Fav}}}$ will be an intersection of three independent events i.e., ${{\mathbf{Fav}}}:={{\mathbf{Dilation}}}\cap {{\mathbf{Barrier}}}\cap {{\mathbf{Boosting}}},$ where the three events on the RHS will be independent. We will use ${{\mathbf{Barrier}}}$ to denote the intersection of the events ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,j_1)$ and ${{\mathbf{Barrier}}}^{\rm{int}}({{\mathfrak{n}}}_0,j_1).$ ${{\mathbf{Boosting}}}$ and ${{\mathbf{Dilation}}}$ will be used to define the edge weights in $\bo{B}_{{{\mathfrak{n}}}_4}(j_1,v,w)$ where $w \in \llbracket 1, \frac{n_1}{n}\rrbracket^2$ and $v\in A$ and $v\in \llbracket 1, 2^{j_1} \rrbracket^2\setminus A$ respectively. We define the event that the passage time on all the edges in $\bigcup_{v\in A,w}\bo{B}_{{{\mathfrak{n}}}_4}(j_1,v,w)$ is in $[b-{\varepsilon}_7,b]$ as ${{\mathbf{Boosting}}}.$
Finally we define the event ${{\mathbf{Dilation}}}$ in the following constructive way: Sample $\left(\frac{n_1}{n}\right)^2$ many independent realizations of ${{\mathbf{Base-event}}}$ which yields environments $$\Pi_1,\Pi_2\ldots, \Pi_{\left(\frac{n_1}{n}\right)^2}$$ such that $\Pi_i\in {{\mathbf{Base-event}}},$ for all $i \in \llbracket 1, \frac{n_1}{n}\rrbracket^2.$ For each $v\in \llbracket 1, 2^{j_1} \rrbracket^2\setminus A$ and $w \in \llbracket 1, \frac{n_1}{n}\rrbracket^2$, let the edge weights on the edges in $\bo{B}_{{{\mathfrak{n}}}_4}(j_1,v,w)$ be the same as the edge weights of $\Pi_w$ in ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_4}(j_1,v)$ where we use the natural identification between $\bo{B}_{{{\mathfrak{n}}}_4}(j_1,v,w)$ and ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_4}(j_1,v).$
Note that the choice of the term ${{\mathbf{Dilation}}}$ to denote the above event is natural, as by using $(\frac{n_1}{n})^2$ copies of ${{\mathbf{Base-event}}}$ we ensure that for any $v \in \llbracket 1, 2^{j_1} \rrbracket^2$, the environments in different $\bo{B}_{{{\mathfrak{n}}}_4}(j_1,v,w)$ are essentially the same. Now the event ${{\mathbf{Barrier}}}$ along with ${{\mathbf{Dilation}}}$ describe the projection of the event ${{\mathbf{Fav}}}$ on all the edges except the edges in $\bigcup_{v\in A, w\in \llbracket 1, \frac{n_1}{n}\rrbracket^2}\bo{B}_{{{\mathfrak{n}}}_4}(j_1,v,w)$ whereas ${{\mathbf{Boosting}}}$ defines those in the latter. Hence $$\label{lb7896}
\P({{\mathbf{Fav}}})=[\P({{\mathbf{Base-event}}})]^{\frac{n_1^2}{n^2}}\nu([b-{\varepsilon}_7,b])^{O(({\varepsilon}_1+{\varepsilon}_6) n_1^2)},$$ where $\nu$ is the passage time distribution satisfying the hypothesis in Theorem \[t:ldp\].
The proof of Proposition \[p:pathconstruct\] will now be complete from the following lemma.
\[verify\] Given ${\varepsilon}_8$ and ${\varepsilon}_9$ there exists choice of parameters in the definition of ${{\mathbf{Base-event}}}$ in Lemma \[proj23\], and ${\varepsilon}_6,{\varepsilon}_7$ in the definition of ${{\mathbf{Fav}}}$, such that $$\frac{\log(\P({{\mathbf{Fav}}}))}{n_1^2}\ge \kappa-{\varepsilon}_8, \text{ and, } {{\mathbf{Fav}}}\subset \sU_{\zeta'}(n_1),$$ where $\zeta'>\zeta-{\varepsilon}_9.$
Note that the lower bound on the probability of ${{\mathbf{Fav}}}$ is a straightforward consequence of and Lemma \[proj23\]. The rest of the discussion is devoted to the proof of the second part which will follow from a series of lemmas. Before stating the lemmas we roughly describe our strategy. The proof involves broadly showing that on the event ${{\mathbf{Fav}}}$ two things occur:
$$\begin{aligned}
\label{part1}
{{\mathbf{PT}}}_{{\mathfrak{B}}_0}(\bo{0},\bo{n_1})& \ge (\mu+\zeta')n_1, \\
\label{part2}
{{\mathbf{PT}}}_{{\mathfrak{B}}_0}(\bo{0},{\mathfrak{B}}_0^c)& \ge b n_1.\end{aligned}$$
Now the proof of both the above bounds is obtained by the same strategy. Keep in mind the two random fields given by ${{\mathbf{Fav}}}$ and ${{\mathbf{Base-event}}}$ on ${\mathfrak{B}}_0$ and ${\mathfrak{B}}_4$ respectively. Recall that the former is a ‘dilation’ of the latter by a factor of $\frac{{{\mathfrak{n}}}_2}{{{\mathfrak{n}}}_4},$ with some additional changes including the setting up of the barriers and the boosting on the unstable tiles.
As outlined in Section \[outline\], given the above, the strategy is to show that for any path $\gamma$ (joining $\bo{0}$ and $\bo{n_1}$) in ${\mathfrak{B}}_0,$ there exists a scaled version $\gamma_{{{\mathbf{S}}}}$ (joining $\bo{0}$ and $\bo{n}$) in ${\mathfrak{B}}_4$ such that $$\label{scale879}
|\gamma|\ge \frac{n_1}{n}(1-o(1))|\gamma_{{{\mathbf{S}}}}|,$$ where the LHS is computed on ${{\mathbf{Fav}}}$ and the RHS is computed on ${{\mathbf{Base-event}}}$. Thus $\gamma$ can be thought of as a path obtained by dilating the path $\gamma_{{{\mathbf{S}}}}$.
Since $\gamma_{{{\mathbf{S}}}}$ is a path in the random field given by ${{\mathbf{Base-event}}},$ it follows by definition of the latter that $|\gamma_{{{\mathbf{S}}}}|\ge (\mu+\zeta)n$ and this yields the sought lower bound of $|\gamma|.$ To make formal we need some regularity properties of the path $\gamma$ which will be obtained by some preprocessing. This is done in the next section.
Preprocessing of Paths {#s:process}
----------------------
Observe that given any path $\gamma$ contained in ${\mathfrak{B}}_0$ it admits a unique decomposition as a concatenation of a number of paths i.e., $\gamma= \alpha_0\chi_0\alpha_1\chi_1\alpha_2\chi_2\ldots \alpha_{L}\chi_{L}\alpha_{L+1}$ with the following properties:
i. Each $\alpha_i$ is contained in some ${{\mathbf{Tile}}}^*_{{{\mathfrak{n}}}_1}(j_1,v_i)$ for some $v_i \in \llbracket 1, 2^{j_1} \rrbracket ^2$; $\alpha_0$ and $\alpha_{L+1}$ could be empty.
ii. Each $\chi_i$ is contained in ${{\mathbf{Corridor}}}^{\rm{ext}}(j_1, {{\mathfrak{n}}}_0)$.
![A schematic diagram describing the preprocessing. (i) illustrates the decomposition of the path $\gamma$, into $\alpha_i$ (blue segments) and $\chi_i$ (red segments). (ii) describes the content of Lemma \[l:outside\] where each red segment is replaced by a regular path. (iii) describes the content of Lemma \[l:excursion\] where if an excursion is small (the part in the north-east tile) then we replace it by a larger excursion without changing the length too much. []{data-label="fig9"}](process.pdf)
Given ${\varepsilon}_6$ as in let us call the paths $\alpha_i$ for $i\in \{1,2,\ldots, L\}$ as **excursions** of $\gamma$ and let us call the above decomposition of $\gamma$ its decomposition into excursions. Let $\mathbf{x}_i$ (resp. $\mathbf{y}_i$) denote the starting (resp. ending) vertex of $\alpha_i$. Let us call the excursion $\alpha_i$ **large** if there exists a vertex $\bz_i$ on $\alpha_i$ such that $\min\{|\bx_i-\bz_i|, |\by_i-\bz_i|\} \geq {\varepsilon}_6^2\frac{{{\mathfrak{n}}}_0}{2^{j_1}}$. Observe that $\alpha_i$ is large if $|\mathbf{x}_i-\mathbf{y}_i|\geq 2{\varepsilon}_6^2\frac{{{\mathfrak{n}}}_0}{2^{j_1}}$.
We shall need to define one more property of a path. Consider a path $\gamma$ with the decomposition into excursions as above. Observe that each $\chi_i$ must start at ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,v)$ and end at some ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,v')$ for some $v=v(\chi_i),v'=v'(\chi_i)\in [2^{j_1}]^2$. We call the path $\gamma$ **regular** if for each $\chi_i$ we have $\|v(\chi_i)-v'(\chi_i)\|_1=1$ (i.e., they are neighbouring vertices) and the starting point of $\chi_i$ lies in the same vertical (if $v$ and $v'$ are on the same vertical line) or horizontal (if $v$ and $v'$ are on the same horizontal line) line as its endpoint. Recall the parameters ${\varepsilon}_6$ and ${\varepsilon}_7$ in the definition of the event ${{\mathbf{Fav}}}.$
\[decomp120\] For any path $\gamma$ starting at $\mathbf{0}$ and ending at $\bo{n_1}$ and contained in ${\mathfrak{B}}_0$, there exists a regular path $\cP$ from $\mathbf{0}$ to $(n_1,0)$ such that
i. All the excursions of $\cP$ are large.
ii. On ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,{j_1})$, we have $|\gamma|\ge (1-O({\varepsilon}_7+{\varepsilon}_6))|\cP|.$
The proof of the above lemma is done in two steps (see Figure \[fig9\] for an illustration). Let $\gamma$ be fixed as in the lemma. Consider its decomposition into excursions: $\gamma=\alpha_0\chi_0\alpha_1\chi_1\alpha_2\chi_2\cdots$. Observe that if we can replace each $\chi_i$ by a regular path with the same endpoints, the resulting path will be regular. The following lemma shows that this can be done without increasing the length of the path by more than a factor of $(1-O({\varepsilon}_7))^{-1}$.
\[l:outside\] Consider a path $\chi$ completely contained in ${{\mathbf{Corridor}}}^{\rm{ext}}(j_1, {{\mathfrak{n}}}_0)$ whose starting and ending points are located at the boundary of ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,v)$ and ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,v')$ respectively. Then there exists a regular path $\cP_{\chi}$ with the same starting and ending point such that on ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,{j_1})$, we have $|\chi|\geq (1-O({\varepsilon}_7))|\cP_{\chi}|$.
Let $\bx$ and $\by$ be the starting and ending point of $\chi$ respectively. Consider the $1-$norm minimizing path from $\bx$ to $\by$ that constitutes of a horizontal path followed by a vertical path (this choice is arbitrary): i.e., for $\bx=(x_1,x_2)$ and $\by=(y_1,y_2)$ consider the piecewise linear curve $\mathbb{L}$ obtained by concatenating the straight line segment obtained by joining $\bx$ to $(y_1,x_2)$ followed by the straight line segment obtained by joining $(y_1,x_2)$ to $\by$. Consider $\L$ as a path on the nearest neighbour graph of $\Z^2$. Observe that there exists points $u_0=\bx, u_1, \ldots, u_{\ell}=\by$ on $\mathbb{L}$ all on boundaries of ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,u)$’s such that the $\mathbb{L}$ restricted between $u_{i}$ and $u_{i+1}$ (called $\mathbb{L}_i$) is either (a) contained in ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,u)$ for some $u$ (type A, say) or (b) is entirely contained in ${{\mathbf{Corridor}}}^{\rm{ext}}(j_1, {{\mathfrak{n}}}_0)$, and further $u_i\in {{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,u), u_{i+1}\in {{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1, u')$ for some $u,u'$ that have $\ell_1$ distance one (type $B$). Observe again that such a decomposition is unique. Now if $\mathbb{L}_i$ is type $A$ let us set $\cP_i$ to be the shortest path between $u_i$ and $u_{i+1}$ contained in ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}({j_1},u)$, and if $\mathbb{L}_i$ is type $B$ we set $\cP_i=\mathbb{L}_i$. Consider the path $\cP_{\chi}=\cP_0\cP_1\cdots \cP_{\ell-1}$ obtained by concatenating $\cP_{i}$’s. It is clear that the path $\cP_{\chi}$ obtained as above is regular, (see Figure \[fig9\] for an illustration) and hence it only remains to show that on ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,{j_1})$, we have $|\chi|\geq (1-O({\varepsilon}_7))|\cP_{\chi}|$. Observe first that, on ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,{j_1})$, we have $|\chi|\geq (b-{\varepsilon}_7)\|\bx-\by\|_1$. It also follows from definitions that $|\cP_i|\leq b\|u_i-u_{i+1}\|_1$. The lemma now follows from observing that $\sum_{i} \|u_i-u_{i+1}\|_1= \|\bx-\by\|_1$.
Lemma \[l:outside\] tells us that for any $\gamma$ as in the statement of Lemma \[decomp120\] one can replace the paths $\chi_i$ in its decomposition by the paths $\cP_{\chi_i}$ as constructed in Lemma \[l:outside\] to end up with a regular path $\cP_{*}$ with the same endpoints such that $|\gamma|\geq (1-O({\varepsilon}_7))|\cP_*|$. The following lemma ensuring the largeness of the excursions, therefore will suffice to complete the proof of Lemma \[decomp120\].
\[l:excursion\] For any regular path $\gamma$ starting at $\mathbf{0}$ and ending at $\bo{n_1}$ and contained in ${\mathfrak{B}}_0$, there exists a regular path $\cP$ from $\mathbf{0}$ to $\bo{n_1}$ such that
i. Each excursion of $\cP$ is large.
ii. On ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,j_1)$, we have $|\gamma|\ge (1-O({\varepsilon}_6))|\cP|.$
Let $\gamma$ be as in the statement of the lemma. Consider its decomposition into excursions $\gamma=\alpha_0\chi_0\alpha_1\chi_1\alpha_2\chi_2\cdots \alpha_{L}\chi_{L}\alpha_{L+1}$. The proof, again will be a step by step procedure, we shall inspect the short excursions one by one, and remove them by modifying the path locally without increasing the lengths too much. Let $\alpha_i$ be contained in ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,v_i)$. We shall establish the following: for any excursion $\alpha_{i}$ that is not large, there exists a path $\alpha'_{i}$ contained in ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,v_i)$ with the same starting and ending point as $\alpha_{i}$ such that: (i) $\alpha'_i$ is a large excursion and (ii) on ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,j_1)$, we have $|\alpha_i \chi_{i}|\geq (1-O({\varepsilon}_{6}))|\alpha'_i \chi_i|$. Before proving this, let us observe that this clearly suffices. Consider the path $\cP=\alpha_0\chi_0\alpha'_1\chi_1\cdots \alpha'_{L}\chi_{L}\alpha_{L+1}$ where $\alpha'_{i}$ is as above if $\alpha_{i}$ is not a large excursion and $\alpha'_{i}=\alpha_{i}$ otherwise. Clearly the above exhibits a decomposition of $\cP$ into excursions which ensures that $\cP$ is regular. The second assertion of the lemma is immediate from the bound on $|\alpha'_i \chi_i|$. It remains to prove the claim.
Consider any excursion $\alpha_i$ that is not large. Let $\bx_i$ and $\by_i$ be its starting and ending points respectively. Fix a vertex $\bz_i$ in ${{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}(j_1,v_i)$ such that $|\bx_i-\bz_i|, |\by_i-\bz_i|\in ({\varepsilon}_6^2\frac{{{\mathfrak{n}}}_0}{2^{j_1}}, 2{\varepsilon}_6^2\frac{{{\mathfrak{n}}}_0}{2^{j_1}})$; clearly such a vertex exists. now consider the path $\alpha'_i=\alpha^{(1)}_i\alpha^{(2)}_i$ where $\alpha^{(1)}_i$ (resp. $\alpha^{(2)}_i$) is the shortest path between $\bx_i$ and $\bz_i$ (resp. $\bz_i$ and $\by_i$) contained in $ {{\mathbf{Tile}}}^{*}_{{{\mathfrak{n}}}_1}({j_1},v_i)$. Clearly $\alpha'_i$ is a large excursion. To get an upper bound on $|\alpha'_i\chi_i|$, observe that $|\alpha'_i|\leq 8b{\varepsilon}_6^2 \frac{{{\mathfrak{n}}}_0}{2^{j_1}}$ and on ${{\mathbf{Barrier}}}^{\rm{ext}}({{\mathfrak{n}}}_0,{j_1})$, by taking ${\varepsilon}_6,{\varepsilon}_7$ sufficiently small we have $|\chi_i|\geq (b-{\varepsilon}_7){\varepsilon}_6 \frac{{{\mathfrak{n}}}_0}{2^{j_1}}$, and $|\alpha_i \chi_{i}|\geq (1-O({\varepsilon}_6))|\alpha'_i \chi_i|$. This completes the proof of the lemma.
Given the regular path $\cP=\alpha_0\chi_0\alpha_1\chi_1\alpha_2\chi_2\ldots \alpha_{L}\chi_{L}\alpha_{L+1}$ from Lemma \[decomp120\], we use essentially the same arguments on each of the excursions $\alpha_i$ as in the proof of Lemma \[decomp120\] to obtain a further decomposition in to excursions, i.e. if $\alpha_i$ is contained in ${{\mathbf{Tile}}}^*_{{{\mathfrak{n}}}_1}(j_1,v_i)$ then the further excursions would be contained in $\bo{B}({{\mathfrak{n}}}_4,v_i,w)$ for some $w\in \llbracket1, \frac{n_1}{n} \rrbracket^2.$ We now work with the obvious adaptations of the terms $\textbf{regular}$ (replacing ${{\mathbf{Tile}}}^*_{{{\mathfrak{n}}}_1}(j_1,v)$ and ${{\mathbf{Tile}}}^*_{{{\mathfrak{n}}}_1}(j_1,v')$ by $\bo{B}({{\mathfrak{n}}}_4,v_i,w)$ and $\bo{B}({{\mathfrak{n}}}_4,v_i,w')$ respectively) and $\textbf{large}$ (replacing ${{\mathfrak{n}}}_0$ by ${{\mathfrak{n}}}_4$).
Using the above altered definitions along with the same argument as before we obtain the following [whose proof we omit]{}.
![The figure illustrates a natural identification between ${\mathfrak{B}}_{0}$ and ${{\mathbf{Box}}}(\sC n_1).$ The top two figures show the effect of ignoring ${{\mathbf{Corridor}}}^{\rm{ext}}(j_1, {{\mathfrak{n}}}_0)$. The bottom two figures zoom in the on the south-east tile and shows the effect locally of ignoring ${{\mathbf{Corridor}}}^{\rm{int}}(j_1, {{\mathfrak{n}}}_0)$.[]{data-label="fig10"}](squishing.pdf)
\[decomp121\] For any $\cP=\alpha_0\chi_0\alpha_1\chi_1\alpha_2\chi_2\ldots$ satisfying the properties listed in Lemma \[decomp120\], for each $\alpha_i, i\ge 1$ there is a regular path $\beta_i$ with the same starting and ending points as $\alpha_i$ and a decomposition into excursions $\beta_i=\beta_{i,1}\chi_{i,1}\beta_{i,2}\chi_{i,2}\ldots$ such that
i. All the excursions of $\beta_i$ are large.
ii. On ${{\mathbf{Barrier}}}^{\rm{int}}({{\mathfrak{n}}}_0,j_1)$, we have $|\alpha_i|\ge (1-O({\varepsilon}_7+{\varepsilon}_6))|\beta_i|.$
Equipped with these results we are now ready to prove . In the next few lemmas we create a scaled version $\cP^{{{\mathbf{S}}}}$ of path $\cP$ mentioned in the above lemma. What we do is rather simple and natural. Consider the $\cP=\alpha_0\chi_0\alpha_1\chi_1\alpha_2\chi_2\ldots,$ and then using Lemma \[decomp121\] let $\cP'=\beta_1\chi_1\beta_2\chi_2\ldots $ and moreover consider the decomposition of each $\beta_i$ as provided by the last lemma.
Now by the regularity of the path $\cP$ and all the $\beta_i$’s there is a natural path one can form in ${{\mathbf{Box}}}(\sC n_1)$ by squishing all the corridors. Formally one can naturally identify $${\mathfrak{B}}_0 \setminus ({{\mathbf{Corridor}}}^{\rm{ext}}(j_1, {{\mathfrak{n}}}_0) \cup {{\mathbf{Corridor}}}^{\rm{int}}(j_1, {{\mathfrak{n}}}_0))$$ with ${{\mathbf{Box}}}(\sC n_1)$. This allows one to identify with the path $\cP',$ a path $\cP^*$ in ${{\mathbf{Box}}}(\sC n_1)$ formed by ignoring all the bridges $\chi_{i,j}$ and $\chi_i$. This is possible since in the above identification of ${\mathfrak{B}}_0 \setminus \left\{{{\mathbf{Corridor}}}^{\rm{ext}}(j_1, {{\mathfrak{n}}}_0) \cup {{\mathbf{Corridor}}}^{\rm{int}}(j_1, {{\mathfrak{n}}}_0)\right\}$ with ${{\mathbf{Box}}}(\sC n_1)$ the endpoints of $\chi_{i,j}$ or $\chi$ map to adjacent points in ${{\mathbf{Box}}}(\sC n_1)$ (see Figure \[fig10\]).
Under the above operation $\cP^*$ admits a decomposition $\cP^*=\cP_1\cP_2\ldots$ where $\cP_i$ belongs to ${{\mathbf{Tile}}}_{\sC n_1}(j_1,v_i)$ where $v_i \in \llbracket 1, 2^{j_1}\rrbracket^2$ is such that $\alpha_i\subset {{\mathbf{Tile}}}^*_{{{\mathfrak{n}}}_1}(j_1,v_i).$ Let the starting and ending points of $\cP_i$ be $x_i$ and $y_i$. Recalling $\ell_2, j_1$ from the definition of ${{\mathbf{Base-event}}},$ let $x^{{{\mathbf{S}}}}_i$ be the closest point in ${{\mathbf{Grid}}}_{{{\mathfrak{n}}}_4}(\ell_2;j_1)$ to $\frac{n}{n_{1}}x_{i}$ and similarly let $y^{{{\mathbf{S}}}}_i$ be the closest point in ${{\mathbf{Grid}}}_{{{\mathfrak{n}}}_4}(\ell_2;j_1)$ to $\frac{n}{n_{1}}y_{i}$ (see Figure \[fig11\]).
Since $x_{i+1}$ and $y_i$ are adjacent it follows that $x^{{{\mathbf{S}}}}_{i+1}$ and $y^{{{\mathbf{S}}}}_i$ are adjacent points in ${{\mathbf{Grid}}}_{{{\mathfrak{n}}}_4}(\ell_2;j_1)$. Now a natural candidate for $\alpha^{{{\mathbf{S}}}}$ in would be to take the shortest path passing through the points $x^{{{\mathbf{S}}}}_{1},y^{{{\mathbf{S}}}}_{1},x^{{{\mathbf{S}}}}_{2},y^{{{\mathbf{S}}}}_{2},\ldots$. However it is a little inconvenient notationally since $x^{{{\mathbf{S}}}}_{i+1}$ and $y^{{{\mathbf{S}}}}_i$ are not necessarily adjacent in ${\mathfrak{B}}_{4}.$ Since $\ell_2$ by our choice of parameters will be much smaller than the distance between $x^{{{\mathbf{S}}}}_{i}$ and $x^{{{\mathbf{S}}}}_{i+1}$ we will in fact ignore $y^{{{\mathbf{S}}}}_{i}$ and define $\tilde x^{{{\mathbf{S}}}}_{i}$ to be the point adjacent to $x^{{{\mathbf{S}}}}_{i+1}$ contained in the tile containing $\tilde x^{{{\mathbf{S}}}}_{i}$ (namely ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_4}(j_1,v_i)$).
![(i) illustrates the $x_i$ and $y_i$ denoted by red and blue colors respectively and (ii) the corresponding scaled picture.[]{data-label="fig11"}](scaling.pdf)
Given the above we now let $\alpha^{{{\mathbf{S}}}}_i$ to be the shortest path between $ x^{{{\mathbf{S}}}}_{i}$ and $\tilde x^{{{\mathbf{S}}}}_{i}$ and define $\alpha^{{{\mathbf{S}}}}$ to be the concatenation of $\alpha^{{{\mathbf{S}}}}_i$ thought of as a sequence of vertices. This is a valid construction as by the above discussion the endpoint of $\alpha^{{{\mathbf{S}}}}_i$ ($\tilde x^{{{\mathbf{S}}}}_{i}$) is adjacent to the starting point of $\alpha^{{{\mathbf{S}}}}_{i+1}$ ($x^{{{\mathbf{S}}}}_{i+1}$).
As a consequence of the above lemma and the approximate convexity statement in Proposition \[conc1\] we have the following key result.
\[lb123\] Given any ${\varepsilon}_{11}>0,$ there exists a choice of the parameters in the definition of ${{\mathbf{Base-event}}}$ and ${{\mathbf{Fav}}}$ such that, deterministically for any $i$. $$|\alpha_i|\ge (1-{\varepsilon}_{11})\frac{n_1}{n}|\alpha^{{{\mathbf{S}}}}_i|,$$ where the LHS is computed on the event ${{\mathbf{Fav}}}$ and the RHS is computed on any environment in ${{\mathbf{Base-event}}}.$ [^3]
Before proving the above lemma we finish the proof of Lemma \[verify\] using the above results.
The proof will clearly follow by showing and . We will show only the former and the latter has an identical proof. Fix any path $\gamma$ from $\bo{0}$ to $\bo{n_1}$ in ${\mathfrak{B}}_{0}.$ Now applying Lemma \[decomp120\], we obtain a path $\cP=\alpha_1\chi_1\alpha_2\chi_2\ldots$ and by the previous result
$$\begin{aligned}
\label{inequality}
|\gamma|\ge (1-O({\varepsilon}_7+{\varepsilon}_6))|\cP|&\ge (1-O({\varepsilon}_7+{\varepsilon}_6))\sum_{i}|\alpha_i|\ge (1-O({\varepsilon}_7+{\varepsilon}_6))(1-{\varepsilon}_{11})\frac{n_1}{n}\sum_i |\alpha^{{{\mathbf{S}}}}_i|,\\
&=(1-O({\varepsilon}_7+{\varepsilon}_6))(1-{\varepsilon}_{11})\frac{n_1}{n}|\alpha^{{{\mathbf{S}}}}|,\end{aligned}$$
where $|\cP|$ is computed on ${{\mathbf{Fav}}}$ and $|\alpha^{{{\mathbf{S}}}}|$[^4] is computed on any environment in ${{\mathbf{Base-event}}}.$ Now by definition, $\alpha^{{{\mathbf{S}}}}$ is a path joining $\bo{0}$ and $\bo{n}(1-O({\varepsilon}_6))$ in ${\mathfrak{B}}_4$ and hence on ${{\mathbf{Base-event}}}$ we have $|\alpha^{{{\mathbf{S}}}}|\ge (\mu+\zeta)n(1-O({\varepsilon}_6))$ and thus we are done by choosing ${\varepsilon}_6,{\varepsilon}_7$ and ${\varepsilon}_{11}$ small enough depending on ${\varepsilon}_9$.
We now prove Lemma \[lb123\] using Lemmas \[decomp120\] and \[decomp121\] and Proposition \[conc1\].
Recall the set of unstable tiles $A$ in the definition of ${{\mathbf{Base-event}}}.$ For the proof let us consider an environment $\Pi_1 \in {{\mathbf{Base-event}}}.$ Let us obtain an altered environment $\Pi_*$ which agrees with $\Pi_1$ on ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_4}(j_1,v)$ for any $v\in \llbracket 1, 2^{j_1}\rrbracket^2\setminus A$ and is $b-{\varepsilon}_7$ (${\varepsilon}_7$ appearing in the definition of the barrier and boosting events.) on the edges in the tiles corresponding to $v\in A.$ Clearly the length of the shortest path between $\bo{0}$ and $\bo{n}$ in $\Pi_*$ is at least $(1-O({\varepsilon}_7))$ times the length of shortest path between $\bo{0}$ and $\bo{n}$ in $\Pi_1$ since pointwise for any edge $e$, $\Pi_*(e)\ge (1-\frac{{\varepsilon}_7}{b})\Pi_1(e).$
Fix any $i.$ Note that by Lemma \[decomp121\] all the $\beta_{i,j}$ are large which means that there is a point (say $a_2$) which is far apart from the end points $a_0$ and $a_1$. Let $\overset{\rightarrow}{w}^{1}_{i,j}$ and $\overset{\rightarrow}{w}^{2}_{i,j}$ be the vectors obtained by taking the difference of $a_2-a_0$ and $a_1-a_2$ respectively. Also let $\overset{\rightarrow}{w}'_{i,j}$ be the vector obtained by taking the difference of of the starting and ending points of $\chi_{i,j}$. Note that by the properties listed in Lemma \[decomp121\], it follows that for all $i,j,$ $$\label{lbnorm}
\min (\|\overset{\rightarrow}{w}^{1}_{i,j}\|_2, \|\overset{\rightarrow}{w}^{2}_{i,j}\|_2, \|\overset{\rightarrow}{w}'_{i,j}\|_2)\ge {\varepsilon}^2_6 \frac{{{\mathfrak{n}}}_4}{2^{j_1}}.$$ where $\|\cdot\|_2$ denotes the usual Euclidean norm. Note that the bound on $ \|\overset{\rightarrow}{w}'_{i,j}\|_2$ follows since $\chi_{i,j}$ is a bridge across the barriers of width ${\varepsilon}_6 \frac{{{\mathfrak{n}}}_4}{2^{j_1}}.$ Now recall that for any $\Pi \in {{\mathbf{Base-event}}},$ and all $v\in \llbracket 1, 2^{j_1}\rrbracket^2\setminus A,$ ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_4}(j_1,v)$ is stable with a certain choice of parameters, and moreover recall the approximate norm $\|\cdot\|_{({j_1},v_i)}$ from the statement in Proposition \[conc1\] where $\alpha_i\in {{\mathbf{Tile}}}_{{{\mathfrak{n}}}_0}({j_1},v_i)$ in case $v_i\notin A.$ Now the following argument is split into two cases:
$\bo{(1)}$ $v_i \notin A.$ In this case the proof is now complete by the following string of inequalities (${\varepsilon}'$ below changes from line to line). $$\begin{aligned}
|\alpha_i| &\ge (1-O({\varepsilon}_7+{\varepsilon}_6))\sum_{j}\left(|\beta_{i,j}|+|\chi_{i,j}|\right) \ge (1-{\varepsilon}')\sum_{j}\left({\left\lVert\overset{\rightarrow}{w}^{1}_{i,j}\right\rVert}_{({j_1},v_i)}+{\left\lVert\overset{\rightarrow}{w}^{2}_{i,j}\right\rVert}_{({j_1},v_i)}+{\left\lVert\overset{\rightarrow}{w}'_{i,j}\right\rVert}_{({j_1},v_i)}\right)\\
& \ge (1-{\varepsilon}'){\left\lVert\sum_{j}\left(\overset{\rightarrow}{w}^1_{i,j}+\overset{\rightarrow}{w}^2_{i,j}+\overset{\rightarrow}{w}'_{i,j}\right)\right\rVert}_{({j_1},v_i)}\ge (1-{\varepsilon}'){\left\lVert\overset{\rightarrow}{\alpha}_i\right\rVert}_{({j_1},v_i)}= (1-{\varepsilon}')\frac{n_1}{n}{\left\lVert\overset{\rightarrow}{\alpha}^{{{\mathbf{S}}}}_i\right\rVert}_{({j_1},v_i)} \\
& \ge (1-{\varepsilon}_{10})\frac{n_1}{n}|\alpha^{{{\mathbf{S}}}}|\end{aligned}$$ where $\overset{\rightarrow}{\alpha}_i$ (resp. $\overset{\rightarrow}{\alpha}^{{{\mathbf{S}}}}_i$) is the vector obtained by joining the end points of the path $\alpha_i$ (resp. $\alpha^{{{\mathbf{S}}}}_i$) and ${\varepsilon}_{10}$ can be made arbitrarily small by choosing the parameters appropriately. The first inequality follows from Lemma \[decomp121\]. Now note that the second inequality follows from the definition of the approximate norm $\|\cdot\|_{(j_1,v_i)},$ along with the fact that ${{\mathbf{Tile}}}_{{{\mathfrak{n}}}_4}(j_1,v_i)$ is stable and most importantly the lower bound on the euclidean norms of the vectors in . The last fact is needed crucially since recall that $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}}$ allows us to relate the passage time to the norm only for pairs of points which are at a distance $\ell_1$ or more apart (see for e.g. and Definition \[tilegrad2proj\]). The third inequality is the content of Proposition \[conc1\] and the fourth inequality again follows from the definition of $\|\cdot\|_{(j_1,v_i)}.$ Again as above the last inequality relating the passage time to $\|\cdot\|_{(j_1,v_i)}$ follows since $|\overset{\rightarrow}{\alpha}^{{{\mathbf{S}}}}_i|$ is large enough by choice.
$\bo{(2)}$ $v_i \in A.$ In this case clearly $$|\alpha_i|\ge (b-{\varepsilon}_7)\|\overset{\rightarrow}{\alpha}_i\|_1 \ge (b-{\varepsilon}_7)\frac{n_1}{n}\|\overset{\rightarrow}{\alpha}^{{{\mathbf{S}}}}_i\|_1\ge \frac{n_1}{n} (1-O({\varepsilon}_7))|\alpha^{{{\mathbf{S}}}}_i|$$
where the last inequality follows from the discussion at the beginning of the proof and $\|\overset{\rightarrow}{\alpha}_i\|_1$ denotes the $1-$norm of the vector $\overset{\rightarrow}{\alpha}_i.$
The next three sections prove the three key technical results, Proposition \[t:stable\], \[conc1\] and \[p:cont\] regarding stability, approximate convexity of the distance function as well as continuity of the rate function. We start with the continuity result.
Continuity of the rate function {#s:cont}
===============================
In this section we prove Proposition \[p:cont\]. Recall that the statement says that for each ${\varepsilon}>0$, there exists ${\varepsilon}'>0$ such that for all $n$ sufficiently large we have $$\frac{\log \P(\sU_{\zeta-{\varepsilon}'}(n))}{n^2} \leq \frac{\log \P(\sU_{\zeta}(n))}{n^2}+{\varepsilon}.$$ This is where the assumption of continuous density of the edge distribution will simplify the proof significantly. Moreover, to avoid introducing new notation, we will use several letters in this section which has been used earlier to denote different quantities. However this section will be completely self contained and hence this should not create any confusion or conflict.
The basic approach is simply to start with an environment $\Pi \in \sU^*_{\zeta-{\varepsilon}'}$ and then increase the weight of ‘all’ the edges slightly to construct an environment $\Pi' \in \sU^*_{\zeta}.$ However a technical issue arises since we have assumed the variables are bounded by a constant $b>0$. Hence the variables in $\Pi$ which are very close to $b$ cannot be increased. Thus the first step is to localize the set of such really high valued edges. In fact we will also localize the set of edges which takes values where the density $f_{\nu}$ is close to zero. To carry this out, for any ${\varepsilon}_1,$ let ${\varepsilon}_2$ be such that $\P(X_{e}\in[b-{\varepsilon}_2,b])\le {\varepsilon}_1$ and moreover we will choose ${\varepsilon}_2$ such that there exists ${\varepsilon}_3>0$ such that $$\inf\{f_{\nu}(x): x\in[b-{\varepsilon}_2,b-{\varepsilon}_2+{\varepsilon}_3]\}\ge {\varepsilon}_3.$$ Now let $\bo{B}=\{x \in [0,b-{\varepsilon}_2]: f_{\nu}(x)\le \frac{{\varepsilon}^3_3}{b}\}.$ Thus by definition $\nu(\bo{B}) \leq {\varepsilon}_3^3.$ Now for any $n,$ recall the notation ${\mathfrak{B}}_4={{\mathbf{Box}}}(\sC n)$ from . We will work with the event $\sU^*_{\zeta-{\varepsilon}'}(n)$ which is a function of the edges on ${{\mathbf{Box}}}(4\sC n)$. However recall that on $\sU^*_{\zeta-{\varepsilon}'}(n)$ any path from $\bo{0}$ which exited ${\mathfrak{B}}_4$ has length bigger than $bn$ thus it would suffice to increase the value of the edges only inside ${\mathfrak{B}}_4.$ Let $\bo{H}_1=\{e \in {\mathfrak{B}}_4 :X_e \in [b-{\varepsilon}_2,b]\}.$ Now by a straightforward union bound over all possible choices of $\bo{H}_1$ (at most $2^{O(n^2)}$), for any ${\varepsilon}_4>0,$ $$\begin{aligned}
\label{localize23}
\P(|\bo{H}_1|\ge {\varepsilon}_4n^2)\le 2^{O(n^2)}{\varepsilon}_1^{{\varepsilon}_4 n^2}&= e^{O(n^2)+{\varepsilon}_4\log(\frac{1}{{\varepsilon}_1})n^2} \text{and hence, }\\
\nonumber
&=o\left(\P(\sU^*_{\zeta})\right) \text{ for all small enough } {\varepsilon}_1,\\
\nonumber
&= o\left(\P(\sU^*_{\zeta-{\varepsilon}'})\right) \text{ for all }{\varepsilon}'>0. \end{aligned}$$ Similarly letting $\bo{H}_2=\{e \in {\mathfrak{B}}_4 :X_e \in \bo{B}\}$ we get $$\begin{aligned}
\label{localize24}
\P(|\bo{H}_2|\ge {\varepsilon}_4n^2)\le 2^{O(n^2)}{\varepsilon}_3^{3 {\varepsilon}_4 n^2}&= e^{O(n^2)+3 {\varepsilon}_4\log(\frac{1}{{\varepsilon}_3})n^2} \text{and hence, }\\
\nonumber
&=o\left(\P(\sU^*_{\zeta})\right) \text{ for all small enough } {\varepsilon}_3,\\
\nonumber
&= o\left(\P(\sU^*_{\zeta-{\varepsilon}'})\right) \text{ for all }{\varepsilon}'>0. \end{aligned}$$
The above allows us to localize $\bo{H}_1$ and $\bo{H}_2,$ without paying too much in the probability. Formally fix some ${\varepsilon}'>0$ (whose value would be specified later). Observe that the total number of subsets of ${\mathfrak{B}}_4$ of size at most ${\varepsilon}_4n^2$ is at most $2^{O({\varepsilon}_5)n^2}$ where ${\varepsilon}_5=-\left[{\varepsilon}_4\log({\varepsilon}_4)+(1-{\varepsilon}_4)\log(1-{\varepsilon}_4)\right]$ goes to zero as ${\varepsilon}_4$ goes to zero. From the above discussion it follows that for any ${\varepsilon}_4,$ by choosing ${\varepsilon}_1,{\varepsilon}_3$ small enough we have $$\P(\{|\bo{H}_1|\le {\varepsilon}_4 n^2\}\cap \{|\bo{H}_2|\le {\varepsilon}_4 n^2\} \cap \sU^*_{\zeta-{\varepsilon}'})\ge \P(\sU_{\zeta-{\varepsilon}'})(1-o(1)).$$ Thus by pigeon-hole principle it follows that there exists subset $A_1,A_2$ of ${\mathfrak{B}}_4$ each of size at most ${\varepsilon}_4 n^2$ such that $$\label{probbound6754}
\P(\{\bo{H}_1=A_1\} \cap \{\bo{H}_2=A_2\} \cap \sU^*_{\zeta-{\varepsilon}'})\ge \P({\sU_{\zeta-{\varepsilon}'}})e^{-O({\varepsilon}_5n^2)}.$$
For easy referencing let us call the event $\{\bo{H}_1=A_1\} \cap \{\bo{H}_2=A_2\} \cap \sU^*_{\zeta-{\varepsilon}'}$ as $\bo{C}.$ We will also use the following consequence of uniform continuity of $f_{\nu}$ on $[0,b]$, and the fact that $\overline{\bo{D}},$ where $\bo{D}=[0,b]\setminus \{\bo{B}\cup [b-{\varepsilon}_2,b]\}$ is compact and more importantly $f_{\nu}$ is uniformly away from zero (at least $\frac{{\varepsilon}^3_3}{b}$) on the former: Given any ${\varepsilon}_6$ there exists ${\varepsilon}_7$ such that for any $x\in \overline{\bo{D}}$ such that $$\label{RN}
\frac{1}{1+{\varepsilon}_6}\le \frac{f_{\nu}(x+{\varepsilon}_7)}{f_{\nu}(x)}.$$ Now let us modify the event $\bo{C}$ to get an event $\bo{C}_{1}$ which will posses the property that $\log(\P(\bo{C}_1))-\log(\P(\bo{C}))=o(n^2)$ and most importantly $\bo{C}_1\subset \sU_{\zeta}.$ Formally for any $\Pi \in \bo{C}$ noting that by definition $A_1$ and $A_2$ are disjoint, $$\begin{aligned}
\bo{C}_1(\Pi)=&\{ \Pi': \Pi'(e)=\Pi(e)\,\, \forall e \in A_1,\\
& \Pi'(e)\in [b-{\varepsilon}_2+\frac{{\varepsilon}_3}{2},b-{\varepsilon}_2+{\varepsilon}_3]\,\, \forall e \in A_2,\\
& \Pi'(e)=\Pi(e)+{\varepsilon}_7 \,\, \forall e\in {\mathfrak{B}}_4\setminus A_1\cup A_2
\}.\end{aligned}$$ Let $\displaystyle{\bo{C}_1=\bigcup_{\Pi\in \bo{C}} \bo{C}_1(\Pi)}.$ We now compute $\P(\bo{C}_1).$ For any $\Pi,$ and subset $B$ of edges in ${\mathfrak{B}}_{4},$ it would be convenient to let $\Pi|_{B}$ be the restriction of $\Pi$ on the edges in $B$; for any event $\bo{E}$ let $\bo{E}(B)=\{\Pi|_{B}:\Pi \in \bo{E}\}$; and let $\displaystyle{f_{\nu}(\Pi|_B)):=\prod_{e\in B}}f_{\nu}(\Pi(e))$ (in case $B={\mathfrak{B}}_4$ we would omit the above notations.). Thus $$\label{prob98534}
\P(\bo{C})=\int_{\bo{C}} f_{\nu}(\Pi) \,\,\mathrm{d}\Pi \le {\varepsilon}_3^{3|A_2|} \int_{\bo{C}({\mathfrak{B}}_4\setminus A_2)} f_{\nu}(\Pi|_{{\mathfrak{B}}_4\setminus A_2})\,\,\mathrm{d}\Pi|_{{\mathfrak{B}}_4\setminus A_2},$$ where the second inequality follows from the definition of $A_2.$ Note that by definition $\bo{C}_1({\mathfrak{B}}_4\setminus A_2)=\bo{C}({\mathfrak{B}}_4\setminus A_2)+\bo{v},$ where $$\bo{v}(e)=\left\{\begin{array}{cc}
0 & \text{ if } e \in A_1 \\
{\varepsilon}_7 & \text{ if } e \in {\mathfrak{B}}_4\setminus \{A_1\cup A_2\}
\end{array}\right..$$ Now observe that $$\begin{aligned}
\label{lbc1}
\P(\bo{C}_1)&\ge {\left(\frac{{\varepsilon}_3}{2}\right)^{2|A_2|}} \int_{\bo{C}_1({\mathfrak{B}}_4\setminus A_2)} f_{\nu}(\Pi|_{{\mathfrak{B}}_4\setminus A_2}),\\
\nonumber
&\ge \left(\frac{{\varepsilon}_3}{2}\right)^{2|A_2|} \left(\frac{1}{1+{\varepsilon}_6}\right)^{|{\mathfrak{B}}_4|} \int_{\bo{C}({\mathfrak{B}}_4\setminus A_2)} f_{\nu}(\Pi|_{{\mathfrak{B}}_4\setminus A_2})\\
&\ge e^{-O({\varepsilon}_6)n^2}\P(\bo{C}),\end{aligned}$$ where the first inequality follows from the definition of $\bo{C}_1$, the second inequality is by and the final equality is by by choosing ${\varepsilon}_3$ and ${\varepsilon}_7$ small enough. Thus by , the proof will now be complete once we show that $\bo{C}_1 \subset \sU_{\zeta}.$ To do this note that for any $\Pi'\in \bo{C}_1$ there exists $\Pi \in \bo{C}$ such that $\Pi'(e)\ge \Pi(e)+\min (\frac{{\varepsilon}_3}{2},{\varepsilon}_7)$ for all ${\varepsilon}\in {\mathfrak{B}}_4 \setminus A_1$ and $\Pi'(e)=\Pi(e)$ for $e\in A_1.$
Note that since $\Pi \in \sU^{*}_{\zeta-{\varepsilon}'}$, any path $\sG$ starting from the origin, which exits ${\mathfrak{B}}_4$ has weight at least $bn$ in $\Pi$ and hence by the above discussion also in $\Pi'$.
Thus to prove the lemma we only consider the path $\sG$ which is the shortest path between $\bo{0}$ and $\mathbf{n}$ lying inside ${\mathfrak{B}}_4$, in the environment $\Pi'.$ We want to show $|\sG|_{\Pi'}\ge (\mu+\zeta)n,$ where $|\sG|_{\Pi},|\sG|_{\Pi'}$ denote the weights of $\sG$ in the environments $\Pi$ and $\Pi'$ respectively.
Now since trivially $|\sG|_{\Pi'}\ge |\sG|_{\Pi},$ there is nothing to show if $|\sG|_{\Pi}>(\mu+\zeta)n.$ Assuming otherwise, it follows that $$|\sG\cap A_1|\le \frac{(\mu+\zeta)n}{b-{\varepsilon}_2}=cn$$ for some $c<1$ for all ${\varepsilon}_2$ small enough since $\mu+\zeta <b$ (by $\sG\cap A_1$ we denote the set of edges in $A_1$ that $\sG$ passes through). Indeed, this is true since each edge in $A_1$ has weight at least $b-{\varepsilon}_2.$ However note that since $\sG$ connects $\bo{0}$ and $\bo{n},$ trivially $\sG$ passes through at least $n$ edges. Thus $|\sG\cap A_1^c|\ge (1-c)n$ and hence $|\sG|_{\Pi'}-|\sG|_{\Pi}\ge {(1-c)}n\min (\frac{{\varepsilon}_3}{2},{\varepsilon}_7)$. By definition $|\sG|_{\Pi} \ge (\mu+\zeta-{\varepsilon}')n$ and hence taking ${\varepsilon}'=\frac{\min(\frac{{\varepsilon}_3}{2},{\varepsilon}_7){(1-c)}}{2}$ implies the sought bound $|\sG|_{\Pi'}\ge (\mu+\zeta)n.$ Thus to finish the proof phrased in terms of the parameters in the statement of Proposition \[p:cont\], given ${\varepsilon}$ we must choose ${\varepsilon}_4$ small enough so that in the $O({\varepsilon}_5)$ term is at most ${\varepsilon}.$ This dictates the choice of ${\varepsilon}_1$ and ${\varepsilon}_3,$ which in turn dictates the choice of ${\varepsilon}_2$. The choice of ${\varepsilon}_6$ and hence ${\varepsilon}_7$ is governed by which then fixes the value of ${\varepsilon}'$.
Approximate convexity properties {#pconc1}
================================
In this section we will prove Proposition \[conc1\], i.e. given $\delta_1,m_1$ and $j_1$, for any ${{\mathbf{Tile}}}_{\sC n}({j_1},v)$ which is $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}}$ where $\ell_1=\frac{n}{2^{j_1+m_1}}$ and $k=2^{2m_1}$, and any set of vectors $\bo{w}_1,\bo{w}_2,\ldots,\bo{w}_t$, if $\bo{w}=\sum_{i=1}^t \bo{w}_i$ then $$\label{dis9876}
\|\bo{w}\|_{(j_1,v)}\le (1+\delta)\sum_{i=1}^t\|\bo{w}_i\|_{(j_1,v)}$$ where $\delta= O(\delta_1+{2^{-\frac{m_1}{16}}}).$ The proof essentially follows by noticing that any set of vectors as above can be scaled down to get a sum of vectors inside ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ followed by application of stability and triangle inequality. To formalize this, we need some notation: For every $\phi\in \mathbb{S}^1(\eta_1),$ (value of $\eta_1$ will be specified later and sufficiently small) $$\cC(\phi)=\{\bo{b} \in \{\bo{w}_1, \bo{w}_2,\ldots, \bo{w}_t\} : {\rm{arg}}(\bo{b}) \in [\phi,\phi+\eta_1) \}$$ where $\bo{w}_1,\bo{w}_2,\ldots \bo{w}_t$ are as in the statement of the proposition, i.e. $\cC(\phi)$ denotes the collection of vectors among $\{\bo{w}_1, \bo{w}_2,\ldots, \bo{w}_t\}$ whose angle with the $x-$axis falls in the interval $[\phi,\phi+\eta_1).$ Let $\bo{w}_{\phi}= \sum_{i=1}^t \bo{w}_i \mathbf{1}(\bo{w}_i \in \cC(\phi))$ be the sum of the vectors in $\cC(\phi).$ Thus by definition $$\sum_{\phi \in \mathbb{S}^{1}(\eta_1)}\bo{w}_{\phi}=\bo{w}.$$ Also note that for every $\bo{w}_i \in \cC(\phi),$ by Lemma \[smoothgradproj\]
$$\begin{aligned}
\label{dis90678}
\|\bo{w}_i\|_{(j_1,v)}&=\|\bo{w}_i\|_2\left(1+O(\delta_1+\eta_1+2^{-m_1/4})\right)\grad_{{{\mathbf{Proj}}}}((j_1,v),\phi) \text{ and hence,}\\
\label{app453}
\sum_{\bo{w}_i\in \cC(\phi)}\|\bo{w}_i\|_{(j_1,v)}&=A_{\phi}\left(1+O(\delta_1+\eta_1+2^{-m_1/4})\right) \grad_{{{\mathbf{Proj}}}}((j_1,v),\phi)\end{aligned}$$
where $A_{\phi}=\sum_{i=1}^t\|\bo{w}_i\|_2\mathbf{1}(\bo{w}_i \in \cC(\phi)).$ In the sequel for brevity we will denote the term $$\label{app4512}
\left(1+O(\delta_1+\eta_1+2^{-m_1/4})\right) \grad_{{{\mathbf{Proj}}}}((j_1,v),\phi)$$ in by $\tilde \grad_{{{\mathbf{Proj}}}}((j_1,v),\phi).$ Now since the angle made by each $\bo{w}_i\in \cC(\phi)$ lies in the interval $[\phi, \phi+\eta_1),$ for all small $\eta_1$ it also follows that $(1-\eta_1^2)A_{\phi}<\|\bo{w}_{\phi}\|_2\le A_{\phi}.$ Thus from the above two expressions it follows that
$$\label{normbd564}
(1-\eta_1^2)A_{\phi}\tilde \grad_{{{\mathbf{Proj}}}}((j_1,v),\phi)\le \|\bo{w}_{\phi}\|_{(j_1,v)}\le A_{\phi}\tilde \grad_{{{\mathbf{Proj}}}}((j_1,v),\phi).$$
For each $\phi$ let us consider the value ${b}_{\phi}=\frac{A_{\phi}}{\|{{\mathbf{w}}}\|_2}$. Now without loss of generality we can assume that there exists a universal constant $C$ such that ${b}_{\phi}\le C$ for all $\phi$ since otherwise we would be done using and the fact that $ \grad_{{{\mathbf{Proj}}}}((j_1,v),\phi)$ is bounded away from zero and infinity for any $\phi.$ We now define the set $
\cB=\{\phi\in \mathbb{S}^1(\eta_1): b_{\phi}\le {\varepsilon}_1\}
$ ( the value of ${\varepsilon}_1$ is specified later) and hence $$\label{ignoresmall}
\|\sum_{\phi \in \cB}\bo{w}_{\phi}\|_2 \le \frac{2\pi {\varepsilon}_1}{\eta_1} \|\bo{w}\|_2.$$ Now let $$\bo{c}_{\phi}=\frac{\bo{w}_{\phi}}{\|\bo{w}\|_2}\frac{\sC n}{100\times 2^{j_1}}, \text{ and let } \bo{c}=\frac{\bo{w}}{\|\bo{w}\|_2}\frac{\sC n}{100\times 2^{j_1}}.$$ Thus we have rescaled $\bo{w}$ to get a vector $\bo{c}$ of length $\frac{\sC n}{100 \times 2^{j_1}}$ and scaled all the $\bo{w}_{\phi}$’s by the same factor to obtain the $\bo{c}_{\phi}$’s. For convenience let $\mathbb{S}^{1}(\eta_1)\setminus \cB=\{\phi_1,\phi_2,\ldots,\}$ We now consider the sequence of points $\bo{v}_0, \bo{v}_1,\bo{v}_2,\ldots$ such that $\bo{v}_{i}-\bo{v}_{i-1}=\bo{c}_{\phi_i}$ and let $\bo{v}_0$ be for concreteness the center point of ${{\mathbf{Tile}}}_{\sC n}(j_1,v).$
Now consider the path $\cP$ (recall from ) obtained by concatenation of paths $\cP_1,\cP_2,\cP_3,\ldots$ where $\cP_i$ is the shortest path between $\bo{v}_{i-1}$ and $\bo{v}_i$.
At this point we make another assumption that none of the points $\bo{v}_i$ is outside ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ [^5]. Now assuming that ${{\mathbf{Tile}}}_{\sC n}(j_1,v)$ is $(\delta_1,\ell_1,k_1)-{{\mathbf{Stable}}}$ as in the hypothesis of the proposition, it follows from Lemma \[smoothgradproj\] that $$|\cP_i|\le \|\bo{c}_{\phi_i}\|_2 \tilde \grad_{{{\mathbf{Proj}}}}((j_1,v),\phi_i),$$ provided that ${\varepsilon}_1 \gtrsim 2^{-\frac{m_1}{4}}$ since by hypothesis as every $\phi_{i} \notin \cB$, $\|\bo{c}_{\phi_i}\|_2\ge \frac{{\varepsilon}_1\sC n}{100 \times 2^{j_1}}$. Thus $$\label{eq221}
|\cP|=\sum_{i}|\cP_i|\le (1+O(\delta_1+\eta_1+2^{-\frac{m_1}{4}}))\sum_{i}\|\bo{c}_{\phi_i}\|_{(j_1,v)}.$$ Now as $\cP$ is a path (not necessarily the shortest) joining $\bo{v}_0$ and $\bo{v}_0+\bo{c}_*$ where $\bo{c}_*=\sum_{i}\bo{c}_{\phi_i}$, by stability we have $$\label{eq222}
|\cP|\ge (1-O(\delta_1+\eta_1+2^{-\frac{m_1}{4}}))\|\bo{c}_*\|_{(j_1,v)}.$$ However note that by , it follows that $\|\bo{c}-\bo{c}_*\|_2 \le O(\frac{{\varepsilon}_1}{\eta_1}) \|\bo{c}\|_2 $ and hence by Lemma \[smoothgradproj\] $$\label{eq223}
\|\bo{c}_*\|_{(j_1,v)}\ge \left(1-O(\delta_1+\frac{{\varepsilon}_1}{\eta_1}+2^{-\frac{m_1}{4}})\right) \|\bo{c}\|_{(j_1,v)}.$$
Putting the above together (letting $1+B=1+O(\delta_1+\eta_1+2^{-m_1/4})$ appearing in ) it follows that $$\begin{aligned}
\sum_{i=1}^t \| \bo{w}_i\|_{(j_1,v)}&\overset{\eqref{normbd564}}{\ge} (1-\eta_1^2) (1-B) \sum_{\phi \in \bS^{1}(\eta_1)}\| \bo{w}_{\phi}\|_{(j_1,v)} \ge (1-\eta_1^2)(1-B)\sum_{\phi \in \bS^{1}(\eta_1)\setminus \cB}\| \bo{w}_{\phi}\|_{(j_1,v)},\\
&=(1-\eta_1^2)(1-B) \sum_{\phi \in \bS^1(\eta_1)\setminus \cB}\| \bo{c}_{\phi}\|_{(j_1,v)}\frac{100 \times 2^{j_1}\|\bo{w}\|_2}{{\sC n}},\\
&\ge (1-\eta_1^2)(1-B)^2(1-B-O(\frac{{\varepsilon}_1}{\eta_1})) \|\bo{c}\|_{(j_1,v)}\frac{100\times 2^{j_1}\|w\|_2}{{\sC n}} \ge (1-O(\delta_1+2^{-\frac{m_1}{16}}))\|\bo{w}\|_{(j_1,v)}\end{aligned}$$ where the second to last inequality follows from , and and the final inequality follows by choosing $\eta_1=2^{-m_1/8}$ and ${\varepsilon}_1\gtrsim 2^{-m_1/4}$ ensuring $\frac{{\varepsilon}_1}{\eta_1}=O(2^{-\frac{m_1}{16}}).$
Stability of the gradient {#s:sproof}
=========================
This section is devoted to proving Proposition \[t:stable\]. It turns out that this property has little to do with the specific details of the first passage percolation metric, rather it is a property of general distance functions on $\R^2$ that are comparable to the Euclidean metric i.e., it satisfies triangle inequality and that for all $x,y\in \R^2$ such that $|x-y|$ is large enough (possibly $n$ dependent) $$\label{e:comp2}
\alpha |\bx-\by|\leq {{\mathbf{PT}}}(\bx,\by) \leq 3b |\bx-\by|,$$ for some $\alpha>0$. For the ease of reading we recall the statement of the proposition and state it as a theorem to highlight the fact that its generality makes it potentially applicable in other problems of metric geometry. Recall our terminology that $\bz\in \R^2$ is $(\delta, \bS^1(\eta),\ell,k)-{{\mathbf{Stable}}}$ if $\bz$ is $(\delta,\theta,\ell,k)-{{\mathbf{Stable}}}$ for each $\theta \in \bS^1(\eta)$.
\[t:stablegen\] Fix $\delta,{\varepsilon},\eta >0,$ and $k\in \N$ and $J_1\in \N$. There exists $J_2\in \N$ such that for all large enough $n$: for $\Pi \in \sU^*_{\zeta}(n)$ (note that holds for all $\bx,\by \in {{\mathbf{Box}}}(10n)$ such that $|\bx-\by|\ge \sqrt n$) there exists $J_1\leq j\leq J_2$ such that $$\#\{\bz\in {{\mathbf{L-Box}}}(n): \bz~\text{is not}~(\delta, \bS^1(\eta),\frac{ n}{2^j},k)-{{\mathbf{Stable}}}\} \leq {\varepsilon}n^2 .$$
Above we have replaced $\sC n$ in the statement of the proposition by $n$ for notational brevity since as the reader will notice the arguments do not depend on the exact value in any way. Moreover from now on without explicitly stating it, we will assume that holds for all pairs of points $\bx, \by \in {{\mathbf{Box}}}(10 n)$ where $|\bx-\by|\ge \sqrt n$ even though we will not explicitly mention the last qualification every time since it will be trivially satisfied in our applications.
A roadmap of the proof
----------------------
As we are not shooting for optimal bounds the proofs will often rely on several crude averaging arguments and applications of the pigeon hole principle along with the bi-Lipschitz nature of the FPP metric. However, there are many technical steps involved and for the sake of exposition we give a brief overview of the argument at this point. Our argument relies on the following observations.
\(1) Fix $\bz\in {{\mathbf{Box}}}(n)$ and $\theta\in \bS^1(\eta)$. Observe that for all $J_2>J_1,$ $$\begin{aligned}
\label{tria786}
{{\mathbf{PT}}}(\bz,\theta,\frac{n}{2^{J_2}},2^{J_2})-{{\mathbf{PT}}}(\bz,\theta, \frac{n}{2^{J_1}},2^{J_1})&= \sum_{j=J_1}^{J_2-1}\left[{{\mathbf{PT}}}(\bz,\theta,\frac{n}{2^{j+1}},2^{j+1})-{{\mathbf{PT}}}(\bz,\theta,\frac{n}{2^j},2^{j})\right].\end{aligned}$$ The LHS in is bounded by $3bn$ and all the terms in the RHS in are positive by triangle inequality (as in Lemma \[monotone\]).
\(2) Thus if $J_2-J_1\ge \frac{1}{{\varepsilon}},$ then by the pigeon-hole principle there must exist one $J_1\le j\le J_2$ such that $\left[{{\mathbf{PT}}}(\bz,\theta,\frac{n}{2^{j+1}},2^{j+1})-{{\mathbf{PT}}}(\bz,\theta,\frac{n}{2^j},2^{j})\right]\le O({\varepsilon}) n$. [As a matter of fact we should find consecutive many such $j$ if $J_2-J_1\gg \frac{1}{{\varepsilon}}$.]{}
\(3) Now for $j$ as in (2) consider the discrete segments $$\begin{aligned}
\sS(\bz,\theta,\frac{n}{2^{j}},2^{j})=[\bz_0,\bz_1,\ldots \bz_{2^{j}}] \text{ and, }\sS(\bz,\theta,\frac{n}{2^{j+1}},2^{j+1})=[\bz_0, \bz_{0,1},\bz_1, \bz_{1,2},\bz_2,\ldots ,\bz_{2^{j}}], \end{aligned}$$ where $\bz_{i,i+1}$ is the mid-point of the line segment joining $\bz_i$ and $\bz_{i+1}.$ Thus the above observation together with the lower bound in suggests that for most $i$, $${{\mathbf{PT}}}(\bz_i,\bz_{i,i+1})+{{\mathbf{PT}}}(\bz_{i,i+1},\bz_{i+1})\le (1+O({\varepsilon})){{\mathbf{PT}}}(\bz_i,\bz_{i+1}).$$ However this is not quite enough to establish stability and in fact we need something along the lines of the following stronger fact (see Lemma \[crudelem43\]): for most $i,$ $${{\mathbf{PT}}}(\bz_i,\bz_{i,i+1})\approx{{\mathbf{PT}}}(\bz_{i,i+1},\bz_{i+1})\approx \frac{1}{2}{{\mathbf{PT}}}(\bz_i,\bz_{i+1}).$$
\(4) Suppose the contrary and without loss of generality assume that $${{\mathbf{PT}}}(\bz_i,\bz_{i,i+1})\ge (\frac{1}{2}+\delta) {{\mathbf{PT}}}(\bz_i,\bz_{i+1}).$$ The contradiction will come from the fact that the above cannot be true for many consecutive scales. Indeed, if it was true for $j'$ many consecutive scales, then recursively picking one half of an interval at each scale in which the above inequality holds leads to an interval $[{{\mathbf{w}}}_1, {{\mathbf{w}}}_2]$ such that $\|{{\mathbf{w}}}_1-{{\mathbf{w}}}_2\|_2=\frac{\|\bz_i-\bz_{i+1}\|_2}{2^{j'}}$ but $${{\mathbf{PT}}}({{\mathbf{w}}}_1,{{\mathbf{w}}}_2)\ge (1+2\delta)^{j'}\frac{{{\mathbf{PT}}}(\bz_i,\bz_{i+1})}{2^{j'}}.$$ Clearly for $j'$ large enough (depending on $\delta$) this contradicts the upper bound in . We now move towards making the above formal.
Recalling the notion of stability from , the following crude lemma will be useful to show the latter.
\[crudelem43\] Given $\delta>0$, $ \theta \in \bS^1$ and $\ell, k \in \N$. Recalling that $\sS(\bz,\theta,\ell,k)=[\bz=\bz_0,\bz_1,\ldots,\bz_k],$ suppose $$\begin{aligned}
\sup_{0 \le i,j \le k-1} \frac{{{\mathbf{PT}}}(\bz_i,\bz_{i+1})}{{{\mathbf{PT}}}(\bz_j,\bz_{j+1})} \le 1+\delta, ~~~\text{and }\quad
\frac{k{{\mathbf{PT}}}(\bz_0,\bz_{1})}{1+\delta}\le {{{\mathbf{PT}}}(\bz_0,\bz_{k})}\le {k{{\mathbf{PT}}}(\bz_0,\bz_{1})}(1+\delta). \end{aligned}$$ Then for each $i\leq k$, and $k'\le k-i,$ $\bz_i$ is $(\delta',\theta,\ell,i+k')-{{\mathbf{Stable}}}$ where $\delta'=O(\delta k)$.
By hypothesis $$\begin{aligned}
\frac{1}{(1+\delta)}k{{\mathbf{PT}}}(\bz_{0},\bz_1)\le {{\mathbf{PT}}}(\bz_0,\bz_k)\le {(1+\delta)}k{{\mathbf{PT}}}(\bz_{0},\bz_1).\end{aligned}$$ Now as for any $i\le k$ and $k'\le k-i$ we have ${{\mathbf{PT}}}(\bz_0,\bz_k)\le {{\mathbf{PT}}}(\bz_i,\bz_{i+k'})+(k-k')(1+\delta){{\mathbf{PT}}}(\bz_{0},\bz_1).$ Thus it follows that $$\begin{aligned}
{{\mathbf{PT}}}(\bz_i,\bz_{i+k'})&\ge\frac{1}{(1+\delta)}k{{\mathbf{PT}}}(\bz_{0},\bz_1)-(k-k')(1+\delta){{\mathbf{PT}}}(\bz_{0},\bz_1)\ge k'{{\mathbf{PT}}}(\bz_{0},\bz_1)(1-O(\delta k)), \\
&\ge k'{{\mathbf{PT}}}(\bz_{i},\bz_{i+1})(1-O(\delta k)).\end{aligned}$$ Moreover note that by triangle inequality and the hypothesis, ${{\mathbf{PT}}}(\bz_i,\bz_{i+k'})\le (1+\delta)k'{{\mathbf{PT}}}(\bz_i,\bz_{i+1}).$
Thus in the sequel to prove stability we will only prove that the hypothesis of Lemma \[crudelem43\] is satisfied. Going back to the proof of Theorem \[t:stablegen\] following the line of argument in the roadmap above, one can deduce the existence of many stable points along a fixed line in a given direction. Further arguments are then necessary to strengthen this to get the full result. We shall first state and prove the weaker version.
Stability on a fixed line
-------------------------
For the weaker version let us consider the discrete segment $\sS(\bz,\theta,\frac{n}{2^j},2^j)$ for some $\bz\in {{\mathbf{Box}}}(n)$ and $\theta\in \bS^1(\eta)$. We shall show most points on this segments are stable for $k$ consecutive intervals.
\[l:weak\] Let $\bz\in {{\mathbf{Box}}}(n), \theta \in \bS^1(\eta), k\in \N,$ and let $\delta_2>0$ and $J_1\in \N$ be fixed. Then there exists ${{\mathfrak{m}}}$ such that for all small enough $\delta_3$ the following holds: there exists $j\in \N$ with $ J_1\le j \le (J_1+ \frac{1}{\delta_3^2})$ for which all but $O(\delta_2)$ fraction of the points $\bz_i$ in the discrete segment $\sS(\bz,\theta,\frac{n}{2^{j{{\mathfrak{m}}}}},2^{j{{\mathfrak{m}}}})$ are $(\delta_2,\theta,\frac{n}{2^{j{{\mathfrak{m}}}}},k)-{{\mathbf{Stable}}}$.
The quantification in the above statement might be a little hard to parse, but it will create some simplification in the notational choices later.
For the moment let us fix a value of ${{\mathfrak{m}}}$ to be specified later. To make formal the outline described in the subsection it will be convenient to associate trees to the the intervals in $\sS(\bz,\theta,\frac{n}{2^{{{\mathfrak{m}}}J_1}},2^{{{\mathfrak{m}}}J_1})=[\bz=\bz_0, \bz_1, \ldots, \bz_{2^{{{\mathfrak{m}}}J_1}}]$. Let $$\label{treecon12}
\cT_1,\cT_2.\ldots, \cT_{2^{{{\mathfrak{m}}}J_1}}$$ be complete $2^{{{\mathfrak{m}}}}-$ary trees of depth $J_2-J_1$ where the value of $J_2$ will be specified to be a large enough number later (for convenience we shall index the levels of these trees by $j=J_1, J_1+1\ldots , J_2$). Let $L^{(i)}_j$ denote the vertices at the $j^{th}$ level of $\cT_{i}$ and let $L_j=\cup_{i}L^{(i)}_j,$ denote the union of the vertices at the $j^{th}$ level. We will identify $\cT_{i}$ with the interval $[\bz_{i-1},\bz_{i}]$. Now for any $J_1\le j\le J_2,$ consider the discrete segment $$\sS(\bz,\theta, \frac{n}{2^{j{{\mathfrak{m}}}}},2^{j{{\mathfrak{m}}}})=[\bz^*_{0},\ldots \bz^*_{2^{(j-J_1){{\mathfrak{m}}}}},\bz^*_{2^{(j-J_1){{\mathfrak{m}}}}+1},\ldots, \bz^*_{2^{(j+1-J_1){{\mathfrak{m}}}}},\ldots \bz^*_{2^{(j-1){{\mathfrak{m}}}}}\ldots \bz^*_{2^{j{{\mathfrak{m}}}}}].$$
Naturally $[\bz^*_{0},\ldots ,\bz^*_{2^{(j-J_1){{\mathfrak{m}}}}}]$ is a discretization of the interval $[\bz_0,\bz_1]$ and hence can be associated to $L^{(1)}_{j}$ where each vertex in $L^{(1)}_{j}$ corresponds to $[\bz^*_h,\bz^*_{h+1}]$ in the natural order. (e.g. the root of $\cT_1$ corresponds to the interval $[\bz_{0},\bz_1]$). The same correspondence holds for the other intervals and trees. See Figure \[fig12\].
![This figure illustrates the various definitions introduced in this section related to the trees in in the toy case ${{\mathfrak{m}}}=1$ where the trees are binary trees.[]{data-label="fig12"}](stable.pdf)
Now for any vertex $v$ in any of the trees let $Y_v:={{\mathbf{PT}}}(\bz',\bz'')$ where $[\bz',\bz'']$ is the discrete segment associated to the vertex $v$. We need some further notation: let $U_{i,j}=\sum_{v\in L^{(i)}_j}Y_v$ and let $U_j=\sum_{i} U_{i,j}$. It will be convenient to frame our arguments using pigeon hole principle as applications of ‘the probabilistic method’, and hence we define a set of random variables. For any $j$ pick uniformly any edge $e_{j+1}$ at the $(j+1)^{th}$ level across all the trees, i.e., connecting $L_{j}$ and $L_{j+1}$ and let $$\label{edge23}
X_{e_{j+1}}:=\frac{2^{{{\mathfrak{m}}}}Y_{{w}}}{Y_{v}}$$ where $e_{j+1}=(v,w)$ and $v$ is closer to the root. For brevity we will identify the set of such edges with the set $L_{j+1}$ using the natural correspondence. Now by triangle inequality (again, as in Lemma \[monotone\]) $$\label{triangle}
\E(X_{e_{j+1}}\mid Y_{v})\ge 1.$$ However the distributions of $X_{e_j}$ across various $j$ will not be independent and the joint distribution can be defined in the following way: pick uniformly a vertex among all the leaf vertices across all the trees (note that it is naturally and uniquely associated with a uniformly chosen simple path from the root to the leaf in an uniformly chosen tree) and label the edges on the path as $(e_{J_1+1},e_{J_1+2},\ldots , e_{J_2})$ where $e_i$ denotes the intersection of the path with the $i^{th}$ level. It is clear that $e_j$ is uniformly distributed among all edges connecting $L_{j-1}$ and $L_j$. Notice that $$\prod_{j=J_1+1}^{J_2}X_{e_j}\overset{d}{=}\frac{2^{(J_2-J_1){{\mathfrak{m}}}}Y_{w}}{Y_{v}}$$ where $Y_v$ and $Y_w$ are the variables attached to the root of a randomly chosen tree $\cT_i$ and a randomly chosen leaf of $L^{(i)}_{J_2-J_1}$ respectively. We now bound the expectation of $X_{e_j}$ for all $J_1< j\le J_2.$ To do this consider the ratio $\frac{U_{j+1}}{U_j}.$ By definition, we have the following: $$\begin{aligned}
\label{expreps}
\frac{U_{j+1}}{U_j}=\frac{\sum_{w\in L_{j+1}}Y_w}{\sum_{v\in L_{j}}Y_v}=\frac{\sum_{v \in L_j}Y_v\E(X_{e_{j+1}}\mid Y_v)}{\sum_{v\in L_j}Y_v}\ge 1\end{aligned}$$ where the second equality follows from and the fact that the trees $\cT_i$ are $2^{{{\mathfrak{m}}}}-$ary and the final inequality follows from . The following lemma completes the proof of Lemma \[l:weak\] under a further assumption that on a sufficiently large interval contained in $[J_1, J_1+\frac{1}{\delta_3^2}]$ the LHS above is also upper bounded by $1+\delta_3$ for some small enough $\delta_3$.
\[l:conditional\] Fix $c>0$. In the setting of Lemma \[l:weak\], suppose there exists an interval $I\subseteq [J_1, J_1+\frac{1}{\delta_3^2}]$ with $|I|\geq \frac{c}{\delta_3}$ such that for all $j\in I$ for some small enough $\delta_3$ depending on ${{\mathfrak{m}}}$ and $\delta_2$ $$\begin{aligned}
\label{smallgap}
0\le \frac{U_{j+1}}{U_j}-1 \le \delta_3.\end{aligned}$$ Then the conclusion of Lemma \[l:weak\] holds.
Without loss of generality for this proof we shall write $I=[J_1,J_2]$ where $I$ is given by the hypothesis. It is a consequence of that for $j\le k$ and $v\in L_j$ and $w \in L_k,$ $$\label{comparable}
\frac{1}{C}\le \frac{Y_v}{2^{(k-j){{\mathfrak{m}}}}Y_w}\le C$$ for some universal constant $C=C(b,\alpha)>1$. Define now the probability measure $\mu_j$ on the $j^{th}$ level vertices given by $\mu_j(v)=\frac{Y_v}{\sum_{v\in L_j} Y_v}$. In particular, implies the Radon-Nikodym derivative of $\mu_j$ with respect to the uniform measure $\mathfrak{u}$ is bounded above and below by $C$ and $C^{-1}$ respectively. Now, , along with , implies $\E_{{\mu_j}}(\E(X_{e_{j+1}}\mid v)-1)\leq \delta_3$. This, together with the above observation, and the fact $\E(X_{e_{j+1}}\mid v)-1)>0$ implies that
$$\begin{aligned}
\label{uniexp}
\E_{\mathfrak{u}}(\E(X_{e_{j+1}}\mid v)-1)&\le C \delta_3.\end{aligned}$$
By , $C$ in can be chosen such that such that deterministically $\frac{1}{C}\le X_{e_j}\le C$ and moreover, $$\begin{aligned}
\label{log}
\frac{1}{C}\le \prod_{j=J_1+1}^{J_2}X_{e_j}\le C, \text{ which implies, } \left|\sum_{j=J_1+1}^{J_2}\E_\mathfrak{u}(\log X_{e_j})\right|& \leq \log C.\end{aligned}$$ Thus it follows that there exists $J_1+1\le j\le J_2$ such that $\E_{\mathfrak{u}}(\log X_{e_j})\ge -c^{-1}(\log C)\delta_3$. Hence we have found a $J_1+1\le j\le J_2$ with the following two properties: $$\begin{aligned}
1\le \E_{\mathfrak{u}}(X_{e_j})\overset{\eqref{uniexp}}{\le} 1+ C\delta_3 \text{ and }\E_{\mathfrak{u}}(\log(X_{e_j}))\ge-c^{-1}(\log C)\delta_3.\end{aligned}$$ Now for any edge $e,$ denoting $X_e-1=y_e,$ the above can be restated as $$\begin{aligned}
0\le \frac{1}{2^{j{{\mathfrak{m}}}}}\sum_{e\in L_j}y_e \le C\delta_3 \text{ and } \frac{1}{2^{j{{\mathfrak{m}}}}}\sum_{e\in L_j}\log (1+y_e) \ge -c^{-1}(\log C)\delta_3.\end{aligned}$$
Now note that by , $y_e\le C$ and hence using Taylor expansion, $\log (1+y_e)\le y_e -C'y_e^2$ for some universal constant $C'$. Using the above inequalities it follows that $$\E_{\mathfrak{u}}(X_{e_j}-1)^2=\frac{1}{2^{j{{\mathfrak{m}}}}}\sum_{e \in L_j } y^2_{e}=O(\delta_3).$$ Thus by Chebyshev inequality, for at least $1-O(\sqrt{\delta_3})$ fraction of $e\in L_j$, we have $|X_{e}-1|\le \delta_3^{1/4}$. Let us call such an edge $e$, a good edge. Now let us consider all $v\in L_{j-1}$ such that all the children of $v$ are good (let us call such $v$ good). A naive bound shows that the fraction of good $v$ is at least $1-O(2^{{{\mathfrak{m}}}}\sqrt \delta_3).$ Now for any good $v$ corresponding to an interval $[{{\mathbf{w}}}_1,{{\mathbf{w}}}_2]$ say, if the discrete segment $[{{\mathbf{w}}}_1={{\mathbf{w}}}^*_0, {{\mathbf{w}}}^*_1, \ldots, {{\mathbf{w}}}^*_{2^{{{\mathfrak{m}}}}}={{\mathbf{w}}}_2]$ corresponds to the $2^{{{\mathfrak{m}}}}$ children then Lemma \[crudelem43\] implies the following: each ${{\mathbf{w}}}^*_i$ for $i\in \llbracket 0,2^{{{\mathfrak{m}}}}-k\rrbracket$ is $(\delta',\theta,\frac{n}{2^{j{{\mathfrak{m}}}}},k)-{{\mathbf{Stable}}},$ where $\delta'=O(2^{{\mathfrak{m}}}\delta_3^{1/4}).$ Thus the total fraction of points on $\sS[\bz,\theta,\frac{n}{2^{j{{\mathfrak{m}}}}},2^{j{{\mathfrak{m}}}}]$ that are not $(\delta',\theta,\frac{n}{2^{j{{\mathfrak{m}}}}},k)-{{\mathbf{Stable}}}$ is at most $O(\frac{k}{2^{{{\mathfrak{m}}}}}+2^{{{\mathfrak{m}}}}\sqrt \delta_3).$ Now choose ${{\mathfrak{m}}}$ large enough and then $\delta_3$ small enough such that $\max (\frac{k}{2^{{{\mathfrak{m}}}}}+2^{{{\mathfrak{m}}}}\sqrt \delta_3, \delta')\le \delta_2$.
It remains to prove that holds for a number of consecutive scales. This is ensured by the following lemma using another pigeon hole argument.
\[l:consec\] In the setting of Lemma \[l:conditional\], there exists $c>0$, and $I\subseteq [J_1, J_1+\frac{1}{\delta_3^2}]$ with $|I|\geq \frac{c}{\delta_3}$ such that for all $j\in I$ $$0\le \frac{U_{j+1}}{U_j}-1 \le \delta_3.$$
For $c>0$ to be specified later, we divide the $\frac{1}{\delta_3^2}$ many scales into consecutive blocks of $\frac{c}{\delta_3}$ many scales each. For $i\in \llbracket 1, \frac{1}{c\delta_3}\rrbracket$, Let $a_i=U_{J_1+\frac{ic}{\delta_3}}-U_{J_1+\frac{(i-1)c}{\delta_3}}$. By the triangle inequality, $a_{i}\ge 0$ for all $i$, and by , there exists a universal constant $C$ such that $U_{J_1+\frac{1}{\delta_3^2}}\le C U_{J_1}.$ As a consequence, $\sum_{i} a_i \leq CU_{J_1}$ and by choosing $c$ sufficiently small it follows there exists some $i\in \llbracket 1,\frac{1}{c\delta_3} \rrbracket$ such that $a_i \le U_{J_1}\delta_3.$ Now this implies that for any $J_1+\frac{(i-1)c}{\delta_3}\le j\le J_1+\frac{ic}{\delta_3}$ we have $\frac{U_{j+1}-U_j}{U_j}\le \frac{a_i}{U_{J_1}}\leq \delta_3$; completing the proof.
Strengthening Lemma \[l:weak\] to Theorem \[t:stablegen\]
---------------------------------------------------------
We now provide the extra ingredients needed to extend the argument of the previous subsection to establish the stronger statement of Theorem \[t:stablegen\]. To avoid repetition, often instead of providing the full formal proof we shall describe the main ideas and present an elaborate sketch. Observe that to establish Theorem \[t:stablegen\], one needs to extend Lemma \[l:weak\] in the following two directions:
1. Get the stability at a point simultaneously at all directions in $\bS^1(\eta)$ at the same scale $j$.
2. Deducing stability of most lattice points from stability of points on a discrete segment (which are not necessarily lattice points).
We describe below how to take care of these two items. To address the issue in (a) note that one cannot naively apply the above argument separately for all $\theta\in \mathbb{S}^1(\eta)$ since a priori one might not end up with the same scale $j$ for all $\theta\in \mathbb{S}^1(\eta)$.
![This figure illustrates the set of parallel lines $\mathfrak{L}_{\theta}:=\{\mathbb{L}^{\theta}_{-K}, \ldots, \mathbb{L}^{\theta}_{-1}, \mathbb{L}^{\theta}_0,\mathbb{L}^{\theta}_1, \ldots, \mathbb{L}^{\theta}_K\}$. The red and blue dots denote the points $\bz^{i,\theta,\ell, j}_{h}$ for $i \in \llbracket-K,K \rrbracket , \theta \in \mathbb{S}^{1}(\eta), \ell \in \llbracket 0,M-1\rrbracket ,$ and $h \in \llbracket 1,2^{(j-J_1){{\mathfrak{m}}}}\rrbracket$. The red and blue colors denote whether the point is $(\delta_2,\theta,\frac{n}{2^{j{{\mathfrak{m}}}}},k_1)-{{\mathbf{Stable}}}$ or not respectively. For each such point we associate a rectangular box with one of the sides parallel to $\mathbb{L}^{\theta}_{0}$, where the point is at the north-west corner of the associated rectangle. A particular example of a point $\bz^{i,\theta,\ell, j}_{h}$ and the associated rectangle $\mathfrak{R}$ and a lattice point $\bf$ inside $\mathfrak{R}$ are marked in the figure. The green boxes are associated to the blue points and the yellow boxes are associated to the red points. []{data-label="fig13"}](tilestable.pdf)
Instead we do the following: for each $\theta\in \bS^1(\eta),$ consider the set of parallel lines $$\mathfrak{L}_{\theta}:=\{\mathbb{L}^{\theta}_{-K}, \ldots, \mathbb{L}^{\theta}_{-1}, \mathbb{L}^{\theta}_0,\mathbb{L}^{\theta}_1, \ldots, \mathbb{L}^{\theta}_K\}$$ where for any $i\in \llbracket-K,K\rrbracket,$ $\mathbb{L}^{\theta}_i$ is a line segment of length $4n$, making angle $\theta$ with the $x-$axis; $\mathbb{L}^{\theta}_0$ is centered at the origin; and $\mathbb{L}^{\theta}_i$ is obtained by translating $\mathbb{L}^{\theta}_{0}$ in the orthogonal direction by $\frac{i n}{2^{J_3{{\mathfrak{m}}}}}$ where $J_3=J_1+\frac{1}{\delta_3^4}$ and $K=3 \times 2^{J_3 {{\mathfrak{m}}}}$, (see Figure \[fig13\]). For each $\theta \in \mathbb{S}^{1}(\eta),$ and each $i\in \llbracket-K,K \rrbracket$ let $\sS_{i,\theta}$ be the discrete line segment formed by the points on $\mathbb{L}^{\theta}_i$ at spacing $\frac{n}{2^{J_1{{\mathfrak{m}}}}}$ (without loss of generality we assume that the starting and ending points of $\mathbb{L}^{\theta}_i$ and $\sS_{i,\theta}$ are the same to avoid rounding issues). Thus $\sS_{i,\theta}=[\bz^{i,\theta}_0,\bz^{i,\theta}_1,\ldots,\bz^{i,\theta}_{M}]$ where $M=4 \times 2^{J_1{{\mathfrak{m}}}}.$ We now create a tree $\cT_{i,\theta,\ell}$ for each $i\in \llbracket-K,K\rrbracket, \theta \in \mathbb{S}^{1}(\eta), \ell \in \llbracket 0,M-1 \rrbracket$ corresponding to the interval $[\bz^{i,\theta}_{\ell},\bz^{i,\theta}_{\ell+1}]$ as in . As before for any $j\ge J_1$, let $L^{i,\theta,\ell}_j$ denote the $j^{th}$ level of the tree $\cT_{i,\theta,\ell}$ and $L_j=\bigcup_{i,\theta,\ell} L^{i,\theta,\ell}_j$.
Running the same argument as before with these trees in place of the ones in now gives us $J_1\le j\le J_1+\frac{1}{\delta_3^2}$ with the following property. If $$\sS_{i,\theta,\ell,j}=[\bz^{i,\theta,\ell, j}_{0},\bz^{i,\theta,\ell, j}_{1},\ldots, \bz^{i,\theta,\ell, j}_{2^{(j-J_1){{\mathfrak{m}}}}}]$$ denotes the discrete segment corresponding to $L^{i,\theta,\ell}_j,$ i.e., the $2^{(j-J_1){{\mathfrak{m}}}}$ vertices in the latter correspond to the intervals $[\bz^{i,\theta,\ell, j}_{h},\bz^{i,\theta,\ell, j}_{h+1}]$ for $i \in \llbracket-K,K \rrbracket , \theta \in \mathbb{S}^{1}(\eta), \ell \in \llbracket 0,M-1\rrbracket ,$ and $h \in \llbracket 1,2^{(j-J_1){{\mathfrak{m}}}}\rrbracket$. Then then for any $k_1$ (to be specified soon and small enough compared to $J_1$) for any $\theta \in \bS^1(\eta)$ except for at most $O(\frac{k_1}{2^{{{\mathfrak{m}}}}}+\frac{2^{{{\mathfrak{m}}}}\sqrt \delta_3}{\eta})$ fraction, all the remaining $\bz^{i,\theta,\ell, j}_{h}$, are $(\delta',\theta,\frac{n}{2^{j{{\mathfrak{m}}}}}, k_1)-{{\mathbf{Stable}}}$ where $\delta'=O(2^{{\mathfrak{m}}}\delta_3^{1/4}).$
Thus by choosing ${{\mathfrak{m}}}$ large enough followed by $\delta_3$ small enough, provides for any $\theta \in \mathbb{S}^1(\eta)$ a dense set of points at spacing $\frac{n}{2^{j{{\mathfrak{m}}}}}$ which are $(\delta_2,\theta,\frac{n}{2^{j{{\mathfrak{m}}}}},k_1)-{{\mathbf{Stable}}}$ and hence addresses the issue in (a).
To address the issue in (b) we will use the above along with Lemma \[stab23\] to imply stability for most points in ${{\mathbf{L-Box}}}(n)$ with slightly worse parameters. Fixing $\theta \in \bS^1(\eta),$ for any $(\delta_2,\theta,\frac{n}{2^{j{{\mathfrak{m}}}}},k_1)-{{\mathbf{Stable}}}$ $\bz^{i,\theta,\ell, j}_{h}$, consider any lattice point ${{\mathbf{w}}}$ in the associated rectangular box $\mathfrak{R}$ as illustrated in Figure \[fig13\]. Thus $|{{\mathbf{w}}}-\bz^{i,\theta,\ell, j}_{h}|\le 2\frac{n}{2^{j{{\mathfrak{m}}}}}$. Hence applying Lemma \[stab23\] (by taking $\ell=\frac{n}{2^{j{{\mathfrak{m}}}}}$, $m=2$, $k=k_1$ and $C=\sqrt{k_1}$) now implies that: $$\#\{\bz\in {{\mathbf{L-Box}}}(n): \bz~\text{is not}~(\delta', \theta,\frac{n\sqrt k_1}{2^{j{{\mathfrak{m}}}}},\sqrt k_1)-{{\mathbf{Stable}}}\} \leq O(\frac{k_1}{2^{{{\mathfrak{m}}}}}+\frac{2^{{{\mathfrak{m}}}}\sqrt \delta_3}{\eta})n^2 ,$$ where $\delta'=\delta_2+O(\frac{1}{\sqrt k_1}).$ By a simple union bound over $\theta \in \mathbb{S}^{1}(\eta)$ it follows that $$\label{lb123456}\#\{\bz\in {{\mathbf{L-Box}}}(n): \bz~\text{is not}~(\delta',\mathbb{S}^1(\eta), \frac{n\sqrt k_1}{2^{j{{\mathfrak{m}}}}},\sqrt k_1)-{{\mathbf{Stable}}}\} \leq O\left(\frac{1}{\eta}(\frac{k_1}{2^{{{\mathfrak{m}}}}}+\frac{2^{{{\mathfrak{m}}}}\sqrt \delta_3}{\eta})\right) n^2 .$$
The statement of Theorem \[t:stablegen\] now follows from choosing $\sqrt{k_1}\gtrsim \max(\frac{1}{\delta},k)$ followed by $\delta_{2}$ small enough to ensure $\delta'\le \delta.$ and then ${{\mathfrak{m}}}$ large enough followed by $\delta_3$ small enough to ensure that $O\left(\frac{1}{\eta}(\frac{k_1}{2^{{{\mathfrak{m}}}}}+\frac{2^{{{\mathfrak{m}}}}\sqrt \delta_3}{\eta})\right)$ is less than ${\varepsilon}.$ Moreover we take the value of $J_2$ to be $\mathfrak{m}(J_1+\frac{1}{\delta_3^2})$. Note that the value of $j$ in Theorem \[t:stablegen\] is the value $j{{\mathfrak{m}}}-\frac{\log k_1}{2}$ appearing in .
[^1]: [For brevity of notation we shall often denote $\ell(\gamma)$ by $|\gamma|$]{}.
[^2]: Although the domain of ${{\mathbf{Proj}}}$ is determined completely by $\ell_1$ in practice we shall mostly apply this function on pairs of points in ${{\mathbf{Grid}}}_n(\ell_1;j_1)$, hence we chose to keep both parameters $\ell_1$ and $j_1$ while specifying ${{\mathbf{Proj}}}$.
[^3]: Note that the points $x_{i}^{{{\mathbf{S}}}}$ are points in ${{\mathbf{Grid}}}_{{{\mathfrak{n}}}_4}(\ell_2;j_1)$ and hence as $\Pi$ varies over ${{\mathbf{Base-event}}}$, the length of the path $\alpha^{{{\mathbf{S}}}}_i$ can at worst change by a multiplicative factor $(1+\eta_1)$ where $\eta_1$ appears in the definition of ${{\mathbf{Base-event}}}$.
[^4]: Note that to be completely precise there is an edge joining $\alpha_i^{{{\mathbf{S}}}}$ and $\alpha_{i+1}^{{{\mathbf{S}}}}$ which we are ignoring in for brevity since it is easily seen that such edges only have a negligible contribution.
[^5]: This is not essential for the proof but is done for convenience. Note that if this assumption is not satisfied this can always be achieved by chopping our vectors $\bo{c}_{\phi_j}$ in to smaller vectors and rearranging the order of the sum $\sum_{j=1}^{i-1}\bo{c}_{\phi_j}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study magnetotransport properties of the electron-doped superconductor [Pr$_{2-x}$Ce$_x$CuO$_{4\pm\delta}$]{} with $x$ = 0.14 in magnetic fields up to 92 T, and observe Shubnikov de-Haas magnetic quantum oscillations. The oscillations display a single frequency $F$=255$\pm$10 T, indicating a small Fermi pocket that is $\sim$ 1% of the two-dimensional Brillouin zone and consistent with a Fermi surface reconstructed from the large hole-like cylinder predicted for these layered materials. Despite the low nominal doping, all electronic properties including the effective mass and Hall effect are consistent with overdoped compounds. Our study demonstrates that the exceptional chemical control afforded by high quality thin films will enable Fermi surface studies deep into the overdoped cuprate phase diagram.'
author:
- 'Nicholas P. Breznay\*'
- 'Ian M. Hayes'
- 'B. J. Ramshaw'
- 'Ross D. McDonald'
- Yoshiharu Krockenberger
- Ai Ikeda
- Hiroshi Irie
- Hideki Yamamoto
- 'James G. Analytis'
title: 'Shubnikov-de Haas quantum oscilations reveal a reconstructed Fermi surface near optimal doping in a thin film of the cuprate superconductor [Pr$_{1.86}$Ce$_{0.14}$CuO$_{4\pm\delta}$]{}'
---
Understanding the ordering phenomena that compete or coexist with superconductivity in the cuprate superconductors remains an outstanding challenge. Central to this effort is identification of Fermi surface (FS) topology and evolution with doping via studies of magnetic quantum oscillations (QO), led by initial observation of QO in hole-doped YBa$_2$Cu$_3$O$_{6.5}$[@doiron2007; @leboeuf2007]. In hole-doped cuprates, QO studies have shown (1) a large cylindrical hole-like FS consistent with band theory in overdoped Tl$_2$Ba$_2$CuO$_{6+\delta}$ [@vignolle2008], (2) FS reconstruction and a complex topology in underdoped YBa$_2$Cu$_3$O$_{6+\delta}$ [@sebastian2014], and (3) strong enhancements in the quasiparticle effective mass [@ramshaw2015]; see Refs. for further reviews. These observations have been interpreted as evidence for competing electronic ordered phases and the influence of quantum critical fluctuations, crossing over to Fermiology consistent with the band-theory picture. Comparable experimental studies of electron-doped cuprates are limited, and precise description of their FS remains an outstanding challenge. Here we report on the FS topology and effective mass in the electron-doped cuprate [Pr$_{2-x}$Ce$_x$CuO$_{4\pm\delta}$]{} (PCCO) with $x=0.14$. We observe Shubnikov-de Haas QO in a PCCO thin film measured in extreme magnetic fields up to 92 T, where magnetotransport data (Fig. \[f:mr\]) show evidence for a small (255 T) FS pocket, a light quasiparticle effect mass $m/m_e = 0.43$, and direct determination of the orbitally averaged Fermi velocity $v_F = 2.4 \times 10^5$ m/s.
![(Color online) Low-temperature magnetoresistance of a superconducting [Pr$_{2-x}$Ce$_x$CuO$_{4\pm\delta}$]{} thin film (pictured) measured to 92 T at temperatures between 2 K and 30 K. Inset (right scale): sample magnetoresistance with a smooth background subtracted, showing magnetic oscillations that are suppressed with increasing temperature.[]{data-label="f:mr"}](PCCO-MR){width="0.9\columnwidth"}
Electron-doped cuprates *R*$_{2-x}$Ce$_x$CuO$_{4\pm\delta}$ (with *R* = La, Nd, Pr, ...) have a square planar $T'$ structure with two CuO$_2$ layers within the body-centered tetragonal unit cell. Bandstructure calculations predict a large, hole-like FS cylinder centered at $(\pi,\pi)$ arising from the CuO$_2$ planes [@massidda_electronic_1989]. (See Ref. for a recent review.) This cylinder (discussed below) should contain $n = 1 - x$ carriers assuming the electron dopant concentration is equal to the Ce content $x$ in [Pr$_{2-x}$Ce$_x$CuO$_{4\pm\delta}$]{}. Photoemission (ARPES) experiments are consistent with a large FS in overdoped $n$-doped materials [@armitage2002; @matsui2007; @song2012], and show evidence for both electron- and hole-like pockets as $x$ decreases below $\sim 0.16$. However, many issues remain unresolved, including the structure of the FS and nature of its reconstruction and evolution with doping, reconciling quantum oscillation measurements with the observation of Fermi arcs [@shen_nodal_2005; @reber_origin_2012] in the pseudogap phase, and identifying the true nature of the competing ground state (whether arising from observed magnetic [@motoyama_spin_2007; @saadaoui2015] charge [@dasilvaneto2015], or predicted d-density wave (DDW) [@chakravarty_hidden_2001]) order). QO have only been observed in a restricted doping range ($x$ = 0.15 - 0.17) in $n$-doped bulk crystals of [Nd$_{2-x}$Ce$_x$CuO$_4$]{} (NCCO) [@helm2009; @helm2010; @helm2015], and found to be consistent with a theoretical picture of DDW-like order [@eun2010j; @eun2011c]. Resolving QO in Ce-doped PCCO thin films opens an exciting avenue for continuously studying FS evolution, in particular because thin films allow doping levels beyond the $x$ = 0.17 solubility limit of conventional bulk synthesis techniques.
We study a superconducting PCCO film (thickness 100 nm) that was grown using molecular beam epitaxy and characterized as described in detail elsewhere [@krockenberger2012]. Cerium content $x = 0.14$ is controlled to within 1 % via an in-situ quartz-crystal thickness monitor and ICP spectroscopy. Hall-bar devices with active area 200 $\times$ 350 $\mu m^2$ were defined using conventional photolithography techniques, as shown in the inset of Fig. \[f:mr\]. We measured in-plane resistivity $\rho_{xx}$ and Hall effect $\rho_{yx}$ using standard four-point lock-in configurations in DC magnetic fields (to 12 T); applied fields were parallel to the crystallographic $c$-axis. Figure \[f:rvst\] shows the temperature dependent resistivity $\rho_{xx}$, Hall coefficient $R_H \equiv \rho_{yx}/B$, and mobility $\mu$ for this [Pr$_{1.86}$Ce$_{0.14}$CuO$_{4\pm\delta}$]{} film; while $\rho_{yx}$(B) is linear in this field range, high field measurements have shown nonlinearity and evidence for multi-band behavior [@li2007]. Film parameters are summarized in Tab. \[tab:params\], below. The film shows a sharp superconducting transition with $T_c$ = 22 K and a zero-temperature upper critical field $H_{c2} = 5$ T. The residual resistance ratio, defined as $\rho(300 K) / \rho(30 K)$ = 10.2, while the residual resistance is $\approx$15 $\mu\Omega$ cm. The carrier mobility estimated both from the Hall effect in a single-band picture $\mu_H = R_H / \rho_{xx}$, and parabolic component of the magnetoresistance ($\Delta R / R \sim (\mu_{\textrm{MR}} B)^2$), is $\approx 0.01$ m$^2$/Vs; this is in good agreement with the mobility extracted from QO analyses (see below) and indicates that the strong-field limit of $\omega_c \tau \sim$ 1 can be accessed in applied fields $\sim 100$ T. (Here $\omega_c = e B/m^*$ is the cyclotron frequency, and $\tau$ the quasiparticle lifetime.)
We measured the high-field magnetoresistance in pulsed magnetic fields to over 92 T at the NHMFL Pulsed Field Facility. Measurements were carried out with the sample in He-4 and He-3 liquid (2 - 4 K) or gas (5 - 30 K), and care was taken to minimize heating effects during field pulses. The data reported here are from the rising field portion of the pulsed measurements; some hysteresis was observed and is reflected in the quoted uncertainties, but the QO analyses are consistent between rising and falling field sweeps. Figure \[f:mr\] shows the central result of this work: [Pr$_{1.86}$Ce$_{0.14}$CuO$_{4\pm\delta}$]{} film magnetoresistance as a function of applied magnetic fields to above 92 T, at temperatures between 2 K and 30 K. In sufficiently strong magnetic fields, the separation between quasiparticle Landau levels can be greater than their lifetime $\tau$ and thermal broadening. With changing field, the resulting oscillations of the density of states lead to Shubnikov de-Haas oscillations in the conductivity, visible in the magnetoresistance data after a smooth background has been subtracted (Fig. \[f:mr\], inset) with an amplitude that increases with increasing field and decreasing temperature. The appearance and nature of these QO allow for direct observation and study of FS properties.
![(Color online) Resistivity and magnetotransport in a $x$ = 0.14 [Pr$_{2-x}$Ce$_x$CuO$_{4\pm\delta}$]{} thin film. (A) $\rho$-T curve; the sample shows a superconducting transition at $T_c$ = 22 K and a residual resistivity ratio $\equiv \rho(300K)/\rho(30K)$ = 10.2. (B) Hall coefficient $R_H = \rho_{yx} / B$ as a function of temperature. (C) Hall ($\mu_{H}$), magnetoresistance ($\mu_{\textrm{MR}}$), and Dingle ($\mu_{D}$) mobilities calculated as described in the text. (D) $\rho$(B) magnetoresistance traces measured in DC fields at temperatures between 1.8 and 30 K.[]{data-label="f:rvst"}](PCCO-RxxRxy){width="1.0\columnwidth"}
![(Color online) (A) After subtracting a polynomial background, magnetic quantum oscillations periodic in inverse field are visible below 0.02 T$^{-1}$, with period 1/255 T$^{-1}$. (Curves have been vertically offset for clarity.) (B) Location of maxima (up triangles) and minima (down triangles) of the QO traces versus inverse field. (C) FFTs of the QO data, showing a single peak near 250 T.[]{data-label="f:invb"}](PCCO-InvB2){width="1.0\columnwidth"}
To analyze the QO visible in the low-temperature magnetoresistance, we fit to and divide by 3rd-order polynomial background $\rho(B) = \rho_{\text{bkgd}}(1 + \textit{f}_{\textrm{QO}}(B))$ to reveal oscillations periodic in inverse field shown in the main panel of Fig. \[f:invb\]. The oscillation frequency $F = 255\pm10$ T is consistent with both Landau level indexing and by computing Fast Fourier transform (FFT) spectra for all temperatures (also plotted in Fig. \[f:invb\]). The frequency $F = \left( \frac{\hbar}{2 \pi e} \right) A_F$ is determined by the extremal FS cross sectional area $A_F$ perpendicular to applied field; here $A_F = 2.4 \times 10^{18}$ m$^{-2}$ or 1.0% of the two-dimensional (2D) Brillouin zone area. In contrast, a band-filling picture using the nominal Ce concentration $x$ predicts a carrier concentration $p = 1 - x \approx$ 0.86 holes per Cu, or 43% of the 2D Brillouin zone. At 2 K $R_H = 0.9 \times 10^{-9}$ $\Omega$m/T; assuming a single parabolic band yields a carrier density per CuO$_2$ layer of $n_{2D}$ = 4.2 $\times 10^{18}$ m$^{-2}$ or 0.66 carriers per Cu. As has been seen for near-optimal doping in both PCCO thin films [@dagan2004; @li2007; @charpentier2010] and bulk crystals of [Nd$_{2-x}$Ce$_x$CuO$_4$]{} [@helm2015], $R_H$ changes sign with temperature, consistent with a multi-band FS and possible onset of competing order. (As $R_H$ also changes sign with $x$ near optimal doping, the simplistic analysis of the Hall effect discussed here will necessarily be incomplete.) Assuming a quasi-2D FS cylinder with parabolic dispersion, the Fermi energy can be directly calculated from $F$ yielding $E_F$ = 24 meV $\sim$ 280 K. Finally, following a model for reconstruction of the large hole-like FS with a ($\pi,\pi$) ordering wavevector [@helm2009], we estimate an energy gap $\Delta \sim $30 meV.
With a circular orbit, $A_F = \pi k_F^2$ and the orbitally averaged Fermi velocity $v_{F} = \hbar k_F / m^* = 2.4 \times 10^{5}$ m/s. Though $v_F$ is somewhat below that determined by ARPES measurements of [Nd$_{2-x}$Ce$_x$CuO$_4$]{} $4.3 \times 10^{5}$ m/s [@armitage2003], it is several times larger than that reported recently in the hole-doped cuprate materials Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$ and YBa$_2$Cu$_3$O$_{6.5}$ near optimal doping, $7.7-8.4\times 10^4$ m/s [@vishik_doping_2010; @jaudet_dehaas_2008]. Interestingly, $v_F$ is within 10% of the speculative “universal” nodal value [@zhou_high_2003] observed in ARPES measurements in many hole-doped cuprate materials across a wide range of doping levels, hinting that these QOs originate from a nodal (hole-like) FS pocket.
The QO evolution with temperature and field are in excellent agreement with fits to the lowest-order Lifshits-Kosevich [@schoenberg] expression for relative change in conductivity $\textit{f}_{\textrm{QO}}(B) \sim \Delta \sigma / \sigma_0$ as a function of temperature T and field B: $$\textit{f}_{\textrm{QO}}(B) = R_D R_T \cos\left( 2\pi F / B\right)
\label{eq:lk}$$ where $R_D = \exp(-\pi / \omega_c \tau_D)$, $\omega_c \equiv e B / m^*$ is the cyclotron frequency, $\tau_D$ is the Dingle lifetime, and $R_T = (2\pi^2 k_B T/\hbar \omega_c) / \sinh\left( 2\pi^2 k_B T/\hbar \omega_c \right)$. We fit the entire data set using a single frequency $F$ = 255 T and lifetime $\tau_D$, plotted as the continuous curves in Fig. \[f:invb\]A.
Quantity Parameter Value Unit
-------------------------------- --------------------- -------------- ------------------------
Ce content x 0.14 -
Residual resistance ratio RRR 10.2 -
Normal state resistivity (2 K) $\rho_{xx}$ 15 $\mu \Omega$ cm
Transition temperature $T_c$ 22 K
Hall coefficient (2 K) $R_H$ 0.9 $10^{-9}$ $\Omega$ m/T
Mobility from ...
- Hall effect (2 K) $\mu_H$ 0.0059 T$^{-1}$
- Magnetoresistance ($<$50 K) $\mu_{\textrm{MR}}$ 0.013 T$^{-1}$
- Dingle formula (2 K) $\mu_D$ 0.014 T$^{-1}$
Dingle temperature (2 K) $T_D$ 44 K
QO frequency $F$ 255$\pm$10 T
QO effective mass $m^*$ 0.43$\pm$.05 $m_e$
Fermi velocity $v_F$ 2.9 $10^5$ m/s
: Properties of the [Pr$_{2-x}$Ce$_x$CuO$_{4\pm\delta}$]{} film studied in this work, obtained by resistivity, Hall effect, and magnetoresistance (MR) measurements and analyses of magnetic quantum oscillations (lower entries).
\[tab:params\]
The large Dingle temperature ($T_D = \hbar / 2 \pi k_B \tau_{D} \approx $ 44 K) and short quasiparticle lifetime $\tau_{D} = 2.8 \times 10^{-14}$ s indicate that QO will only be visible on small FS pockets. The 255 T pocket cyclotron orbit size $\ell_c = 2 k_F/ \hbar e B$ is 12 nm at 100 T, while the mean free path $\ell_o = (\tau_D \hbar/ m^*) \sqrt{A_F/ \pi}$ is 8.2 nm. The cyclotron orbit size for an $\sim 11$ kT orbit, corresponding to either the low-temperature Hall effect in this PCCO film (see discussion below) or the large pockets predicted by band theory [@massidda_electronic_1989; @helm2009] would be a factor of 6-7 larger and not resolvable at 100 T in this film.
![(Color online) Decay of the quantum oscillation amplitude as a function of temperature, evaluated both using the FFT amplitude (circles) and resistance at fixed fields $\Delta R(T)$ (triangles); together they indicate a quasiparticle effective mass $m^* = 0.43 \pm 0.05 m_e$ (continuous and dashed lines).[]{data-label="f:fft"}](PCCO-mStar){width="0.9\columnwidth"}
At fixed magnetic field the decrease in QO oscillation amplitude with increasing temperature is a direct measure of the quasiparticle effective mass $m^*$, a quantity that can be enhanced in proximity to a quantum critical point [@ramshaw2015]. The QO amplitude is plotted versus temperature in Fig. \[f:fft\], along with a fit yielding $m^* = 0.43 \pm 0.05$ m$_e$ consistent with both FFT spectrum amplitudes (red data) and analyses at fixed magnetic field (black data); also shown are $\pm1 \sigma$ error bars to the best fit (dashed lines). Here $F$ is comparable to that observed in NCCO with $x$ = 0.17 ($F_{\textrm{NCCO}} \approx 250$ T) in studies of bulk crystals [@helm2015] that show evidence for a quantum critical point at a doping $x$ = 0.145. $R_H$ is comparable to that seen in $x$ = 0.17 PCCO films ($\approx 0.8 \times 10^{-9}$ $\Omega$-m/T) grown and oxygen reduced using other techniques [@li2007; @gauthier2007]. Differences in $m^*$ and $T_c$ that separate this film and PCCO and NCCO materials near $x$ = 0.17 will require further systematic study; given the typical decrease in charge carrier mobility with increasing $x$, we propose that films with fixed Ce concentration may, via suitable oxygen annealing, be used to continuously study the FS evolution towards overdoping. QCPs have been reported in PCCO near $x$ = 0.165 near the region of AFM and SC coexistence between $x$ = 0.12 and $x$ = 0.15 [@dagan2004; @yu2007; @charpentier2010], as well as in La$_{2-x}$Ce$_x$CuO$_4$ through analysis of scaling phenomena [@jin2011b; @butch2012], and it is likely that variation in nominal Ce concentation associated with this point in the phase diagram derives from differences in the synthesis and oxygen reduction processes.
![(Color online) Schematic hole-like PCCO Fermi surface according to band theory for $x = 0.14$ (left), reconstruction picture at and below optimal doping (center), and the 2D FS areas (right) shown by magnetic quantum oscillation ($A_{F}$) and estimated using Hall effect measurements ($A_{\textrm{Hall}}$) as discussed in the text.[]{data-label="f:fs"}](PCCO-FS){width="1.0\columnwidth"}
The nature of the electronic order driving the reconstructed FS observed here remains to be settled. A comparison of prevailing FS schematics is shown in Fig. \[f:fs\], including the large hole-like cylinder centered at ($\pi,\pi$) at left, and the reconstructed FS consistent with photoemission measurements and displaying two hole-like (red) and one electron-like (blue) pocket around the reduced (antiferromagnetic) Brillouin zone (dashed lines). As in NCCO [@helm2009], the QO frequency seen here is consistent with the hole pockets of the reconstructed FS, and not from the much larger electron pockets at ($\pi$, 0) (middle of Fig. \[f:fs\]). The absence of oscillations from the electron pockets remains a puzzle, especially in NCCO where a breakdown orbit along the full reconstructed FS is reported. The size of the observed pocket seen in these measurements is shown at right in Fig. \[f:fs\] (small region), to scale with the FS schematics. Also shown is an estimated FS area using the 2 K Hall effect according to Luttinger’s theorm (large region). While the Hall effect is likely to be more complex in the presence of multiple bands within a reconstructed FS, ongoing studies of PCCO[@dagan2004; @charpentier2010; @lin_theory_2005] and hole-doped cuprates[@badoux2016] suggest a direct link between the evolution of $R_H$ and the FS.
The small FS pocket indicating a reconstruction in PCCO similar to that seen in NCCO, coupled with similar observations of Ce-free [Pr$_2$CuO$_4$]{} films[@breznay_2015], suggest that small $F\sim$250-350 T FS pockets are a universal feature of superconducting “electron-doped” materials. We cannot rule out the possibility of a field-induced FS reconstruction beginning at or below $\approx$50 T; although convincing evidence now exists for field-induced FS reconstruction arising from charge order in the hole-doped cuprates[@wu_magnetic_2011; @gerber_three_2015; @leboeuf_thermodynamic_2012], the situation in the electron-doped materials[@dasilvaneto2015] remains less clear. The presence of a carrier density consistent with an unreconstructed FS pocket, along with a small pocket showing evidence for a reconstructed and multi-component FS, suggests that continued improvement in thin film materials quality may connect to the the overdoped, Fermi-liquid region of the phase diagram and provide direct insight into the fate of the electron ground state with doping.
We gratefully acknowledge the scientific and support staff of the Los Alamos National High Magnetic Lab Pulsed Field Facility for their technical assistance on this project, in particular the 100 T operations team. We also appreciate fruitful conversations with Toni Helm. This work was supported by the Laboratory Directed Research and Development Program of Lawrence Berkeley National Laboratory under the U.S. Department of Energy Contract No. DE-AC02-05CH11231. A portion of this work was performed at the National High Magnetic Field Laboratory, which is supported by National Science Foundation Cooperative Agreement No. DMR-1157490 and the State of Florida. Portions of this work were also supported by the Gordon and Betty Moore Foundation’s EPiQS Initiative through Grant GBMF4374, and the U.S. Department of Energy Office of Basic Energy Sciences “Science at 100 T” program.
[10]{}
Nicolas Doiron-Leyraud, Cyril Proust, David LeBoeuf, Julien Levallois, Jean-Baptiste Bonnemaison, Ruixing Liang, D A Bonn, W N Hardy, and Louis Taillefer, , 447:565–568, 2007.
David LeBoeuf, Nicolas Doiron-Leyraud, Julien Levallois, R Daou, J-B Bonnemaison, N E Hussey, L Balicas, B J Ramshaw, Ruixing Liang, D A Bonn, W N Hardy, S Adachi, Cyril Proust, and Louis Taillefer, , 450:533–536, 2007.
B. Vignolle, A. Carrington, R. A. Cooper, M. M. J. French, a. P. Mackenzie, C. Jaudet, D. Vignolles, Cyril Proust, and N. E. Hussey, , 455(7215):952–955, 2008.
Suchitra E. Sebastian, N. Harrison, F. F. Balakirev, M. M. Altarawneh, P. A. Goddard, Ruixing Liang, D. A. Bonn, W. N. Hardy, and G. G. Lonzarich, , 511:61–64, 2014.
B. J. Ramshaw, S. E. Sebastian, R. D. McDonald, James Day, B. S. Tan, Z. Zhu, J. B. Betts, Ruixing Liang, D. A. Bonn, W. N. Hardy, and N. Harrison, , 348:317–320, 2015.
Baptiste Vignolle, David Vignolles, David LeBoeuf, Stéphane Lepault, Brad Ramshaw, Ruixing Liang, D. A. Bonn, W. N. Hardy, Nicolas Doiron-Leyraud, A. Carrington, N. E. Hussey, Louis Taillefer, and Cyril Proust, , 12:446–460, 2011.
Suchitra E Sebastian, Neil Harrison, and Gilbert G Lonzarich, , 75:102501, 2012.
S. Massidda, N. Hamada, Jaejun Yu, and A.J. Freeman, , 157:571–574, 1989.
N. P. Armitage, P. Fournier, and R. L. Greene, , 82:2421–2487, Sep 2010.
N. P. Armitage, F. Ronning, D. H. Lu, C. Kim, A. Damascelli, K. M. Shen, D. L. Feng, H. Eisaki, Z.-X. Shen, P. K. Mang, N. Kaneko, M. Greven, Y. Onose, Y. Taguchi, and Y. Tokura, , 88:257001, 2002.
D. Song, S. R. Park, C. Kim, Y. Kim, C. Leem, S. Choi, W. Jung, Y. Koh, G. Han, Y. Yoshida, H. Eisaki, D. H. Lu, Z. X. Shen, and C. Kim, , 86(14):144520, 2012.
H. Matsui, T. Takahashi, T. Sato, K. Terashima, H. Ding, T. Uefuji, and K. Yamada, , 75:224514, 2007.
Shen, Kyle M. and Ronning, F. and Lu, D. H. and Baumberger, F. and Ingle, N. J. C. and Lee, W. S. and Meevasana, W. and Kohsaka, Y. and Azuma, M. and Takano, M. and Takagi, H. and Shen, Z.-X, , 307:901–904, 2005.
T. J. Reber, N. C. Plumb, Z. Sun, Y. Cao, Q. Wang, K. McElroy, H. Iwasawa, M. Arita, J. S. Wen, Z. J. Xu, G. Gu, Y. Yoshida, H. Eisaki, Y. Aiura and D. S. Dessau, , 8:606–610, 2012.
E. M. Motoyama, G. Yu, I. M. Vishik, O. P. Vajk, P. K. Mang, and M. Greven, , 445(7124):186–189, 2007.
H. Saadaoui, Z. Salman, H. Luetkens, T. Prokscha, A. Suter, W. A. MacFarlane, Y. Jiang, K. Jin, R. L. Greene, E. Morenzoni, and R. F. Kiefl, , 6:6041, 2015.
E. H. [da Silva Neto]{}, R. Comin, F. He, R. Sutarto, Y. Jiang, R. L. Greene, G. A. Sawatzky, and A. Damascelli, , 347:282–285, 2015.
Sudip Chakravarty, R. B. Laughlin, Dirk K. Morr, and Chetan Nayak, , 63:094503, 2001.
T. Helm, M. V. Kartsovnik, M. Bartkowiak, N. Bittner, M. Lambacher, A. Erb, J. Wosnitza, and R. Gross, , 103:157002, 2009.
T. Helm, M. V. Kartsovnik, I. Sheikin, M. Bartkowiak, F. Wolff-Fabris, N. Bittner, W. Biberacher, M. Lambacher, A. Erb, J. Wosnitza, and R. Gross, , 105:247002, 2010.
T. Helm, M. V. Kartsovnik, C. Proust, B. Vignolle, C. Putzke, E. Kampert, I. Sheikin, E.-S. Choi, J. S. Brooks, N. Bittner, W. Biberacher, A. Erb, J. Wosnitza, and R. Gross, , 92:094501, 2015.
J. Eun , X. Jia, and S. Chakravarty, , 82:094515, 2010.
J. Eun, and S. Chakravarty, , 84:094506, 2011.
Y. Krockenberger, H. Yamamoto, A. Tsukada, M. Mitsuhashi, and M. Naito, , 85:184502, May 2012.
Y. Dagan, M. M. Qazilbash, C. P. Hill, V. N. Kulkarni, and R. L. Greene, , 92:167001, Apr 2004.
Pengcheng Li, F. F. Balakirev, and R. L. Greene, , 99:047003, Jul 2007.
S. Charpentier, G. Roberge, S. Godin-Proulx, X. Béchamp-Laganière, K. D. Truong, P. Fournier, and P. Rauwel, , 81:104509, Mar 2010.
N. P. Armitage, D. H. Lu, C. Kim, A. Damascelli, K. M. Shen, F. Ronning, D. L. Feng, P. Bogdanov, X. J. Zhou, W. L. Yang, Z. Hussain, P. K. Mang, N. Kaneko, M. Greven, Y. Onose, Y. Taguchi, Y. Tokura, and Z.-X. Shen, , 68:064517, Aug 2003.
I. M. Vishik, W. S. Lee, F. Schmitt, B. Moritz, T. Sasagawa, S. Uchida, K. Fujita, S. Ishida, C. Zhang, T. P. Devereaux, and Z. X. Shen, , 104:207002, May 2010.
Cyril Jaudet, David Vignolles, Alain Audouard, Julien Levallois, D. LeBoeuf, Nicolas Doiron-Leyraud, B. Vignolle, M. Nardone, A. Zitouni, Ruixing Liang, D. A. Bonn, W. N. Hardy, Louis Taillefer, and Cyril Proust, , 100:187005, May 2008.
X. J. Zhou, T. Yoshida, A. Lanzara, P. V. Bogdanov, S. A. Kellar, K. M. Shen, W. L. Yang, F. Ronning, T. Sasagawa, T. Kakeshita, T. Noda, H. Eisaki, S. Uchida, C. T. Lin, F. Zhou, J. W. Xiong, W. X. Ti, Z. X. Zhao, A. Fujimori, Z. Hussain, and Z.-X. Shen, , 423:398, 2003.
D. Shoenberg, . Cambridge Univ. Press, Cambridge, UK, 1984.
J. Gauthier, S. Gagné, J. Renaud, M.-È. Gosselin, P. Fournier, and P. Richard, , 75:024424, Jan 2007.
W. Yu, B. Liang, P. Li, S. Fujino, T. Murakami, I. Takeuchi, and R. L. Greene, , 75:020503, Jan 2007.
K. Jin, N. P. Butch, K. Kirshenbaum, J. Paglione, and R. L. Greene, , 476:73–75, 2011.
Nicholas P. Butch, Kui Jin, Kevin Kirshenbaum, Richard L. Greene, and Johnpierre Paglione, , 109(22):8440–8444, 2012.
Jie Lin and A. J. Millis, , 72:214506, 2005.
S. Badoux, W. Tabis, F. Laliberté, G. Grissonnanche, B. Vignolle, D. Vignolles, J. Béard, D. A. Bonn, W. N. Hardy, R. Liang, N. Doiron-Leyraud, Louis Taillefer, and Cyril Proust, [*Nature*]{}, 531:210–214, 2016.
Tao Wu, Hadrien Mayaffre, Steffen Krämer, Mladen Horvatic, Claude Berthier, W. N. Hardy, Ruixing Liang, D. A. Bonn and Marc-Henri Julien, , 477:191–194, 2011.
S. Gerber, H. Jang, H. Nojiri, S. Matsuzawa, H. Yasumura, D. A. Bonn, R. Liang, W. N. Hardy, Z. Islam, A. Mehta, S. Song, M. Sikorski, D. Stefanescu, Y. Feng, S. A. Kivelson, T. P. Devereaux, Z.-X. Shen, C.-C. Kao, W. S. Lee, D. Zhu and J.-S. Lee, , 350:949–952, 2015.
David LeBoeuf, S. Krämer, W. N. Hardy, Ruixing Liang, D. A. Bonn and Cyril Proust. , 9:79–83, 2013.
Nicholas P. Breznay, Ross D. McDonald, Yoshiharu Krockenberger, K. A. Modic, Zengwei Zhu, Ian M. Hayes, Nityan L. Nair, Toni Helm, Hiroshi Irie, Hideki Yamamoto, and James G. Analytis, arXiv:1510.04268.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
While it seems possible that quantum computers may allow for algorithms offering a computational speed-up over classical algorithms for some problems, the issue is poorly understood. We explore this computational speed-up by investigating the ability to de-quantise quantum algorithms into classical simulations of the algorithms which are as efficient in both time and space as the original quantum algorithms.
The process of de-quantisation helps formulate conditions to determine if a quantum algorithm provides a real speed-up over classical algorithms. These conditions can be used to develop new quantum algorithms more effectively (by avoiding features that could allow the algorithm to be efficiently classically simulated), as well as providing the potential to create new classical algorithms (by using features which have proved valuable for quantum algorithms).
Results on many different methods of de-quantisations are presented, as well as a general formal definition of de-quantisation. De-quantisations employing higher-dimensional classical bits, as well as those using matrix-simulations, put emphasis on entanglement in quantum algorithms; a key result is that any algorithm in which the entanglement is bounded is de-quantisable. These methods are contrasted with the stabiliser formalism de-quantisations due to the Gottesman-Knill Theorem, as well as those which take advantage of the topology of the circuit for a quantum algorithm.
The benefits of the different methods are contrasted, and the importance of a range of techniques is emphasised. We further discuss some features of quantum algorithms which current de-quantisation methods do not cover.
author:
- 'Alastair A. Abbott and Cristian S. Calude\'
bibliography:
- 'alastairBib.bib'
title: 'Understanding the Quantum Computational Speed-up via De-quantisation'
---
Introduction
============
Since Feynman first introduced the concept of a quantum computer [@Feynman:1982aa] and noted the apparent exponential cost to simulate general quantum systems with classical computers there has been much interest in the power of quantum computation, in particular the possibility of using quantum physics to develop algorithms which are more efficient than classical ones. Many quantum algorithms (e.g. Deutsch’s algorithm) have been claimed to be superior to any classical one solving the same problem, only to be discovered later that this was not the case. In order to construct good quantum algorithms it is important to know what features are necessary for a quantum algorithm to be better than a classical one. Many quantum algorithms have a trivial classical counterpart: with care, all the operations in the matrix mechanical formulation of quantum mechanics can be computed by classical means [@Ekert:1998aa]. In this paper we review the ability to *de-quantise* a quantum algorithm to obtain a classical algorithm which is not exponentially slower in time (or larger in space) compared to the quantum algorithm, and explore when such a de-quantisation is possible.
A Preliminary Example
=====================
The Deutsch-Jozsa Problem {#sec:Deutsch-Jozsa}
-------------------------
The standard formulation of the Deutsch-Jozsa problem is as follows. Let $f : \{0,1\}^n \to \{0,1\}$, and suppose we are given a black-box computing $f$ with the guarantee that $f$ is either constant (i.e. for all $x \in \{0,1\}^n$ and some $a \in \{0,1\}: f(x) = a$) or balanced (i.e. $f(x) = 0$ for exactly half of the possible inputs $x \in \{0,1\}^n$). Such a function $f$ is called *valid*. The Deutsch-Jozsa problem is to determine if $f$ is constant or balanced in as few black-box calls as possible. A typical classical algorithm would require $2^{n-1}$ black-box calls, while the quantum solution requires only one.
The special case of $n=1$ was first considered by [@Deutsch:1985aa] and is called the Deutsch problem; this was de-quantised by Calude [@Calude:2007aa].[^1]
It is important to note that unlike Deutsch’s problem, where there are exactly two balanced and two constant functions $f$, the distribution of constant and balanced functions is asymmetrical in the Deutsch-Jozsa problem. In general, there are $N=2^n$ possible input strings, each with two possible outputs ($0$ or $1$). Hence, for any given $n$ there are $2^N$ possible functions $f$. In this finite class, exactly two functions are constant and $\binom{N}{N/2}$ are balanced. Evidently the probability that a valid function $f: \{0,1\}^n \to \{0,1\}$ is constant tends towards zero very quickly (recall that in Deutsch-Jozsa problem, $f$ is guaranteed to be valid). Furthermore, the probability that any randomly chosen function of the $2^N$ possible functions is valid is $(\binom{N}{N/2}+2)\cdot 2^{-N}$, which again tends to zero as $n$ increases. This is clearly not an ideal problem to work with, however even in this case we can gain, via de-quantisation, useful information.
### Quantum Solution
The quantum black-box we are given takes as input three qubits and is represented by the following unitary operator $U_f$, just as it was for $n=1$: $$U_f \ket{x} \ket{y} = \ket{x} \ket{y \oplus f(x)},$$ where $x \in \{0,1\}^2$. There are sixteen possible Boolean functions. Two of these are constant, another six are balanced and the remaining eight are not valid. All these possible functions are listed in Table \[table:bFuncs\].
$f(x)$
------------ --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- ---
$f(00) = $ 0 1 0 1 0 1 1 0 1 0 1 0 1 0 0 1
$f(01) = $ 0 1 0 1 1 0 0 1 1 0 1 0 0 1 1 0
$f(10) = $ 0 1 1 0 1 0 1 0 1 0 0 1 1 0 1 0
$f(11) = $ 0 1 1 0 0 1 0 1 0 1 1 0 1 0 1 0
: All possible Boolean functions $f : \{0,1\}^2 \to \{0,1\}$.[]{data-label="table:bFuncs"}
Evidently, half of these functions are simply the negation of another. If we let $f'(x) = f(x) \oplus 1$ and define $\ket{\pm} = {\frac{1}{\sqrt{2}}}(\ket{0} \pm \ket{1})$, we have: 01 $$\begin{aligned}
U_{f'}\ket{x}\ket{-} &= (-1)^{f'(x)}\ket{x}\ket{-}\\
&= -\left((-1)^{f(x)}\ket{x}\ket{-} \right)\\
&= - U_f \ket{x}\ket{-}.
$$ $$U_{f'}\ket{x}\ket{-} = (-1)^{f'(x)}\ket{x}\ket{-}\\
= -\left((-1)^{f(x)}\ket{x}\ket{-} \right)\\
= - U_f \ket{x}\ket{-}.$$ In this case the result obtains a global phase factor of $-1$. Since global phase factors have no physical significance to measurement (a result is obtained with probability proportional to the amplitude squared), the outputs of $U_f$ and $U_{f'}$ are physically indistinguishable.
We will present a revised form of the standard quantum solution in which we emphasise separability of the output state. We initially prepare our system in the state $\ket{00}\ket{1}$, and then operate on it with $H^{\otimes 3}$ to get: $$\label{eqn:equalSuperpos}
H^{\otimes 3}\ket{00}\ket{1} = \frac{1}{2}\sum_{x\in \{0,1\}^2}\ket{x}\ket{-} = \ket{++}\ket{-}.$$ In the general case, after applying the $f$-cNOT gate $U_f$ we have $$\label{eqn:n2fcNot}
\begin{split}
U_f \sum_{x\in \{0,1\}^2}c_x\ket{x}\ket{-} & = \left[ (-1)^{f(00)}c_{00}\ket{00} + (-1)^{f(01)}c_{01}\ket{01} + (-1)^{f(10)}c_{10}\ket{10}\right.
\\ & \left. \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad\qquad + (-1)^{f(11)}c_{11}\ket{11} \right]\ket{-}.
\end{split}$$ From the well known rule (see e.g [@Jorrand:2003aa]) for the separability of 2-qubit states, we know that this state is separable if and only if $$(-1)^{f(00)}(-1)^{f(11)}c_{00}c_{11} = (-1)^{f(01)}(-1)^{f(10)}c_{01}c_{10}.$$ While there are various initial superpositions of 2-qubit states which satisfy this condition, we only need to consider the equal superposition (shown in Equation \[eqn:equalSuperpos\]) that is used in this algorithm. The situation is further simplified by noting that the mapping $$(-1)^{f(a)}(-1)^{f(b)} \leftrightarrow f(a) \oplus f(b)$$ is a bijection. In this case, the separability condition reduces to $f(00) \oplus f(11) = f(01) \oplus f(10)$. By looking back at Table \[table:bFuncs\] it is clear this condition holds for all balanced or constant functions $f$ for $n=2$.
We can now rewrite Equation \[eqn:n2fcNot\] as follows: $$\label{eqn:n2Uf}
U_f \ket{++} \ket{-} = \frac{\pm 1}{2} \left( \ket{0} + (-1)^{f(00) \oplus f(10)} \ket{1} \right) \left( \ket{0} + (-1)^{f(10) \oplus f(11)} \ket{1} \right) \ket{-}.$$
01 Indeed, $$\begin{aligned}
& \quad (-1)^{f(00)}\ket{00} + (-1)^{f(01)}\ket{01} + (-1)^{f(10)}\ket{10} + (-1)^{f(11)}\ket{11} \\
&= (-1)^{f(00)}\ket{00} + (-1)^{f(00)\oplus f(10) \oplus f(11))}\ket{01} + (-1)^{f(10)}\ket{10} + (-1)^{f(11)}\ket{11} \\
&= (-1)^{f(00)}\left( \ket{00} + (-1)^{f(10)\oplus f(11)}\ket{01} + (-1)^{f(00)\oplus f(10)}\ket{10} + (-1)^{f(00)\oplus f(11)}\ket{11} \right)\\
&= \pm \left( \ket{0} + (-1)^{f(00) \oplus f(10)} \ket{1} \right) \left( \ket{0} + (-1)^{f(10) \oplus f(11)} \ket{1} \right),
$$ as desired.
By applying a final 3-qubit Hadamard gate to project this state onto the computational basis we obtain $$\begin{gathered}
$$\frac{\pm 1}{2} H^{\otimes 3} \left( \ket{0} + (-1)^{f(00) \oplus f(10)} \ket{1} \right) \left( \ket{0} + (-1)^{f(10) \oplus f(11)} \ket{1} \right) \ket{-} = \pm \ket{f(00) \oplus f(10)} \\ \otimes \ket{f(10) \oplus f(11)} \ket{1}.$$\end{gathered}$$ By measuring both the first and second qubits we can determine the nature of $f$: if both qubits are measured as $0$, then $f$ is constant, otherwise $f$ is balanced. This is correct with probability-one.
### De-quantising the Quantum Solution {#dqqs}
Because the quantum solution contains no entanglement, the problem can be de-quantised by embedding classical bits in complex numbers [@Calude:2007aa; @Abbott:2009Dissertation]. The set $ \{1,i=\sqrt{-1}\}$ acts as a computational basis in the same way that $\{\ket{0},\ket{1}\}$ does for quantum computation.[^2] A complex number may be written as $z=a+bi$, so $z$ is a natural superposition of the basis in the same way that a qubit is.
01 We are now given a classical black-box that computes our function $f$. In order to measure the output, we need a way to project our complex numbers back on to the computational basis. This is easily done by multiplying by the input so the output is either purely imaginary or purely real.
If $z=1+i$ (an equal superposition of basis states), $$\frac{1}{2}z \times C_f(z) =
\begin{cases}
\frac{\pm 1}{2}z^2 =\pm i & \text{if $f$ is constant,}\\
\frac{\pm 1}{2}z \overline{z} = \pm 1 & \text{if $f$ is balanced.}\\
\end{cases}$$ In this manner, if the output is imaginary then $f$ is constant, if it is real then $f$ is balanced. Importantly, this is a deterministic result, and in fact the sign of the output allows us to identify *which* balanced or constant function $f$ is. This is something the quantum algorithm provably cannot do [@Mermin:2007aa].
We are now given a classical black-box that computes the function $f$. Similarly to $U_f$, the black-box operates on two complex numbers, $C_f : {{\mathbb{C}}}^2 \to {{\mathbb{C}}}^2$. Let $z_1$, $z_2 $ be complex numbers, $$\label{deQuant:eqn:2bitCf}
C_f{\begin{pmatrix}z_1\\#2\end{pmatrix}} = C_f{\begin{pmatrix}a_1+b_1i\\#2\end{pmatrix}} = (-1)^{f(00)} {\begin{pmatrix}a_1 + (-1)^{f(00) \oplus f(10)}b_1i\\#2\end{pmatrix}}\raisebox{.7mm}{.}$$ Just as in the quantum case where the output of the black-box was two qubits that can be independently measured, the output of $C_f$ is two complex numbers that can be independently manipulated, rather than the complex number resulting from their product. Note, however, that in a quantum system it is impossible to measure entangled qubits independently of each other.
To simulate a Hadamard gate we multiply each of the complex numbers that the black-box outputs by their respective inputs. If we let $z_1 = z_2 = 1+i$, we obtain the following: $$\frac{(1+i)}{2}\times C_f {\begin{pmatrix}z_1\\#2\end{pmatrix}} = \frac{(-1)^{f(00)}}{2}\times
\begin{cases}
{\begin{pmatrix}(1+i)(1+i)\\#2\end{pmatrix}} = {\begin{pmatrix}i\\#2\end{pmatrix}} & \\ &\text{if $f$ is constant,}\\[-2ex]
{\begin{pmatrix}(1+i)(1-i)\\#2\end{pmatrix}} = {\begin{pmatrix}1\\#2\end{pmatrix}}\\ & \\[-2ex]
{\begin{pmatrix}(1+i)(1+i)\\#2\end{pmatrix}} = {\begin{pmatrix}i\\#2\end{pmatrix}} \\ & \text{if $f$ is balanced.}\\[-2ex]
{\begin{pmatrix}(1+i)(1-i)\\#2\end{pmatrix}} = {\begin{pmatrix}1\\#2\end{pmatrix}} &
\end{cases}$$
By measuring both resulting complex numbers, we can determine whether $f$ is balanced or constant with [*certainty*]{}. If both complex numbers are imaginary then $f$ is constant, otherwise it is balanced. In fact, the ability to determine if the output bits are negative or positive allows us to determine the value of $f(00)$ and thus which Boolean function $f$ is.
Because the quantum solution is *separable*, it is possible to write the output of the black-box as a list of two complex numbers, and hence we can find a solution equivalent to the one obtained via a quantum computation. Writing the output in this form would not have been possible if the state was not separable, and finding a classical solution in this fashion would have required a list of complex numbers exponential in the number of input qubits. Interestingly, this de-quantisation is equivalent [@Abbott:2009aa] to the ‘physical de-quantisation’ using classical photon polarisations described by Arvind in [@Arvind:2001aa].
### Implementing the De-quantised Solution {#implement}
01 An alternative classical approach can be presented using two photons. If a transformation on two qubits can be written as a transformation on each qubit independently (e.g. $H\otimes H$) then the transformation is trivially implemented classically. It only remains to show that the 2-qubit transformation $U_f$ can be implemented classically on two photons. Equation \[eqn:n2Uf\] shows that the quantum black-box $U_f$ can be written as a product of two 1-qubit gates:[^3] $$\begin{aligned}
U_f^{(1)} \ket{+} = \frac{1}{\sqrt{2}}\left(\ket{0} + (-1)^{f(00)\oplus f(10)}\ket{1}\right),\\
U_f^{(2)} \ket{+} = \frac{1}{\sqrt{2}}\left(\ket{0} + (-1)^{f(10)\oplus f(11)}\ket{1}\right).
$$ Each of these are valid unitary operators, and the transformation describing the black-box may be written $U_f = U_f^{(1)}\otimes U_f^{(2)}$. This means that the operation of $f$ can be computed by applying a 1-qubit operation (implemented as wave-plates) to each photon independently, and thus a classical solution is easily found. The photons need not interact with each other at any point during the algorithm, not even inside the black-box implementation.
This classical, optical method is equivalent to both the quantum solution and the previously described classical solution. The difference is in how it is represented, bringing emphasis on the fact that the quantum solution does not take advantage of uniquely quantum behaviour and is thus classical in nature. Further, it shows that the solution can be obtained without any interaction or sharing of information between qubits.
It is only natural to ask the question: how efficiently can we physically implement the de-quantised solution presented in Section \[dqqs\]? There are different possible approaches, but we will discuss only one.
Nuclear magnetic resonance (NMR) spectroscopy [@Levitt:2008fk] exploits the spin dynamics of nuclear spin systems; it involves placing a sample in a strong external magnetic field and hence a splitting of the nuclear spin energy levels (Zeeman splitting). The corresponding resonance frequencies (Larmor frequencies) are typically of the order of hundreds of MHz. NMR spectroscopy is a rich source of information as spectra reflect interactions of nuclear spins with their electronic environment as well as interactions (couplings) between nuclear spins themselves.
In particular, solution-state NMR has been extensively examined as a possible implementation platform for quantum computations. This approach relies on couplings between spins within molecules and the manipulation of such finite-sized spin systems with appropriate pulse sequences. For example, Shor’s algorithm has been successfully implemented in a 7-qubit NMR quantum computer [@Vandersypen:2001aa].
A new approach proposed in [@Rosello:2009] uses NMR as a classical computing substrate, where interactions between spins play no role and where the dynamics of these isolated spins can be fully described by a classical vector model. The technical difficulties of instability and decoherence present in quantum computation with NMR are less of an issue in this classical approach as their major source (internuclear couplings) is absent. Three different implementations have been demonstrated to simulate logic gates and other more complicated classical circuits. By making suitable choices of input and output parameters from the parameter space describing the NMR experiment, one can achieve different types of classical computations. The available parallelism, stability and ease in implementing two-dimensional classical bits (e.g. based on the three-dimensional vector model, or using two different spin species) makes NMR a well-suited substrate for implementations of de-quantised solutions of quantum algorithms. Work in progress of the groups in York (UK) and Auckland (NZ) involves NMR implementations of the de-quantised algorithms for the Deutsch-Josza problem described in Section \[sec:Deutsch-Jozsa\].
Benefits
========
The above example allows us to enumerate a few ‘immediate’ benefits of de-quantisation as well as some long-term possible benefits:
- an example of a problem previously thought to be classically impossible to solve, was solved by ‘de-quantising’ a quantum solution;\
- the solution is not uniform, so not ideal. It seems hard to analyse the complexity (asymptotically) of the de-quantised solution;\
- the de-quantised solution is stronger than the original quantum one: it is deterministic and it can distinguish between functions not only classes (balanced/constant);\
- via de-quantisation, a new classical computational technique was proposed;\
- the lack of entanglement[^4] ‘allowed’ this type of de-quantisation;\
- de-quantisation is not only theoretical: it can lead to efficient implementations.
De-quantisation can be one technique (among others) used to gain a better understanding of complexity in quantum computation, which can help to:
- understand the power and need for quantum computation;\
- more clearly see where quantum speed ups potentially come from;\
- develop new quantum algorithms.
De-quantisation
===============
Until now we have used the term ‘de-quantisation’ in an intuitive sense, so it is time to propose a more formal definition.
In the most general sense, a quantum circuit $C_n$ for a computation operating on an $n$-qubit input can be considered a sequence of gates $G=G_{T(n)}\dots G_1$, where each gate is either a unitary gate chosen from a fixed, finite set of gates $\mathcal{G}$, or a measurement gate. We can define a quantum algorithm in a similarly general sense. A quantum algorithm $\mathcal{A}$ is an infinite, uniformly generated, sequence of quantum circuits $(C_0, C_1, \dots)$. We say the algorithm runs in time $T(n)$ if $C_n$ contains $T(n)$ gates.
Many well known algorithms fall into the class BQP, which where $T(n) = \text{poly}(n)$ and $C_n = M_{T(n)}U_{T(n)-1}\dots U_1$, where the $U_i \in \mathcal{G}$ are unitary and $M_{T(n)}$ is a measurement gate [@Gruska:1999aa]. In other words, measurement is the last step of the algorithm. However, the definition of a quantum computation is more general than this, and any de-quantisation should be equally able to handle intermediate measurements and any other reasonable requirements.
A classical algorithm is a program for a probabilistic Turing machine or any other computationally equivalent model of classical computation. The random access program machine is a particularly useful variation which operates with an infinite set of distinguishable, numbered, but unbounded registers each of which can contain an integer. Such a program has the capability for indirect addressing (i.e.the contents of a register can be used as an address to specify another register), thus allowing for optimisations based on memory indices [@BoolosJeffrey:2007].
A quantum algorithm $(C_0, C_1, \dots)$ running in time $T(n)$, with output probability distribution $\mathcal{P}$, given by a classical Turing machine that computes $C_{n}$ in time $\text{poly}(n)$ can be [*de-quantised*]{} if there is a probabilistic universal Turing machine $U$ such that for every computable real $\gamma >0$ there (effectively) exists a probability distribution $\mathcal{P'}$ with $|\mathcal{P'} - \mathcal{P}|< \gamma$ such that $U$ sampling from $\mathcal{P'}$ runs in time $\text{poly}(T(n), \log (1/\gamma))$.
De-quantisation Techniques
==========================
Entanglement Based Methods
--------------------------
One of the simplest approaches of de-quantisation arises from simulating the matrix-mechanical formalism of the state evolution. While the quantum mechanical state vector for $n$ qubits contains, in general, $2^n$ components, under certain conditions it is possible to find compact representations for the state vector which are polynomial in $n$, and this can lead to de-quantisations.
The simplest such case is, as in the Deutsch-Jozsa example of Section \[sec:Deutsch-Jozsa\], when the state vector remains separable throughout the computation. In these situations, the mathematics of the quantum algorithm can be directly simulated in an efficient manner, because both the state vector and any transformations scale polynomially in the number of qubits $n$, and thus also in classical resources. This type of de-quantisation is simple to understand and implement classically, as mentioned for the Deutsch-Jozsa problem, but is too restrictive since most quantum algorithms make use of entanglement.
However, the conditions requiring separability can be loosened. Jozsa and Linden [@Jozsa:2003aa] and Vidal [@Vidal:2003kx] studied the situation where entanglement is bounded throughout the computation, and the primary result is Theorem \[thm:JozsaLinden\]. It was also noted [@Vidal:2003kx] that these results are applicable to the simulation of continuous time quantum dynamics in some many-body systems.
\[thm:JozsaLinden\] Suppose $\mathcal{A}$ is a polynomial time quantum algorithm with the property that at each step in the computation on an input of $n$-qubits, no more than $p_n$ qubits are entangled. If $p_n$ is $O(\log{n})$, i.e. the entanglement grows no faster than logarithmically in the input size, then the quantum computation is de-quantisable.
This is an important result for de-quantisations, but it is not directly applicable to algorithms such as those which solve the Deutsch-Jozsa or Simon’s [@Simon:1997aa] problems, where the algorithm must make use of a black-box. Since this ‘quantum oracle’ is not usually in $\mathcal{G}$ or efficiently decomposable into gates from $\mathcal{G}$, we further require that the entanglement of the quantum state is bounded both before and after the application of the black-box [@Abbott:2009aa]—this allows the equivalent classical black-box to be represented in an efficient form, preserving the ability to de-quantise.
\[thm:Abbott\] Suppose $\mathcal{A}$ is a black-box algorithm which makes use of a black-box $U_f$. If $\mathcal{A}$ satisfies the conditions for Theorem \[thm:JozsaLinden\] with the gate set $\mathcal{G'}=\mathcal{G} \cup U_f$, then it is de-quantisable.
These results require good quantum algorithms to necessarily utilise unbounded entanglement if they are to have any benefit over classical algorithms, and while this was already suspected by many, the ability to utilise these results to de-quantise known algorithms can lead to surprising classical results. Another example of such an instance is with the quantum Fourier transform (QFT). While it often creates unbounded entanglement, for certain classes of input states this is not the case and the computation remains separable [@Abbott:2010aa]. It is conceivable that in various problems there may be natural constraints which enforce such conditions and allow a simple de-quantisation.
Circuit Topology Methods
------------------------
The study of de-quantising the QFT has led to another class of de-quantisations which, rather than focusing on the mathematical form of the operators and states, exploits various properties of the structure of the quantum circuit for the algorithm. One of the simplest such results is that of Arahanov, Landau and Makowsky [@Aharonov:2007aa]. They show that a slightly modified version of the QFT circuit can be expressed in a form with logarithmic bubblewidth, a visual measure closely related to treewidth.[^5] This leads to a polynomial time classical simulation computing the QFT.
In a similar fashion, both Markov and Shi [@Markov:2008aa] and Jozsa [@Jozsa:2006uq] have explored de-quantisation of circuits by working with tensor networks and treewidth. A tensor network for a circuit associates a tensor with every operator or end of wire in the quantum circuit, and distinct indices are used for different wire segments in the circuit. The network is simulated by contracting tensors together, and results focus around the ability to do so efficiently. While the input state must be separable in order to be simulated, this formalism has the notable advantage that it will work even if entanglement is present in the algorithm. The main result [@Markov:2008aa] is Theorem \[thm:MarkovShi\].
\[thm:MarkovShi\] Quantum circuits with $T$ gates and treewidth $d$ can be simulated in time polynomial in $T$ and exponential in $d$ by the method of tensor contraction for product state inputs. Hence, polynomial size circuits with logarithmic treewidth are de-quantisable for product state inputs.
Jozsa further extended [@Jozsa:2006uq] the set of de-quantisable circuits to those which could be arranged so that for every qubit $i$, there are only logarithmically many 2-qubit gates applied to qubits $j$ and $k$ with $j \le i \le k$.
These results, along with a few others [@Yoran:2006kx; @Valiant:2002aa], provide the basis of the circuit topological de-quantisations. By dealing with circuits they are able to make use of the extensive graph theoretic literature relating to properties such as the treewidth. These results have been applied to the QFT [@Yoran:2007aa], complementing the de-quantisation using entanglement based techniques. These results have the advantage that they can simulate the circuit on arbitrary product state inputs, but unlike the bounded entanglement simulations can only sample from the probability distributions; in many cases this is reasonable, but it makes understanding the role of the QFT as a ‘quantum subroutine’ in other algorithms more difficult [@Yoran:2007aa].
It is further worth nothing that the structural methods generally produce more complicated de-quantised algorithms. This is evident in the comparison of the different types of QFT de-quantisations [@Abbott:2010aa], and is a result of being overly faithful to the quantum construction which must conform to the restrictions of avoiding measurement and locality. Another example of this is the de-quantisation result of Browne [@Browne:2007aa], who realised that Niu and Griffiths’ semiclassical QFT [@Griffiths:1996aa] can be easily turned into a completely classical de-quantised algorithm with no loss in efficiency. This method is different from the other structural approaches as rather than primarily focusing on the internal structure, it is more a result of the ability to measure or ‘sample’ a qubit once all transformations involving it are completed, and using this to condition the next qubit’s transformations.
Operator Methods
----------------
At the other end of the spectrum from the de-quantisation techniques which follow the evolution of the state vector, are the methods which follow the evolution of the operators acting on the state—very much as is the case in the Heisenberg representation in quantum mechanics, as opposed to the Schrödinger representation in which the states evolve. This approach led to the well known Gottesman-Knill Theorem [@Gottesman:1999aa; @Aaronson:2004aa], which provides a de-quantisation result for algorithms using only the controlled-NOT, Hadamard and Phase gates, which are generators for the Clifford group.
\[thm:Gottesman-Knill\] Any quantum computation which uses only gates from the Clifford group (possibly conditioned on classical bits) and measurements on the computational basis, can be de-quantised.
While the Clifford gates are not universal, this result is in some sense surprising because it allows de-quantisation of algorithms which contain unbounded entanglement. This result is a complement to Theorem \[thm:JozsaLinden\], as it indicates that a good quantum algorithm must not permit a compact description of the state *or* the operators. This counters the notion that it is entanglement which provides the quantum computational advantage. The Gottesman-Knill Theorem has further been extended by Van den Nest [@Nest:2009aa] by reducing them to a simplified normal form and showing that all circuits consisting of Toffoli and diagonal gates only, followed by a basis measurement, are de-quantisable.
Given the advantages of the two complementary (state and operator based) de-quantisations, it is natural to ask if there is some further relation between these methods. This is an area in need of more research, and understanding the relationships between de-quantisation techniques will help understand quantum computation better. *It is not unreasonable to consider next an interaction picture type de-quantisation, making the best use of compact descriptions of state and operators simultaneously*.
Levels of De-quantisation
=========================
It is interesting to note that certain de-quantisation techniques appear to be ‘stronger’ than others [@Nest:2010aa]. Since quantum computation is inherently probabilistic, the goal of the de-quantisation is primarily to classically sample from the same probability distribution. However, the sampling techniques such as the entanglement-based techniques and the Gottesman-Knill method are somewhat artificial. In these cases, the probability distribution is calculated, and then a sample is taken by classical probabilistic methods at the end of the computation. This is in contrast to tensor-network de-quantisations, in which the de-quantised algorithm is inherently probabilistic, and the probability distribution is never calculated, only sampled. While this is sufficient for de-quantisation, the amount of work being done is somewhat different. In [@Nest:2010aa] it is shown that there exist circuits for which this ‘weaker’ sampling based form of de-quantisation is possible in polynomial time, but calculating the probability distribution is $\# P$-complete and thus at least as hard as an NP-complete problem.
This result suggests we should focus our attention on sample-based de-quantisations, but this is perhaps a little premature. Even though they may be less general, the ‘strong’ de-quantisations have the advantage that they are trivial to compose together (unlike the ‘weak’ methods [@Yoran:2007ab]), easier to implement classically, and if the de-quantised algorithm is one where the quantum solution is correct with probability-one, such as the Deutsch-Jozsa problem, the de-quantised algorithm can be made deterministic rather than probabilistic. Examining which type of de-quantisation is possible for an algorithm gives further insight into, and distinction between the power of different quantum algorithms. On the other side of the picture, this sample-based approach to de-quantisation shows much promise to be extended, and alternative probabilistic de-quantisations are being explored [@Nest:2010aa].
‘Where to Next?’ Is the Resounding Question
===========================================
As we have seen, a range of de-quantisation techniques with different advantages and disadvantages have been developed. These techniques give us necessary, but not sufficient, conditions which a quantum algorithm must have in order to pose a benefit over a classical algorithm. For example, we know that a good quantum algorithm must lack both a concise description of the state and the operators. However, there may exist many other properties which allow de-quantisation, and all such properties must be absent from a good quantum algorithm [@Jozsa:2003aa]. Extending these conditions to necessary conditions is the final, optimistic goal, as this would allow us to understand the relation between quantum and classical complexity classes.
However, since this has proven to be extremely difficult, searching for new, different properties which allow de-quantisation is a rewarding and realistic goal. Such properties are beneficial as they deepen our understanding of the power of quantum computation, and the more insight we have to this, the more effectively we can develop quantum algorithms.
In order to find new de-quantisation techniques, it is worth exploring other types of quantum algorithms. Current techniques have focused around the standard algorithms which primarily consist of Fourier transforms and interference. However, alternative classes of algorithms, such as those based on quantum random walks have been studied [@Aharanov:1993aa; @Shenvi:2003aa]. Exploring de-quantisation in these different settings could lead to new results in this area.
Comparing the complexities of a quantum and a classical algorithm solving the same problem is not easy. For example, a polynomial-time classical algorithm is stronger than a polynomial-time quantum algorithm solving the same problem. For ‘oracle-type’ problems, like the Deutsch-Jozsa problem, to compare complexities means to compare the classical and quantum black-boxes. Why? Let us recall that in the Deutsch-Jozsa problem the input is a classical black-box computing a function $f : \{0,1\}^n \to \{0,1\}$. The quantum solution [*embeds*]{} the classical black-box into a (more powerful) quantum black-box, capable of computing with superposition states. Formally, we have changed the problem, as we do not operate with the given data, the classical black-box, but with a modified version of this black-box. The new black-box computes the function in a higher-dimension than the original classical one. Indeed, one could argue that we could create a classical black-box which takes as input a classical version of the ‘equal superposition’ (which is separable), and output the suitable solution to the problem. Intuitively this is cheating, as all of the complexity has been hidden within the black-box. However, it is not clear how to take into consideration this black-box complexity, and at what point we are no longer solving the same problem.
The root of the proposed de-quantised solution lies in the fact that the embedding can be done as efficiently classically as it can be quantum-mechanically. To compare the complexities of the quantum and de-quantised solutions we ought to compare the costs/resources necessary for performing these ‘embeddings’. In order to understand the cost of the embedding, it seems necessary take into consideration its physical feasibility. Consider the following: by realising that the quantum black-box is a physical object, it must take, as input, a physical resource. If the black-box could be suitably isolated and embedded into the quantum computational system, since all physics is inherently quantum mechanical, the classical black-box could reasonably be transformed into a quantum one. It is not clear to see how the same can be done to embed the black-box in a de-quantised solution. For example, the embedding in the NMR implementation is somewhat artificial as we are able to ‘create’ the classical black-box. However, mathematically the quantum and de-quantised algorithms are identical and this apparent difference cannot be readily evaluated. So an important question is: how de we take into account the physical cost of the embedding in order to truly evaluate the complexity of the classical and de-quantised solutions?
Conclusion
==========
We have reviewed the ability to *de-quantise* a quantum algorithm to obtain a classical algorithm which is not exponentially slower in time compared to the quantum algorithm. The main ideas involved in de-quantisation have been illustrated with the Deutsch-Jozsa problem: from re-visiting the quantum solution to the construction of the de-quantised algorithm, the identification of the ‘ingredient’ allowing de-quantisation, a physical implementation of the de-quantised algorithm, to benefits and open questions. A formal definition of de-quantisation was proposed and the main techniques for de-quantisation have been briefly reviewed. Finally, the discussion of open problems has ended with the main unsolved problem related to the de-quantisation of the Deutsch-Jozsa problem: how to compare the classical and quantum black-boxes.
Acknowledgement {#acknowledgement .unnumbered}
===============
We thank Matthias Bechmann, Sonny Datt and Angelika Sebald for comments that improved this paper. This work was in part supported by UoA Summer 2010 Fellowship (Abbott) and UoA FRDF Grant 2010 (Calude).
[^1]: Apparently, in this article the term ‘de-quantisation’ was used for the first time.
[^2]: While we are not labelling the basis bits ‘0’ and ‘1’, they represent the classical bits 0 and 1 in the same way that $\ket{0}$ and $\ket{1}$ do.
[^3]: So far we have been considering the case where $U_f$ operates on $n$ input qubits and one auxiliary qubit, $\ket{-}$. It has been shown (see [@Collins:1998aa]) that the auxiliary qubit is not necessary if we restrict ourselves to the subspace spanned by $\ket{-}$. We have presented the algorithm with the auxiliary qubit present because it is more intuitive to think of the input-dependent phase factor being an eigenvalue of the auxiliary qubit which is kicked back. The de-quantised solutions, however, bear more resemblance to this reduced version of $U_f$ operating only on $n$ qubits.
[^4]: Just one type of many features which leads to de-quantisation;
[^5]: The bubblewidth and treewidth differ by no more than poly-logarithmic factors. See [@Aharonov:2007aa] for further details.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We define a new topos, the *Herbrand topos*, inspired by the modified realizability topos and our earlier work on Herbrand realizability. We also introduce the category of Herbrand assemblies and characterise these as the $\lnot\lnot$-separated objects in the Herbrand topos. In addition, we show that the category of sets is included as the category of $\lnot\lnot$-sheaves and prove that the inclusion functor preserves and reflects validity of first-order formulas.'
author:
- Benno van den Berg
bibliography:
- 'ast.bib'
date: 'April 18, 2013'
title: The Herbrand Topos
---
Introduction
============
In [@bergbriseidsafarik12] the author, together with Eyvind Briseid and Pavol Safarik, hit upon a new realizability interpretation in an attempt to find computational content in arguments performed in nonstandard analysis. This new interpretation, which was a variant of modified realizability, was dubbed Herbrand realizability. Our investigations in [@bergbriseidsafarik12] were entirely proof-theoretic; the question was whether it would be possible to understand Herbrand realizability from a semantic point of view as well. This paper shows that that is indeed the case.
To develop this semantics we use topos theory (for which see [@maclanemoerdijk92; @johnstone02a; @johnstone02b]). This choice was motivated by the fact that the notion of a topos is the most comprehensive notion of model for a constructive system we have available, incorporating topological, sheaf and Kripke models, as well as various realizability and functional interpretations. In addition, it shows that these interpretations can be made to work for full higher-order arithmetic. The starting point for this paper was the theory of realizability toposes (beginning with [@hyland82] and surveyed in [@vanoosten08]): indeed, the topos most closely related to the topos we will introduce here is the modified realizability topos (for which, see [@vanoosten97; @vanoosten08]).
In order to arrive at the modified realizability topos, one has to abstract considerably from Kreisel’s original definition [@kreisel59]. First of all, one fixes the hereditarily effective operations (HEO) as a model of Gödel’s $T$. Then a type gets identified with a certain inhabited set of codes and a set of realizers of that type will simply be subset of that set. The step that Grayson took in [@grayson81] was to take as truth values any pair $(A_0, A_1)$ where $A_0$ and $A_1$ are two sets of codes (often called the actual realizers and the potential realizers, respectively) with $A_0 \subseteq A_1$ and $A_1$ containing a fixed element. One can build a tripos around such pairs and in the associated topos the finite types will be interpreted as the hereditarily effective operations.
In order to define the Herbrand topos, we make a similar move. Recall from [@bergbriseidsafarik12] that in order to realize $$({\exists^{\operatorname{st}{}}\!}n \in \NN) \, \varphi(n)$$ it suffices to supply a finite list of natural numbers $(n_1, \ldots, n_k)$ such that $\varphi(n_i)$ is realized for some $i \leq k$. Abstracting away from the details, this means that potential realizers are finite lists of natural numbers, while the actual realizers are those finite lists $(n_1, \ldots, n_k)$ which contain an $n_i$ which works (this is similar to the idea of Herbrand disjunctions in proof theory; hence the name). Abstracting even further, we say that truth values in the Herbrand topos are pairs of sets of codes $(A_0, A_1)$ such that $A_0$ consists of finite sequences all whose elements belong to $A_1$ and which is closed upwards (by this we mean that it is closed under supersets, if we regard finite sequences as representatives for their set of components; see also Lemma 5.4 in [@bergbriseidsafarik12]). This paper shows that on the basis of these pairs one can construct a tripos, whose associated topos we will call the Herbrand topos.
The Herbrand topos turns out to have several features in common with other realizability toposes. It has an interesting subcategory consisting of the $\lnot\lnot$-separated objects (we will call these the Herbrand assemblies) and the category of sets is included as a subtopos via the $\lnot\lnot$-topology. What is very unusual, however, is that this inclusion functor, which we will call $\nabla$, preserves and reflects the validity of first-order logic; in fact, $\nabla$ preserves and reflects the structure of a locally cartesian closed pretopos. In particular, $\nabla 2 = 2$ in the Herbrand topos.
This is a striking illustration of the fact that in the Herbrand topos disjunction has essentially no constructive content. Indeed, in order for a disjunction $\varphi \lor \psi$ to be realized it is sufficient that one of the two disjuncts is realized; but a realizer for $\varphi \lor \psi$ need not say which disjunct it is that is actually realized. This fact explains many of the features of the Herbrand topos: why it believes in the law of excluded middle for $\Pi^0_1$-formulas, and why it does not believe in Church’s thesis or in continuity principles.
However, arithmetic in the Herbrand topos is not classical. This is due to the fact that existential quantifiers still have some constructive content. Admittedly, this content is less than is usually the case, but it is still strong enough to rule out Markov’s principle.
Finally, there are two other properties of the Herbrand topos which are worth mentioning. First of all, because $\nabla$ preserves and reflects first-order logic, $\nabla A$ will be a nonstandard model of arithmetic in the Herbrand topos for any nonstandard model $A$ (actually, it will also be a nonstandard model when $A$ is the standard model). This is interesting, because realizability toposes are unfavourable terrain for nonstandard models of arithmetic (see [@mccarty87]).
A proof-theoretic feature of the Herbrand topos which is worth stressing is that in it the Fan Theorem holds (see below); for this it is not necessary to assume its validity in the metatheory. The proof should look very familiar to anyone who is aware of the bounded modified realizability interpretation and its properties (for which, see [@ferreiranunes06]).
The contents of the paper are as follows. In Section 2 we will explain the notation that we will use, in so far as it is not standard. We define the Herbrand tripos and topos in Section 3. Section 4 introduces the Herbrand assemblies and proves that they form a locally cartesian closed regular category with finite stable colimits and a natural numbers object. Then, in Section 5, we characterise these Herbrand assemblies as the $\lnot\lnot$-separated objects in the Herbrand topos; in addition, we show that the category of sets is included as the $\lnot\lnot$-sheaves and that the inclusion functor preserves and reflects the structure of a locally cartesian closed pretopos. In Section 6, we characterise the projectives in the Herbrand assemblies and show that the natural numbers object is projective in the category of Herbrand assemblies, but not in the Herbrand topos. Section 7 studies the logical properties of the Herbrand topos and, finally, in Section 8 we discuss some further developments.
In this paper we assume familiarity with the theory of triposes and partial combinatroy algebras; for the necessary background information, we recommend [@vanoosten08].
We would like to thank Jaap van Oosten, Wouter Stekelenburg and the referee for useful comments. The author was supported by a grant from the Netherlands Organisation for Scientific Research (NWO).
Notation
========
Throughout this paper, $\pca{P}$ will be a fixed nontrivial partial combinatory algebra (pca). It is well known that we can code finite sequences of elements in $\pca{P}$ as elements of $\pca{P}$. If $n$ is such a code, then we write $|n|$ for the length of the sequence it codes and $n_i$ for the $i$th projection (where $n_1$ is the first element of the sequence and $n_{|n|}$ the last). We will write $<n_1, \ldots, n_k>$ for the code of the sequence $(n_1, \ldots, n_k)$ and $<>$ for the code of the empty sequence. Moreover, $*$ will denote the operation of concatenation of coded sequences and $a_1 * a_2 * \ldots * a_k$ will stand for the result of concatenating $a_1$ till $a_k$. If $k = 0$ (i.e., if we are taking the empty concatenation), then the result should be the code of the empty sequence $<>$.
If $A \subseteq \pca{P}$ is some subset of $\pca{P}$, then we will write $!A$ for the set of codes of finite sequences all whose elements belong to $A$ (note that this will always include the empty sequence). For our purposes it will be important that $!A$ carries a preorder structure with $m \preceq n$ if $$(\forall i \leq |m|) \, (\exists j \leq |n|) \, m_i = n_j.$$
We will write ${\bf p}$ for the pairing operator and denote the $n$th Church numeral simply by $n$. In addition, if $A$ and $B$ are subsets of $\pca{P}$, we will write $$\begin{aligned}
A \& B & = & \{ {\bf p}0a \, : \, a \in A \} \cup \{ {\bf p}1b \, : \, b \in B \}, \\
A \otimes B & = & \{ {\bf p}ab \, : \, a \in A, b \in B \}.\end{aligned}$$ Observe that there is an exponential isomorphism $$!(A \& B) \cong !A \otimes !B,$$ where both the map itself and and its inverse are order-preserving and represented by elements in the pca, independent of the specific $A$ and $B$. We will often implicitly use this isomorphism and regard elements of $!(A \& B)$ as pairs ${\bf p}xy$ with $x \in !A$ and $y \in !B$.
The Herbrand tripos
===================
The Herbrand topos will be obtained from the *Herbrand tripos*. To construct this tripos first put $$\Sigma = \{ (A_0, A_1) \, : \, A_0, A_1 \subseteq \pca{P}, A_0 \subseteq !A_1 \mbox{ and } A_0 \mbox{ is upwards closed in } !A_1 \},$$ where $A_0$ being upwards closed in $!A_1$ means that $m \in A_0, n \in !A_1$ and $m \preceq n$ imply $n \in A_0$. For $X \in \Sigma$, we will denote the result of projecting on the first and second coordinate by $X_0$ and $X_1$. As for modified realizability, we will refer to the elements of $X_0$ as the *actual realizers* and the elements of $!X_1$ as the *potential realizers*.
If $X$ is any set, then we preorder $\Sigma^X$ as follows: $(\phi: X \to \Sigma) \leq (\psi: X \to \Sigma)$ if there is an element $r \in \pca{P}$ such that for all $x \in X$ and $n \in \pca{P}$ $$\mbox{if } n \in !\phi(x)_1 \mbox{, then } r \cdot n \downarrow \mbox{ and } r \cdot n \in !\psi(x)_1$$ and $$\mbox{if } n \in \phi(x)_0 \mbox{, then } r \cdot n \in \psi(x)_0.$$ Finally, if $f: X \to Y$ is any function, then reindexing $f^*: \Sigma^Y \to \Sigma^X$ is given simply by precomposition.
[Herbrandtripos]{} The indexed preorder $\Sigma^X$ defined above is a tripos.
The associated topos we will call the Herbrand topos and denote by $\HT[\pca{P}]$.
We first verify that $\Sigma$ has the structure of a Heyting prealgebra. The top and bottom element are $(!\pca{P}, \pca{P})$ and $(\emptyset, \emptyset)$, respectively. The conjunction $(C_0, C_1) = (A_0, A_1) \land (B_0, B_1)$ is given by $$\begin{aligned}
C_1 & = & A_1 \& B_1, \\
C_0 & = & \{ {\bf p}ab \, : \, a \in A_0 \mbox{ and } b \in B_0 \}\end{aligned}$$ (making use of the exponential isomorphism). The disjunction $(C_0, C_1) = (A_0, A_1) \lor (B_0, B_1)$ is given by $$\begin{aligned}
C_1 & = & A_1 \& B_1, \\
C_0 & = & \{ {\bf p}ab \, : \, a \in A_0 \mbox{ or } b \in B_0 \}\end{aligned}$$ (again making use of the exponential isomorphism). To see that this works, suppose that $(A_0, A_1) \leq (D_0, D_1)$ is tracked by $r$ and $(B_0, B_1) \leq (D_0, D_1)$ is tracked by $s$. Then $(C_0, C_1) \leq (D_0, D_1)$ is tracked by the following map $t$: $t({\bf p}ab) = ra * sb$. This works, because for ${\bf p}ab \in C_0$ we have $a \in A_0$ or $b \in B_0$: in the former case, we have $ra \in D_0$, in the latter $sb \in D_0$. In either case we have $ra * sb$ because $D_0$ is upwards closed.
The implication $(C_0, C_1) = (A_0, A_1) \to (B_0, B_1)$ is given by $$\begin{aligned}
C_1 & = & \{ c \in \pca{P} \, : \, (\forall m \in !A_1) \, c \cdot m \downarrow \mbox{ and } c \cdot m \in !B_1 \}, \\
C_0 & = & \uparrow_{!C_1} \{ <c> \, : \, c \in C_1 \mbox{ and } (\forall m \in A_0) \, c \cdot m \in B_0 \}.\end{aligned}$$ To see that this works, note that the evaluation map $(C_0, C_1) \land (A_0, A_1) \to (B_0, B_1)$ is tracked by the function which maps ${\bf p}(<c_1,\ldots, c_k >, a)$ to $$c_1(a) * \ldots * c_k(a).$$ It is clear that this Heyting prealgebra structure lifts to each $\Sigma^X$.
As said, the reindexing functors are given by precomposition. We now check that these have both adjoints satisfying the Beck-Chevalley condition. So suppose we have $\phi: X \to \Sigma, \chi: Y \to \Sigma$ and $f: X \to Y$. The existential quantifier can be defined as follows: $$\begin{aligned}
\exists_f(\phi)(y)_1 & = & \bigcup_{x \in f^{-1}(y)} !\phi(x)_1, \\
\exists_f(\phi)(y)_0 & = & \uparrow_{!\exists_f(\phi)(y)_1} \{ <n> \, : \, (\exists x \in f^{-1}(y)) \, n \in \phi(x)_0 \}.\end{aligned}$$ To see this, note that
1. we have $\exists_f(\phi) \leq \chi$ if there is an $r \in \pca{P}$ sending for each $y \in Y$ elements from $!\exists_f(\phi)(y)_1$ to elements in $!\chi(y)_1$, in such a way that if $m \in \exists_f(\phi)(y)_0$, then $r \cdot m \in \chi(y)_0$.
2. And that we have $\phi \leq f^*(\chi)$, if there is an $s \in \pca{P}$ sending for each $x \in X$ elements in $!\phi(x)_1$ to elements in $!\chi(fx)_1$, in such a way that if $m \in \phi(x)_0$, then $s \cdot m \in \chi(fx)_0$.
To construct such an $s$ from an $r$, put $$s(m) = r(<m>)$$ and to construct such an $r$ from an $s$, put $$r(<m_1, \ldots, m_k>) = s(m_1) * \ldots s(m_k).$$
The universal quantifier is constructed as follows: $$\begin{aligned}
\forall_f(\phi)(y)_1 & = & \{ a \in \pca{P} \, : \, (\forall x \in f^{-1}(y)) \, (\forall b \in \pca{P}) \, a \cdot b \downarrow \mbox{ and } a \cdot b \in !\phi(x)_1 \}, \\
\forall_f(\phi)(y)_0 & = & \uparrow_{!\forall_f(\phi)(y)_1} \{ <a> \, : \, a \in \forall_f(\phi)(y)_1 \mbox{ and } \\
& & (\forall x \in f^{-1}(y)) \, (\forall b \in \pca{P}) \, a \cdot b \in \phi(x)_0 \}.\end{aligned}$$ To see this, note that
1. we have $f^*(\chi) \leq \phi$ if there is an $r \in \pca{P}$ sending for each $x \in X$ elements from $!\chi(fx)_1$ to elements in $!\phi(x)_1$, in such a way that if $m \in \chi(fx)_0$, then $r \cdot m \in \phi(x)_0$.
2. And that we have $\chi \leq \forall_f(\phi)$, if there is an $s \in \pca{P}$ sending for each $y \in Y$ elements in $!\chi(y)_1$ to elements in $!\forall_f(\phi)(y)_1$, in such a way that if $m \in \chi(y)_0$, then $s \cdot m \in \forall_f(\phi)(y)_0$.
To construct such an $s$ from an $r$, put $$s = \lambda p.<\lambda q.r \cdot p>$$ and to construct such an $r$ from an $s$, put $$r = \lambda p.\big[ \, (s \cdot p)_0(k) * \ldots *(s \cdot p)_{|s \cdot p|}(k)\, \big].$$ It is easy to see that the universal quantifier, and hence also the the existential quantifier, satisfies Beck-Chevalley.
Finally we need to construct a generic element; but that can quite straightforwardly be taken to be the identity map $\id: \Sigma \to \Sigma$.
Herbrand assemblies
===================
In this section we introduce the Herbrand assemblies and prove that they form a locally cartesian closed regular category with stable colimits. Later we will characterise them as the $\lnot\lnot$-separated objects in the Herbrand topos. Of course, it follows from this that the category is a regular locally cartesian closed category with stable colimits, but we need an explicit description of this structure later; in addition, such a description is, we believe, of independent interest.
[herbrassemblies]{} A *Herbrand assembly* over $\pca{P}$ is a triple $(A, {\cal A}, \alpha)$ in which
- $A$ is a set,
- ${\cal A}$ is a subset of $\pca{P}$, and
- $\alpha: A \to {\rm Pow}^{upcl}_i(!{\cal A})$ is a function whose codomain ${\rm Pow}^{upcl}_i(!{\cal A})$ consists of the subsets $X$ of $!{\cal A}$ that are inhabited and upwards closed in $!\pca{A}$.
A morphism $f: (B, {\cal B}, \beta) \to (A, {\cal A}, \alpha)$ of Herbrand assemblies is a function $f: B \to A$ for which there is an $n \in \pca{P}$ such that
- for all $m \in !{\cal B}$, the expression $n \cdot m$ is defined and its value belongs to $!{\cal A}$,
- and if $b \in B$ and $m \in \beta(b)$, then $n \cdot m \in \alpha(fb)$.
We will say that such an $n$ *tracks $f$* or is a *tracking of $f$*. This clearly defines a category: we will denote it by $\HAsm[\pca{P}]$.
[HAsmcartesian]{} The category $\HAsm[\pca{P}]$ has finite limits.
The terminal object is $(1, {\cal C}, \gamma)$ with ${\cal C} = \pca{P}$ and $\gamma(*) = !\pca{P}$.
Equalizers as in can be computed by putting $C = \{ b \in B \, : \, fb = gb \}$, ${\cal C} = {\cal B}$ and $\gamma = \beta \upharpoonright C$.
A product $(C, {\cal C}, \gamma) = (A, {\cal A}, \alpha) \times (B, {\cal B}, \beta)$ can be constructed by putting $$\begin{aligned}
C & = & A \times B, \\
{\cal C} & = & {\cal A} \& {\cal B}, \\
\gamma(a, b) & = & \{ \, {\bf p}ab \, : \, a \in \alpha(a), b \in \beta(b) \, \}.\end{aligned}$$
[HAsmregular]{} The category $\HAsm[\pca{P}]$ is regular.
We claim that a map $f: (B, {\cal B}, \beta) \to (A, {\cal A}, \alpha)$ is monic precisely when the underlying function $f$ is injective (this should be clear), and is a cover precisely when there is an element $r \in \pca{P}$ such that
- for all $n \in !{\cal A}$ the expression $r \cdot n$ is defined and belongs to $!{\cal B}$, and
- there is for any $a \in A$ and $n \in \alpha(a)$ an element $b \in f^{-1}(a)$ with $r \cdot n \in \beta(b)$.
Let us call maps which have these two properties *super epis*. It is not hard to check that super epis are covers and that they are stable under pullback, so the proof will be finished once we show that every map can be factored as a super epi followed by a mono. But if $f: (B, {\cal B}, \beta) \to (A, {\cal A}, \alpha)$ is any map, then it factors through $(C, {\cal C}, \gamma)$ with $C = {\rm Im}(f)$, ${\cal C} = {\cal B}$ and $\gamma(a) = \bigcup_{b \in f^{-1}(a)} \beta(b)$. Moreover, the obvious maps $(B, {\cal B}, \beta) \to (C, {\cal C}, \gamma)$ and $(C, {\cal C}, \gamma) \to (A, {\cal A}, \alpha)$ are a super epi and a mono, respectively.
[HAsmlextensive]{} The category $\HAsm[\pca{P}]$ has stable sums and coequalizers.
The initial object is $(0, \pca{P}, \gamma)$ with $\gamma$ the empty map.
The sum $(C, {\cal C}, \gamma) = (A, {\cal A}, \alpha) + (B, {\cal B}, \beta)$ is given by $C = A + B$, ${\cal C} = {\cal A} \& {\cal B}$, with $\gamma(a) = \{ {\bf p}xy \, : \, x \in \alpha(a), y \in !\pca{B} \} $ and $\gamma(b) = \{ {\bf p}xy \, : \, x \in !\pca{A}, y \in \beta(b) \}$.
A coequalizer is computed by letting $C$ be the coequalizer in the category of sets, ${\cal C} = {\cal A}$ and $\gamma(c) = \bigcup_{a \in q^{-1}(c)} \alpha(a)$.
[HAsmlccc]{} The category $\HAsm[\pca{P}]$ is locally cartesian closed.
Assume $f: (B, {\cal B}, \beta) \to (A, {\cal A}, \alpha)$ and $g: (S, {\cal S}, \sigma) \to (B, {\cal B}, \beta)$ are two morphisms of Herbrand assemblies. The object $\prod_f(g) = (T, {\cal T}, \tau)$ is computed as follows: $$\begin{aligned}
T & = & \{ (a \in A, t: B_a \to S) \, : \, gt = \id_{B_a} \mbox{ and } t \mbox{ is tracked} \, \}, \\
{\cal T} & = & {\cal A} \, \& \, \{ n \in \pca{P} \, : \, (\forall m \in !{\cal B}) \, n \cdot m \downarrow \mbox{ and } n \cdot m \in !{\cal S} \, \}, \\
\tau(a, t) & = & \{ \, {\bf p}(m, <n_1, \ldots, n_k>) \, : \, m \in \alpha(a) \mbox{ and } \\
& & \mbox{there is an } n_i \mbox{ tracking } t \, \}.\end{aligned}$$ The evaluation map $f^*\prod_f(g) \to g: ((a, t), b) \mapsto t(b)$ is tracked by a code for the function $L$ defined by $$L({\bf p}(<n_1, \ldots, n_k>, m)) = n_1(m) * \ldots * n_k(m).$$
[HAsmnno]{} The category $\HAsm[\pca{P}]$ has a natural numbers object (nno).
The nno is given by $({\mathbb{N}}, {\cal N}, \nu)$ with ${\cal N}$ the collection of Church numerals in $\pca{P}$ and $\nu(n) = \uparrow_{!\cal N} <n>$. For suppose a structure $1 \to (A, {\cal A}, \alpha) \to (A, {\cal A}, \alpha)$ is given and the first map is tracked by $p$ and the second by $q$. Define $s$ by recursion as $s0 = p0$ and $s(n+1) = q(s(n))$, and $t$ as $t(n) = s(n_0) * \ldots *s(n_{|n|})$. Then the canonical map ${\mathbb{N}}\to A$ will be tracked by $t$.
To summarise:
[ModHerbrandtopos]{} The category $\HAsm[\pca{P}]$ is a locally cartesian closed regular category with nno and stable colimits.
Gamma and nabla
===============
In this section we show that some of the theory developed for modified realizability also applies to Herbrand realizability. In particular, we show that the facts established on pages 281 and 282 of Van Oosten’s paper on the modified realizability topos [@vanoosten97] hold for the Herbrand topos as well. (Warning: We follow the notation of that paper, rather than that of [@vanoosten08].)
First of all, note that $\HAsm[\pca{P}]$ is a full subcategory of $\HT[\pca{P}]$, because every triple $(A, \pca{A}, \alpha)$ can be considered as an object $(A, =)$ of $\HT[\pca{P}]$ as follows: $$\begin{aligned}
\llbracket a = a' \rrbracket & = & \left\{ \begin{array}{cl}
(\alpha(a), \pca{A}) & \mbox{if } a = a' \\
(\emptyset, \pca{A}) & \mbox{otherwise}
\end{array} \right.\end{aligned}$$ In addition, we have a functor $\nabla: \Sets \to \HAsm[\pca{P}]$ which sends a set $X$ to the Herbrand assembly $(X, \pca{P}, \phi)$ with $\phi(x) = !\pca{P}$ for all $x \in X$. Taking the composition of these two functors we obtain a functor $\Sets \to \HT[\pca{P}]$ which we will also denote by $\nabla$.
[warningonnabla]{} This is *not* the constant objects functor as defined in [@vanoosten08] (and denoted $\nabla$ there).
[existence]{} The functor $\nabla$ has a finite limit preserving left adjoint $$\Gamma: \HT[\pca{P}] \to \Sets.$$ Moreover, $\Gamma\nabla \cong 1$.
As usual, we send an object $(X, =)$ to $X_0/\sim$, where $X_0 = \{ x \in X \, : \, \llbracket x = x \rrbracket_0 \mbox{ inhabited} \}$ with $x \sim x'$ if $\llbracket x = x' \rrbracket_0$ inhabited. In addition, the transpose of a function $f: \Gamma(X, =) \to Y$ is the function $(X, =) \to \nabla Y$ represented by: $$\begin{aligned}
F(x, y) & = & \left\{ \begin{array}{cl}
\llbracket x = x \rrbracket & \mbox{if } x \in X_0 \mbox{ and } f([x]) = y, \\
(\emptyset, \llbracket x = x \rrbracket_1) & \mbox{otherwise.} \end{array} \right.\end{aligned}$$ Note that therefore the unit $\eta: (X, =) \to \nabla \Gamma (X, =)$ is represented by $H: X \times X_0 \to \Sigma$ with $H(x, [x']) = \llbracket x = x \rrbracket$ if $x \in [x']$, and $(\emptyset, \llbracket x = x \rrbracket_1)$ otherwise.
[doublenegation]{} Let $(A_0, A_1) \in \Sigma$. Then $A_0$ is inhabited iff $(\lnot\lnot(A_0, A_1))_0$ is inhabited; in which case, $<\lambda p. <>> \in (\lnot\lnot(A_0, A_1))_0$.
This follows from the fact that $\lnot(A_0, A_1) = (C_0, C_1)$, where $C_1$ is the set of codes of functions which map elements from $!A_1$ to the empty sequence, and $C_0 = !C_1 - \{ <> \}$ if $A_0$ is empty, and $C_0 = \emptyset$ otherwise. Hence $\lambda p.<> \in C_1$ always and $<\lambda p.<>> \in C_0$ if $A_0$ is empty.
[charHerbrandassemblies]{} For an object $(X, =)$ in $\HT[\pca{P}]$ the following are equivalent:
1. $\eta_{(X, =)}$ is a monomorphism.
2. $(X, =)$ is $\lnot\lnot$-separated.
3. $(X, =)$ is isomorphic to a Herbrand assembly.
$1 \Rightarrow 2$: Suppose $\eta$ is mono; so $$H(x, [z]) \land H(x', [z]) \to x = x'$$ holds. Suppose $<a_1, \ldots, a_n>$ is an actual realizer for this. Furthermore, suppose $b \in \llbracket x = x \rrbracket_0, c \in \llbracket x' = x' \rrbracket_0$ and $d \in \lnot\lnot \llbracket x = x' \rrbracket_0$. Then $b \in H(x, [x])_0, c \in H(x', [x'])_0$ and $[x] = [x']$ by the previous lemma, so $$a_1({\bf p}bc) * \ldots * a_n({\bf p}bc) \in \llbracket x = x' \rrbracket_0 ;$$ similar for potential realizers. So $$x = x \land x' = x' \land \lnot\lnot(x = x') \to x = x'$$ holds and $(X, =)$ is $\lnot\lnot$-separated.
$2 \Rightarrow 3$: Suppose $t$ is an actual realizer of $x = x \land x' = x' \land \lnot\lnot(x = x') \to x = x'$. Define $(A, \pca{A}, \alpha)$ by $$\begin{aligned}
A & = & X_0, \\
\pca{A} & = & \bigcup_{x, x' \in X} !\llbracket x = x' \rrbracket_1, \\
\alpha(a) & = & \uparrow_{!\pca{A}} \{ < s > \, : \, (\exists x, x' \in a) \, s \in \llbracket x= x' \rrbracket_0 \}.\end{aligned}$$ One easily sees that there is a map $F: (X, =) \to (A, \pca{A}, \alpha)$ defined by $$\begin{aligned}
F(x, a) & = & \left\{ \begin{array}{cl}
\llbracket x = x \rrbracket \land (\alpha(a), \pca{A}) & \mbox{if } x \in a, \\
\llbracket x = x \rrbracket \land (\emptyset, \pca{A}) & \mbox{otherwise.} \end{array} \right.\end{aligned}$$ To see that $F(x, a) \land F(x', a) \to x = x'$ is realized (i.e., that $F$ is monic), one uses that if there are actual for realizers $F(x, a)$ of $F(x',a)$, then $\llbracket x = x' \rrbracket_0$ must be inhabited; but then one can compute an element in this set using $t$ and the previous lemma.
To show that $a = a \to (\exists x \in X) \, F(x, a)$ is realized (i.e., to show that $F$ is epic), one needs an element $s \in \pca{P}$ which uniformly in $a \in A$ sends elements from $!\pca{A}$ to elements in $$!\bigcup_{x \in X} !F(x, a),$$ in such a way that if $m \in \alpha(a)$, then $$s \cdot m \in \uparrow \{ <n> \, : \, (\exists x \in X) \, n \in F(x, a)_0 \}.$$ This can be done: for if $k = <k_1, \ldots, k_n> \in !\pca{A}$, then each $k_i$ belongs to some $!\llbracket x = x' \rrbracket_1$. Since this equality is symmetric and transitive, we can compute from $k_i$ an element $tk_i \in !\llbracket x = x \rrbracket_1$. Then ${\bf p}(tk_i, k) \in !F(x, a)_1$ and hence $<{\bf p}(tk_i, k)>_i \in ! \bigcup_{x \in X} !F(x, a)$. If $<k_1, \ldots, k_n>$ also belongs to $\alpha(a)$, then some $k_i$ belongs to $\llbracket x = x' \rrbracket_0$ with $x, x' \in a$. Then ${\bf p}(tk_i, k) \in F(x, a)_0$ and hence $s \cdot m \in \uparrow \{ <n> \, : \, (\exists x \in X) \, n \in F(x, a)_0 \}$, as desired.
The implication $3 \Rightarrow 1$ is left to the reader.
[charnablas]{} For an object $(X, =)$ in $\HT[\pca{P}]$ the following are equivalent:
1. $\eta_{(X, =)}$ is an isomorphism.
2. $(X, =)$ is isomorphic to an object of the form $\nabla Z$.
3. $(X, =)$ is a $\lnot\lnot$-sheaf.
$1 \Rightarrow 2$ is trivial.
$2 \Rightarrow 3$: One easily checks by hand that if $f: (X, =) \to (Y, =)$ is a dense mono and $g: (X, =) \to \nabla Z$ is any map, then the relation $g \circ f^{-1}$ is actually a function.
$3 \Rightarrow 1$: If $(X, =)$ is a $\lnot\lnot$-sheaf, then it is certainly separated. Hence $\eta_{(X,=)}$ is monic, by the previous proposition; as it is also dense, it follows that it has a left inverse. But then it is not hard to see that it must be an isomorphism.
We take a closer look at the objects in the image of $\nabla$.
[useofzero]{} If $(A, \pca{A}, \alpha)$ is a Herbrand assembly and there is an element $e \in \pca{P}$ such that $e \in \pca{A}$ and $<e> \in \alpha(a)$ for all $a \in A$, then $(A, \pca{A}, \alpha) \cong \nabla A$.
Obvious.
[nablaandlogic]{} The functor $\nabla: \Sets \to \HT[\pca{P}]$ preserves and reflects the structure of a locally cartesian closed pretopos. In particular, it preserves and reflects validity of first-order formulas.
One easily checks by hand that $\nabla$ preserves and reflects quotients of equivalence relations. Therefore it suffices to prove that the functor $\nabla: \Sets \to \HAsm[\pca{P}]$ preserves and reflects the structure of a locally cartesian closed regular category with sums, because this structure is preserved and reflected by the inclusion $\HAsm[\pca{P}] \to \HT[\pca{P}]$. But then the result follows immediately from the constructions we gave in Section 4 and the previous lemma.
We have just seen that $\nabla: \Sets \to \HT[\pca{P}]$ preserves and reflects first-order logic. But it does not preserve the natural numbers object, because:
[nnoinHerbrandtopos]{} The Herbrand assembly $(\NN, {\cal N}, \nu)$ is the natural numbers object in the Herbrand topos.
Since the successor map $s: (\NN, {\cal N}, \nu) \to (\NN, {\cal N}, \nu)$ is monic in the Herbrand topos, the nno there will be the smallest subobject of $(\NN, {\cal N}, \nu)$ containing 0 and closed under this map (see, for example, [@johnstone02b Corollary D5.1.3]). Because any subobject of a separated object is automatically separated, this smallest subobject is the object $(\NN, {\cal N}, \nu)$ itself.
Therefore $\nabla \NN$ is a nonstandard model of arithmetic in the Herbrand topos. It is this model which underlies the realizability interpretation of nonstandard arithmetic defined in Section 5 of [@bergbriseidsafarik12].
Projectives
===========
In this section we study the projective objects in the Herbrand assemblies and the Herbrand topos.
[partherbassembly]{} We call a Herbrand assembly $(A, {\cal A}, \alpha)$ *partitioned*, if for every $a \in A$ there is an element $g_a \in \pca{P}$ such that $$\alpha(a) = \uparrow_{!\cal A} <g_a>.$$
We first note that:
[easyfactsaboutpartassemblies]{} We have:
1. The natural numbers object $(\NN, {\cal N}, \nu)$ is partitioned.
2. Every object of the form $\nabla Z$ is isomorphic to a partioned Herbrand assembly and hence every Herbrand assembly is a subobject of a partitioned one.
3. Partitioned Herbrand assemblies are closed under finite limits.
4. Every Herbrand assembly can be covered by a partitioned one.
5. Every retract of a partitioned Herbrand assembly is isomorphic to a partitioned Herbrand assembly.
Items 1 and 2 are obvious, so we concentrate on the others.
It is clear that the terminal object in the category of Herbrand assemblies is partitioned and that partitioned Herbrand assemblies are closed under equalizers. Moreover, if $(A, {\cal A}, \alpha)$ and $(B, {\cal B}, \beta)$ are partitioned Herbrand assemblies, then the following partitioned Herbrand assembly is a product in the category of Herbrand assemblies: $$\begin{aligned}
C & = & A \times B, \\
{\cal C} & = & {\cal A} \otimes {\cal B}, \\
\gamma(a, b) & = & \uparrow_{!({\cal A} \otimes {\cal B})} <{\bf p}g_ag_b>.\end{aligned}$$
A Herbrand assembly $(A, {\cal A}, \alpha)$ can be covered by the partioned Herbrand assembly $(A', {\cal A'}, \alpha')$, where $$\begin{aligned}
A' & = & \{ (a, n) \in A \times \pca{P} \, : \, a \in A, n \in \alpha(a) \}, \\
{\cal A'} & = & !{\cal A} \mbox{ and } \\
\alpha'(a, n) & = & \uparrow_{!!{\cal A}} <n>,\end{aligned}$$ via the projection. The easy details are left to the reader.
If $(A, {\cal A}, \alpha)$ is a retract of the partitioned Herbrand assembly $(B, {\cal B}, \beta)$ via maps $f: A \to B$ and $g: B \to A$ with $gf = \id$, then $(A, {\cal A}, \alpha)$ is isomorphic to the partioned assembly $(A, {\cal B}, \beta f)$.
[partHassemblproj]{} Partitioned Herbrand assemblies are projective in the category of Herbrand assemblies, both internally and externally.
We first show that partitioned Herbrand assemblies are externally projective. So suppose $p: (B, {\cal B}, \beta) \to (A, {\cal A}, \alpha)$ is a cover between Herbrand assemblies and $(A, {\cal A}, \alpha)$ is partitioned. Then there is an element $r \in \pca{P}$ such that $r$ is defined on all elements of $!{\cal A}$ and then yields values in $!{\cal B}$, and $$(\forall a \in A) \, (\forall n \in \alpha(a)) \, (\exists b \in p^{-1}(a)) \, r \cdot n \in \beta(b).$$ In particular, $$(\forall a \in A) \, (\exists b \in p^{-1}(n)) \, r \cdot <g_a> \in \beta(b).$$ Using choice in the metatheory, this means that there is a function $f: A \to B$ such that $pf = \id$ and $$(\forall a \in A) \, r \cdot <g_a> \in \beta(fa).$$ Therefore the function $f$ is a section of $p$ tracked by $$s \cdot <m_1, \ldots, m_k> = r \cdot <m_1> * \ldots * r \cdot <m_k>.$$ Since the terminal object is partitioned (item 3 of the previous lemma) and therefore externally projective, this implies that externally projective objects are also internally projective.
[regexcompl]{} Up to isomorphism, the partitioned Herbrand assemblies are the projective objects in the category of Herbrand assemblies. In particular, this category is equivalent to the reg/lex-completion of its full subcategory on the partitioned Herbrand assemblies.
The first statement follows from the previous proposition and items 4 and 5 of the previous lemma; the second follows from the characterisation of reg/lex-completions in [@carboni95].
[nnoprojectiveinassemblies]{} The natural numbers object is both internally and externally projective in the category of Herbrand assemblies.
In contrast, it is not generally true that the natural numbers object is externally projective in the entire topos. Indeed, this fails if we take as our pca $\pca{P}$ Kleene’s first pca $K_1$. To define this pca we need to fix an enumeration of the partial recursive functions satisfying some properties (for the precise properties that one needs, see [@vanoosten08 pages ix and 15]). Then $K_1$ has as underlying set the natural numbers and the partial application $n \cdot m$ is defined to be the result of applying the $n$th partial recursive function to the natural number $m$.
[nnonotprojectiveinHerbrandtopos]{} If $\pca{P} = K_1$, then the natural numbers object is not externally projective in the Herbrand topos.
Let $A_0$ and $A_1$ be two recursively inseparable subsets of $\NN$ and let $(\NN \times \{ 0, 1 \}, E)$ be the object of the Herbrand topos which has no actual or potential realizers for nontrival equalities, and whose existence predicate $E = \llbracket - = - \rrbracket$ is given by: $$\begin{aligned}
E(n, i)_1 & = & \NN \otimes \{ i \}, \\
E(n, i)_0 & = & \left\{ \begin{array}{cl}
\emptyset & \mbox{if } n \in A_{1-i}, \\
\uparrow_{!E(n,i)_1} <{\bf p}ni> & \mbox{otherwise.}
\end{array} \right.\end{aligned}$$ Now consider the projection $p: (\NN \times \{0,1\}, E) \to (\NN, {\cal N}, \nu)$ given by $$p((n, i), m) = \left\{ \begin{array}{cl}
E(n,i) \land (\nu(m), {\cal N}) & \mbox{if } n =m, \\
\emptyset & \mbox{otherwise.}
\end{array} \right.$$ This map is surjective, because its surjectivity is realized by the element $s \in \pca{P}$ satisfying $$s \cdot <n_1, \ldots, n_k> = <<{\bf p}n_10>, <{\bf p}n_11>, \ldots, <{\bf p}n_k0>, <{\bf p}n_k1>>.$$ But, on the other hand, this map does not have a section, because a realizer $r$ for such a map would allow one to recursively separate the sets $A_0$ and $A_1$ (for $n \in \NN$ compute the second projection of $(r \cdot <n>)_1$; this always yields either 0 or 1 and for $n \in A_i$ it yields $i$).
[noexreg]{} If $\pca{P} = K_1$, then not every object in the Herbrand topos can be covered by a Herbrand assemby. In particular, the Herbrand topos is not the ex/reg-completion of the category of Herbrand assemblies.
Logical features of the Herbrand topos
======================================
In this section we investigate the validity in the Herbrand realizability of some significant logical principles. Such questions are pca dependent and here we restrict attention to the Herbrand topos based on the pca $K_1$. More information on the principles we consider can be found in [@troelstravandalen88a Chapter 4].
[herbrandcomputable]{} A function $f: \NN \to \NN$ is tracked (as a morphism $(\NN, {\cal N}, \nu) \to (\NN, {\cal N}, \nu)$ in the Herbrand topos) iff it is bounded by a computable function. Indeed, from a tracking one can compute a bound and vice versa.
Note that it is necessary and sufficient for a function $f: (\NN, {\cal N}, \nu) \to (\NN, {\cal N}, \nu)$ to be tracked that one can compute for every $n$ a sequence $r(n) = <n_1, \ldots, n_k>$ such that $f(n) = n_i$ for some $i \leq k$. Since one can compute maxima, this implies that $f$ is bounded by a computable function $g$. If, on the other hand, $g$ is a computable function bounding $f$, then $r(n) = <0, \ldots, g(n)>$ shows that $f$ is tracked.
[boundedCT]{} In the Herbrand topos the following bounded form of Church’s Thesis holds: $$(\forall x \in \NN) \, (\exists y \in \NN) \, \varphi(x, y) \to (\exists e \in \NN) \, (\forall x \in \NN) \, \big( e \cdot x \downarrow \land \, (\exists y \leq e \cdot x) \, \varphi(x, y) \big).$$
Follows from the previous lemma and .
[someprinciples]{} In the Herbrand topos, the weak law of excluded middle $$\lnot \varphi \lor \lnot \lnot \varphi$$ is valid. Hence the De Morgan laws hold, as does: $$\label{formofLEM} (\forall x \in \NN) \, ( \, P(x) \lor \lnot P(x) \,) \to (\forall x \in \NN) \, P(x) \lor \lnot(\forall x \in \NN) \, P(x).$$ In particular, the law of excluded middle is valid for $\Pi_1^0$-formulas. But Markov’s principle fails and so does Church’s thesis.
It follows from the proof of that $<\lambda p. <>>$ is an actual realizer of $\lnot \varphi$ or of $\lnot \lnot \varphi$ (depending on whether $\varphi$ has an actual realizer or not). Hence ${\bf p}(<\lambda p. <>>,<\lambda p. <>>)$ is an actual realizer of $\lnot \varphi \lor \lnot \lnot \varphi$. The De Morgan laws (in particular, $\lnot (\varphi \land \psi) \to \lnot \varphi \lor \lnot \psi$), the principle in (\[formofLEM\]) and the law of excluded middle for $\Pi^0_1$-formulas are immediate consequences of this.
Nevertheless, not all classical arithmetic is valid in the Herbrand topos: this follows from the previous corollary. Therefore Markov’s Principle must fail (because Markov’s Principle together with (\[formofLEM\]) implies full classical arithmetic). Alternatively, one can argue directly for the failure of Markov’s Principle by showing that a realizer for it would allow one to solve the halting problem (along the same lines as for modified realizability; see [@bergbriseidsafarik12 Proposition 5.9]).
Church’s Thesis in the form ${\rm CT}^{\lor}_0$ is incompatible with the law of excluded middle for $\Pi_1^0$-formulas (again, use two recursively inseparable r.e. subsets of $\NN$, or see [@troelstravandalen88a Section 4.3]). Also ${\rm CT}$ fails: in fact, it follows from that every bounded function $f: \NN \to \NN$ is tracked.
[continuity]{} Continuity principles, like every function $f: {\mathbb{R}}\to {\mathbb{R}}$ is continuous, fail in the Herbrand topos.
Because these are incompatible with the law of excluded middle for $\Pi^0_1$-formulas (see, for example, [@troelstravandalen88a Proposition 4.6.4]).
[notordinaryrealizability]{} The Herbrand topos is not equivalent to a realizability topos over an (order-)pca.
Because in such toposes Markov’s Principle holds.
[twovaluednotboolean]{} The Herbrand topos is two-valued, but not boolean.
The terminal object has only two subobjects in the category of Herbrand assemblies; since the terminal object is separated and separated objects are closed under subobjects, the same applies to the Herbrand topos. Hence the Herbrand topos is two-valued. Nevertheless, it is not boolean, since Markov’s Principle fails.
[tree]{} (See [@troelstravandalen88a page 186].) In the following we will mean by a *tree* an inhabited and decidable set of finite sequences of natural numbers, closed under predecessors. A tree will be called *finitely branching* if $$(\forall n \in T) \, (\exists z \in \NN) \, (\forall x\in \NN) \, ( \, n * <x> \in T \to x \leq z).$$ Finally, by an *infinite path* in a tree $T$ we will mean a function $\alpha: \NN \to \NN$ such that $\overline{\alpha}n = < \alpha0, \ldots, \alpha(n-1)> \in T$ for all $n \in \NN$.
[fansuniform]{} The collection of infinite paths in a finitely branching tree is a uniform object. More precisely, if $T$ is a finitely branching tree, then one can compute from a realizer for the statement that $T$ is a finitely branching tree an element $n \in \pca{P}$ such that $<n>$ is a common realizer for all infinite paths in the tree.
From a realizer for the statement that $T$ is finitely branching tree one can compute a bound $f(n)$ on the value of $\alpha(n)$ for every infinite path $\alpha$. This is sufficient in view of .
[koenig]{} In the Herbrand topos König’s Lemma holds.
Recall that König’s Lemma says that every finitely branching tree which is infinite contains an infinite path. Of course, König’s Lemma is true, but the question is whether we can compute a realizer for an infinite path. The previous lemma, however, guarantees that we can.
[fan]{} In the Herbrand topos the Fan Theorem holds.
Suppose we have realizers for the statements that $T$ is a finitely branching tree, that $A(x)$ is a property of its nodes, inherited by successors, and that for every infinite path $\alpha$ there is a natural number $n \in \NN$ such that $A(\overline{\alpha}n)$ holds. From a realizer for the first statement we compute a common realizer for all infinite paths. Then from this and a realizer for the third statement we compute a finite sequence $<n_1, \ldots, n_k>$ such that for each infinite path there is an $i$ such that $A(\overline{\alpha}n_i)$ holds. But as $A$ is inherited by successors, we must have $A(\overline{\alpha}n)$ for all infinite paths $\alpha$, if $n = {\rm max}(n_1, \ldots, n_k)$.
Conclusion
==========
We have introduced a new topos and established some of its basic properties. Since this paper was written, a number of connections to other toposes have been uncovered: Jaap van Oosten has shown that the Herbrand topos is a subtopos of the corresponding modified realizability topos, while Peter Johnstone has shown that it is the Gleason cover of the corresponding realizability topos [@johnstone13]. In addition, we have shown [@berg13] that there is a topos for the nonstandard functional interpretation developed in [@bergbriseidsafarik12] which is related to the modified Diller-Nahm topos (see [@streicher06] and [@biering08]) in the same way as the Herbrand topos is related to the modified realizability topos. But we expect that much more can be said.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, we revisit the problem of finding the longest systematic-length $k$ for a linear minimum storage regenerating (MSR) code with optimal repair of only systematic part, for a given per-node storage capacity $l$ and an arbitrary number of parity nodes $r$. We study the problem by following a geometric analysis of linear subspaces and operators. First, a simple quadratic bound is given, which implies that $k=r+2$ is the largest number of systematic nodes in the *scalar* scenario. Second, an $r$-based-log bound is derived, which is superior to the upper bound on log-base $2$ in the prior work. Finally, an explicit upper bound depending on the value of $\frac{r^2}{l}$ is introduced, which further extends the corresponding result in the literature.
[**Key Words:**]{} MSR Codes, Systematic-Length, Linear Subspaces, Upper Bounds.
author:
- 'Kun Huang, Udaya Parampalli'
- Ming Xian
title: 'Improved Upper Bounds on Systematic-Length for Linear Minimum Storage Regenerating Codes [^1] '
---
INTRODUCTION
============
Distributed storage systems (DSSs) are efficient storage systems designed for coping with the tremendously growing data. They have become an indispensable component of many distributed applications such as cloud storage, social networking and peer to peer networking. Through introducing redundancy in the form of replication or erasure coding, DSSs provide data storage services with high reliability [@Re:A.G.Dimakis1]. Compared with the systems that employ replication, erasure codes achieve higher reliability for the same level of redundancy [@Re:H.Weatherspoon]. With regard to repair bandwidth as a classical measurement of reliable DSSs, Dimakis et al. [@Re:A.Dimakis] propose a novel class of codes termed *regenerating codes* that are particularly advantageous in the repair efficiency.
Regenerating Codes
------------------
Regenerating codes [@Re:A.Dimakis] are a family of maximal distance separable (MDS) codes. In this framework, a $B$-sized original data file over a finite field $\mathbb{F}$ is encoded into $n\alpha$ symbols that are distributed across $n$ nodes with each node storing $\alpha$ symbols. Data symbols stored in any $k$ out of $n$ nodes are sufficient to recover the original data file, and meanwhile, any single node failure can be repaired by downloading $\beta$ symbols from each of $d$ helper nodes out of the remaining $n-1$ nodes. It is shown in [@Re:A.Dimakis] that regenerating codes with a parameter set $\{n,k,d,\alpha,\beta,B\}$ have to comply with the following constraint (the tradeoff curve): $$\label{cut}
B\leq \sum_{i=1}^{k}\min\{\alpha,(d-i+1)\beta\},$$ where codes that achieve the above curve are called *optimal* regenerating codes. This tradeoff curve is equipped with two extreme points, namely, the minimum bandwidth regenerating (MBR) point and the minimum storage regenerating (MSR) point, respectively representing optimal regenerating codes with the least repair bandwidth and ones with the least per node storage. The corresponding parameters of the two points are determined by $$\left\{\begin{aligned}
&(\alpha_{\mathbf{MSR}},\beta_{\mathbf{MSR}})=(\frac{B}{k}, \frac{B}{k(d-k+1)})\\
&(\alpha_{\mathbf{MBR}},\beta_{\mathbf{MBR}})=(\frac{2dB}{k(2d-k+1)}, \frac{2B}{k(2d-k+1)}).
\end{aligned}\right.$$
There are three repair models considered in literature: functional repair, exact repair and exact repair of systematic nodes [@Re:A.G.Dimakis1]. Compared to the functional repair, the exact repair is highly favorable, since exact repair can restore the exact replicas of the lost data [@Re:C.Huang]. In the scenario of exact repair, Shah et al. in [@Re:Shah.N.B] demonstrate that it is not achievable for most interior points on the tradeoff curve. For those possibly achievable interior points, there are few constructions of codes [@Re:C.Tian; @Re:T.Ernvall]. Hereinafter, we restrict our focus to exact repair of linear MSR codes.
Minimum Storage Regenerating Codes
----------------------------------
Apart from the concept of regenerating codes, array codes [@Re:Blaum] are extensively studied over several years and have been widely deployed in DSSs [@Re:K.Rashmi; @Re:C.Suh; @Re:Y.Wu1; @Re:K.V.Rashmi; @Re:Rashmi; @Re:N.B.Shah; @Re:Z.Wang; @Re:D.S.Papailiopoulos; @Re:V.R; @Re:I.Tamo; @Re:Z.Wang1; @Re:Z.Wang2; @Re:V.R.Cadambe; @Re:V.R.Cadambe1; @Re:G.K.Agarwal; @Re:B.S; @Re:Y.S.Han; @Re:J.Li; @Re:A.S; @Re:N.Raviv; @Re:M.Ye; @Re:Sasidharan]. An $\{n,k,l\}$-array code is built from an $(l\times n)$ matrix, where each column is the codeword stored in the $i$-th node. The notation $l$ known as the *sub-packetization size* [@Re:Cadambe] basically represents the same meaning with the parameter $\alpha$ of linear regenerating codes [@Kun]. In this sense, as for optimal-repair MDS array codes with $\{d=n-1\}$ that reside in the MSR scenario, the corresponding parameter notations can be rewritten as $\{n,k,\alpha=l,\beta=\frac{l}{r},B=kl\}$ where $r=n-k$. Additionally, any optimal-repair systematic MDS array code obviously belongs to the scenario of MSR codes, since the systematic structure of a code signifies its property of minimum storage.
From the perspective of coding theory and technique for MSR codes, a lot of progresses have been made. Collectively, MSR codes are categorized into *scalar* MSR codes with $\{\beta=1\}$ [@Re:K.Rashmi; @Re:C.Suh; @Re:Y.Wu1; @Re:K.V.Rashmi; @Re:Rashmi; @Re:N.B.Shah] and *vector* MSR codes with $\{\beta>1\}$ [@Re:Z.Wang; @Re:D.S.Papailiopoulos; @Re:V.R; @Re:I.Tamo; @Re:Z.Wang1; @Re:Z.Wang2; @Re:V.R.Cadambe; @Re:V.R.Cadambe1; @Re:G.K.Agarwal; @Re:B.S; @Re:Y.S.Han; @Re:J.Li; @Re:A.S; @Re:N.Raviv; @Re:M.Ye; @Re:Sasidharan]. When $d=n-1$, scalar MSR codes are formed by $\{l=r\}$ and the majority of these vector MSR codes are provided with $\{\beta=(n-k)^{x}\leftrightarrow l=r^{x+1}, x\geq1\}$. Most of current constructions are established on the technique of *interference alignment* [@Re:N.B.Shah] that is necessary for constructing linear MSR codes. Moreover, it is verified in [@Re:N.B.Shah] that there only exist scalar linear MSR codes within the low rate regime $\{\frac{k}{n}\leq\frac{1}{2}\}$. The known product-matrix-based MSR code [@Re:K.Rashmi] is scalar for $\{2k-2\leq d\leq n-1\}$. Another design for scalar MSR codes with $\{d=n-1\geq 2k-1\}$ is given in [@Re:C.Suh] that entirely stems from interference alignment. In the high rate regime $\{\frac{k}{n}>\frac{1}{2}\}$, vector MSR codes can be leveraged to allow arbitrarily high rates to be attained. However, many of them allow optimal repair of only systematic part [@Re:V.R; @Re:I.Tamo; @Re:Z.Wang1; @Re:Z.Wang2; @Re:V.R.Cadambe; @Re:V.R.Cadambe1; @Re:G.K.Agarwal; @Re:B.S; @Re:Y.S.Han; @Re:J.Li; @Re:A.S; @Re:N.Raviv] that are also referred to as systematic-repair[^2] MSR codes for brevity, such as Permutation codes [@Re:V.R] and Zigzag codes [@Re:I.Tamo]. Vector MSR codes permitting optimal repair of parity nodes as well are proposed in [@Re:Z.Wang; @Re:D.S.Papailiopoulos; @Re:M.Ye], where the code given in [@Re:Z.Wang] is a variant of Zigzag code. On the other hand, Cadambe et al. in [@Re:Cadambe] prove that with $l$ (or $\alpha$) scaling to infinity, there exist high-rate MSR codes with parameter set $\{n,k,d\}$ of any value. Goparaju et al. in [@Re:S.G] provide an non-explicit construction of systematic-repair MSR codes for all $\{k\leq d\leq n-1\}$ that can be used to meet the high-rate requirement. For $d<n-1$, authors in [@Re:Sasidharan] present a special $\{l=\frac{1}{2}\cdot r^{\frac{n}{r}}\}$ vector MSR code with an explicit coupled-layer construction that optimally repairs all nodes, which in fact is built on the design of [@Re:B.S] and also can be adjusted to have $\frac{k}{n}$ as close to $1$ as desirable high rate.
Existing Upper Bounds on Systematic-Length
------------------------------------------
Centering on the case of systematic-repair MSR codes parameterized by $\{n=k+r,k,l,d=n-1\}$, we revisit the problem considered in [@Re:I.Tamo1; @Re:S.Goparaju1] “the relationship between the node storage capacity $l$ and the number of systematic nodes $k$ for some constant $r$". To be precise, the concern is as follows: what is the longest systematic-length $k$ for which there exists a linear systematic-repair MSR code, for a given sub-packetization level $l$ and a given number of parity nodes $r$?
Tamo et al. in [@Re:I.Tamo1] initially bring forward this question. They first consider systematic-repair MSR codes with two properties of particular interest, i.e., *optimal access* and *optimal update*, where an optimal-access code transmits only the symbols it accesses and an optimal-update one updates exactly once in each parity node. Restricted to the condition of optimal update, they demonstrate that the code with optimal access has the same longest systematic-length $k$ as the one with optimal bandwidth only, where $k_{max}=\log_rl$. For instance, Zigzag codes [@Re:I.Tamo] are of optimal-update that can attain this value. Without this restriction, they further show that $k=r\log_rl$ is the longest systematic-length for a systematic-repair MSR code with optimal access. The scheme for optimal-access and systematic-repair MSR code that achieves this optimal value of systematic-length is initially proposed in [@Re:V.R.Cadambe1]. Recently, authors in [@Re:N.Raviv] also present an explicit construction for systematic-repair MSR codes with optimal access, where $l=r^{\frac{k}{r}}$ reaches the optimal value. Another construction of systematic-repair MSR codes with optimal access given in [@Re:M.Ye] requires that $l=r^{\lceil\frac{n}{r}\rceil}=r^{\lceil\frac{k}{r}\rceil+1}$, which however is unattainable and differs from $l_{min}=r^{\frac{k}{r}}$ by $r^2$ approximately.
In regard to the general situation (namely the case of non-optimal update as well as non-optimal access), Tamo et al. only provide a loose upper bound $$k\leq l{{l}\choose{\frac{l}{r}}},$$ and conjecture that $k$ is of the order of $\log_rl$. In [@Re:Z.Wang1] and its extended version [@Re:Z.Wang2], authors design a general systematic-repair MSR code, wherein $l=r^{\frac{k}{r+1}}$ or $k=(r+1)\log_rl$ that is consistent with this conjecture. Nevertheless, it still remains open to explore better upper bounds on systematic-length $k$ and simultaneously construct tighter systematic-repair MSR codes for the general case [@Re:Z.Wang2].
Aiming at the above general situation, Goparaju et al. in [@Re:S.Goparaju1] transform it into a linear algebraic problem involving a set of linear subspaces and operators, where *repair subspaces*[^3] are the main research target herein and the operators in fact are a series of invertible square matrices. Through operating these square matrices on repair subspaces for purpose of satisfying the conditions of subspace properties, they sequentially analyze and demonstrate the linear independence among these operators, where the subspace properties actually are established on the technique of interference alignment as shown in [@Re:I.Tamo1]. In particular, they mainly work on the case of $r=2$ based upon the subspace properties for any distinct $i,j\in[1,k]$ as follows: $$\left\{\begin{aligned}
&\mathbf{S}_{i}\simeq\mathbf{S}_{i}\mathbf{\Phi_j}\\
&\mathbf{S}_{i}\mathbf{\Phi_i}\oplus \mathbf{S}_{i} \simeq \mathbb{F}^l,
\end{aligned}\right.$$ where $\mathbf{S}_{i}$ is the repair subspace of the $i$-th systematic node with dimension $\frac{l}{2}$, $\mathbf{\Phi_i}$ is the corresponding operator of order $l$, $``\oplus"$ represents the direct sum and $``\simeq"$ stands for a symbol for identical space[^4]. Briefly speaking, they take advantage of the restricted geometric properties of subspaces and operators, by which they construct linearly independent operators of special features and finally derive three new upper bounds on $k$ owing to the limited dimension of matrices space $\mathbb{F}^{l\times l}$ that is equal to $l^2$.
They first present a simple systematic-length bound $k\leq l^2$ that does not depend on the number of parity nodes $r$. For the special case of $2$ parity nodes, they prove that $k\leq 4l+1$. Subsequently, employing the method of *partition*[^5] for systematic part $[1,k]$ where repair subspaces within each partition span the entire space $\mathbb{F}^l$, they design $2^{\frac{k}{\lambda}}$ linearly independent matrices and thereby derive $$k\leq 2\lambda\log_2l,$$ where $\lambda$ is the size of each partition. Applying the geometric analysis of repair subspaces, they prove that $\lambda\leq\left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$, which means that if the partition contains $\big(\left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1\big)$ number of repair subspaces, they necessarily span the whole space $\mathbb{F}^l$. As a matter of fact, the second case is included herein, since it is easy to verify that $2\log_2l\big(\left \lfloor\log_{2}l\right \rfloor+1\big)< 4l+1$.
\[exa\] Consider a tiny storage disk of size that equals with $2^{8}$ bits, i.e., $l=2^{8}$. As for the case of $r=16$ parity nodes, from the latest upper bound on $k$, we have that $k\leq 2\big(\log_22^8\big)\big(\log_{\frac{16}{15}}2^8+1\big)\backsimeq1391$. With our new upper bound present in this paper, we will prove that $k$ cannot exceed $348$ (see the next section for details).
Our Contribution
----------------
In this work, we continue to investigate the aforementioned general situation, i.e., for simplicity, seek tighter upper bounds on systematic-length $k$ for general systematic-repair MSR codes with a given $l$ and $r$. We study by means of further exploiting the geometric analysis of linear subspaces and operators applied in [@Re:S.Goparaju1] for the case of arbitrary number of parity nodes $r$.
By virtue of the basic principle of interference alignment and some equivalent translations, we begin with formalizing two crucial derivative extended properties of subspaces and operators: (i) it is closed for operations on repair subspaces by the addition and multiplication of encoding matrices, and (ii) any vectors coming from different linear subspaces that mutually intersect trivially are linearly independent. Using these two properties, we sequentially find the linear independence among the existing encoding matrices and further proceed to create new linearly independent matrices, which both naturally bound the systematic-length $k$ due to the dimensional restriction of square matrices.
First of all, we prove that all the encoding matrices are linearly independent, with which we obtain a simple quadratic bound $$k\leq \frac{l^2-1}{r-1}.$$ This simple bound implies that $k=r+2$ is the longest systematic-length in the scalar scenario.
Secondly, we follow the idea of partition employed in [@Re:S.Goparaju1], where we construct new matrix by “*adding*" matrices of the same row within each partition. The reason for adding matrices instead of multiplying them as in [@Re:S.Goparaju1] is that matrix’s addition possesses the commutative law while multiplication of matrix is non-commutative. Afterwards, through multiplying these constructed matrices from different partitions in sequence, we further strike out $r^{\frac{k}{\lambda}}$ new matrices, where $\lambda$ is the size of a partition. We demonstrate that these $r^{\frac{k}{\lambda}}$ new matrices are all non-zero matrices and linearly independent as well, from which we derive an $r$-based-log upper bound $$k\leq 2\lambda\log_rl.$$
Finally, we consider estimating the partition size $\lambda$, where Goparaju et al. in [@Re:S.Goparaju1] show that $\lambda\leq \left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$. Leveraging our previous work in [@Kun], we introduce an explicit result as follows $$\lambda:\left\{\begin{array}{ll}
=r, &\textrm{when} \quad t=r;\\
\leq t+1+\left \lfloor\log_{\frac{r}{r-1}}\frac{(r-t)l}{r}\right \rfloor, &\textrm{when} \quad 1\leq t<r,
\end{array}\right.$$ where $t=\lceil\frac{r^2}{l}\rceil$, i.e., $t$ is an integer such that $\frac{r}{t}\leq \frac{l}{r}<\frac{r}{t-1}$ or $\frac{r^2}{l}\leq t<\frac{r^2}{l}+1$. It indicates that $\lambda=r$ if $t=r$ and $\lambda\leq\left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$ for $t=1$, which extend the corresponding result in [@Re:S.Goparaju1]. Furthermore, it can be verified that $r\leq \lambda\leq \left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$. To be more visible, when setting $l=2^8$ and $r=16$ as in Example \[exa\], we have that $t=\lceil\frac{16^2}{2^8}\rceil=1$ and therewith our $r$-based-log upper bound leads to that $k\leq 2\big(\log_{16}2^8\big)\big(\log_{\frac{16}{15}}2^8+1\big)\backsimeq348$.
Table \[Tab:Comparision\] illustrates the research progresses on systematic-length $k$ for linear systematic-repair MSR codes[^6] including upper bounds and the scheme [^7] with the longest known systematic-length [@Re:Z.Wang2].
------------------------------------------------------------------------
------------------------------------------------------------------------------------------------------------------------------------------------------------------
Citation Corresponding Results
------------------------------------- ----------------------------------------------------------------------------------------------------------------------------
Tamo et al. [@Re:I.Tamo1] $k\leq l{{l}\choose{\frac{l}{r}}}$
Goparaju et al. [@Re:S.Goparaju1] $ k\leq l^2$ $\textbf{\&}$ $k\leq 2\lambda\log_2l$, where $\lambda\leq\left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$
Wang et al.[@Re:Z.Wang2] the longest known scheme with systematic-length $k=(r+1)\log_rl$
This Paper $k\leq \frac{l^2-1}{r-1}$ $\textbf{\&}$ $k\leq 2\lambda\log_rl$, where $\lambda:\left\{\begin{array}{ll}
=r, &\textrm{when} \quad t=r;\\
\leq t+1+\left \lfloor\log_{\frac{r}{r-1}}\frac{(r-t)l}{r}\right \rfloor, &\textrm{when} \quad 1\leq t<r.
\end{array}\right.$
------------------------------------------------------------------------------------------------------------------------------------------------------------------
Organization
------------
Section 2 gives preliminaries including the basic system setting, the technique of interference alignment and two useful properties related to subspaces. Section 3 presents the detailed proof of our new results for upper bounds on systematic-length. Section 4 concludes this paper.
PRELIMINARIES
=============
In this section, we first describe the system setting about the basic construction of linear systematic-repair MSR codes. Then, the technique of interference alignment is introduced, where some equivalent transformations are applied. Last, we formalize two properties of subspaces and operators, which will be used to find new linearly independent matrices.
System Setting
--------------
Consider a linear systematic-repair MSR code with $\{n=k+r,k,l\}$, where each of $n$ nodes stores $l$ data symbols over a finite field $\mathbb{F}$ such that any $k$ nodes are sufficient to recover the $k\cdot l$ original data symbols. In light of its systematic feature, we let $(W_1,\cdots,W_k)$ denote these original data symbols stored in the $k$ systematic nodes. Each of the remaining $r$ nodes, as a parity node, stores a linear combination of $(W_1,\cdots,W_k)$. Formally speaking, the $l$ encoded symbols of $W_{k+u}$ stored in parity node $u$ for each $u\in[1,r]$, can be expressed as $$\label{encode}
W_{k+u}=\sum_{j=1}^{k}\mathbf{C}_{u,j}W_j,$$ where $\mathbf{C}_{u,j}$ is a square encoding matrix of order $l$ corresponding to the parity node $u$ and the systematic node $j\in[1,k]$. Hence, any linear systematic-repair MSR code can be uniquely represented as follows $$\label{encode}
\mathbb{C}=(\mathbf{C}_{u,j})_{\{u\in[1,r],j\in[1,k]\}}=\left[
\begin{array}{ccccc}
\mathbf{C}_{1,1} &\cdots & \mathbf{C}_{1,k} \\
\vdots &\ddots &\vdots \\
\mathbf{C}_{r,1} &\cdots &\mathbf{C}_{r,k} \\
\end{array}
\right].$$ The MDS property of systematic-repair MSR codes requires that any $r$ node failures can be recovered, which is equivalent to that any block submatrix of (\[encode\]) has to be invertible. Consequently, each encoding matrix $\mathbf{C}_{u,j}$ also must be invertible.
In order to optimally repair any systematic node $i\in[1,k]$, the remaining $n-1$ nodes are required to send a fraction $\frac{l}{r}$ of their data stored, i.e., each helper node $\{\nu\neq i|\nu\in[1,n]\}$ sends $\frac{l}{r}$ repair data symbols represented by $\mathbf{S}_{i,\nu}W_\nu$, where $\mathbf{S}_{i,\nu}$ is a matrix over $\mathbb{F}^{\frac{l}{r}\times l}$. To this end, each repair data made of $\frac{l}{r}$ data symbols sent from helper node $\nu$ to systematic node $i$ can be regarded as the projection of $W_\nu$ onto a subspace of dimension $\frac{l}{r}$, which corresponds to the subspace spanned by the rows of $\mathbf{S}_{i,\nu}$ that is sequentially called repair subspace.
\[note1\] To avoid confusion of the subsequent discussion, we let $``\sum"$ denote the normal sum of matrices, $``\biguplus"$ be the sum of subspaces (spanned by the rows of the corresponding matrices) and $``\bigoplus"$ represent the direct sum of certain subspaces that mutually intersect trivially. Besides, $\mathbf{S}_{i}$ and $\mathbf{S}_{i}\mathbf{C}_{u,j}$ can represent either a $(\frac{l}{r}\times l)$ matrix or a subspace of dimension $\frac{l}{r}$. Furthermore, we let $``0"$ denote the scalar zero, zero-matrix or zero-subspace, which will be clear from the context.
Interference Alignment
----------------------
As shown in [@Re:Z.Wang1], a linear systematic-repair MSR code has to meet the requirement of interference alignment. Optimal repair of a systematic node $i$ is possible if and only if there exist $\big\{\mathbf{S}_{i,\nu}|\nu\neq i,\nu\in[1,n]\big\}$ satisfying, for any $u\in[1,r]$ and $j\in[1,k]$ with $j\neq i$, $$\label{equ}
\left\{\begin{aligned}
&\mathbf{S}_{i,j}\simeq\mathbf{S}_{i,k+u}\mathbf{C}_{u,j}\\
&\biguplus_{u=1}^{r}\mathbf{S}_{i,k+u}\mathbf{C}_{u,i} \simeq \mathbb{F}^l,
\end{aligned}\right.$$ where $``\simeq"$ means that the subspaces spanned by the rows of the corresponding matrices are identical. Besides, the sum of the subspaces $``\biguplus"$ must be a direct sum $``\bigoplus"$, since the dimension of each subspace $\{\mathbf{S}_{i,\nu}|\nu\neq i,\nu\in[1,n]\}$ is $\frac{l}{r}$ and each matrix $\mathbf{C}_{u,i}$ is invertible.
As explained in [@Re:I.Tamo1], each encoding matrix for one of the parity nodes in a linear systematic-repair MSR code can be assumed to be the identity matrix $\mathbf{I}$, i.e., $\mathbf{C}_{1,i}=\mathbf{I}$ for any $i\in[1,k]$. Thereafter, the authors prove that if there exists a $\{k+r+1,k+1,l\}$ linear systematic-repair MSR code, then there also exists a $\{k+r,k,l\}$ linear systematic-repair MSR code where the repair subspaces are independent of the helper nodes. It indicates that, for a given $i\in[1,k]$, $\mathbf{S}_{i,\nu}$ stays unchanged for any $\nu\in[1,n]$ and $\nu\neq i$, which can be replaced by $\mathbf{S}_{i}$ accordingly.
Henceforth, the subspace conditions (\[equ\]) almost can be equivalently transformed into: $$\label{interference}
\left\{\begin{aligned}
&\mathbf{S}_{i}\simeq\mathbf{S}_{i}\mathbf{C}_{u,j}\\
&\bigoplus_{u=2}^{r}\mathbf{S}_{i}\mathbf{C}_{u,i}\oplus \mathbf{S}_{i} \simeq \mathbb{F}^l,
\end{aligned}\right.$$ where $i,j\in[1,k]$ with $j\neq i$ and $\mathbf{S}_{i}$ is the repair matrix (or viewed as the independent repair subspace) for systematic node $i$.
Extended Properties of Subspaces
--------------------------------
From the subspace conditions (\[interference\]), two derivative extended properties are formally presented as follows, which can be counted as a generalization of some results derived in [@Re:S.Goparaju1].
\[premise1\] Given an $\{n=k+r,k,l\}$ linear systematic-repair MSR code with independent repair subspaces, for any $u_1,u_2\in[1,r]$ and any $i,j_1,j_2\in[1,k]$ such that $j_1\neq i$ and $j_2\neq i$, it must be that $$\left\{\begin{aligned}
&\mathbf{S}_{i}(\mathbf{C}_{u_1,j_1}+\mathbf{C}_{u_2,j_2})\preceq\mathbf{S}_{i};\\
&\mathbf{S}_{i}\mathbf{C}_{u_1,j_1}\mathbf{C}_{u_2,j_2}\preceq\mathbf{S}_{i},
\end{aligned}\right.$$ where $``\preceq"$ represents a symbol of inclusion for subspaces ($x\preceq y$ means that the subspace $x$ lies in the subspace $y$).
According to the first item in (\[interference\]), $``\mathbf{S}_{i}\simeq\mathbf{S}_{i}\mathbf{C}_{u,j}"$ signifies that the subspace spanned by the $\frac{l}{r}$ row vectors of $\mathbf{S}_{i}$ is same as the one spanned by the $\frac{l}{r}$ row vectors of $\mathbf{S}_{i}\mathbf{C}_{u,j}$. Equivalently speaking, there exists an invertible $(\frac{l}{r}\times \frac{l}{r})$ matrix $\mathbf{P}_{u,i,j}$ such that $\mathbf{S}_{i}\mathbf{C}_{u,j}=\mathbf{P}_{u,i,j}\mathbf{S}_{i}$.
Thereby, we have $$\left\{\begin{aligned}
&\mathbf{S}_{i}(\mathbf{C}_{u_1,j_1}+\mathbf{C}_{u_2,j_2})=(\mathbf{P}_{u_1,i,j_1}+\mathbf{P}_{u_2,i,j_2})\mathbf{S}_{i};\\
&\mathbf{S}_{i}\mathbf{C}_{u_1,j_1}\mathbf{C}_{u_2,j_2}=\mathbf{P}_{u_1,i,j_1}\mathbf{P}_{u_2,i,j_2}\mathbf{S}_{i}.
\end{aligned}\right.$$ From the above equation, it is clear that $\mathbf{S}_{i}\mathbf{C}_{u_1,j_1}\mathbf{C}_{u_2,j_2}\simeq\mathbf{S}_{i}$, since $\mathbf{P}_{u_1,i,j_1}\mathbf{P}_{u_2,i,j_2}$ is an invertible matrix. Moreover, because $\mathbf{P}_{u_1,i,j_1}+\mathbf{P}_{u_2,i,j_2}$ also is a $(\frac{l}{r}\times \frac{l}{r})$ matrix, it apparently leads to that $(\mathbf{P}_{u_1,i,j_1}+\mathbf{P}_{u_2,i,j_2})\mathbf{S}_{i}\preceq\mathbf{S}_{i}$.$\hfill\blacksquare$
Lemma \[premise1\] indicates the closure of operations on $\mathbf{S}_{i}$ by the addition and multiplication of $\mathbf{C}_{u,j}$ for any $j\neq i$ and $u\in[2,r]$. Consequently, any subspace generated from operating on $\mathbf{S}_{i}$ by a matrix made up of arbitrary times of addition and multiplication of matrices $\mathbf{C}_{u,j}$ still lies in $\mathbf{S}_{i}$, for $j\neq i$.
\[premise2\] Given an $\{n=k+r,k,l\}$ linear systematic-repair MSR code with independent repair subspaces, for any $i\in[1,k]$, if there exist $r$ matrices $\{\mathbf{\Theta}_1,\cdots,\mathbf{\Theta}_r\}$ of size $(\frac{l}{r}\times l)$ such that $\mathbf{\Theta}_u\preceq\mathbf{S}_{i}\mathbf{C}_{u,i}$ for each $u\in[1,r]$ and $\sum_{u=1}^r\mathbf{\Theta}_u=0$, then it must be that $$\mathbf{\Theta}_u=0.$$
According to the second item in (\[interference\]), $``\bigoplus_{u=2}^{r}\mathbf{S}_{i}\mathbf{C}_{u,i}\oplus \mathbf{S}_{i} \simeq \mathbb{F}^l"$ indicates that all the $r$ subspaces $\{\mathbf{S}_{i}\mathbf{C}_{u,i}|u\in[1,r]\}$ where $\mathbf{C}_{1,i}=\mathbf{I}$, exactly span the entire space $\mathbb{F}^l$. That is to say, the total $l$ row vectors of $\{\mathbf{S}_{i}\mathbf{C}_{u,i}|u\in[1,r]\}$ are linearly independent. Thus, it leads to that any $r$ non-zero vectors coming from $r$ different subspaces $\mathbf{S}_{i}\mathbf{C}_{u,i}$ for $u\in[1,r]$ are also linearly independent.
Hereby, we assume there are non-zero matrices $\mathbf{\Theta}_u$ for some $u\in[1,r]$. However, the condition $\sum_{u=1}^r\mathbf{\Theta}_u=0$ implies that all non-zero vectors of the same row within these non-zero matrices $\mathbf{\Theta}_u$ are not linearly independent. Since these non-zero vectors lie in the corresponding subspaces $\mathbf{S}_{i}\mathbf{C}_{u,i}$, contradiction arises.$\hfill\blacksquare$
Lemma \[premise2\] in fact is a derivative property of direct sum of linear subspaces, which means that those vectors from different linear subspaces that mutually intersect trivially are linearly independent.
Subsequently, we use Lemma \[premise1\] and Lemma \[premise2\] to find new linearly independent matrices, which will naturally offer the upper bounds on the systematic-length $k$.
IMPROVED SYSTEMATIC-LENGTH UPPER BOUNDS
=======================================
In this section, we will present two new upper bounds on systematic-length. The first one is a simple quadratic bound, which stems from the linear independence of encoding matrices. The second one is an $r$-based-log upper bound, which is based on the partition of systematic nodes $[1,k]$, where repair subspaces within each standard partition span the entire space $\mathbb{F}^l$. Furthermore, we estimate the size of the standard partition and introduce an explicit result depending on the value of $\frac{r^2}{l}$, which can be integrated into the $r$-based-log bound.
A Simple Quadratic Bound
------------------------
\[scalar\] For any $\{n=k+r,k,l\}$ linear systematic-repair MSR code with independent repair subspaces, the following upper bound holds $$k\leq \frac{l^2-1}{r-1}.$$
We proceed to show that the $k(r-1)+1$ number of matrices $\Big\{\mathbf{I},\mathbf{C}_{u,i}|u\in[2,r],i\in[1,k]\Big\}$ are linearly independent.
Assume there exist $k(r-1)+1$ coefficients $a_{u,i}$ and $b$ in $\mathbb{F}$ such that $$\label{nondependent}
\sum_{i=1}^{k}\sum_{u=2}^{r} a_{u,i}\mathbf{C}_{u,i}+b\mathbf{I}=0,$$ where $``0"$ represents the zero-matrix.
Then, operating the above equation on $\mathbf{S}_{i}$ for each $i\in [1,k]$, we have $$\label{independent matrix}
\sum_{u=2}^{r} a_{u,i}\mathbf{S}_{i}\mathbf{C}_{u,i}+(\sum_{j\neq i}\sum_{u=2}^{r}a_{u,j}\mathbf{S}_{i}\mathbf{C}_{u,j}+b\mathbf{S}_{i})=0,$$ where Lemma \[premise1\] leads to that $(\sum_{j\neq i}\sum_{u=2}^{r}a_{u,j}\mathbf{S}_{i}\mathbf{C}_{u,j}+b\mathbf{S}_{i})\preceq\mathbf{S}_{i}$. With Lemma \[premise2\], we know that, for any $u\in[2,r]$ and $i\in[1,k]$, $$a_{u,i}=0.$$ Combining with equation (\[nondependent\]), we further derive $$b=0,$$ which exactly means that the $k(r-1)+1$ matrices $\Big\{\mathbf{I},\mathbf{C}_{u,i}|u\in[2,r],i\in[1,k]\Big\}$ are linearly independent[^8].
Since they all lie in the $l^2$-dimensional space of matrices $\mathbb{F}^{l\times l}$, we have $$k(r-1)+1\leq l^2\Leftrightarrow k\leq \frac{l^2-1}{r-1}.$$$\hfill\blacksquare$
In [@Re:S.Goparaju1], Goparaju et al. obtain that $k\leq l^2$ under the case of two parity nodes. Here, we look into the case of arbitrary number of parity nodes $r$ and derive that $k\leq \frac{l^2-1}{r-1}$.
In addition, we find that when $l=r$, it has to be that $k\leq r+1$. Without the premise that the repair subspaces are independent of the helper nodes, we further have $k\leq r+2$ that is exactly consistent with the corresponding result in [@Re:N.B.Shah], i.e., scalar linear systematic-repair MSR codes only exist when $d\geq 2k-3$ for $d=n-1$. That is to say, in the scalar linear systematic-repair MSR scenario, $k=r+2$ is the longest systematic-length.
An $r$-Based-Log Bound
----------------------
\[log\] For any $\{n=k+r,k,l\}$ linear systematic-repair MSR code with independent repair subspaces, the following upper bound holds $$k\leq 2\lambda\log_rl,$$ where $\lambda$ as defined in Footnote 4 represents the size of each partition.
We follow the idea of partition considered in [@Re:S.Goparaju1], while we employ the addition of matrices within each partition instead of matrices’ multiplication therein. The basic difference is that matrix’s addition has the commutative law while multiplication of matrix is non-commutative. We analyze two situations as follows.
**1. The situation when $\lambda$ divides $k$.**
Divide the indices of systematic nodes $[1,k]$ into $\frac{k}{\lambda}$ disjoint partitions $\{\mathcal{X}_1,\mathcal{X}_2,\cdots,\mathcal{X}_{\frac{k}{\lambda}}\}$, where $\mathcal{X}_i=[1+(i-1)\lambda,\cdots,i\lambda]$. Accordingly, partition $\{\mathbf{S}_1,\cdots,\mathbf{S}_{k}\}$ and $\Big\{\mathbf{I},\mathbf{C}_{u,i}|u\in[2,r],i\in[1,k]\Big\}$ as follows $$\left[
\begin{array}{ccc;{3pt/3pt}ccc;{3pt/3pt}c;{3pt/3pt}ccc}
\quad \mathbf{S}_1 &\cdots & \mathbf{S}_{\lambda}\quad & \quad \mathbf{S}_{\lambda+1} & \cdots & \mathbf{S}_{2\lambda} \quad & \quad \cdots \quad & \quad\mathbf{S}_{k-\lambda+1} &\cdots &\mathbf{S}_{k}\quad \\ \hline
\quad \mathbf{I} & \cdots & \mathbf{I} \quad& \quad \mathbf{I} & \cdots & \mathbf{I} \quad &\quad\cdots\quad &\quad\mathbf{I} &\cdots & \mathbf{I}\quad \\ \hdashline[2pt/2pt]
\quad \mathbf{C}_{2,1} & \cdots & \mathbf{C}_{2,\lambda} \quad & \quad \mathbf{C}_{2,\lambda+1} & \cdots & \mathbf{C}_{2,2\lambda} \quad &\quad\cdots\quad &\quad\mathbf{C}_{2,k-\lambda+1} &\cdots & \mathbf{C}_{2,k}\quad \\\hdashline[2pt/2pt]
\quad \vdots & \cdots & \vdots \quad & \quad \vdots & \cdots & \vdots\quad &\quad\cdots\quad &\quad\vdots &\cdots & \vdots \quad\\\hdashline[2pt/2pt]
\quad \mathbf{C}_{r,1} & \cdots & \mathbf{C}_{r,\lambda} \quad & \quad \mathbf{C}_{r,\lambda+1} & \cdots & \mathbf{C}_{r,2\lambda}\quad &\quad\cdots\quad &\quad\mathbf{C}_{r,k-\lambda+1} &\cdots & \mathbf{C}_{r,k} \quad\\
\end{array}
\right],$$ where all entries of the second row are set to be the identity matrix $\mathbf{I}$.
**Setting**: For each $u\in [1,r]$ and $i\in[1,\frac{k}{\lambda}]$, define $$\mathbf{\Gamma}_{i,u}=\left\{\begin{aligned}
&\mathbf{I},~~ if ~~ u=1;\\
&\sum_{j\in \mathcal{X}_i}\mathbf{C}_{u,j}, ~~ if~~ u\in[2,r].
\end{aligned}\right.$$
Then, for any $u_i\in[1,r]$ and $i\in[1,\frac{k}{\lambda}]$, we define $$\mathbf{\Delta}_{u_1u_2\cdots u_{\frac{k}{\lambda}}}=\prod_{i=1}^{\frac{k}{\lambda}}\mathbf{\Gamma}_{i,u_i}.$$ Thus, we totally obtain $r^{\frac{k}{\lambda}}$ square matrices formed as $\mathbf{\Delta}_{u_1u_2\cdots u_{\frac{k}{\lambda}}}$ for $u_i\in[1,r]$, which in fact all are non-zero matrices (It is proved in **Appendix**). In the following, we further show that these $r^{\frac{k}{\lambda}}$ square matrices are also linearly independent.
**Induction claim**: Assume for some $s\in [1,\frac{k}{\lambda}]$, the following $r^s$ number of square matrices $$\left\{\mathbf{\Delta}_{u_1u_2\cdots u_s}=\prod_{i=1}^{s}\mathbf{\Gamma}_{i,u_i}\mid u_i\in[1,r],i\in[1,s]\right\}$$ are linearly independent.
**Base case**: For $s=1$, we have $$\label{basematrix}
\Big\{\mathbf{\Delta}_{u_1}\mid u_1\in[1,r]\Big\}=\Big\{\mathbf{\Gamma}_{1,u_1}\mid u_1\in[1,r]\Big\}
=\left\{\mathbf{I},\sum_{j=1}^{\lambda}\mathbf{C}_{2,j},\cdots,
\sum_{j=1}^{\lambda}\mathbf{C}_{r,j}\right\}.$$ Assume there exist $r$ coefficients $(a_1,\cdots,a_r)$ in $\mathbb{F}$ such that $$\label{base}
a_1\mathbf{I}+a_2\sum_{j=1}^{\lambda}\mathbf{C}_{2,j}+\cdots+a_r\sum_{j=1}^{\lambda}\mathbf{C}_{r,j}=0.$$ Operating equation (\[base\]) on $\mathbf{S}_i$ for any $i\in[1,\lambda]$, we have $$\sum_{u=2}^r a_u\mathbf{S}_i\mathbf{C}_{u,i}+(\sum_{j\neq i}\sum_{u=2}^r a_u\mathbf{S}_i\mathbf{C}_{u,j}+a_1\mathbf{S}_i)=0,$$ where $\sum_{j\neq i}\sum_{u=2}^r a_u\mathbf{S}_i\mathbf{C}_{u,j}+a_1\mathbf{S}_i\preceq\mathbf{S}_i$ following from Lemma \[premise1\]. According to Lemma \[premise2\], we can similarly derive $$a_1=a_2=\cdots=a_r=0,$$ which means the $r$ matrices in (\[basematrix\]) are linearly independent.
**Inductive step**: Let the inductive claim hold for some $s$, then it is true for $s+1$. Otherwise, we have $$\mathbf{\Psi}_1^{(s)}+\mathbf{\Psi}_2^{(s)}\mathbf{\Gamma}_{s+1,2}+\cdots+\mathbf{\Psi}_r^{(s)}\mathbf{\Gamma}_{s+1,r}=0,$$ where $\mathbf{\Psi}_u^{(s)}$ for each $u\in[1,r]$ are linear combinations of elements formed as $\mathbf{\Delta}_{u_1u_2\cdots u_s}$[^9].
Operating the above equation on $\mathbf{S}_i$ for each $i$ in $\mathcal{X}_{s+1}=[s\lambda+1,(s+1)\lambda]$, we have $$\label{equiv}
\mathbf{S}_i\mathbf{\Psi}_1^{(s)}+\mathbf{S}_i\mathbf{\Psi}_2^{(s)}(\sum_{j\in\mathcal{X}_{s+1}}\mathbf{C}_{2,j})+\cdots
+\mathbf{S}_i\mathbf{\Psi}_r^{(s)}(\sum_{j\in\mathcal{X}_{s+1}}\mathbf{C}_{r,j})=0,$$ which can be rearranged to $$\mathbf{S}_i\Big\{\mathbf{\Psi}_1^{(s)}+\mathbf{\Psi}_2^{(s)}(\sum_{j\neq i}\mathbf{C}_{2,j})+\cdots
+\mathbf{\Psi}_r^{(s)}(\sum_{j\neq i}\mathbf{C}_{r,j})\Big\}+\mathbf{S}_i\mathbf{\Psi}_2^{(s)}\mathbf{C}_{2,i}+\cdots
+\mathbf{S}_i\mathbf{\Psi}_r^{(s)}\mathbf{C}_{r,i}=0.$$ Similarly, Lemma \[premise1\] leads to that $$\left\{\begin{aligned}
&\mathbf{S}_i\Big\{\mathbf{\Psi}_1^{(s)}+\mathbf{\Psi}_2^{(s)}(\sum_{j\neq i}\mathbf{C}_{2,j})+\cdots+\mathbf{\Psi}_r^{(s)}(\sum_{j\neq i}\mathbf{C}_{r,j})\Big\}\preceq\mathbf{S}_i;\\
&\mathbf{S}_i\mathbf{\Psi}_{u}^{(s)}\mathbf{C}_{u,i}\preceq\mathbf{S}_i \mathbf{C}_{u,i}, ~~for~~u\in[2,r].
\end{aligned}\right.$$
By Lemma \[premise2\], we know for each $u\in[2,r]$ and $i\in\mathcal{X}_{s+1}$, $$\mathbf{S}_i\mathbf{\Psi}_{u}^{(s)}\mathbf{C}_{u,i}=0,$$ from which we derive $\mathbf{S}_i\mathbf{\Psi}_{u}^{(s)}=0$ for $\mathbf{C}_{u,i}$ is invertible. Combining with equation (\[equiv\]), we further know $\mathbf{S}_i\mathbf{\Psi}_1^{(s)}=0$. Thus, for each $u\in[1,r]$ and $i\in\mathcal{X}_{s+1}$, we have $\mathbf{S}_i\mathbf{\Psi}_{u}^{(s)}=0$.
Due to the condition that $\biguplus_{i\in\mathcal{X}_{s+1}}\mathbf{S}_{i}\simeq \mathbb{F}^l$, it has to be that $\mathbf{\Psi}_u^{(s)}=0$ for each $u\in[1,r]$, which contradicts the induction assumption.
**Conclusion**: That is to say, the $r^{\frac{k}{\lambda}}$ square matrices formed by $\mathbf{\Delta}_{u_1u_2\cdots u_{\frac{k}{\lambda}}}$ are linearly independent. Because they all lie in the $l^2$-dimensional space of matrices $\mathbb{F}^{l\times l}$, we have $$r^{\frac{k}{\lambda}}\leq l^2\Leftrightarrow k\leq 2\lambda\log_rl.$$
**2. The situation when $\lambda$ does not divide $k$.**
Assume $p\lambda<k<(p+1)\lambda$, i,e., $p=\left \lfloor\frac{k}{\lambda}\right \rfloor$ and $p+1=\left\lceil\frac{k}{\lambda}\right \rceil$. Then, divide $[1,k]$ into $p+1$ disjoint sets as $\mathcal{X}_1=[1,\cdots,k-p\lambda]$ and $\mathcal{X}_i=[k-(p+2-i)\lambda+1,\cdots,k-(p+1-i)\lambda]$, where $|\mathcal{X}_1|=k-p\lambda$ and $|\mathcal{X}_i|=\lambda$ for each $i\in[2,p+1]$. It should be noted here that $\mathcal{X}_1$ is not a standard partition and repair subspaces within it cannot span the whole space $\mathbb{F}^l$, because $|\mathcal{X}_1|<\lambda$.
As in the first situation, we only need to consider the base case, since we apply the same method of induction. In this case, by operating on $\mathbf{S}_i$ for any $i\in[1,k-p\lambda]$, it is trivial to verify that $$\Big\{\mathbf{\Gamma}_{1,u_1}\mid u_1\in[1,r]\Big\}
=\left\{\mathbf{I},\sum_{j=1}^{k-p\lambda}\mathbf{C}_{2,j},\cdots,
\sum_{j=1}^{k-p\lambda}\mathbf{C}_{r,j}\right\}$$ are also linearly independent. Thus, we totally obtain $r^{p+1}$ linearly independent square matrices $$\left\{\mathbf{\Delta}_{u_1u_2\cdots u_{p+1}}=\prod_{i=1}^{p+1}\mathbf{\Gamma}_{i,u_i}\mid u_i\in[1,r],i\in[1,p+1]\right\}.$$
Therefore, we derive $$r^{\left\lceil\frac{k}{\lambda}\right \rceil}\leq l^2\Rightarrow k< 2\lambda\log_rl.$$$\hfill\blacksquare$
In [@Re:S.Goparaju1], Goparaju et al. derive that $k\leq 2\lambda\log_2l$, where $\lambda\leq\left\lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$. Here, we improve the upper bound from log-base $2$ to log-base $r$ and obtain that $k\leq2\lambda\log_rl$.
In subsequent, we estimate the size of each standard partition $\lambda$, with which we will further give an explicit upper bound.
An Explicit $\frac{r^2}{l}$-Dependent Bound
-------------------------------------------
In the following, we use our previous work [@Kun] to estimate the size of standard partition $\lambda$, combining which with Theorem \[log\] we will further present an explicit bound depending on the value of $\frac{r^2}{l}$.
### 3.3.1 Estimation of $\lambda$.
In [@Re:S.Goparaju1], Goparaju et al. apply the geometric analysis of the invariant subspace and derive that for any $m\in [1,k]$, $$\label{pre}
\dim(\biguplus_{i=1}^{m}\mathbf{S}_i)=\dim(\mathbf{S}_1\uplus\mathbf{S}_2\uplus\cdots\uplus\mathbf{S}_m)\geq \big(1-(\frac{r-1}{r})^{m}\big)l,$$ from which they obtain that $\lambda\leq\left\lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$, i.e., if the size of a standard partition equals with $\left\lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$, the repair subspaces within it necessarily span the entire space $\mathbb{F}^l$. In addition, they find that formula (\[pre\]) is useful for studying the secrecy capacity of linear MSR codes in [@Re:S.Goparaju]. We also focus on this secrecy issue in [@Kun] and ultimately present an explicit result on secrecy capacity in the linear MSR scenario that extends the result given in [@Re:S.Goparaju]. As shown in Theorem $6$ of [@Kun], it closely depends on the value of $\beta$ that is equal to $\frac{l}{r}$ herein. The following theorem can be looked upon as a direct corollary of formula $(72)$ of Theorem $6$ given in [@Kun].
\[class\] Given an $\{n=k+r,k,l\}$ linear systematic-repair MSR code with independent repair subspaces, for any set $F\subsetneq[1,k]$ with $|F|=m$, we have $$\label{part}
\dim(\biguplus_{i\in F}\mathbf{S}_i):\left\{\begin{array}{ll}
=m\frac{l}{r}, &\textrm{if} \quad m\leq t;\\
\geq t\frac{l}{r}+\frac{l}{r}(r-t)[1-(\frac{r-1}{r})^e]= l-\frac{l}{r}(r-t)(\frac{r-1}{r})^{e}, &\textrm{if} \quad m=t+e,
\end{array}\right.$$ where $t=\lceil\frac{r^2}{l}\rceil$ and $e\geq 1$.
As stated in Remark $11$ of [@Kun], all results therein in fact also apply to MSR codes with optimal repair of systematic nodes only, although the general case for MSR codes with optimal repair of all nodes is the study object in [@Kun]. In this paper, we let a linear systematic-repair MSR code represented by $\{n=k+r,k,l\}$, where $n=d+1$, $l=\alpha$ and $\frac{l}{r}=\beta$. Through changing the expression of related parameters, formula $(72)$ of Theorem $6$ in [@Kun] can be equivalently transformed into as follows $$\label{part}
\dim(\biguplus_{i\in F}\mathbf{S}_i):\left\{\begin{array}{ll}
=m\frac{l}{r}, & if \quad m\leq t, \frac{l}{r}<\frac{r}{t-1};\\
\geq t\frac{l}{r}+\frac{l}{r}(r-t)[1-(\frac{r-1}{r})^e]= l-\frac{l}{r}(r-t)(\frac{r-1}{r})^{e}, & if \quad m=t+e, \frac{r}{t}\leq \frac{l}{r}<\frac{r}{t-1},
\end{array}\right.$$ where $1\leq t\leq r$ and $e\geq 1$. Since $t$ is an integer satisfying the condition that $\Big\{\frac{r}{t}\leq \frac{l}{r}<\frac{r}{t-1}\Leftrightarrow\frac{r^2}{l}\leq t<\frac{r^2}{l}+1\Big\}$, we can rewrite it as $t=\lceil\frac{r^2}{l}\rceil$, where it is trivial that $1\leq t\leq r$.$\hfill\blacksquare$
\[verify\] It has been verified in [@Kun] that, $$\left\{\begin{aligned}
&t\frac{l}{r}+\frac{l}{r}(r-t)[1-(\frac{r-1}{r})^e]<t\frac{l}{r}+\frac{l}{r}(r-t)[e(1-\frac{r-1}{r})]
=t\frac{l}{r}+e\frac{l}{r}(1-\frac{t}{r})<(t+e)\frac{l}{r}=m\frac{l}{r};\\
&t\frac{l}{r}+\frac{l}{r}(r-t)[1-(\frac{r-1}{r})^e]=l[1-(\frac{r-t}{r})(\frac{r-1}{r})^e]
\geq l[1-(\frac{r-1}{r})^t(\frac{r-1}{r})^e]=[1-(\frac{r-1}{r})^{m}]l,\\
\end{aligned}\right.$$ which implies that when $\dim(\biguplus_{i\in F}\mathbf{S}_i)\leq l$, $$\label{compare}
[1-(\frac{r-1}{r})^{m}]l\leq\dim(\biguplus_{i\in F}\mathbf{S}_i)\leq m\frac{l}{r}.$$
As assumed in [@Re:S.Goparaju1], we also make hypothesis that $\lambda< k$, i.e., there exists at least one standard partition properly included in $[1,k]$. Otherwise, we have $k\leq\lambda$. Next, we apply Theorem \[class\] to estimate $\lambda$, the smallest value of $m$ satisfying $\dim(\biguplus_{i\in F}\mathbf{S}_i)=l$.
\[partition\] Given an $\{n=k+r,k,l\}$ linear systematic-repair MSR code with independent repair subspaces, we have $$\lambda:\left\{\begin{array}{ll}
=r, &\textrm{if} \quad t=r;\\
\leq t+1+\left \lfloor\log_{\frac{r}{r-1}}\frac{(r-t)l}{r}\right \rfloor, &\textrm{if} \quad 1\leq t<r.
\end{array}\right.$$ where $t=\lceil\frac{r^2}{l}\rceil$.
There are two cases analyzed as follows.
**1. The case when $t=r$.**
According to the first item of Theorem \[class\], we know when $m=r$, it is clear that $\dim(\biguplus_{i\in F}\mathbf{S}_i)=l$. Thus, we have $\lambda=r$.
**2. The case when $1\leq t<r$.**
According to the second item of Theorem \[class\], we know when $m=t+e$, then $\dim(\biguplus_{i\in F}\mathbf{S}_i)\geq l-\frac{l}{r}(r-t)(\frac{r-1}{r})^{e}$. Hence, when $\frac{l}{r}(r-t)(\frac{r-1}{r})^{e}<1$ or $e>\log_{\frac{r}{r-1}}\frac{(r-t)l}{r}$, it has to be that $\dim(\biguplus_{i\in F}\mathbf{S}_i)=l$. Thus, we obtain that $\lambda\leq t+e=t+1+\left \lfloor\log_{\frac{r}{r-1}}\frac{(r-t)l}{r}\right \rfloor$.$\hfill\blacksquare$
From Theorem \[partition\], we know when $t=1$, it is clear that $\lambda\leq2+\left \lfloor\log_{\frac{r}{r-1}}\frac{(r-1)l}{r}\right \rfloor=2+\left \lfloor\log_{\frac{r}{r-1}}l-1\right \rfloor=1+\left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor$, which is consistent with the corresponding result in [@Re:S.Goparaju1]. Besides, one can check that, for $1<t<r$, $$r<t+1+\left \lfloor\log_{\frac{r}{r-1}}\frac{(r-t)l}{r}\right \rfloor<1+\left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor,$$ which basically corresponds to the inequality (\[compare\]).
### 3.3.2 Final Explicit Result.
Eventually, combining Theorem \[log\] and Theorem \[partition\], we derive the final explicit bound as follows.
\[final\] For any $\{n=k+r,k,l\}$ linear systematic-repair MSR code with independent repair subspaces, the following upper bound holds $$k\leq 2\lambda\log_rl,$$ wherein $$\lambda:\left\{\begin{array}{ll}
=r, &\textrm{when} \quad t=r;\\
\leq t+1+\left \lfloor\log_{\frac{r}{r-1}}\frac{(r-t)l}{r}\right \rfloor, &\textrm{when} \quad 1\leq t<r,
\end{array}\right.$$ where $t=\lceil\frac{r^2}{l}\rceil$.
Through literature survey, MSR codes for $\{d=n-1\}$ known so far are generally divided into scalar MSR codes with $\{l=r\}$ and vector MSR codes with $\{l=r^{x+1},x\geq1\}$, which correspond to $t=r$ and $t=1$ respectively. As stated in [@Kun], it remains open for constructing vector MSR codes with $\{r<l<r^2\}$.
When $l=r$, Theorem \[scalar\] implies that $k=r+1$ is the longest systematic-length for the scalar linear systematic-repair MSR codes with independent repair subspaces. According to Theorem \[final\], we know that when $l=r$, it has to be that $\lambda=t=r$, following which we have $k\leq 2r$. Essentially, when $k=r+1$, it is evident that $\lambda=r$ does not divide $k$ and thus it should be that $k<2r$ following from Theorem \[log\]. The reason is that, for $k=r+1$, $k$ systematic nodes can only be split into one standard partition as well as an nonstandard partition, where the size of the nonstandard partition equals with $1$ that is smaller than $r$. In some sense, the $r$-based-log upper bound basically coincides with the real largest number of systematic nodes in the scalar scenario. In other words, the longest systematic-length $k=r+1$ is indeed included in the actual case of $k<2r$ derived from the $r$-based-log upper bound for $l=r$.
When $l\geq r^2$, it is clear that $t=1$. For example in Table \[Tab:Comparision\], the systematic-repair MSR code with the longest systematic-length known so far given in [@Re:Z.Wang2] is with $\big\{k=(r+1)\log_rl\big\}$. However, Theorem \[final\] leads to that $k\leq 2\log_rl\big(1+\left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor\big)$. By comparison, we find that $r+1$ is strictly less than $2\big(1+\left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor\big)$ even for $l=r^2$ and $r=2$, which means that the code in [@Re:Z.Wang2] cannot reach the derived upper bound. That is to say, although we do improve the upper bound from previous log-base $2$ to current log-base $r$, it is still unknown yet whether the upper bound in Theorem \[final\] can be achieved for some unexplored vector linear systematic-repair MSR code. Nevertheless, our $r$-based-log upper bound is nearly in line with the conjecture that $k$ is of the order of $\log_rl$ proposed by Tamo et al. in [@Re:I.Tamo1].
Further Discussions
-------------------
As stated in Remark \[verify\], our Theorem \[class\], Theorem \[partition\] and Theorem \[final\] all are based on the assumption $\lambda<k$, i.e., systematic part at least properly includes one standard partition, while Theorem \[log\] also applies to the case of $k=\lambda$ as the systematic part $[1,k]$ exactly constitutes a standard partition. Thereby, the situation $k\leq\lambda$ is left open.
When $k=\lambda$, it is clear that $k\leq 2\lambda\log_rl$ from Theorem \[log\], since it is trivial that $r\leq l$ for $r$ divides $l$. When $k<\lambda$, the systematic part cannot satisfy the condition for the formation of a standard partition, which implies the notation $\lambda$ is meaningless at present. In this case, we cannot obtain the log-based upper bound similarly. Hence, we put forward two questions as follows.
\[1\] *What value of $l$ and $r$ taken can we make the largest number of systematic node $k$ equal with $\lambda$ in the linear systematic-repair MSR scenario? What is the exact value of $k$ for this case?*
Technically, under the condition (\[interference\]), we assume $k_{\{l,r\}}$ is the longest systematic-length for given $l$ and $r$. Then, what values of $l$ and $r$ taken will make $k_{\{l,r\}}$ satisfy
$$\left\{\begin{aligned}
&\dim(\biguplus_{i=1}^{k_{\{l,r\}}} \mathbf{S}_{i})=l;\\
&\dim(\biguplus_{j=1}^{k_{\{l,r\}}-1} \mathbf{S}_{i_j})<l, for~ any~ distinct~~ i_j\in[1,k_{\{l,r\}}],
\end{aligned}\right.$$
where $\dim(\mathbf{S}_{i})=\frac{l}{r}$ for each $i\in[1,k_{\{l,r\}}]$.
\[2\] *What value of $l$ and $r$ taken will make the systematic part cannot form a standard partition in the linear systematic-repair MSR scenario? What is the largest number of systematic nodes $k$ now?*
With the same definition of $k_{\{l,r\}}$ as above, what range of values for $l$ and $r$ will make $k_{\{l,r\}}$ satisfy
$$\dim(\biguplus_{i=1}^{k_{\{l,r\}}} \mathbf{S}_{i})<l,$$
where $\dim(\mathbf{S}_{i})=\frac{l}{r}$ for each $i\in[1,k_{\{l,r\}}]$.
Question \[1\] and Question \[2\] are our next direction of study. It may help further improve the upper bound on systematic-length, since our $r$-based-log upper bound are built on the assumption $\dim(\biguplus_{i=1}^{k_{\{l,r\}}}\mathbf{S}_{i})>l$. The intuition establishes on the fact that the length $k_{\{l,r\}}$ of the systematic part herein is not larger than $\lambda$ the least number of systematic nodes for supporting a standard partition.
CONCLUSION
==========
Finding the exact systematic-length upper bound for a systematic-repair MSR code with $r$ parity nodes and storage capacity $l$ is an open problem. Following the method of geometric analysis on a set of subspaces and operators, we formalize two helpful derivative properties of subspaces under the case of arbitrary number of parity nodes. With them, we first demonstrate the linear independence among all encoding matrices and then design new linearly independent matrices based on the idea of partition, which naturally bound the systematic-length $k$. Finally, we derive a simple quadratic bound and an $r$-based-log bound, which both are superior to the previous results. Moreover, leveraging our prior work [@Kun], we estimate the size of a standard partition and further present an explicit bound depending on the value of $\frac{r^2}{l}$.
APPENDIX
========
In Theorem \[log\], we define $$\mathbf{\Delta}_{u_1u_2\cdots u_{\frac{k}{\lambda}}}=\prod_{i=1}^{\frac{k}{\lambda}}\mathbf{\Gamma}_{i,u_i},$$ where $$\mathbf{\Gamma}_{i,u}=\left\{\begin{aligned}
&\mathbf{I},\quad if ~ u=1;\\
&\sum_{j\in \mathcal{X}_i}\mathbf{C}_{u,j}, \quad if~ u\in[2,r],
\end{aligned}\right.$$ for $u_i\in[1,r]$ and $i\in[1,\frac{k}{\lambda}]$. In fact, all of these $r^{\frac{k}{\lambda}}$ matrices are non-zero.
We still employ the proof of contradiction.
**Induction claim**: Assume for some $s\in [1,\frac{k}{\lambda}]$, the following square matrix $$\mathbf{\Delta}_{u_1u_2\cdots u_s}=\prod_{i=1}^{s}\mathbf{\Gamma}_{i,u_i}\neq 0,$$ where $u_i\in[1,r]$.
**Base case**: For $s=1$, we have $$\mathbf{\Delta}_{u_1}=\mathbf{\Gamma}_{1,u_1},$$ where it is clear from the definition that $\mathbf{\Gamma}_{1,u_1}=\mathbf{I}\neq 0$ if $u_1=1$.
If $u_1\in[2,r]$, $\mathbf{\Gamma}_{1,u_1}=\sum_{j\in \mathcal{X}_1}\mathbf{C}_{u_1,j}$. According to Theorem \[scalar\], we know that the matrices $\{\mathbf{C}_{u_1,j}\mid j\in \mathcal{X}_1\}$ are linearly independent, from which it is apparent that $\mathbf{\Gamma}_{1,u_1}=\sum_{j\in \mathcal{X}_1}\mathbf{C}_{u_1,j}\neq 0$ for any $u_1\in[2,r]$.
**Inductive step**: Let the inductive claim hold for some $s$, then it is true for $s+1$. Otherwise, we have $$\mathbf{\Delta}_{u_1u_2\cdots u_{s+1}}=\prod_{i=1}^{s+1}\mathbf{\Gamma}_{i,u_i}=0.$$
1\. When $u_{s+1}=1$, we have $\mathbf{\Gamma}_{s+1,1}=\mathbf{I}$. Thus, it has to be that $\mathbf{\Delta}_{u_1u_2\cdots u_{s+1}}=\mathbf{\Delta}_{u_1u_2\cdots u_{s}}=0$, which contradicts the induction assumption.
2\. When $u_{s+1}\in[2,r]$, $\mathbf{\Gamma}_{s+1,u_{s+1}}=\sum_{j\in \mathcal{X}_{s+1}}\mathbf{C}_{u_{s+1},j}$, with which we have $$\mathbf{\Delta}_{u_1u_2\cdots u_{s+1}}=\mathbf{\Delta}_{u_1u_2\cdots u_{s}}(\sum_{j\in \mathcal{X}_{s+1}}\mathbf{C}_{u_{s+1},j})=0.$$ Operating the above equation on $\mathbf{S}_i$ for each $i$ in $\mathcal{X}_{s+1}=[s\lambda+1,(s+1)\lambda]$, we obtain $$\mathbf{S}_i\mathbf{\Delta}_{u_1u_2\cdots u_{s}}\mathbf{C}_{u_{s+1},i}+\mathbf{S}_i\mathbf{\Delta}_{u_1u_2\cdots u_{s}}(\sum_{j\neq i}\mathbf{C}_{u_{s+1},j})=0,$$ where $\mathbf{S}_i\mathbf{\Delta}_{u_1u_2\cdots u_{s}}\mathbf{C}_{u_{s+1},i}\preceq\mathbf{S}_i \mathbf{C}_{u_{s+1},i}$ and $\mathbf{S}_i\mathbf{\Delta}_{u_1u_2\cdots u_{s}}(\sum_{j\neq i}\mathbf{C}_{u_{s+1},j})\preceq\mathbf{S}_i$ from Lemma \[premise1\].
Further by Lemma \[premise2\], we derive, for each $i\in\mathcal{X}_{s+1}$, $$\mathbf{S}_i\mathbf{\Delta}_{u_1u_2\cdots u_{s}}\mathbf{C}_{u_{s+1},i}=0,$$ from which we have $\mathbf{S}_i\mathbf{\Delta}_{u_1u_2\cdots u_{s}}=0$, since $\mathbf{C}_{u_{s+1},i}$ is invertible. Under the condition $\biguplus_{i\in\mathcal{X}_{s+1}}\mathbf{S}_{i}\simeq \mathbb{F}^l$, it must be that $\mathbf{\Delta}_{u_1u_2\cdots u_{s}}=0$, which also contradicts the induction assumption.
**Conclusion**: That is to say, any square matrix formed as $\mathbf{\Delta}_{u_1u_2\cdots u_{\frac{k}{\lambda}}}$ is non-zero.$\hfill\blacksquare$
[12]{}
A. G. Dimakis, K. Ramchandran, Y. Wu, and C. Suh, “A survey on network codes for distributed storage," *Proc. IEEE*, vol. 99, no. 3, pp. 476–489, Mar. 2011.
M. Blaum, P. G. Farell, and H. van Tilborg, “Array codes," in *Handbook of Coding Theory*, V. Pless and W. C. Huffman, Eds. Elsevier Science, 1998, vol. II, ch. 22, pp. 1855–-1909.
H. Weatherspoon and J. D. Kubiatowicz, “Erasure coding vs. replication: A quantitiative comparison," in *Proc. Int. Workshop. Peer–Peer Syst.*, 2002, pp. 328–338.
A. G. Dimakis, P. B. Godfrey, Y. Wu, M. J. Wainwright, and K. Ramchandran, “Network coding for distributed storage systems," *IEEE Trans. Inf. Theory*, vol. 56, no. 9, pp. 4539–4551, Sep. 2010.
C. Huang, H. Simitci, Y. Xu, A. Ogus, B. Calder, P. Gopalan, J. Li, and S. Yekhanin, “Erasure coding in windows azure storage," presented at the USENIX Annu. Tech. Conf., Boston, MA, USA, Jun. 2012.
N. B. Shah, K. V. Rashmi, and P. V. Kumar, and K. Ramachandran, “Distributed storage codes with repair-by-transfer and nonachievability of interior points on the storage-bandwidth tradeoff," *IEEE Trans. Inf. Theory*, vol. 58, no. 3, pp. 1837–1852, Mar. 2012.
C. Tian, B. Sasidharan, V. Aggarwal, V. A. Vaishampayan, and P. V. Kumar, “Layered exact-repair regenerating codes via embedded error correction and block designs," *IEEE Trans. Inf. Theory*, vol. 61, no. 4, pp. 1933–1947, Apr. 2015.
T. Ernvall, “Codes between MBR and MSR points with exact repair property," *IEEE Trans. Inf. Theory*, vol. 60, no. 11, pp. 6993–7005, Nov. 2014.
K. V. Rashmi, N. B. Shah, and P. V. Kumar, “Optimal exact-regenerating codes for distributed storage at the MSR and MBR points via a product-matrix construction," *IEEE Trans. Inf. Theory*, vol. 57, no. 8, pp. 5227–5239, Aug. 2011.
C. Suh and K. Ramchandran, “Exact-repair MDS code construction using interference alignment," *IEEE Trans. Inf. Theory*, vol. 57, no. 3, pp. 1425–1442, Mar. 2011.
Y. Wu and A. G. Dimakis, “Reducing repair traffic for erasure coding-based storage via interference alignment," in *Proc. Int. Symp. Inf. Theory. (ISIT)*, Jul. 2009, pp. 2276–2280.
K. V. Rashmi, N. B. Shah, P. V. Kumar, and K. Ramchandran, “Explicit construction of optimal exact regenerating codes for distributed storage," in *Proc. 47th Annu. Allerton Conf. Commun., Control, Comput.*, Urbana-Champaign, Sep. 2009, pp. 1243–1249.
K. V. Rashmi, N. B. Shah, K. Ramchandran, and P. V. Kumar, “Regenerating codes for errors and erasures in distributed storage," in *Proc. Int. Symp. Inf. Theory. (ISIT)*, Jul. 2012, pp. 1202–1206.
N. B. Shah, K. V. Rashmi, P. V. Kumar, and K. Ramchandran, “Interference alignment in regenerating codes for distributed storage: Necessity and code constructions," *IEEE Trans. Inf. Theory*, vol. 58, no. 4, pp. 2134–2158, Apr. 2012.
Z. Wang, I. Tamo, and J. Bruck, “On codes for optimal rebuilding access," in *Proc. 49th Annu. Allerton Conf. Commun., Control, Comput.*, Monticello, IL, USA, Sep. 2011, pp. 1374–1381.
D. S. Papailiopoulos, A. G. Dimakis, and V. R. Cadambe, “Repair optimal erasure codes through hadamard designs," *IEEE Trans. Inf. Theory*, vol. 59, no. 5, pp. 3021–3037, May. 2013.
V. R. Cadambe, C. Huang, and J. Li, “Permutation code: Optimal exact-repair of a single failed node in MDS code based distributed storage systems," in *Proc. Int. Symp. Inf. Theory (ISIT)*, Jul. 2011, pp. 1225–-1229.
I. Tamo, Z. Wang, and J. Bruck, “Zigzag codes: MDS array codes with optimal rebuilding," *IEEE Trans. Inf. Theory*, vol. 59, no. 3, pp. 1597–1616, Mar. 2013.
Z. Wang, I. Tamo, and J. Bruck, “Long MDS codes for optimal repair bandwidth," in *Distrib. Inf. Syst. Group, Dept. Elect. Eng., California Inst. Technol*, Pasadena, CA, USA, Tech. Rep. ETR116, Jul. 2012, pp. 1182–1186.
Z. Wang, I. Tamo, and J. Bruck, “Explicit minimum storage regenerating codes," *IEEE Trans. Inf. Theory*, vol. 62, no. 8, pp. 4466–4480, Aug. 2016.
V. R. Cadambe, C. Huang, S. A. Jafar, and J. Li. (Jun. 2011). “Optimal repair of MDS codes in distributed storage via subspace interference alignment." \[Online\]. Available: http://arxiv.org/abs/1106.1250
V. R. Cadambe, C. Huang, J. Li, and S. Mehrotra, “Polynomial length MDS codes with optimal repair in distributed storage," in *Proc. Conf. Rec. Signals, Syst. Comput. 45th Asilomar Conf.*, Nov. 2011, pp. 1850–1854.
G. K. Agarwal, B. Sasidharan, and P. V. Kumar, “An alternate construction of an access-optimal regenerating code with optimal subpacketization level," in *Proc. Nat. Conf. Commun. (NCC)*, Feb. 2015, pp. 1–6.
B. Sasidharan, G. K. Agarwal, and P. V. Kumar, “A high-rate MSR code with polynomial sub-packetization level," in *Proc. Int. Symp. Inf. Theory (ISIT)*, Jul. 2015, pp. 2051–2055.
Y. S. Han, H. T. Pai, R. Zheng, and P. K. Varshney, “Update-efficient regenerating codes with minimum per-node storage," in *Proc. Int. Symp. Inf. Theory (ISIT)*, Jul. 2013, pp. 1436–1440.
J. Li, X. Tang, and U. Parampalli, “A framework of constructions of minimal storage regenerating codes with the optimal access/update property," *IEEE Trans. Inf. Theory*, vol. 61, no. 4, pp. 1920-–1932, Apr. 2015.
A. S. Rawat, O. O. Koyluoglu, and S. Vishwanath, “Progress on high-rate MSR codes: Enabling arbitrary number of helper nodes." \[Online\]. Available: http://arxiv.org/abs/1601.06362
N. Raviv, N. Silberstein, and T. Etzion, “Access-optimal MSR codes with optimal sub-packetization over small fields." \[online\]. Available: http://arxiv.org/abs/1505.00919
M. Ye and A. Barg, “Explicit constructions of optimal-access MDS codes with nearly optimal sub-packetization." \[online\]. Available: http://arxiv.org/abs/1605.08630
B. Sasidharan, M. Vajha, and P. V. Kumar, “An explicit, coupled-layer construction of a high-rate MSR code with low sub-packetization level, small field size and all-node repair." \[online\]. Available: http://arxiv.org/abs/1607.07335
S. Goparaju, A. Fazeli, and A. Vardy, “Minimum storage regenerating codes for all parameters," in *Proc. Int. Symp. Inf. Theory (ISIT)*, Jul. 2016, pp. 76–80.
V. R. Cadambe, S. Jafar, H. Maleki, K. Ramchandran, and C. Suh, “Asymptotic Interference Alignment for Optimal Repair of MDS Codes in Distributed Storage," *IEEE Trans. Inf. Theory*, vol. 59, no. 5, pp. 2974–2987, May. 2013.
I. Tamo, Z. Wang, and J. Bruck, “Access vs. bandwidth in codes for storage," *IEEE Trans. Inf. Theory*, vol. 60, no. 4, pp. 2028–2037, Apr. 2014.
S. Goparaju, I. Tamo, and R. Calderbank, “An improved sub-packetization bound for minimum storage regenerating codes," *IEEE Trans. Inf. Theory*, vol. 60, no. 5, pp. 2770–2779, May. 2014.
S. Goparaju, S. El Rouayheb, R. Calderbank, and H. V. Poor, “Data secrecy in distributed storage systems under exact repair," in *Proc. IEEE Int. Symp. Netw. Coding. (NETCOD)*, Calgary, AB, Canada, Jun. 2013, pp. 1–6.
K. Huang, U. Parampalli, and M. Xian, “On secrecy capacity of minimum storage regenerating codes," *IEEE Trans. Inf. Theory*, vol. 63, no. 3, pp. 1510–1524, Mar. 2017.
[^1]: $^1$Kun Huang and Ming Xian are with State Key Laboratory of Complex Electromagnetic Environment Effects on Electronics and Information System, National University of Defense Technology, Changsha, 410073, China .\
$^2$ Udaya Parampalli is with Department of Computing and Information Systems, University of Melbourne, VIC 3010, Australia .\
[^2]: The two concepts of systematic MSR codes and systematic-repair MSR codes have generated differences in the definition of MSR codes. As described in [@Re:A.Dimakis], the formal description of an MSR code requires all nodes be optimally repairable. But, many references in the literature also term a code as an MSR code even if it allows optimal repair of systematic nodes only. In this case, we will distinguish between two types of MSR codes by calling them as systematic-repair MSR codes and systematic MSR codes with optimal repair of all nodes (or just systematic MSR codes for simplicity) respectively. When the context is clear or the distinction is unnecessary, they both are named by MSR codes.
[^3]: The formal description of repair subspace stems from [@Re:S.Goparaju1], that can be also clarified from the phrases on top of Note \[note1\] and Equation (\[interference\]) in this paper. Simply put, repair subspaces correspond to a set of matrices over $\mathbb{F}^{\frac{l}{r}\times l}$ that, in conjunction with stored data vectors, are used to generate repair data symbols.
[^4]: In subsequent section, we give the extended subspace properties for the case of arbitrary number of parity nodes $r$. With these extended properties, we will then present two improved upper bounds on $k$.
[^5]: The method of partition is introduced from [@Re:S.Goparaju1] that basically stands for dividing the systematic part $[1,k]$ in a certain way, which can be technically described as follows. To be specific, we denote the size of each partition by the notation $\lambda$, i.e., the smallest integer value such that there exist $\lambda$ number of systematic nodes $\{i_1,\cdots,i_{\lambda}\}$ satisfying $\biguplus_{\nu=1}^{\lambda} \mathbf{S}_{i_\nu}\simeq \mathbb{F}^l$ $\big($or $\dim(\biguplus_{\nu=1}^{\lambda} \mathbf{S}_{i_\nu})=l$$\big)$ for $i_\nu\in[1,k]$, where $\mathbf{S}_{i_\nu}$ is the repair subspace of the $i_\nu$-th systematic node and $``\biguplus"$ represents the sum of subspaces as defined in Note \[note1\]. Thereupon, it signifies that every collection of $\lambda$ repair subspaces among $\{\mathbf{S}_1,\ldots,\mathbf{S}_k\}$ can span the whole space, i.e., repair subspaces within each partition sized by $\lambda$ span the entire space $\mathbb{F}^l$.
[^6]: In subsequent description, construction of linear systematic-repair MSR codes has to satisfy the condition required for the interference alignment, which however aims at designing MSR codes with only optimally repairing systematic nodes. Since interference alignment does not apply to systematic MSR codes (with optimally repairing all nodes), it is evident that upper bounds on $k$ for systematic MSR codes should be tighter than the one for systematic-repair MSR codes. For the sake of clarity, we stress again that linear systematic-repair MSR codes are our research object in this paper.
[^7]: Although the construction scheme by Wang et al. [@Re:Z.Wang2] is a general case with the longest systematic-length $k=(r+1)\log_rl$, it is obviously shorter than our new upper bound $2\lambda\log_rl$ where $r\leq \lambda\leq \left \lfloor\log_{\frac{r}{r-1}}l\right \rfloor+1$. That is to say, the longest one so far given in [@Re:Z.Wang2] is still not achievable.
[^8]: As for an arbitrary MDS code, the linear independence among the matrices $\Big\{\mathbf{I},\mathbf{C}_{u,i}|u\in[2,r],i\in[1,k]\Big\}$ cannot always hold. The reason is that unlike systematic-repair MSR code, an arbitrary MDS code is not necessarily equipped with the property of interference alignment and as a consequence cannot meet the conditions required in Lemma \[premise1\] and Lemma \[premise2\].
[^9]: As assumed in Lemma \[premise2\], we let $\mathbf{C}_{u,j}=\mathbf{I}$ when $u=1$ and thereby we can rewrite that $\mathbf{\Gamma}_{i,u}=\sum_{j\in \mathcal{X}_i}\mathbf{C}_{u,j}$ for any $u\in[1,r]$. In this case, $\mathbf{\Psi}_u^{(s)}$ can be explicitly expressed as $\mathbf{\Psi}_u^{(s)}=\sum_{\{u_1\in[1,r],\cdots,u_s\in[1,r]\}} d_{u_1u_2\cdots u_s}\mathbf{\Delta}_{u_1u_2\cdots u_s}=\sum d_{u_1u_2\cdots u_s}\prod_{i=1}^{s}\mathbf{\Gamma}_{i,u_i}=\sum d_{u_1u_2\cdots u_s}\prod_{i=1}^{s}\big\{\sum_{j\in \mathcal{X}_i}\mathbf{C}_{u_i,j}\big\}=\sum d_{u_1u_2\cdots u_s}\prod_{i=1}^{s}\big\{\sum_{j\in [1+(i-1)\lambda,\cdots,i\lambda]}\mathbf{C}_{u_i,j}\big\}=\sum_{\{u_1\in[1,r],\cdots,u_s\in[1,r]\}}d_{u_1u_2\cdots u_s}\sum_{\{j_1\in \mathcal{X}_1,j_2\in \mathcal{X}_2,\cdots,j_s\in \mathcal{X}_s\}}\mathbf{C}_{u_1,j_1}\mathbf{C}_{u_2,j_2}\cdots\mathbf{C}_{u_s,j_s}$, where some of the coefficients $d_{u_1u_2\cdots u_s}$ maybe equal to zero. Henceforth, each non-zero uniterm of $\mathbf{\Psi}_u^{(s)}$ is formed as $d_{u_1u_2\cdots u_s}\mathbf{C}_{u_1,j_1}\mathbf{C}_{u_2,j_2}\cdots\mathbf{C}_{u_s,j_s}$ for $j_{\tau}\in\mathcal{X}_{\tau}=[1+(\tau-1)\lambda,\tau\lambda]$ and $\tau\in[1,s]$. As a result, it is evident from Lemma \[premise1\] that $\mathbf{S}_i\mathbf{C}_{u_1,j_1}\mathbf{C}_{u_2,j_2}\cdots\mathbf{C}_{u_s,j_s}\preceq\mathbf{S}_i$ for $i\in\mathcal{X}_{s+1}=[s\lambda+1,(s+1)\lambda]$, and therewith we further have that $\mathbf{S}_i\mathbf{\Psi}_u^{(s)}\preceq\mathbf{S}_i$.
| {
"pile_set_name": "ArXiv"
} |
[P. Binétruy]{}\
[*LPTHE[^1], Université Paris-XI, Bâtiment 211, F-91405 Orsay Cedex, France*]{}\
[E. Dudas]{}\
[*CEA-SACLAY, Service de Physique Théorique*]{}\
[*F-91191 Gif-sur-Yvette Cedex, France*]{}\
[**Abstract**]{}
We study gaugino condensation in presence of an anomalous $U(1)$ gauge group and find that global supersymmetry is dynamically broken. An example of particular interest is provided by effective string models with 4-dimensional Green-Schwarz anomaly cancellation mechanism. The structure of the hidden sector is constrained by the anomaly cancellation conditions and the scale of gaugino condensation is shifted compared with the usual case. We explicitly compute the resulting soft supersymmetry breaking terms.
Introduction
============
Among the scenarios for breaking supersymmetry, the condensation of gauginos in a hidden sector remains a favourite one [@Nilles]. Gaugino condensation is indeed central to the idea of dynamical supersymmetry breaking in the context of gauge symmetries. But in its explicit realizations in superstring models [@DRSW], it suffers from a number of drawbacks. The degeneracies associated with the flat directions of the scalar potential are lifted but the corresponding degrees of freedom such as the dilaton are not stabilized and the true ground state is found at infinite field values where supersymmetry is restored. In order to overcome this problem, so-called “racetrack” or multicondensate models [@racetrack] have been proposed where two terms of different orders conspire in order to stabilise the field and provide a supersymmetry-breaking minimum.
Many of the superstring models have an anomalous $U(1)_X$ gauge symmetry [@DSW] which could play a important role in issues such as fermion mass hierarchies [@IR; @ramond], cosmology [@cosmo] and ... gaugino condensation. The latter connection was recently stressed by Banks and Dine [@BD].[^2] Indeed, the anomalous $U(1)_X$ has mixed anomalies with the other gauge symmetries –those of the standard model as well as the hidden sector–, anomalies which are cancelled through a 4-dimensional Green-Schwarz mechanism [@GS] using the couplings of the dilaton superfield to the gauge superfields. It is therefore not surprising that the whole issue of supersymmetry breaking through gaugino condensation is deeply modifed in such models. Moreover, because in the Green-Schwarz mechanism all the mixed anomalies are non-vanishing and proportional to one another, there must exist fields charged under $U(1)_X$ in the observable as well as in the hidden sector. The $U(1)_X$ gauge symmetry thus serves as a messenger interaction competitive with the gravitational interaction. In this paper, we wish to stress the modifications that the presence of such an anomalous $U(1)_X$ symmetry is bringing to the scenario of a dynamical supersymmetry breaking through gaugino condensation.
To be more specific, the relevant couplings of the dilaton superfield $S$ to the gauge superfields $V_a, V_X$ (of respective gauge invariant field strengths $W^{\alpha}_a, W^{\alpha}_X$) of the groups $G_a$ and $U(1)_X$ read in the global limit: $$\begin{aligned}
{\cal L}_{S,V} &=& - \int d^4 \theta \ln (S+ S^+ -
\delta_{GS}V_X) \nonumber \\
& & + \int d^2\theta \left[{S \over 4} (\sum_a k_a{\rm Tr}
W^{\alpha}_a W_{a\alpha} + k_X {\rm Tr} W^{\alpha}_X W_{X\alpha}) +
{\rm h.c.}\right] \label{eq:L} \end{aligned}$$ where $\delta_{GS}$ is the Green-Schwarz coefficient and $k_a$ ($k_X$) is the Kac-Moody level of the group $G_a$ ($U(1)_X$). Under a $U(1)_X$ gauge transformation ($A^X_\mu \rightarrow A^X_\mu
+ \partial_\mu \alpha$), $S$ is shifted as $$S \rightarrow S + {i \over 2} \delta_{GS} \alpha(x). \label{eq:shift}$$ The complete Lagrangian is invariant provided the mixed $U(1)_X[G_a]^2$ anomaly coefficients $C_a$ satisfy the condition $$\delta_{GS} = {C_a \over k_a} = {C_X \over k_X} = {C_g \over k_g},
\label{eq:Ck}$$ where $C_g$ is the mixed gravitational anomaly proportional to ${\rm Tr} X$. Indeed a string computation yields $$\delta_{GS} = {1 \over 192 \pi^2} {\rm Tr} X.
\label{eq:deltaGS}$$ The mixing between $S$ and $V_X$ in the Kähler potential (\[eq:L\]) gives rise to a D-term in the scalar potential: $$V_D = {g^2_X \over 2} \left(\sum_A X_A K_A \phi^A + {1 \over 4}
k_X g^2_X \delta_{GS} M^2_P \right)^2, \label{eq:VD}$$ where $M_P$ is the Planck scale, $$k_X g^2_X = {2 \over S+S^+},$$ and $K_A$ is the derivative of the Kähler potential $K$ with respect to the field $\phi^A$. The presence of the Fayet-Iliopoulos term induced by the $U(1)_X$ anomaly usually induces a non-zero vacuum expectation value for one (or more) field of $X$ charge of sign opposite to $\delta_{GS}$.
In the case where there is no anomalous $U(1)$, a non-anomalous R-symmetry under which the $S$ superfield undergoes a translation similar to (\[eq:shift\]) imposes that the scale $\Lambda$ at which the hidden sector gauge coupling becomes strong (which sets the scale for the corresponding gaugino condensates) behaves as $$\Lambda \sim M_P e^{-{kS \over 2 b_0}}, \label{eq:Lambda}$$ where $b_0$ is the one-loop beta function coefficient for the hidden sector gauge group and $k$ is its Kac-Moody level. This is obviously not invariant under the $U(1)_X$ transformation (\[eq:shift\]) of $S$. This shows that the presence of an anomalous $U(1)_X$ necessarily modifies the standard discussion of gaugino condensation.
This study goes beyond the case of effective superstring models. Indeed, it is not unusual that, in the course of sequential gauge symmetry breaking, appears an anomalous $U(1)_X$ abelian gauge symmetry whose mixed anomalies are cancelled through a Green-Schwarz type mechanism using an effective degree of freedom with dilaton-axion couplings. Thus we will consider in what follows the general case of a gauge model with symmetry group $SU(N_c)
\times U(1)_X$ with a non-vanishing Fayet-Iliopoulos term and the following matter content: $N_f \le N_c$ flavors, a dilaton-axion superfield and a chiral supermultiplet which breaks the anomalous $U(1)_X$ symmetry and whose vacuum expectation value helps to cancel the $U(1)_X$ D-term.
In section $2$, we analyze in detail this model with global supersymmetry. It is shown, using an effective lagrangian approach, that global supersymmetry is dynamically broken.
In section $3$, we compute the resulting soft breaking terms in the observable sector and discuss their phenomenological consequences. We end with some comments.
Gauge group $G=SU(N_c) \times U(1)_X$ with $N_f \le N_c$ flavors
================================================================
The model that we consider is an extension of SUSY-QCD based on the gauge group $SU(N_c)$ with $N_f \le N_c$ flavors of “quarks” $Q^i$ of $U(1)_X$ charge $q$ in the fundamental of $SU(N_c)$ and “antiquarks” ${\tilde Q}_{{{\bar{\imath}}}}$ of charge $\tilde q$ in the antifundamental of $SU(N_c)$.
Since we want to avoid $SU(N_c)$ breaking in the $U(1)_X$ flat direction (\[eq:VD\]), we require that the charges $q$ and $\tilde q$ are positive (this is in fact not restrictive: see the comment in the footnote below). We then need at least one field of negative charge in order to cancel the D-term (\[eq:VD\]). For simplicity we will introduce a single field $\phi$ of $U(1)_X$ charge normalized to $-1$.
The classical lagrangian compatible with the symmetries is ${\cal L} =
{\cal L}_{kin} + {\cal L}_{couplings}$, where we assume a flat Kähler potential for the matter fields: $${\cal L}_{kin} = \int d^4 \theta \left[ Q^+ e^{2q V_X + V_N} Q
+ {\tilde Q} e^{2{\tilde q} V_X - V_N} {\tilde Q}^+ + \phi^+
e^{-2V_X} \phi \right] + {\cal L}_{S,V} $$ and $${\cal L}_{couplings} = \int d^2 \theta ({\phi \over M_P})^{q +
{\tilde q}} m_i^{{{\bar{\imath}}}} Q^i {\tilde Q}_{{{\bar{\imath}}}} + h.c. \ .
\label{eq:coupling}$$
As stressed in the introduction, the model can be studied [*per se*]{} or be used as an illustrative example of a hidden sector where supersymmetry is dynamically broken in presence of an anomalous $U(1)$. In the former case, $M_P$ is the scale of the underlying non-anomalous theory (say the mass of some heavy fermions we have integrated upon), in the latter case it is the Planck scale.
The mixed anomaly $U(1)_X [SU(N_c)]^2$ which will fix, through (\[eq:Ck\]), all the mixed anomalies in the model is given by $$C_N = {1 \over 4 \pi^2} N_f (q + {\tilde q}) = k_N \delta_{GS}.
\label{eq:CN}$$ We thus require $q + {\tilde q} > 0$, which in turn justifies the presence of the superpotential term (\[eq:coupling\]).[^3]
One may note, using (\[eq:shift\]), that the following combination $$f = k_N S - {N_f \over 8 \pi^2} (q+ \tilde q) \ln {\phi \over M_P}
\label{eq:f}$$ is invariant under $U(1)_X$. Such a gauge kinetic function would be obtained by integrating over the hidden matter degrees of freedom, assuming unbroken supersymmetry. It could then be used to determine the gaugino masses. We will see however that supersymmetry is broken, which makes matters less straightforward.
The two scales present in the problem are:
- the scale at which the anomalous $U(1)_X$ symmetry is broken which is set by $$\xi = {1 \over 2} k^{1/2}_X g_X \delta^{1/2}_{GS} M_P. \label{eq:ksi}$$
- the scale at which the gauge group $SU(N_c)$ enters in a strong coupling regime: $$\Lambda = M_P e^{-8\pi^2 k_N S/(3N_c-N_f)}, \label{eq:newLambda}$$ where we have used (\[eq:Lambda\]) with $b_0=(3N_c-N_f)/(16\pi^2)$. Notice that, by using the transformation (\[eq:shift\]), we find that the dynamical scale $\Lambda$ has a charge $q_{\Lambda} = N_f (q + {\tilde q})/(3 N_c -
N_f)$.
From now on, we will suppose that $\Lambda << \xi$. We could write the effective theory below the scale $\xi$ and study within this theory the strongly coupled $SU(N_c)$ theory. It is however simpler to keep the complete theory down to the scale $\Lambda$ since most of the nontrivial effects we will obtain result from an interplay between the scales $\Lambda$ and $\xi$.
Below the scale $\Lambda$ the appropriate degrees of freedom for $N_f < N_c$ are the field $\phi$ and the mesons $M_{{{\bar{\imath}}}}^i = Q^i
{\tilde Q}_{{{\bar{\imath}}}}$. The effective superpotential is fixed uniquely by the global symmetries [@TVY; @ADS] as follows $$W = (N_c - N_f) {\Lambda^{3N_c - N_f \over N_c - N_f} \over
{(det M)}^{1 \over N_c - N_f}} + ({\phi \over M_P})^{q +
{\tilde q}} m_i^{{{\bar{\imath}}}} M_{{{\bar{\imath}}}}^i \ \label{eq:sup}$$ and is seen to be automatically $U(1)_X$ invariant. Similarly, the gaugino condensation scale $${<\lambda \lambda>}
= \left( \Lambda^{3N_c-N_f}/det M \right)^{1 \over N_c - N_f}
\label{eq:gauginocond}$$ is also $U(1)_X$ gauge invariant, as it should be.
The gauge contributions to the scalar potential can be computed along the $SU(N_c)$ classical flat directions. The result is $$V_D = {g_X^2 \over 2} \left[ (q + {\tilde q}) Tr (M^+ M)^{1/2}
- \phi^+ \phi + \xi^2 \right]^2 \ . \label{eq:pot}$$
The auxiliary fields, computed from (\[eq:sup\]) and (\[eq:pot\]) are $${\bar F}_{S^+} = -{8 \pi^2 \over M_P} k_N (S+S^+)^2 {\Lambda^{3N_c -
N_f \over N_c - N_f} \over {(det M)}^{1 \over N_c - N_f}}
\label{eq:FS}$$ and $$\begin{aligned}
{(\bar F_{M^+})}^{{{\bar{\imath}}}}_i &=& 2 \left[ -{(M^{-1})}_i^{\bar \jmath}
{\Lambda^{3N_c - N_f \over N_c - N_f} \over (det M)^{1 \over N_c -
N_f}} + ({\phi \over M_P})^{q + {\tilde q}} m_i^{\bar \jmath} \right]
[(M^+ M)^{1/2}]^{{{\bar{\imath}}}}_{\bar \jmath}, \nonumber \\
\bar F_{\phi^+} &=& {q + {\tilde q} \over M_P} ({\phi \over
M_P})^{q + {\tilde q}-1} Tr (m M) \ , \nonumber \\
D_X &=& g_X^2 \left[(q + {\tilde q}) Tr (M^+ M)^{1/2} - \phi^+ \phi +
\xi^2 \right]\ . \label{eq:aux}\end{aligned}$$
In the limit $S \rightarrow \infty$, the scale $\Lambda$ vanishes and we can choose $M^i_{{{\bar{\imath}}}}=0$ and $\phi=\xi$ to cancel all these auxiliary fields. This is the usual global supersymmetry minimum at infinite values of $S$ which leads to the dilaton stabilization problem. We will assume that the dilaton is stabilized at some finite value $S_0$, possibly through some extra $S$-dependent term in the superpotential (we will therefore refrain from using (\[eq:FS\])), and that $F_S (S_0) = 0$. Indeed we are going to show that, even in this unfavorable case (supersymmetry conserving groundstate for $S$), the other fields present in the theory yield supersymmetry breaking because of the anomalous behavior of $U(1)_X$. From now on, we will therefore restrict our attention to the auxiliary fields (\[eq:aux\]) associated with $M^i_{{{\bar{\imath}}}}$, $\phi$ and the $U(1)_X$ gauge degree of freedom.
It is readily seen that the system of equations $F_M = F_{\phi}
= D_X=0$ has no solution as long as $\xi \not= 0$ that is $q +
\tilde q \not= 0$. Therefore supersymmetry is dynamically broken. As usual, the origin of supersymmetry breaking is chiral: the non-abelian gauge group content is vector-like but $Q^i$ and $\tilde
Q_{{{\bar{\imath}}}}$ do not transform in a vectorlike fashion under $U(1)_X$: $q
\not= - \tilde q$; this is precisely what drives the anomaly, which is therefore the source of chirality in the model.
We will now minimize the scalar potential in terms of $M^i_{{{\bar{\imath}}}}$, $\phi$ at fixed $S=S_0$ value such that $F_S(S_0)=0$: $$V = {1 \over (S+ S^+)^2} {\bar F}_{S^+} F_S + {\bar F}_{\phi^+} F_\phi
+ {1 \over 2} (\bar F_{M^+})^{{{\bar{\imath}}}}_i [(M^+ M)^{-1/2}]^{\bar
j}_{{{\bar{\imath}}}}(F_M)^i_{\bar j} + V_D.$$ In order to be able to give analytic solutions, we make a few simplifying asumptions. First, we linearize the minimization procedure by looking for a minimum in the vicinity of:
a\) $\phi_0 = \xi$, the field value which minimizes $V_D$ in the absence of condensates;
b\) $(M_0)^i_{{{\bar{\imath}}}}$, the solution of $({\bar F}_{M^+})^{{{\bar{\imath}}}}_i =
0$: $$(M_0)^i_{{{\bar{\imath}}}} = (m^{-1})^i_{{{\bar{\imath}}}} (det m)^{1/N_c} \Lambda^{{3 N_c -
N_f \over N_c}} \left({\xi \over M_P}\right)^{{N_f - N_c \over
N_c}(q+ \tilde q)} \; \; \; \; .$$ One can make a field transformation in order to have a diagonal matrix $m^{{{\bar{\imath}}}}_i$, in which case $(M_0)^i_{{{\bar{\imath}}}}$ is also diagonal. Since we are only interested in orders of magnitude, we will make the assumption that $m^{{{\bar{\imath}}}}_i = m
\delta^{{{\bar{\imath}}}}_i$ and search for solutions $M^i_{{{\bar{\imath}}}} = M
\delta^i_{{{\bar{\imath}}}}$ of the equations of motion.
The minimum is obtained by making around the field configuration $(M_0)^i_{{{\bar{\imath}}}} \equiv M_0
\delta^i_{{{\bar{\imath}}}}$ an expansion in the parameter $$\epsilon \equiv {M_0 \over \xi^2} = \left( {\Lambda \over \xi}
\right)^{{3 N_c - N_f \over N_c}} \left[ {m \over M_P} \left({\xi
\over M_P}\right)^{q+ \tilde q-1} \right]^{{N_f - N_c \over N_c}}
\; \; \; . \label{eq:epsilon}$$ One obtains $$\begin{aligned}
<\phi^+ \phi> &=& \xi^2 \left[ 1 + \epsilon N_f (q+ \tilde q)
+ \epsilon^2 N_f^2 (q+ \tilde q)^2 \left( -{(N_c -
N_f)(2N_c-2N_f) \over 2 N_c^2} (q+ \tilde q) \right. \right.
\nonumber \\
&+& \left. \left. {1 \over g^2_X} {\hat m}^2 \left[ 1 - {N_f \over
N_c} (q+ \tilde q)\right] \right) + O(\epsilon^3) \right], \\
<M> &=& M_0 \left[ 1 -
{\epsilon \over 2} {N_f (N_c - N_f) (2 N_c-N_f) \over N^2_c} (q+
\tilde q)^2 + O(\epsilon^2) \right], \end{aligned}$$ where we have introduced the scale $$\hat m = m \left( {\xi \over M_P} \right)^{q+ \tilde q} \ .$$ The value of the auxiliary terms at this ground state are, to leading order: $$\begin{aligned}
<D_X> &=& -\epsilon^2 {\hat m}^2 N_f^2 (q+ \tilde q)^2 \left[ 1 -
{N_f \over N_c} (q + \tilde q)\right], \nonumber\\
<F_\phi> &=& \epsilon {\hat m} \xi N_f (q+ \tilde q), \nonumber \\
<F_M> &=& -\epsilon^2 {\hat m} \xi^2 {N_f (N_c-N_f) \over N_c} (q +
\tilde q)^2 .
\label{eq:DF} \end{aligned}$$ One may note that $<D_X^{1/2}>$, $<F_\phi / \phi>$ and $<F_M / M>$ are all of the same order $\epsilon \tilde m$. This will have definite consequences for the soft terms, as we will see in the next section. Also, using (\[eq:epsilon\]), one checks that this order of magnitude goes to zero as $\Lambda \rightarrow 0$, as well as $\xi
\rightarrow \infty$ (at fixed value of $\xi / M_P$). This shows once again the mixed role that the two scales $\Lambda$ and $\xi$ play as far as supersymmetry breaking is concerned in this model.
A similar analysis can be performed when $N_f=N_c \equiv N$. In this case, new degrees of freedom must be introduced in the low energy effective lagrangian [@Seiberg] $$B =\epsilon_{i_1 \cdots i_N} Q^{i_1} \cdots Q^{i_N},
\; \; \tilde B =\epsilon^{{{{\bar{\imath}}}}_1 \cdots {{{\bar{\imath}}}}_N} {\tilde
Q}_{{{{\bar{\imath}}}}_1} \cdots Q_{{{{\bar{\imath}}}}_N}.$$ The effective superpotential compatible with all the symmetries reads $$W = U \ln {detM - B \tilde B \over \Lambda^{2N}} + \left( {\phi \over
M_P} \right)^{q + \tilde q} m_i^{{{\bar{\imath}}}} M^i_{{{\bar{\imath}}}},$$ where $U$ is a Lagrange multiplier, physically interpreted as the gauge composite superfield $U = {\rm Tr} W^{\alpha}_N W_{N \alpha}$. As in the $N_f < N_c$ case, the system of equations $F_U=F_M=F_B=F_{\tilde B}=D_X=0$ has no solution and global supersymmetry is broken.
The properties of the model do not depend either on the assumption of a single $\phi$ field. Similar conclusions can be reached when one considers for example a vector-like pair of such fields.
Soft terms in the observable sector
===================================
We now use the results of the preceding section to determine the order of magnitude of the soft terms in the observable low energy sector of quarks, leptons and gauginos.
The magnitude of the soft terms in the observable sector is fixed by the auxiliary fields $F_\phi$, $F_M$ and $D_X$. At the tree level of the Lagrangian of our model, we find soft scalar masses $\tilde
m_i^2$ and trilinear soft terms $A_{ijk}$ given by the expressions[^4]: $$\tilde
m_i^2 = X_i <D_X>, \;\;\; A_{ijk} = (X_i + X_j + X_k) {F_\phi \over
\phi}, \label{eq:mA}$$ where $X_i$ is the $U(1)_X$ charge of the corresponding field $\Phi^i$. Gaugino masses in the hidden sector are also induced: $m_\lambda \sim N_f <F_M/M>$. The gaugino masses in the observable sector are absent at tree level and are induced by standard gauge loops. Notice that, because supersymmetry is broken, gaugino masses are not simply given by the gauge invariant kinetic function (\[eq:f\]).
In the preceding section we obtained, in the limit $\Lambda \ll \xi$, the following relation among the auxiliary fields $<F_M / M> \sim
<F_\phi / \phi> \sim <D_X^{1/2}>$. Consequently, all the soft breaking terms induced at tree level are of the same order. By using (\[eq:epsilon\]) and (\[eq:DF\]), we obtain $$\tilde m \sim N_f (q+ \tilde q) {\Lambda^3
\over \xi^2} \left[ {m \over \Lambda} \left( {\xi \over M_P}
\right)^{q+ \tilde q} \right]^{N_f / N_c} = N_f (q +
\tilde q) {<\lambda \lambda> \over \xi^2},
\label{eq:softorder}$$ where $\tilde m$ generically denotes a soft-breaking term (\[eq:mA\]) and we have used (\[eq:gauginocond\]) in order to derive the last relation. This relation is indeed central to the kind of models described here and stresses the connected role of the relevant scales: $\xi$ as the scale of messenger interaction and the gaugino condensate as the seed of supersymmetry breaking (although, as stressed earlier, the chiral nature of the $U(1)_X$ plays an important role: $q \not= -\tilde q$).
We now restrict our attention to hidden sector models where the messengers of supersymmetry breaking are the anomalous $U(1)_X$ gauge [*and*]{} gravitational interactions. The scale $M_P$ is therefore the Planck scale. Eq. (\[eq:softorder\]) should be compared with the gravitationally-induced soft terms of order $$\tilde m|_{grav.} \sim {<\lambda \lambda> \over M_P^2},$$ which must be included if the supergravity interactions are switched on. Since $${1 \over N_f (q + \tilde q)} \left( {\xi \over M_P} \right)^2
\label{eq:ratio}$$ is a small number in the context of superstring models, the supergravity soft terms can be neglected and the soft terms computed above can be viewed as phenomenological predictions of these models. It is worth noting that, using (\[eq:CN\]) and (\[eq:ksi\]), the factor $N_f (q
+ \tilde q)$, which may be large and is model dependent, drops out out of the ratio (\[eq:ratio\]). One is left with a pure number which depends on the gauge coupling –as a genuine one-loop effect– and the Kac-Moody levels.
If $\xi \sim M_P$, phenomenologically interesting soft terms $\tilde m \le 1 \ {\rm TeV}$ call for $\Lambda
\ll \Lambda_0$, where $\Lambda_0$ is a typical intermediate condensation scale: $\Lambda_0 \sim 10^{13} \ {\rm TeV}$.[^5] For $\xi \ll M_P$ the required values of $\Lambda$ depend on the parameters of the hidden sector $N_f$, $N_c$, $q + \tilde q$. A generic prediction in this case is a rather light gravitino: $$m_{3/2} \sim {<W> \over M_P^2} \sim
{N_c \over N_f (q+ \tilde q)} \left( {\xi \over M_P} \right)^2 \tilde
m.$$ It is useful to notice that generally $N_f (q+ \tilde q) > N_c$. Then, by using (\[eq:DF\]), we find $D_X>0$, which means that scalar particles with positive $U(1)_X$ charges will acquire positive squared masses. This is a very welcome feature of the model since, by using the mixed anomaly conditions (\[eq:Ck\]), (\[eq:deltaGS\]), we know that the fields in the observable sector must have predominantly positive charges.
This is also consistent with the interpretation of the $U(1)_X$ as an horizontal symmetry which may explain the low energy mass spectrum [@IR; @JS; @ramond; @DPS; @Nir]. In this context, positive charges are required in the model presented above to account for the observed fermion masses and mixings. The anomalous symmetry also plays an important role in constraining the soft terms [@LNS; @DGPS; @KK].
A general feature of the models presented is the large scale of gauge symmetry breaking. Despite the fact that the messengers of supersymmetry breaking are mostly gauge interactions, the resulting soft terms have much lower values, as a result of a conspiracy between the two scales $\Lambda$ and $\xi$. This is to be contrasted with the standard gauge-mediated scenarios where the scale of supersymmetry breaking is much lower [@gaugemed]. In particular, as we have seen, in the limit where the two scales are very far away $\Lambda \rightarrow 0$ or $\xi
\rightarrow \infty$, supersymmetry is restored.
There are several aspects which we did not address in this short paper and which we reserve for further studies: one is the question of dilaton stabilization since the scalar potential has a complicate $S$ dependence in its $F$-term through the scale $\Lambda$ and its $D$-term through the coupling $g_X$. Another is the generalization to other gauge structures with a chiral content and explicit superstring realizations.
: We wish to thank the Aspen Center for Physics where this work was completed for its hospitality and the participants of the workshop “Flavor and Gauge Hierarchies”, in particular Pierre Ramond, for the lively discussions and the interesting questions they raised.
[0]{} H.P. Nilles, [*Phys. Lett.*]{} [**B115**]{} (1982) 193; S. Ferrara, L. Girardello and H.P. Nilles, [*Phys. Lett.*]{} [**B125**]{} (1983) 457. J.P. Derendinger, L.E. Ibáñez and H.P. Nilles, [*Phys. Lett.*]{} [**B155**]{} (1985) 65; M Dine, R. Rohm, N. Seiberg and E. Witten, [*Phys. Lett.*]{} [**B156**]{} (1985) 55. N. V. Krasnikov, [*Phys. Lett.*]{} [**B193**]{} (1987) 37. M. Dine, N. Seiberg and E. Witten, [*Nucl. Phys.*]{} [**B289**]{} (1987) 589; J. Atick, L. Dixon and A. Sen, [*Nucl. Phys.*]{} [**B292**]{} (1987) 109; M. Dine, I. Ichinose and N. Seiberg, [*Nucl. Phys.*]{} [**B293**]{} (1988) 253. L. Ibáñez and G.G. Ross, [*Phys. Lett.*]{} [ **B332**]{} (1994) 100. P. Binétruy and P. Ramond, [*Phys. Lett.*]{} [**B350**]{} (1995) 49; P. Binétruy, S. Lavignac and P. Ramond, LPTHE-Orsay 95/54, UFIFT-HEP-96-1, hep-ph/9601243 (to be published in Nuclear Physics). J.A. Casas and C. Munoz, [*Phys. Lett.*]{} [**B216**]{} (1989) 37; J.C. Casas, J. M. Moreno, C. Munoz and M. Quiros, [*Nucl. Phys.*]{} [**B328**]{} (1989) 272; P. Binétruy and G. Dvali, preprint CERN-TH/96-149, LPTHE-ORSAY 96/40. T. Banks and M. Dine, preprint RU-95-51, SCIPP 95/41. M. Green and J. Schwarz, [*Phys. Lett.*]{} [**B149**]{} (1984) 117. T.R. Taylor, G. Veneziano and S. Yankielowicz, [*Nucl. Phys.*]{} [**B218**]{} (1983) 493. I. Affleck, M. Dine and N. Seiberg, [*Nucl. Phys.*]{} [**B256**]{} (1985) 557. N. Seiberg, [*Phys. Rev.*]{} [**D49**]{} (1994) 6857. V. Jain and R. Schrock, [*Phys. Lett.*]{} [**B352**]{} (1995) 83. E. Dudas, S. Pokorski and C.A. Savoy, [*Phys. Lett.*]{} [**B356**]{} (1995) 45. Y. Nir, [*Phys. Lett.*]{} [**B354**]{} (1995) 107. M. Leurer, Y. Nir and N. Seiberg, [*Nucl. Phys.*]{} [**B398**]{} (1993) 319 and [**B309**]{} (1993) 337. E. Dudas, S. Pokorski and C.A. Savoy, [*Phys. Lett.*]{} [**B369**]{} (1996) 255; E. Dudas, C. Grojean, S. Pokorski and C.A. Savoy, Saclay T96/065, hep-ph/9606383. Y. Kawamura and T. Kobayashi, DPSU-96-2, hep-ph/9601365. M. Dine and A. Nelson, [*Phys. Rev.*]{} [**D48**]{} (1993) 1277; M. Dine, A. Nelson and Y. Shirman, [*Phys. Rev.*]{} [**D51**]{} (1995) 1362; M. Dine, A. Nelson, Y. Nir and Y. Shirman, [*Phys. Rev.*]{} [**D53**]{} (1996) 2658.
[^1]: Laboratoire associé au CNRS-URA-D0063.
[^2]: We wish to thank Luis Ibáñez for drawing our attention to this paper and the whole issue.
[^3]: Alternatively, since from (\[eq:CN\]) $q + \tilde q$ and $\delta_{GS}$ have the same sign, we would still avoid $SU(N_c)$ breaking in the $U(1)_X$ flat direction with $q+\tilde q <0$. The field $\phi$ which cancels the D-term would then be chosen with charge $+1$.
[^4]: We assume the presence in the superpotential of terms of the form $(\phi / M_P)^{X_i + X_j + X_k} \Phi_i \Phi_j
\Phi_k$, as allowed by the $U(1)_X$ symmetry.
[^5]: We assume here that $m\sim M_P$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Reconstructing past population size from present day genetic data is a major goal of population genetics. Recent empirical studies infer population size history using coalescent-based models applied to a small number of individuals. Here we provide tight bounds on the amount of exact coalescence time data needed to recover the population size history of a single, panmictic population at a certain level of accuracy. In practice, coalescence times are estimated from sequence data and so our lower bounds should be taken as rather conservative.
[**Keywords:**]{} population size; estimation; coalescent.
author:
- 'Junhyong Kim [^1]'
- 'Elchanan Mossel [^2]'
- 'Miklós Z. Rácz [^3]'
- 'Nathan Ross [^4]'
bibliography:
- 'Density.bib'
title: '**[^5]**'
---
Introduction {#sec:intro}
============
Reconstructing the past size and structure of the population of a species is a major goal of population genetics with applications in, for example, ecology, epidemiology [@heled2008bayesian], and paleoanthropology [@Li2011]. It is also important for understanding relationships between different evolutionary parameters, e.g., the dynamics of different parts of the genome or how demography affects selection [@Li2012].
Inference is based on sequence data from individuals sampled from the population under consideration. Under a given population history, the coalescent is a model that provides likelihoods of observed genetic data and is one of the main tools used to infer population history. But the space of population histories typically considered is huge and so maximum likelihood estimation requires approximation techniques [@Bhaskar2014; @Excoffier2013; @harris2013inferring; @Li2011; @Nielsen2000; @palamara2012length; @Sheehan2013] which lack theoretical guarantees; the same statement applies to Bayesian methods [@Drummond2005; @heled2008bayesian]. (These methods are discussed in greater detail in Section \[sec:back\] below.)
Here we provide provable information-theoretic lower bounds on the amount of coalescence data needed to estimate, up to some specified accuracy, events in a population’s past history (see Theorem \[thm:main\_bd\] below). Our bounds are asymptotically tight as shown by analysis of a simple inference algorithm which recovers the history given slightly more data than required by the lower bounds.
Before stating our results in more detail, we provide a brief introduction to inference using the coalescent, as well as a summary of existing literature in this area.
Inference using the coalescent {#sec:back}
------------------------------
Let $N(t)$ be the size of a single panmictic haploid population at time $t$ “generations" in the past[^6] and call $N = \left\{N\left(t\right) \right\}_{t {\geqslant}0}$ the *shape of the population size history*, or simply the *population shape*. Given $N$, Kingman’s coalescent (see [@Tavare2004] for background) is a random genealogy on $n$ sampled individuals from the present day population. The basic description is that the rate of coalescence between any two individuals/lineages at time $t$ in the past is $1/N(t)$ and so given $k$ lineages at time $t$, the rate of coalescence is $\binom{k}{2}/N(t)$. We focus primarily on the case where data comes from pairs of individuals, i.e., $n=2$, just as in, e.g., [@Li2011].
The population shape $N = \left\{ N\left(t \right) \right\}_{t {\geqslant}0}$ determines a distribution ${\mathbb{P}}_{N}$ over coalescent trees; and in particular, ${\mathbb{P}}_{N}$ determines the distribution of coalescent trees of any finite number of individuals, at any number of independent loci. The first step to infer $N$ using the coalescent is to ensure that the distribution over coalescent trees uniquely determines the shape of a population history, i.e., that $N \neq N'$ implies that ${\mathbb{P}}_{N} \neq {\mathbb{P}}_{N'}$. This is indeed true: if we know ${\mathbb{P}}_N$, then we also know the rate of coalescence of two arbitrary individuals at any time $t$, which is just $1/N(t)$. Thus with an infinite amount of coalescence time data, the population shape can be reconstructed.
Considering sequence data, the model assumes that for $n$ individuals in a population, each genomic site follows an $n$-coalescent tree. Two sites have the same coalescent tree if there is no recombination breakpoint between them. At each site, mutations occur on top of the trees according to a Poisson process with small mutation rate and so, in principle, likelihoods of statistics of sequence data can be derived. Unfortunately, recombination is a complicated process and even under simplifying assumptions, likelihood functions are typically intractable, both analytically and computationally. Thus inexact methods must be developed, which we now describe.
Given whole genome data, likelihoods of various population parameters can be estimated across the parameter space by MCMC [@Nielsen2000; @Excoffier2013]. Using a simplified model of recombination [@McVean2005], simpler likelihood functions arise; however, these must still be analyzed using approximation schemes [@Li2011; @Sheehan2013]. Sequence data can also be used to infer the lengths of nonrecombinant blocks [@jasmine2014ibd], the distribution of which can be used to infer various aspects of the population history [@palamara2012length].
The problem simplifies when it is assumed that all loci in a given sequence are linked, that is, not separated by recombination events or in parts of the genome where no recombination occurs (such as mitochondrial DNA). In such cases, the coalescent trees at each of the sites are *identical*. For such data, given the coalescent tree, the number of segregating sites (where mutations have occurred) follows a Poisson distribution and analytic (though intractable) expressions for likelihoods can be derived.
If all sites are unlinked, one can use inference tools involving the population allele frequency spectrum [@Bhaskar2014] or Bayesian approaches such as the “Bayesian Skyline” [@Drummond2005; @heled2008bayesian] for both single and multi-locus data. Outside of the coalescent framework, the allele frequency spectrum and its diffusion approximation [@gutenkunst2009inferring; @Lukic2011] (which is derived from the underlying Wright-Fisher dynamics that also drive the coalescent) can also be used, though the same computational caveats as above apply. The allele frequency spectrum suffers from identifiability issues in general [@myers2008can], though not under biologically realistic assumptions [@bhaskar2013identifiability].
Results overview and applications {#sec:low}
---------------------------------
We provide lower bounds on the amount of *exact* coalescence time data necessary to infer past population history events. The assumption that our data are exact coalescence times is unrealistic but idealized: for a single, panmictic population, the rate of coalescence $t$ generations in the past determines the population size at that time and so the most direct route to estimating the population history is through the coalescence times. Since our lower bounds on the amount of samples are for idealized data, the bounds should also be taken to apply to methods which use sequence data (and should be considered as underestimates for such methods). In fact, all the previously mentioned coalescent-based methods used to infer population history based on sequence data also infer the coalescence times along the way (usually implicitly).
The following theorem provides bounds on the probability of correctly distinguishing between two population histories that differ only on an interval $(T, T+S)$ over which each is constant, given coalescence times between pairs of individuals at $L$ independent loci. See Figure \[fig:bottle\] for an illustration of two such histories.
\[thm:main\_bd\] Let $a, b$, and $S$ be positive constants and let $T {\geqslant}0$. Consider the following hypothesis testing problem: under both hypotheses the population sizes are equal in the intervals $[0,T)$ and $[T+S,\infty)$, given by some function $N(\cdot)$, but under $H_1$ the population size is constant $aN(0)=:aN_0$ in the interval $[T,T+S)$ while under $H_2$ the population size during the interval $[T,T+S)$ is constant $bN_0$. If $L$ independent coalescence times are observed from either $H_1$ or $H_2$, with prior probability $1/2$, then the Bayes error rate for any classifier is at least $(1-\operatorname{\mathcal{E}})/2$, where $$\begin{aligned}
\operatorname{\mathcal{E}}^2 &{\leqslant}2L \exp \left( - \int_0^{T} 1 / N\left( t \right) dt \right)\left(1-e^{-\frac{S}{2N_0} \frac{a+b}{ab}}\right)\frac{\left(\sqrt{a}-\sqrt{b} \right)^2}{a+b} \label{eq:main_bd1} \\
&{\leqslant}2 L \exp \left( - \int_0^{T} 1 / N\left( t \right) dt \right) \min \left\{ \frac{S}{2N_0}, \frac{ab}{a+b} \right\} \frac{\left( \sqrt{a} - \sqrt{b} \right)^2}{ab}. \label{eq:main_bd2}\end{aligned}$$ In other words, for any classification procedure, the chance of correctly determining whether the samples came from $H_1$ or $H_2$, is at most $(1+\operatorname{\mathcal{E}})/2$.
; coordinates [ (1,2) (2,2) ]{}; ;
; coordinates [ (1,1) (2,1) ]{}; ;
The main features of the bound of the theorem above are that if $L \ll \left( S/N_0 \right)^{-1}$, or if $L\ll 1/(\sqrt{a}-\sqrt{b})^2$, or if $T$ is large enough, then the chance of distinguishing between the two histories will be near $1/2$. Consequently, given $a,b,S$, and $T$, the theorem provides a lower bound on the number $L$ of independent coalescence times necessary in order to distinguish between the two histories with a given probability.
To understand the bound in more concrete settings and to compare Theorem \[thm:main\_bd\] to previous work, consider Li and Durbin [@Li2011], who apply the pairwise sequentially Markovian coalescent model (PSMC) to the complete diploid genome sequences of seven individuals in order to infer human population size history, with one of their main goals being to infer the timing of the out-of-Africa event which caused a bottleneck in East Asian and European populations. To validate their model, they apply PSMC to simulated data where the population histories consist of a sharp out-of-Africa bottleneck followed by a population expansion. They note that the simulations “reveal a limitation of PSMC in recovering sudden changes in effective population size.” We use Theorem \[thm:main\_bd\] to quantitatively show that *every method* must suffer from this to a certain extent.
Take the population history considered in [@Li2011 Fig. 2a], reproduced in the left panel of Fig. \[fig:LiDur\] below. Here the present day effective population size is $N_0:=N(0) = 2.732 \times 10^4$; the effective population size is $N_0$ in the time interval $[0,2.732 \times 10^4)$ back in time (measured in years, assuming $25$ years per generation), it is $0.05\times N_0$ in the time interval $[2.732 \times 10^4, 1.0245 \times 10^5)$ back in time, it is $0.5 \times N_0$ in the time interval $[1.0245 \times 10^5, 3.415 \times 10^6)$ back in time, and it is $N_0$ in the time interval $[3.415 \times 10^6, \infty)$ back in time. We apply Theorem \[thm:main\_bd\] to obtain bounds on the amount of data needed to estimate the timing of the bottleneck at approximately $100$ kyr to a given accuracy. We have $a = 0.05$, $b=0.5$, and $N_0 = 2.732 \times 10^4$. Assuming $25$ years per generation, we have $T = 0.15 N_0$ and $\int_0^{T} 1 / N\left( t \right) dt = \int_0^{0.04 N_0} 1/N_0 dt + \int_{0.04 N_0}^{0.15 N_0} 1/(0.05 N_0) dt = 2.24$.
coordinates [ (8000,2.732) (27320,2.732) (27320,2.732/20) (102450,2.732/20) (102450,2.732/2) (3415000,2.732/2) (3415000,2.732) (10\^7,2.732) ]{};
coordinates [ (8000,2.732) (27320,2.732) (27320,2.732/20) (802450,2.732/20) (802450,2.732/2) (3415000,2.732/2) (3415000,2.732) (10\^7,2.732) ]{};
Theorem \[thm:main\_bd\] tells us that, given coalescence times from $L$ independent loci, in order to distinguish between the two histories considered in Figure \[fig:LiDur\] with probability at least $0.95$, it is necessary that $\operatorname{\mathcal{E}}^2 {\geqslant}0.81$, so using and plugging in the numbers above, it is necessary that $$\label{eq:LiDurbinBD}
2L e^{-2.24} \left( 1 - e^{-\frac{S}{54640} \times \frac{0.55}{0.025}} \right) \frac{\left( \sqrt{0.5} - \sqrt{0.05} \right)^2}{0.55} {\geqslant}0.81.$$ From this we immediately see that there is no solution for $S$ when $L {\leqslant}8$, i.e., when the number of independent loci is too small, the length of the 95% “uncertainty interval”[^7] is infinite. When $L {\geqslant}9$, is equivalent to $$\label{eq:LiDurbinBD2}
S {\geqslant}\frac{27320}{11} \log \left( 1+ \frac{\frac{891 \times e^{2.24}}{4000 \times \left( \sqrt{0.5} - \sqrt{0.05} \right)^2}}{L - \frac{891 \times e^{2.24}}{4000 \times \left( \sqrt{0.5} - \sqrt{0.05} \right)^2}} \right),$$ i.e., the length of the 95% “uncertainty interval” is inversely proportional to $L$ for large $L$. Table \[table:bottle\] collects the numerical values of some of the lower bounds on the lengths of the 95% uncertainty intervals given by and (note that in and the unit of $S$ is generations, while in Table \[table:bottle\] the unit of time is years, where we assume $25$ years per generation). These estimates are in line with the simulation results of [@Li2011], where in the PSMC reconstruction of the population history the sudden drop in population is spread out over several tens of thousands of years.
Number of loci ${\leqslant}8$ 10 20 30 50
------------------------------------------- ---------------- ------------------- ------------------- ------------------- -------------------
Lower bound on interval length (in years) $\infty$ $1.3 \times 10^5$ $3.6 \times 10^4$ $2.1 \times 10^4$ $1.2 \times 10^4$
: Lower bounds on the lengths of the 95% uncertainty intervals for determining the timing of a bottleneck given a sample of $L$ independent loci in the scenario depicted in Figure \[fig:LiDur\].[]{data-label="table:bottle"}
Similarly, Theorem \[thm:main\_bd\] also provides bounds on the amount of coalescence data needed to estimate the time of the final jump in the population to $N_0$ in this same scenario. We may consider two population histories, one the same as Fig. 2a in [@Li2011], and the other a modified version where the final jump in the population to $N_0$ occurs at some time in the interval $[1.0245\times 10^5, 3.415\times 10^6]$ years in the past; see Fig. \[fig:LiDur2\].
coordinates [ (8000,2.732) (27320,2.732) (27320,2.732/20) (102450,2.732/20) (102450,2.732/2) (3415000,2.732/2) (3415000,2.732) (10\^7,2.732) ]{};
coordinates [ (8000,2.732) (27320,2.732) (27320,2.732/20) (102450,2.732/20) (102450,2.732/2) (602450,2.732/2) (602450,2.732) (10\^7,2.732) ]{};
Here we thus have $a=0.5$, $b=1$, $25(T+S)=3.415\times 10^6$, and $25T \in [1.0245\times 10^5, 3.415\times 10^6]$. For each such $T$ we have $$\begin{aligned}
\int_0^T 1/N(t) dt & =\int_0^{.04 N_0} 1/N_0 dt +\int_{.04 N_0}^{.15 N_0} 1/(.05 N_0) dt + \int_{.15N_0}^{T} 1/(. 5 N_0) dt =1.94 + 2(T/N_0).\end{aligned}$$ Plugging these expressions into of Theorem \[thm:main\_bd\] and taking $ab/(a+b)$ in the minimum, we find that, in order to recover the true population history with probability at least $0.95$ given $L$ independent samples, we must have $$\label{eq:lastjump}
T {\leqslant}13660 \log \left( L / 49.2 \right).$$ Notice again that recovery with $95\%$ chance is impossible when $L {\leqslant}49$, so a considerable amount of coalescence time data is required for accurate inference. When $L {\geqslant}50$, Table \[table:bd4\] summarizes lower bounds on the length of the $95\%$ uncertainty intervals implied by (again converted to years).
Number of loci $100$ $200$ $500$ $10^3$ $10^4$
---------------------------------------- ------------------- ------------------- ------------------- ------------------- -------------------
Lower bound on interval length (in yr) $3.1 \times 10^6$ $2.9 \times 10^6$ $2.6 \times 10^6$ $2.3 \times 10^6$ $1.6 \times 10^6$
: Lower bounds on the lengths of the 95% uncertainty intervals for determining the timing of the last population size change in the scenario depicted in Figure \[fig:LiDur2\], given a sample of $L$ independent loci.[]{data-label="table:bd4"}
The PSMC reconstruction of the population history ends at approximately 5 Myr, so the lengths of their uncertainty intervals on the timing of the last population size change are unclear, but appear to be at least a few Myr. Our results are therefore in line with these simulation results, and show that no method can perform substantially better than PSMC.
Organization of the paper
-------------------------
The layout of the paper is as follows: in the next section we describe a procedure that infers a population history given slightly more data than required by the lower bounds implied by Theorem \[thm:main\_bd\]. We prove our results on lower bounds, including Theorem \[thm:main\_bd\], in Section \[sec:lower\], and then prove results about the inference procedure in Section \[sec:algo\]. We support our results by simulations presented in Section \[sec:sim\], and we end with a summarizing discussion section with some open problems.
A Simple Reconstruction Algorithm {#sec:results}
=================================
To complement the results on lower bounds detailed above, we describe a simple estimation procedure (analyzed in Section \[sec:algo\]) that takes coalescence time data and returns an estimate for the population shape. The analysis of this procedure shows that the amount of data it requires almost matches the lower bounds stated in our results above.
The procedure takes i.i.d. pairwise coalescence times $\mathbf{t}^L = \left\{ t_1, \dots, t_L \right\}$ and returns a piecewise constant estimate $\widehat{N} = \left\{ \widehat{N} \left( t \right) \right\}_{t {\geqslant}0}$ for the population shape. The procedure involves a single parameter, ${\varepsilon}$, which controls the length of the time intervals where our estimate is constant, and which then also affects the accuracy of our estimate in each time interval. Assume that there are $N_0$ individuals initially, at time $0$.
1. Partition time backwards in time into intervals of length ${\varepsilon}N_0$, i.e., let $I_1 = \left[ 0, {\varepsilon}N_0 \right]$, $I_2 = \left[ {\varepsilon}N_0, 2 {\varepsilon}N_0 \right], \dots, I_K = \left[ \left( K - 1 \right) {\varepsilon}N_0, K {\varepsilon}N_0 \right]$. ($K$ is the minimum integer such that the interval $[0, K {\varepsilon}N_0]$ covers the data and we do not provide estimates past time $K {\varepsilon}N_0$.)
2. For $k = 1, \dots, K$, denote the fraction of data points lying in the time interval $I_k$ by $$\widehat{X}_k := \frac{1}{L} \# \left\{ i : t_i \in I_k \right\},$$ and furthermore let $\widehat{S}_0 = 0$ and $\widehat{S}_k = \sum_{i=1}^{k} \widehat{X}_i$, the fraction of data points lying in the time interval $\left[ 0, k {\varepsilon}N_0 \right]$.
3. Our estimate $\widehat{N}_k$ in the time interval $I_k$ is $$\label{eq:est}
\widehat{N}_k := \frac{{\varepsilon}N_0}{- \log \left( 1 - \frac{\widehat{X}_k}{1-\widehat{S}_{k-1}} \right)},$$ provided that $\widehat{X}_k > 0$, i.e., we have at least one data point in the time interval $I_k$. If $\widehat{X}_k = 0$, then we do not give an estimate.
The estimate is motivated by the fact that $$\frac{{\mathbb{P}}\left( t_1 \in I_k \right)}{{\mathbb{P}}\left( t_1 \notin \left[ 0, \left( k - 1 \right) {\varepsilon}N_0 \right] \right)} = 1 - \exp \left( - \int_{\left( k -1 \right) {\varepsilon}N_0}^{k{\varepsilon}N_0} \frac{1}{N \left( t \right)} dt \right).$$
In Step 1 above, we partition time into intervals of equal length. This is done solely to make the subsequent analysis and discussion as simple as possible. Depending on the specific application, it might be of interest to consider other choices of partitions, for instance, choosing intervals whose lengths grow exponentially backwards in time. Our estimation procedure (and also the subsequent analysis) works in an analogous way: $\widehat{X}_k$ and $\widehat{S}_k$ can be defined in the same way in Step 2, and the only change in the estimate is to replace ${\varepsilon}N_0$ in the numerator of the fraction with the length of the appropriate interval, $\left| I_k \right|$.
In order to state the properties of this procedure, define for $k {\geqslant}1$ the “effective constant population size in the time interval $I_k$” by $$\widetilde{N}_k := \frac{{\varepsilon}N_0}{\int_{\left( k - 1 \right) {\varepsilon}N_0}^{k{\varepsilon}N_0} \frac{1}{N\left( t \right)} dt};$$ the $\widetilde{N}_k $ give a natural piecewise constant approximation of the population shape $N(t)$ that is directly comparable to the piecewise estimate $\widehat{N}$. Let $$E_k := \sup_{t \in I_k} \left| \log N\left( t \right) - \log \widehat{N}_k \right|$$ be the absolute error of our estimate $\widehat{N}$ on a logarithmic scale for each time interval. When estimating the error $E_k$, there are two types of errors to consider. One is the inherent error coming from the fact that we are approximating the shape with a piecewise constant function; the other error comes from the finite sample size $L$. By the triangle inequality we can bound the error $E_k$ by the sum of these two errors: $$E_k {\leqslant}E_{k,1} + E_{k,2},$$ where $$E_{k,1} := \sup_{t \in I_k} \left| \log N \left( t \right) - \log \widetilde{N}_k \right|$$ is the error coming from approximating the shape in the time interval $I_k$ with a constant, and $$E_{k,2} := \left| \log \widetilde{N}_k - \log \widehat{N}_k \right|$$ is the error coming from the finite sample size. Ignoring the error $E_{k,1}$ for now, we can use concentration inequalities to derive the following finite sample estimate for the accuracy of our estimator:
\[prop:k\_cond\_intro\] Given that $\ell$ samples “survived” the first $k-1$ intervals, the probability that $\log \widetilde{N}_k$, is in the (random) interval $$\label{eq:k_cond_int_intro}
\left[ \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( \left( 1 - \frac{L}{\ell}\widehat{X}_k - c \right) \vee 0 \right) \right), \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( \left( 1 - \frac{L}{\ell}\widehat{X}_k + c \right) \wedge 1 \right) \right) \right]$$ is at least $1 - 2 \exp \left( - 2 c^2 \ell \right)$ for all $c {\geqslant}0$.[^8]
Note that the interval in the proposition contains the estimate $\log \widehat{N}_k$.
To understand in what sense Proposition \[prop:k\_cond\_intro\] and Theorem \[thm:main\_bd\] are matching bounds, first consider the following easy corollary of Theorem \[thm:main\_bd\] that better matches the setting of Proposition \[prop:k\_cond\_intro\].
\[lem:bd4\_intro\] Let $a, b$, and $S$ be positive constants and $T {\geqslant}0$. Consider the following hypothesis testing problem: Under $H_1$ the population size is constant $aN(0)=:aN_0$ in the interval $[T,T+S)$ while under $H_2$ the population size during the interval $[T,T+S)$ is constant $bN_0$. Assume the data are independent coalescence times and let $\ell$ be the number of pairs that have not coalesced by time $T$. If the true history is given by either $H_1$ or $H_2$, each with prior probability $1/2$, then the Bayes error rate for any classifier is at least $(1 - \Delta)/2$, where $\Delta$ satisfies: $$\label{eq:bd4_intro}
\Delta^2 {\leqslant}2\ell \frac{\left( \sqrt{b}-\sqrt{a}\right)^2}{a+b}. $$ In other words, for any classification procedure, the chance of correctly determining whether the samples came from $H_1$ or $H_2$, is at most $(1+\Delta)/2$.
Writing $b=a(1+\eta)$ for $\eta>0$, the bound of the theorem becomes $2\ell (1-2\sqrt{1+\eta}/(2+\eta)){\leqslant}\ell \eta^2/4$ and so we see that if $\eta \ll \ell^{-1/2}$, then *no procedure* will distinguish between the two histories given by $H_1$ and $H_2$ with good probability. On the other hand, Proposition \[prop:k\_cond\_intro\] implies that for a fixed confidence $\alpha$, $$1 - 2 \exp \left( - 2 c^2 \ell \right)=\alpha,$$ and the constant $c$ is of order $\ell^{-1/2}$ as $\ell$ becomes large, and thus the width of the interval in Proposition \[prop:k\_cond\_intro\] is of order $\log(1+C \ell^{-1/2})$ where $C$ is some constant. To summarize, if $\eta \gg \ell^{-1/2}$, then our method will distinguish between the histories with high probability; but if $\eta \ll \ell^{-1/2}$, Theorem \[lem:bd4\_intro\] shows that *no procedure* will distinguish between the two histories with good probability. On a conceptual level, this last statement is the main purpose of the paper: a significant amount of data is needed to infer past population size, especially in deep history where there is likely to be little coalescence information.
For illustration, we implement our estimation procedure on simulated data in Section \[sec:sim\], where we find a good general performance, matching our theoretical results.
Related theoretical work {#sec:related}
------------------------
Our reconstruction algorithm is a special case of the following problem: given $n$ i.i.d. copies of the *first point* of a Poisson point process on $[0,\infty)$ with intensity ${\varphi}(t)$, what is a good estimate of ${\varphi}(t)^{-1}$? Poisson process intensity estimation has a large literature, see for example [@Birge2007; @Reynaud-Bouret2003; @Willett2007] and references therein, but the (natural) data assumed in this area is one realization of the point process, or the point process observed up to some fixed time, or i.i.d. copies of such data, which does not fit our framework.
For another perspective to this question, define the hazard rate for a positive random variable $X$ with density $f$ and distribution function $F$ to be $$-\frac{d}{dt} \log(1-F(t))=f(t)/(1-F(t)). \label{222}$$ A simple calculation shows that the time of the first point of a Poisson process with intensity ${\varphi}(t)$ has the same distribution as a positive random variable with hazard rate ${\varphi}(t)$. Due largely to their importance in applications in, e.g., insurance, medicine, and reliability theory [@Lawless2003 Section 1.1], hazard rate estimation is well studied; some seminal papers are [@Rice1976; @Sethuraman1981; @Yandell1983] and see the recent [@Cheng2006] and references there. Without embellishments specific to lifetime data (such as censoring where some lifetimes are only known to be at least some value), the main technique to estimating (which also applies to its inverse) is to adapt estimators of $f$ and $F$.
Indeed, our reconstruction algorithm is essentially an adaptation of the histogram estimate of the density and distribution function to our setting. Other popular density estimation techniques such as those in the introduction of [@Silverman1986] can be adapted to our setting through the use of ; for example see [@Wang2005] for a survey of kernel smoothing methods for hazard function estimation. Our particular estimation procedure was chosen due to its simplicity and explicitness; in particular, we mention two points. The first is that we desire results like Proposition \[prop:k\_cond\_intro\] with explicit non-asymptotic confidence intervals. Asymptotic confidence intervals can be obtained and used as estimates for smoothed density estimators, but with error depending on unknown quantities related to the underlying density which can lead to poor coverage accuracy [@Hall1992]. Secondly, smoothed density estimators have improved performance only when the underlying density is itself smooth (expressed as differentiability and continuity conditions). A major purpose of estimating past population size is to discover drastic changes in population size such as bottlenecks [@harris2013inferring; @Li2011; @palamara2012length; @Sheehan2013], when it is not clear such smoothness assumptions are appropriate.
Proof of Lower Bounds {#sec:lower}
=====================
In this section we prove Theorem \[thm:main\_bd\], as well as derive some other lower bounds for the amount of data needed for a given accuracy of estimating the population shape. This is done by formulating hypothesis tests deciding between two population shapes, and proving upper bounds on the probability of correctly inferring the population shape.
Background on probability metrics {#sec:prob_metrics}
---------------------------------
We first recall a few metrics between probability distributions (see [@gibbs2002choosing] for a survey). Let $P$ and $Q$ be two probability measures that are absolutely continuous with respect to a third probability measure $\lambda$. Write $f_P = \frac{dP}{d\lambda}$ and $f_Q = \frac{dQ}{d\lambda}$ for the respective Radon-Nikodym derivatives. The square of the Hellinger distance between $P$ and $Q$ is then defined as $$d_H^2 \left( P, Q \right) := \frac{1}{2} \int \left( \sqrt{f_P} - \sqrt{f_Q} \right)^2 d\lambda.$$ The definition does not depend on the choice of $\lambda$. A nice property of the Hellinger distance is that for product measures $P = P_1 \times P_2$, $Q = Q_1 \times Q_2$, we have that $$1 - d_H^2 \left( P, Q \right) = \left( 1 - d_H^2 \left( P_1, Q_1 \right) \right) \left( 1 - d_H^2 \left( P_2, Q_2 \right) \right),$$ which immediately implies that $$d_H^2 \left( P, Q \right) {\leqslant}d_H^2 \left( P_1, Q_1 \right) + d_H^2 \left( P_2, Q_2 \right).$$ Another commonly used metric is the total variation distance: $$d_{TV} \left( P,Q \right) := \sup_{A \in \operatorname{\mathcal{F}}} \left| P \left( A \right) - Q \left( A \right) \right|,$$ or, equivalently: $$d_{TV} \left( P,Q \right) = \frac{1}{2} \int \left| f_P - f_Q \right| d\lambda.$$ We use the following well-known fact:
With the notation above we have $$d_{TV} {\leqslant}\sqrt{2} d_H.$$
This follows from the identity $f_P - f_Q = \left( \sqrt{f_P} - \sqrt{f_Q} \right) \left( \sqrt{f_P} + \sqrt{f_Q} \right)$, the Cauchy-Schwarz inequality, and the inequality $\left( \sqrt{f_P} + \sqrt{f_Q} \right)^2 {\leqslant}2 \left( f_P + f_Q \right)$.
A lower bound on the amount of data needed to recover a constant history {#sec:con_hist}
------------------------------------------------------------------------
We start in a simpler setting than Theorem \[thm:main\_bd\] where we are trying to differentiate with good probability between two populations of *constant* size. For this simple setup we assume our data are $L$ i.i.d. copies of coalescent trees on $n$ individuals from a constant population, and we want to estimate the size of the population. We derive lower bounds on the amount of data needed for recovery.
\[lem:bdm\_intro\] Consider the following hypothesis testing problem: $H_1$ states that the population size during the interval $[0 ,\infty)$ is constant $N$, while $H_2$ states that the population size during the interval $[0,\infty)$ is the constant $(1+\eta)N$, where $\eta > 0$ is fixed. If $L$ i.i.d. coalescent trees on $n$ individuals are observed from either $H_1$ or $H_2$, each with prior probability $1/2$, then the Bayes error rate for any classifier is at least $(1-\Upsilon)/2$, where $\Upsilon$ satisfies: $$\Upsilon^2 {\leqslant}2L \left(1-\left(\frac{2\sqrt{1+\eta}}{2+\eta}\right)^{n-1}\right) {\leqslant}\frac{L(n-1)\eta^2}{4}.$$ In other words, for any classification procedure, the chance of correctly determining whether the samples come from $H_1$ or $H_2$, is at most $(1+\Upsilon)/2$.
The interpretation of the theorem is that if $\eta \ll (n L)^{-1/2}$, then *no procedure* will distinguish between the two histories given by $H_1$ and $H_2$ with good probability. In other words, we need $L = \Omega \left( 1 / \left( n \eta^2 \right) \right)$ samples to differentiate between the two histories $N_1$ and $N_2$.[^9] We reiterate that these bounds hold knowing exact rather than estimated coalescence times, and so should be considered as underestimates in more realistic data settings.
We set up for the proof of Theorem \[lem:bdm\_intro\]; the same paradigm will be used to prove Theorem \[thm:main\_bd\]. Consider the following hypothesis testing problem. Let $\eta > 0$, and let $N_1 \left( \cdot \right) \equiv N$ and $N_2 \left( \cdot \right) \equiv \left( 1 + \eta \right) N$ be two population size histories. Let $\kappa$ be uniform in $\left\{ 1, 2 \right\}$, and, given $\kappa$, let $\mathbf{R}^{\kappa, L} = \left\{ R_1^\kappa, \dots, R_L^\kappa \right\}$ be a collection of $L$ i.i.d. coalescence trees on $n$ individuals drawn from the distribution induced by the population size history $N_{\kappa}$. The problem is to infer $\kappa$ from $\mathbf{R}^{\kappa,L}$.
The probability of correctly inferring $\kappa$ using the optimal reconstruction strategy is clearly at least $1/2$; denote this probability by $\left( 1 + \Upsilon \right) / 2$ (here $\Upsilon = \Upsilon \left( L,n,N, \eta \right)$). The reconstruction method which gives the largest probability of correctly inferring $\kappa$ is maximum likelihood: let $\widehat{\kappa} = 1$ if ${\mathbb{P}}\left( \kappa = 1 \, \middle| \, \mathbf{R}^{\kappa, L} \right) {\geqslant}{\mathbb{P}}\left( \kappa = 2 \, \middle| \, \mathbf{R}^{\kappa, L} \right)$ and $\widehat{\kappa} = 2$ otherwise. Then we have $$\Upsilon = {\mathbb{P}}\left( \widehat{\kappa} = \kappa \right) - {\mathbb{P}}\left( \widehat{\kappa} \neq \kappa \right) = d_{TV} \left( \mathbf{R}^{1, L}, \mathbf{R}^{2, L} \right).$$
By the facts in Section \[sec:prob\_metrics\] we have $$\label{2201}
\Upsilon(L,n,N,\eta)^2 = d_{TV}^2\left( \mathbf{R}^{1,L}, \mathbf{R}^{2,L}\right){\leqslant}2 d_{H}^2\left( \mathbf{R}^{1,L}, \mathbf{R}^{2,L}\right){\leqslant}2 L d_H^2 \left( R_1^1, R_1^2 \right).$$ Since the increasing sequence of times of coalescence of the trees $R_1^i$, denoted by $\mathbf{s}^i=(s_1^i, \ldots, s_{n-1}^i)$, are sufficient statistics for $R_1^i$, we have $d_H^2 \left( R_1^1, R_1^2 \right)=d_H^2(\mathbf{s}^1,\mathbf{s}^2)$. We can directly compute the density $f_i(\mathbf{x})$ of $\mathbf{s}^i$ as $$f_i(\mathbf{x})=\prod_{j=1}^{n-1} \exp\left\{-\binom{n-j+1}{2}(x_j-x_{j-1})/N_i \right\}\frac{\binom{n-j+1}{2}}{N_i} \mathbf{1}_{\left\{ 0 < x_1 < \dots < x_{n-1} \right\}},$$ where we have set $N_1:=N$, $N_2:=(1+\eta)N$, and $x_0=0$. Using these densities in the definition of the Hellinger distance and noting especially that since $f_i$ is a density, we have for any $\alpha>0$ that $$\int_{0<x_1<\ldots< x_{n-1}} \prod_{j=1}^{n-1} \exp\left\{-\alpha \binom{n-j+1}{2}(x_j-x_{j-1}) \right\}\binom{n-j+1}{2}d\mathbf{x}=1/\alpha^{n-1},$$ a calculation shows that $$\label{eq:dH1}
d_H^2 \left( s_1^1, s_1^2 \right) = 1 - \left(\frac{2 \sqrt{1+\eta}}{2 + \eta}\right)^{n-1}.$$ Plugging this into yields the first bound of the result. When $\eta > 0$, we can upper bound the right hand side of by $(n-1)\eta^2 / 8$ to get the simpler bound $\Upsilon^2 {\leqslant}L (n-1) \eta^2 / 4$.
Proof of Theorem \[thm:main\_bd\] {#sec:main_bd}
---------------------------------
We prove Theorem \[thm:main\_bd\] using the same strategy as that of Section \[sec:con\_hist\]. For $i=1,2$, let $N_i(\cdot)$ be the history corresponding to hypothesis $H_i$. Let $\kappa$ be uniform in $\left\{ 1, 2 \right\}$, and, given $\kappa$, let $\mathbf{t}^{\kappa, L} = \left\{ t_1^\kappa, \dots, t_L^\kappa \right\}$ be a collection of $L$ i.i.d. coalescence times of pairs of individuals drawn from the distribution induced by the population size history $N_{\kappa}$. The problem is to infer $\kappa$ from $\mathbf{t}^{\kappa,L}$.
As above, the chance that we infer $\kappa$ correctly from $\mathbf{t}^{\kappa,L}$ is bounded above by $(1+\operatorname{\mathcal{E}}(L,a,b,T,S))/2$ where $$\operatorname{\mathcal{E}}(L,a,b,T,S)^2=d_{TV}^2\left( \mathbf{t}^{1,L}, \mathbf{t}^{2,L}\right){\leqslant}2 L d_H^2 \left( t_1^1, t_1^2 \right). \label{2200}$$ Writing $a_1:=a$ and $a_2:=b$ to shorten formulas, a straightforward calculation shows that the density of $t_1^i$ is $$f_i(x)=
\begin{cases}
\exp\left(-\int_0^x\frac{1}{N(s)} ds\right) \frac{1}{N(x)}, & x < T,\\
\exp\left(-\int_0^T\frac{1}{N(s)} ds\right) \exp\left(- \frac{x-T}{a_i N_0}\right) \frac{1}{a_i N_0}, & T{\leqslant}x<T+S,\\
\exp\left(-\int_0^T\frac{1}{N(s)} ds\right)\exp\left(- \frac{S}{a_i N_0}\right) \exp\left(-\int_{T+S}^x\frac{1}{N(s)} ds\right) \frac{1}{N(x)}, & T+S {\leqslant}x.
\end{cases}$$ Using these densities in the definition of the Hellinger distance, we find after some simple calculations that $$d_H^2 \left( t_1^1, t_1^2 \right) =
\exp \left( - \int_0^{T} 1 / N\left( t \right) dt \right)\left(1-e^{-\frac{S}{2N_0} \frac{a+b}{ab}}\right)\frac{\left(\sqrt{a}-\sqrt{b} \right)^2}{a+b}.$$ Finally, using the inequality $1 - e^{-x} {\leqslant}\min \left\{x, 1 \right\}$, simplifying, and plugging the result into implies the bound of the theorem.
Estimating the population shape {#sec:algo}
===============================
Recall our setting of the estimation procedure for the population shape $N = \left\{ N\left(t \right) \right\}_{t {\geqslant}0}$ from the i.i.d. coalescence times $\mathbf{t}^L = \left\{ t_1, \dots, t_L \right\}$. In this section we analyze our piecewise constant population shape estimator $\widehat{N} = \left\{ \widehat{N} \left( t \right) \right\}_{t {\geqslant}0}$ introduced in Section \[sec:results\]. We consider the absolute error of our estimate $\widehat{N}$ on a logarithmic scale for each time interval, i.e., for $k{\geqslant}1$ we consider $$E_k := \sup_{t \in I_k} \left| \log N\left( t \right) - \log \widehat{N}_k \right|.$$ Recall also that $E_k {\leqslant}E_{k,1} + E_{k,2}$, where $E_{k,1}$ and $E_{k,2}$ are also both defined in Section \[sec:results\]. To bound the error $E_{k,1}$ it is necessary to make an additional assumption on the population shape. We introduce an additional parameter, $\delta$, which controls how much the population size can vary within a time interval, and we make the following assumption.
\[ass:delta\] We assume that in each time interval the population size can increase by a factor of at most $e^{\delta {\varepsilon}}$, and can decrease by a factor of at most $e^{- \delta {\varepsilon}}$.
Using this assumption, it is simple to bound the first type of error.
Given Assumption 1, we have that $E_{k,1} {\leqslant}2 \delta {\varepsilon}$.
Assumption 1 implies that $$\log N \left( \left( k - 1 \right) {\varepsilon}N_0 \right) - \delta {\varepsilon}{\leqslant}\log N \left( t \right) {\leqslant}\log N \left( \left( k - 1 \right) {\varepsilon}N_0 \right) + \delta {\varepsilon}$$ for all $t \in I_k$, and consequently also that $$\log N \left( \left( k - 1 \right) {\varepsilon}N_0 \right) - \delta {\varepsilon}{\leqslant}\log \widetilde{N}_k {\leqslant}\log N \left( \left( k - 1 \right) {\varepsilon}N_0 \right) + \delta {\varepsilon}.$$ These inequalities then imply that $E_{k,1} {\leqslant}2 \delta {\varepsilon}$.
To estimate the second type of error, $E_{k,2}$, it is not necessary to make any assumptions. We use concentration results for sums of i.i.d. random variables, and, in particular, we use the following simple corollary of the Chernoff bound.
\[thm:Chernoff\] Let $Y_1, \dots, Y_n$ be i.i.d. Bernoulli($p$) random variables, and let $Y = \sum_{i=1}^n Y_i$. Then for any $\lambda > 0$ we have $$\label{eq:Ch1}
\left.
\begin{aligned}
&{\mathbb{P}}\left( Y {\leqslant}np - \lambda \right)\\
&{\mathbb{P}}\left( Y {\geqslant}np + \lambda \right)
\end{aligned}
\right\}
{\leqslant}\exp \left( - \frac{2 \lambda^2}{n} \right).$$
The bounds in Theorem \[thm:Chernoff\] imply the following concentration bound.
\[cor:chernoff\] For any $k {\geqslant}1$ and $\lambda > 0$ we have $$\label{eq:conc2}
{\mathbb{P}}\left( \left| \widehat{X}_k - {\mathbb{E}}\left( \widehat{X}_k \right) \right| {\geqslant}\lambda \right) {\leqslant}2 \exp \left( - 2 \lambda^2 L \right).$$
In the following we present two bounds on the error $E_{k,2}$. We first present a bound for the first interval (i.e., when $k = 1$), which then also implies conditional bounds for general intervals, by conditioning on the number of data points that have not coalesced by a given time.
Bounds for the first interval {#sec:bds1}
-----------------------------
\[prop:k1-simple\] For any $c {\geqslant}0$, with probability at least $1 - 2 \exp \left( - 2 c^2 L \right)$, the logarithm of the effective constant population size in $I_1$, $\log \widetilde{N}_1$, is in the interval $$\label{eq:k1_conf_int}
\left[ \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( \left( 1 - \widehat{X}_1 - c \right) \vee 0 \right) \right), \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( \left( 1 - \widehat{X}_1 + c \right) \wedge 1 \right) \right) \right].$$
Note that the interval in is an interval around our estimate $\log \widehat{N}_1$.
The inequality for $k=1$ can be rephrased as $${\mathbb{P}}\left( {\mathbb{E}}\widehat{X}_1 \in \left[ \left( \widehat{X}_1 - c \right) \vee 0, \left( \widehat{X}_1 + c \right) \wedge 1 \right] \right) {\geqslant}1 - 2 \exp \left( - 2 c^2 L \right).$$ By algebraic manipulation, ${\mathbb{E}}\widehat{X}_1 \in \left[ \left( \widehat{X}_1 - c \right) \vee 0, \left( \widehat{X}_1 + c \right) \wedge 1 \right]$ is equivalent to $\log \widetilde{N}_1$ being contained in the interval in .
This bound is useful because we can immediately determine a confidence interval for our estimate. To achieve a confidence level of $1-\alpha$, we can choose $c = c \left( \alpha, L \right)$ to satisfy $2 \exp \left( - 2 c^2 L \right) = \alpha$, i.e., choose $$\label{eq:c}
c = \sqrt{\frac{\log \left( 2 / \alpha \right)}{2L}}.$$ Then the interval in with $c$ given by has a confidence level of $1- \alpha$.
Conditional bounds {#sec:bds_cond}
------------------
Next, we present conditional bounds: given the number of samples that did not coalesce in the time interval $\left[0,\left( k-1 \right) {\varepsilon}N_0 \right]$, what is the error we make when estimating the population size in the time interval $I_k$? The following result is the same as Proposition \[prop:k\_cond\_intro\] but worded more precisely.
\[prop:k\_cond\] For any $c {\geqslant}0$, the probability conditioned on $L \left( 1 - \widehat{S}_{k-1} \right) = \ell$ (i.e., that $\ell$ samples “survived” the first $k-1$ intervals) that the logarithm of the effective constant population size, $\log \widetilde{N}_k$, is in the interval $$\label{eq:k_cond_int}
\left[ \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( \left( 1 - \frac{L}{\ell}\widehat{X}_k - c \right) \vee 0 \right) \right), \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( \left( 1 - \frac{L}{\ell}\widehat{X}_k + c \right) \wedge 1 \right) \right) \right]$$ is at least $1 - 2 \exp \left( - 2 c^2 \ell \right)$.
Note that the interval in is an interval around our estimate $\log \widehat{N}_k$, given $L \left( 1 - \widehat{S}_{k-1} \right) = \ell$.
Let $\widehat{Y}_k := \frac{L}{\ell} \widehat{X}_k$. Given $L \left( 1 - \widehat{S}_{k-1} \right) = \ell$, $\widehat{Y}_k$ is the average of $\ell$ i.i.d. indicator variables. Therefore Chernoff’s bound gives that $${\mathbb{P}}\left( \left| \widehat{Y}_k - {\mathbb{E}}\widehat{Y}_k \right| {\geqslant}c \, \middle| \, L \left( 1 - \widehat{S}_{k-1} \right) = \ell \right) {\leqslant}2 \exp \left( - 2 c^2 \ell \right).$$ In other words, $${\mathbb{P}}\left( {\mathbb{E}}\widehat{Y}_k \in \left[ \left( \widehat{Y}_k - c \right) \vee 0, \left( \widehat{Y}_k + c \right) \wedge 1 \right] \, \middle| \, L \left( 1 - \widehat{S}_{k-1} \right) = \ell \right) {\geqslant}1 - 2 \exp \left( - 2 c^2 \ell \right).$$ Just as in the proof of Proposition \[prop:k1-simple\], by algebraic manipulation, given $L \left( 1 - \widehat{S}_{k-1} \right) = \ell$, ${\mathbb{E}}\widehat{Y}_k \in \left[ \left( \widehat{Y}_k - c \right) \vee 0, \left( \widehat{Y}_k + c \right) \wedge 1 \right]$ is equivalent to $\log \widetilde{N}_k$ being contained in the interval in .
Again, this bound is useful because we can immediately determine a confidence interval for our estimate. To achieve a confidence level of $1-\alpha$, we can choose $c = c \left( \alpha, \ell \right)$ to satisfy $2 \exp \left( - 2 c^2 \ell \right) = \alpha$, i.e., choose $$\label{eq:c_conf_int}
c = \sqrt{\frac{\log \left( 2 / \alpha \right)}{2\ell}}.$$ Then the interval in with $c$ given by has a confidence level of $1- \alpha$.
Simulations {#sec:sim}
===========
We illustrate our estimation procedure on simulated data for the following settings: (1) constant size population, (2) piecewise constant size population, and (3) a population experiencing recent exponential growth; the last setting being germane to recent human population history (see, e.g., [@tennessen2012evolution] for a study on how recent accelerated population growth, together with weak purifying selection, can lead to an excess of rare functional variants). In each case, we simulate $L$ independent coalescence times and apply our estimation procedure described in Section \[sec:algo\] with a given ${\varepsilon}$ to the data; the outcome is summarized in Figures \[fig:constant\]–\[fig:reconstruct\] below. Each figure plots $\log \left( N\left( t \right) / N \left( 0 \right) \right)$ versus $t / N \left( 0 \right)$, i.e., we scale time according to the coalescent timescale, and we plot the population size on a logarithmic scale.
The true history is the blue line, the estimates over each interval are the red lines, and the confidence intervals at the 95 percent level are given in pink. Recall that the logarithm of our estimate is $$\log \widehat{N}_k = \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( 1 - \frac{\widehat{X}_k}{1 - \widehat{S}_{k-1}} \right) \right),$$ and our confidence interval for confidence level $1 - \alpha$ is given by : $$\mbox{
\small
$
\left[ \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( \left( 1 - \frac{\widehat{X}_k}{1 - \widehat{S}_{k-1}} - c \right) \vee 0 \right) \right), \log \left( {\varepsilon}N_0 \right) - \log \left( - \log \left( \left( 1 - \frac{\widehat{X}_k}{1 - \widehat{S}_{k-1}} + c \right) \wedge 1 \right) \right) \right],
$
}$$ where $$c = \sqrt{\frac{\log \left( 2 / \alpha \right)}{2L \left( 1 - \widehat{S}_{k-1} \right)}}.$$ In the case that $\widehat{X}_k=0$ we do not give an estimate but can sometimes still obtain a lower bound on the confidence interval. The intervals where the pink extends to the upper or lower margin of the graphing area represent a confidence bound that is infinite or zero, i.e., where the minimum or maximum are taken to be one or zero in the expressions for the confidence bounds above.
The error $E_{k,1}$ is not represented in the plots, but would add $\delta$ to each side of each confidence interval. Alternatively, we can view our statistic as an estimate of the effective constant population size over the given interval.
The plots in Figure \[fig:reconstruct\] compare our lower bounds to the confidence intervals of our estimation procedure. The black lines are the 95 percent uncertainty intervals: in a given time interval, populations within the interval given by the black lines cannot be distinguished from the true blue line population with the amount of data in hand with probability $.95$. Thus if the red line is within the interval given by the black lines, then our estimate is in some sense the best that can be achieved. Note that when an interval has no upper black line, then there is not enough data to distinguish between a history of the blue line size in the interval and *any* larger size population with 95% certainty.
Even though our assumed data is idealized and unrealistic (exact coalescent data at thousands of independent sites), these simulations allow us to make some general qualitative observations. The major determining factor of the performance of our procedure in a time interval is the number of coalescence times that have survived to that interval. So having more data (i.e., larger $L$) leads to more accurate estimates, and the estimates lose accuracy moving back in time as the number of data points decreases. Moreover, there is a rare event effect when there are few coalescences in an interval—having no coalescences in an interval is not very informative—and this leads to the consistent underestimates in the deepest part of the histories. Possibly this effect would be lessened by lengthening the widths of the intervals as they go back in time, although this would smooth out big features in the history. Also note that in the presence of a bottleneck, i.e., a time period where the population becomes small, there are many coalescences due to the increased rate. In turn this decreases the number of available data deeper in history. For example, compare the accuracy of the estimates at time $t/N(0)=4$ in the constant population of Figure \[fig:constant\] to that of the piecewise constant population of Figure \[fig:piece\] where there is a bottleneck starting around time $t/N(0)=1$. Finally, note that in Figure \[fig:reconstruct\] the confidence intervals of our procedure and the lower bound black lines are rather tight in the presence of a significant amount of data, but loosen as the number of data points decreases.
Discussion {#sec:discussion}
==========
An assortment of methods have been developed to infer a population’s history from (an ever increasing amount of) genetic data [@Bhaskar2014; @Drummond2005; @Excoffier2013; @harris2013inferring; @heled2008bayesian; @Li2011; @Nielsen2000; @palamara2012length; @Sheehan2013]. These methods are necessarily computational and approximate and so the quality of the outputs of these methods cannot be rigorously justified. However, understanding the theoretical limitations of inferring past population history [@bhaskar2013identifiability; @myers2008can] is of the utmost importance, since it is becoming increasingly common for such analyses to be used as the main tool for inference, with less emphasis on external verification (e.g., fossil or paleontological record) [@Bos2014].
Here we have provided lower bounds on the amount of idealized data needed to infer a population history to a given accuracy. Our bounds should be considered as underestimates of the amount of data necessary for inference in methods which use sequence data, so they can be used as a guide when performing such analyses. We end with some further avenues of study and open problems.
Open Problems {#sec:open}
-------------
**$n$-coalescence trees.** With the exception of Theorem \[lem:bdm\_intro\], we assume that our data are $L$ i.i.d. coalescence times between pairs of individuals. If instead our data are $L$ i.i.d. coalescence trees among $n$ individuals then how does this affect the bounds? In the setting of Theorem \[lem:bdm\_intro\] when comparing two constant populations, increasing the number of individuals $n$ in the coalescent tree is as good as increasing the number of independent loci $L$. This shouldn’t hold true in general since adding individuals does not greatly increase the depth of the tree. For moderate values of $n$, we expect that estimates of the deep history will not be greatly affected since the time of coalescence for the final two lineages is roughly half the length of the coalescent tree started from *infinitely* many individuals. On the other hand, explosive growth in the near history should be estimated better using more individuals since the amount of coalescing in the near past will increase. It would be interesting to better understand how increasing the number of individuals in the tree affects the lower bounds of Section \[sec:low\] and the upper bounds provided by some generalization of our inference algorithm of Section \[sec:results\]. For some discussion on the affect of increasing the size of the tree versus increasing independent loci, see [@heled2008bayesian].
**Estimation from sequence data.** The assumption that we know exact coalescence times is unrealistic. These times need to be estimated from sequence data at independent loci with good accuracy. How do we estimate coalescence times from sequence data with quantitative upper and lower bounds analogous to those here?
**Population substructure.** Assume we want to estimate a population that is not only changing over time, but also has sub-populations that merge and split, and which may have migration rates between them. Are there analogs of our results in this setting? Note that identifiability can be an issue here since, for example, a constant population that splits at a given point in the past has the same distribution of coalescence times among two individuals as a single population that grows exponentially at a specific rate (backward in time) starting at the time of the split.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Anand Bhaskar, Luke Gandolfo, Jasmine Nirody, Sara Sheehan, and Yun Song for helpful discussions and relevant references. A portion of the work for this project was completed when NR was at University of California, Berkeley with support from NSF grants DMS-0704159, DMS-0806118, DMS-1106999 and ONR grant N00014-11-1-0140.
[0.4]{} ![Estimating a constant population size.[]{data-label="fig:constant"}](L13e05con.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a constant population size.[]{data-label="fig:constant"}](L14e05con.pdf "fig:"){width="\textwidth"}
\
[0.4]{} ![Estimating a constant population size.[]{data-label="fig:constant"}](L13e10con.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a constant population size.[]{data-label="fig:constant"}](L14e10con.pdf "fig:"){width="\textwidth"}
\
[0.4]{} ![Estimating a constant population size.[]{data-label="fig:constant"}](L13e20con.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a constant population size.[]{data-label="fig:constant"}](L14e20con.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a piecewise constant population size.[]{data-label="fig:piece"}](L13e05piece1.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a piecewise constant population size.[]{data-label="fig:piece"}](L14e05piece1.pdf "fig:"){width="\textwidth"}
\
[0.4]{} ![Estimating a piecewise constant population size.[]{data-label="fig:piece"}](L13e10piece1.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a piecewise constant population size.[]{data-label="fig:piece"}](L14e10piece1.pdf "fig:"){width="\textwidth"}
\
[0.4]{} ![Estimating a piecewise constant population size.[]{data-label="fig:piece"}](L13e20piece1.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a piecewise constant population size.[]{data-label="fig:piece"}](L14e20piece1.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a population history with piecewise exponential change. []{data-label="fig:piex"}](L13e05piex.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a population history with piecewise exponential change. []{data-label="fig:piex"}](L14e05piex.pdf "fig:"){width="\textwidth"}
\
[0.4]{} ![Estimating a population history with piecewise exponential change. []{data-label="fig:piex"}](L13e10piex.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a population history with piecewise exponential change. []{data-label="fig:piex"}](L14e10piex.pdf "fig:"){width="\textwidth"}
\
[0.4]{} ![Estimating a population history with piecewise exponential change. []{data-label="fig:piex"}](L13e20piex.pdf "fig:"){width="\textwidth"}
[0.4]{} ![Estimating a population history with piecewise exponential change. []{data-label="fig:piex"}](L14e20piex.pdf "fig:"){width="\textwidth"}
[0.45]{} ![Constant and piecewise constant population histories with uncertainty intervals. []{data-label="fig:reconstruct"}](LBCon1k "fig:"){width="\textwidth"}
[0.45]{} ![Constant and piecewise constant population histories with uncertainty intervals. []{data-label="fig:reconstruct"}](LBCon10k "fig:"){width="\textwidth"}
\
[0.45]{} ![Constant and piecewise constant population histories with uncertainty intervals. []{data-label="fig:reconstruct"}](LBPiece1k "fig:"){width="\textwidth"}
[0.45]{} ![Constant and piecewise constant population histories with uncertainty intervals. []{data-label="fig:reconstruct"}](LBPiece10k "fig:"){width="\textwidth"}
[^1]: University of Pennsylvania; `[email protected]`.
[^2]: University of California, Berkeley; `[email protected]`; supported by NSF grants DMS 1106999 and CCF 1320105 and by DOD ONR grant N000141110140.
[^3]: University of California, Berkeley; `[email protected]`; supported by NSF grant DMS 1106999 and by DOD ONR grant N000141110140.
[^4]: University of Melbourne; `[email protected]`.
[^5]: Our title is inspired by the famous paper of Mark Kac [@kac1966can]; analogously, we study the theoretical limits to inferring a population size history.
[^6]: Throughout the paper the unit of time is generations.
[^7]: For a statistical hypothesis $H_1: \theta=\theta_1$ on a real parameter $\theta$ and a given set of data, we define the $95\%$ *uncertainty interval* to be the set of values $\theta_2$ such that for any classification procedure, the chance of correctly determining whether the samples came from $H_1$ or $H_2: \theta=\theta_2$ is no greater than $0.95$.
[^8]: Here and in the following we use the notation $a\vee b = \max \left\{ a, b \right\}$ and $a \wedge b = \min \left\{ a, b \right\}$.
[^9]: We use the standard asymptotic notation $\Omega$, which means “at least on the order of”. Formally, if $a_n$ and $b_n$ are two sequences such that there exists a positive constant $c$ and an integer $n_0$ such that for every $n {\geqslant}n_0$, $a_n {\geqslant}c \times b_n$, then $a_n = \Omega \left( b_n \right)$ as $n \to \infty$. Similarly, if $f$ and $g$ are two functions such that there exist positive constants $c$ and $x_0$ such that for every $x \in \left( 0, x_0 \right)$, $f \left( x \right) {\geqslant}c \times g \left( x \right)$, then $f\left( x \right) = \Omega \left( g \left( x \right) \right)$ as $x \searrow 0$. Equivalently, $f\left( x \right) = \Omega \left( g \left( x \right) \right)$ if and only if $g\left( x \right) = O \left( f \left( x \right) \right)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give a short update of our research program on nonequilibrium statistical field theory applied to quantum processes in the early universe and black holes, as well as the development of stochastic gravity theory as an extension of semiclassical gravity and an intermediary in the ’bottom-up’ approach to quantum gravity.'
author:
- |
B. L. Hu$^a$[^1], Albert Roura$^a$, Sukanya Sinha$^b$, E. Verdaguer$^c$\
$^a$[Department of Physics, University of Maryland, College Park, Maryland 20742-4111, U.S.A.]{}\
$^b$[Indian Statistical Institute, Bangalore Centre, 8th Mile, Mysore Road, Bangalore-560059, INDIA]{}\
$^c$[Departament de Fisica Fonamental and C.E.R. in Astrophysics, Particles and Cosmology, Universitat de Barcelona, Av. Diagonal 647, 08028 Barcelona, Spain]{}
date: '(UMD Physics Preprint number: umdpp 03-045, April 12, 2003)'
title: Quantum Noise and Fluctuations in Gravitation and Cosmology
---
c i ł ø u Ł Ø ¶
plus 1000pt minus 1000pt \#1 \#1[= to]{} \#1[= to]{}
\#1[Nucl. Phys. \#1]{} \#1[Phys. Lett. \#1]{} \#1[Phys. Rev. Lett. \#1]{} \#1[Phys. Rev. [**D**]{} \#1]{} \#1[Phys. Rev. [**A**]{} \#1]{} \#1[Prog. Theor. Phys. \#1]{} Å\#1[Astron. and Astrophys. \#1]{} \#1[$\sp{#1)}$]{} \#1
=8.5in =6.5in =-0.5in =0.in =0.in
addtoreset[equation]{}[section]{}
*Invited Talk given by BLH at the First International Symposium on Fluctuations and Noise*
*Sponsored by SPIE, 1-4 June 2003, Santa Fe, New Mexico. Paper number 5111-46*
Prologue {#prologue .unnumbered}
========
(BLH) “It is no secret that the message of this conference is ‘NOISE IS GOOD’. In this talk I want to show that [*not only is noise good, it is absolutely essential*]{}... ” . These words were not uttered in this conference, as it would have been acausal, but ten years ago, at a workshop devoted to the nascent yet fascinating subject of Fluctuations and Order [@HMLA]. After a decade, the tenor and objectives of this statement are even closer to reality, as witnessed by the strong focus and great variety of this symposium.
I substantiated this claim then by an enumeration of the many processes in gravitation and cosmology where quantum noise and fluctuations play an active role:
1\. Particle creation as parametric amplification of vacuum fluctuations\
2. Thermal radiance from accelerated observers and black holes as fluctuation-dissipation phenomena\
3. Entropy generation from quantum stochastic and kinetic processes\
4. Phase transitions in the early universe as noise-induced processes\
5. Galaxy formation from primordial quantum fluctuations\
6. Anisotropy dissipation from particle creation as backreaction processes\
7. Dissipation in quantum cosmology and the issue of the initial state\
8. Decoherence, backreaction and the semiclassical limit of quantum gravity\
9. Stochastic spacetime and continuum limit, gravity as an effective theory\
10. Topology change in spacetime and loss of quantum coherence problems\
11. Gravitational entropy, singularity and time asymmetry\
12. ‘Birth’ of the universe as a spacetime fluctuation and tunneling phenomenon\
The above list was prepared for cosmological issues. Further developments in the last ten years focusing on the effects of noise and fluctuations using the statistical field theory approach we have developed (for a review, see [@Banff; @CHM]) include, related to Topic 1 : Preheating in post-inflationary cosmology [@RH1; @RH2]; Topic 2: Thermal and near-thermal radiance in detectors, moving mirrors, black holes and cosmology [@RHA; @RHK; @KHMR; @KMH]; Topic 3: Correlation entropy [@CH00; @CHcorent]; Topic 4: Defect formation [@SCHR], tunneling induced by quantum and thermal noise fluctuations [@CRVopensys; @CRVtunnel]. Topic 5 is presently under pursuit [@RouVer03a] and partially summarized in the last section of this paper. Topic 9 includes preliminary work on wave propagation in stochastic spacetimes [@HuShi] and mesoscopic fluctuations [@Tomeso].
In black hole physics, focusing again on fluctuations, one can mention topics on black hole fluctuations and backreaction [@HRS; @SRH] (see references therein for other related work) and the energetics and dynamics of black hole phase transition [@GregBH; @GPY; @WhiYor]. The wish list would include: applications of stochastic gravity to the statistical mechanical definition of entropy and statistical field theory of black hole nonequilibrium thermodynamics.
Using statistical quantum field theory , Topics 1, 2, 3, 6 are quite well understood, 7, 8 only partially, and Topics 4, 5, 9 are currently under pursuit. Certain concepts in Topics 10-12 may not even be well-defined (e.g., what does the ‘birth’ of the universe mean?), but the newly established theory of stochastic gravity (for a review, see [@stogra; @HVErice; @CQGrev; @LivRev], an ongoing development, see [@RouVer03b]) may offer alternative ways to address these issues, as well as provide an intermediary towards quantum gravity (our definition is different from the ordinary, see [@KinQG]).
In this talk I will focus on the theory of noise in quantum fields and how quantum fluctuations could have played an active and even decisive role in many fundamental processes in cosmology and gravitation, especially near the Planck time ($10^{-43}$ sec from the Big Bang). I will describe how stochastic gravity theory can be understood easily from an open system conceptual framework, and give two examples of its applications: fluctuations and backreaction in black holes, and structure formation in the early universe.
The Planck time is the time when many familiar features of spacetime depicted by Einstein’s theory of general relativity give way to an as-yet-unknown quantum theory of gravity depicting the microstructure of spacetime. (Some theoreticians believe the superstring theory is the answer.) Just below the Planck energy we believe the universe can be adequately described by a semiclassical theory of gravity [@DeW75; @BirDav; @scg], where quantized matter fields coexist with a classical spacetime. Many qualitative changes are believed to have taken place at this energy scale, amongst them the formation of spacetime depictable as a manifold, the emergence of time, the creation of particle pairs from the vacuum, the growth of fluctuations as seeds for galaxies, and possible phase transitions and the ensuing entropy generation processes. It is also the cross-over point of quantum to classical and micro to macroscopic transitions.[^2]
In Section I, we first explain the origin and nature of noise in quantum systems interacting with an environment, using the influence functional method. (For non-Ohmic bath at low temperatures, colored noise and nonlocal dissipation would appear; and for nonlinear coupling multiplicative noise is generally expected. A generalized fluctuation-dissipation relation for these systems can be proven, and the stochastic (master, Langevin and Fokker-Planck) equations derived, depicting the dissipative dynamics of the open system under the influence of noise. We then discuss stochastic gravity theory in Sec. 2 and the two examples in Sections 3 and 4.
Quantum Fluctuations in Open Systems
====================================
We begin by describing how the concepts of quantum open systems can be of use in the treatment of statistical mechanical problems involving quantum fields. We start with two subsystems A and B. When the precise information of subsystem B is not required, but only its averaged effect on subsystem A is of interest, one can [*coarse-grain*]{} B and include its averaged effect on A, which involves finding the [*backreaction*]{} on A. In so doing A is rendered an open system, with B acting as its environment. For the analysis of open systems with backreaction from the environments the influence functional (IF) formalism of Feynman and Vernon [[@if; @qbm]]{} proves useful. Let us first use a simple example from quantum mechanics to illustrate the idea and the method. We then show how we can use this framework to address issues in gravitation and cosmology.
Stochastic Effective Action in Quantum Open Systems {#sec:open}
---------------------------------------------------
We begin with a brief schematic summary of the IF formalism as applied to a simple system. Consider a system $S$, described by the degrees of freedom $x$, interacting with an environment $E$, described by the degrees of freedom $q$.[^3] The full closed quantum system $ S + E$ is described by a density matrix $\rho (x,q;x',q',t)$. If we are interested only in the state of the system as influenced by the overall effect, but not the precise state of the environment, i.e, the dynamics of the open system, then the reduced density matrix $\rho_r(x,x',t)=\int~dq~\rho (x,q;x',q,t)$ would provide the relevant information. (The subscript $r$ stands for reduced.) Assuming that the action of the coupled system decomposes as $S=S_s[x]+S_e[q]+S_{int}[x,q]$, and that the initial density matrix factorizes (i.e., takes the tensor product form), $\rho
(x,q;x',q',t_i)=\rho_s(x,x',t_i)\rho_e(q,q',t_i)$, the reduced density matrix is given by \_r(x,x’,t) =\_[-]{}\^[+]{}dx\_i\_[-]{}\^[+]{}dx’\_i \_[x\_i]{}\^[x\_f]{} Dx \_[x’\_i]{}\^[x’\_f]{} Dx ’ e\^[i(S\_s\[x\]-S\_s\[x’\]+S\_[IF]{}\[x,x’,t\])]{} \_r(x\_i,x’\_i,t\_i ) \[pathint\] where $S_{IF}$ is the influence action related to the influence functional $\cal F$ defined by \[x, x’\] e\^[iS\_[IF]{}\[x,x’,t\]]{} dq\_f dq\_i dq\_i’ \_[q\_i]{}\^[q\_f]{}Dq \_[q’\_i]{}\^[q\_f]{}Dq’ e\^[i(S\_e\[q\]+S\_[int]{}\[x ,q\]-S\_e\[q’\]-S\_[int]{}\[x ’,q’\])]{} \_e(q\_i,q’\_i,t\_i). \[SIF\] $S_{IF}$ in general is complex. Retaining only quadratic terms (an approximation which covers many of the interesting applications that we will consider later), we may write S\_[IF]{}(x,x’)= dt dt’ { [12]{}(x-x’)(t)D(t,t’)(x+x’)(t’) +[i2]{}(x-x’)(t)N(t,t’)(x-x’)(t’)} \[SIFquad\] where $D$ and $N$ stand for the real dissipation and noise kernels respectively. Note that in this quadratic order approximation, the influence action $ S_{IF}(x,x')$ is related to the closed-time-path(CTP) or in-in effective action (for details on the CTP effective action see [@ctp]) $\Gamma_{CTP}[x,x']$ through \_[CTP]{}\[x,x’\] = S\[x\] - S\[x’\] + S\_[IF]{}\[x,x’\]. \[CTPIFconn\] The equation of motion obtained from the CTP effective action for the expectation values is clearly seen to be real and causal [@ctp]. It reads .[x(t)]{}\_[CTP]{}\[x, x’\]|\_[x’ = x = [|x]{}]{} = 0 \[CTPem\] From the influence functional a Langevin equation for the system dynamics may be derived by a formal procedure, first introduced by Feynman and Vernon [@if], which consists of introducing a Gaussian stochastic source $\xi(t)$ with $\langle\xi(t)\rangle_{\xi}
= 0$ and $\langle\xi(t)\xi(t')\langle = N(t,t')$ and defining an improved or stochastic effective action as S\_[eff]{}\[x,x’;\] = S\_s\[x\] - S\_s\[x’\] + S\_[IF]{}\[x,x’\] + (x-x’) \[Seffxi\] such that $\left < e^{iS_{eff}[x,x';\xi]}\right >_{\xi} =
e^{i\Gamma_{CTP}[x,x']}$. This leads to equations of motion with a stochastic force: .[S\_[eff]{}\[x,x’;\]x]{}|\_[x =x’]{} = 0 .[\_[CTP]{}\[x,x’;\]x]{}|\_[x =x’]{} = 0. \[stochem\] The equation of motion obtained from (\[stochem\]) using (\[Seffxi\]) is + dt’ (t,t’) [dx(t’)dt’]{}= \[langevin\] where $D(t,t')=-\partial_{t'}\gamma (t,t')$. Being now in the form of a Langevin equation, the physical meaning of the $\gamma$ and $N$ kernels in Eq. (\[langevin\]) becomes clearer. Both the terms involving $\gamma$ and $\xi$ represent the backreaction of the environment on the system. However, $\gamma$ (or more properly the odd part of $\gamma$) is associated with dissipation and $\xi$ is a stochastic noise term associated with random fluctuations of the system, exactly as the terms are interpreted in the context of Brownian motion. Averaging (\[langevin\]) over the noise using the appropriate probability distribution will give the semiclassical equation of motion for the mean value of $x$. Since the noise and dissipation arise by considering a subsystem within a closed system (as is done here, as opposed to being put in by hand), they are in general related by a set of generalized fluctuation-dissipation relations (FDR), which can be represented by a linear, non-local relation of the following form, provided that the Hamiltonian for the environment and the system-environment interaction are time independent and the initial state of the environment is stationary with respect to the Hamiltonian of the environment: N(t-t’) = d(s -s’)K(t-t’,s-s’)(s -s’) \[FDR\] To keep the discussion simple, we have written the noise and dissipation kernels in terms of single scalar functions. However, the method is general enough to encompass multiple noise and dissipation kernels and cases where the kernels are tensorial, as in the stochastic gravity theory discussed later.
Stochastic Gravity and Metric Fluctuations {#sec:stogra}
==========================================
Stochastic semiclassical gravity [@ELE] of the 90’s is a theory naturally evolved from semiclassical gravity [@scg] of the 80’s and quantum field theory in curved spacetimes [@DeW75; @BirDav] of the 70’s. Whereas semiclassical gravity is based on the semiclassical Einstein equation with sources given by the expectation value of the stress-energy tensor of quantum fields, stochastic semiclassical gravity is based on the Einstein-Langevin equation, which has in addition stochastic sources with correlation functions characterized by the noise kernel. the noise kernel. The noise kernel is the vacuum expectation value of the (operator-valued) stress-energy bi-tensor which describes the fluctuations of quantum matter fields in curved spacetimes.
From Semiclassical to Stochastic Gravity {#sec1}
-----------------------------------------
The first stage in the road to stochastic gravity begins with [*quantum field theory in curved spacetime*]{}, which describes the behavior of quantum matter fields propagating in a specified (not dynamically determined by the quantum matter field as a source) background gravitational field. For a scalar field $\phi$ it obeys the wave equation $ (\Box + m^2) \phi(x) = 0 $ where $\Box$ is the Laplace-Beltrami operator, which contains the imprint of the curvature of the background spacetime. In this theory the gravitational field is given by the classical spacetime metric determined from classical sources by the classical Einstein equations, and the quantum fields propagate as test fields in such a spacetime. For time dependent spacetime geometry it may not be possible to define a physically meaningful vacuum state for the quantum field at all times. Assuming that one defines a vacuum state at some initial time, the vacuum state at a latter time will differ from that defined initially because particles are created in the intervening time. An important process described by quantum field theory in curved spacetime is indeed particle creation from the vacuum (and effects of vacuum fluctuations and polarizations) in the early universe [@Hawking] and Hawking radiation in black holes [@Hawking].
The second stage in the description of the interaction of gravity with quantum fields is [*back-reaction*]{}, i.e., the effect of quantum fields on the spacetime geometry. The dynamic classical spacetime metric creates particles of the quantum field and these in turn provide a backreaction on the spacetime metric which alters its dynamics in response. One assumes a general class of spacetime where the quantum fields live in and act on, and seek a solution which satisfies simultaneously the Einstein equation for the spacetime and the field equations for the quantum fields. The Einstein equation which has the expectation value of the stress-energy operator of the quantum matter field as the source is known as the [*semiclassical Einstein equation*]{}: \[semi\] G\_ (g\_) = \_ \_q where $\hat{T}_{\m\n} $ is the stress-energy tensor operator of, say, a free scalar field $\phi$, $G_{\mu\nu}$ is the Einstein tensor, $\kappa = 8\pi G_N$ and $G_N$ is Newton’s constant. Here $\langle\,\rangle_q$ denotes the expectation value taken with respect to some quantum state compatible with the symmetries of the background spacetime, a classical object. The theory obtained from a self-consistent solution of the geometry of the spacetime and the quantum field is known as [*semiclassical gravity*]{}. Incorporating the backreaction of the quantum matter field on the spacetime is thus the central task in semiclassical gravity.
Studies of the semiclassical Einstein equation for the backreaction problems have been carried out in the last two decades by many authors for cosmological and black hole spacetimes. A well-known example of semiclassical gravity is the damping of anisotropy in Bianchi universes (which is the basis of chaotic cosmology in the 70’s) by the backreaction of vacuum particle creation, and inflationary cosmology [@infcos; @KolTur90; @Linde90] of the 80’s driven by a constant vacuum energy density source such as the expectation value of a Higgs field. In analogy with the open system dynamics described in Section \[sec:open\], Eq.(\[semi\]) is equivalent to Eq.(\[CTPem\]) where the degrees of freedom $x$ of the system are identified with the metric $g_{\alpha\beta}$ and those of the environment $q$ are identified with the scalar field $\phi(x)$. However, from the discussion in the last section it is also clear that Eq. (\[CTPem\]) , and hence also the semiclassical Einstein Eq. (\[semi\]) results on averaging the full Langevin-type Eq. (\[stochem\]) over noise. Thus the semiclassical Einstein equation incorporates the dissipation but misses out the fluctuation aspect of the backreaction. The recognition of this crucial point [@HuPhysica] ushered in a new theory known as [*stochastic semiclassical gravity*]{}, (or in short, stochastic gravity, as there is no confusion in this context as to where the stochasticity originates). Aided by the concept of open systems and the techniques of the influence functional and the Closed Time Path (CTP) effective action, stochastic gravity is the new framework for the consideration of backreaction because it encompasses fluctuations and dissipation (from particle creation and other quantum field processes) on the same footing. Spacetime dynamics is now governed by a stochastic generalization of the semiclassical Einstein equation known as the Einstein-Langevin equation, the analog of Eq. (\[langevin\]) in the context of semiclassical gravity (SCG). Schematically the Einstein-Langevin equation takes on the form $$\tilde G_{\mu\nu}(x)
= \k \left( T_{\mu\nu}^{\rm c} + T_{\mu\nu}^{\rm qs}
\right), \quad
T_{\mu\nu}^{\rm qs}
\equiv \langle \hat{T}_{\mu\nu} \rangle_{\rm q} +
T_{\mu\nu}^{\rm s} \label{eq:effective stress tensor}$$ Here, $T_{\mu\nu}^c$ is due to classical matter or fields, $\langle \hat{T}_{\mu\nu} \rangle_q$ is the expectation value of the stress tensor of the quantum field, and $T_{\mu\nu}^{\rm qs}$ is a new stochastic term which is related to the quantum fluctuations of $T_{\mu\nu}$ for the state of the field under consideration. Taking the average of (\[eq:effective stress tensor\]) with respect to the noise distribution will lead to the conventional semiclassical Einstein equation. It is in this context that SCG is regarded as a mean field theory. The fundamentals of this new theory were developed via two approaches: the axiomatic and the functional. The axiomatic approach is useful to see the structure of the theory from the framework of semiclassical gravity, showing the link from the mean value of the stress-energy tensor to its correlation functions. The functional approach uses the Feynman-Vernon influence functional and the Schwinger-Keldysh closed-time-path effective action methods which are convenient for computations. It also brings out the open systems concepts and the statistical and stochastic contents of the theory such as dissipation, fluctuations, noise and decoherence. There was also theoretical work on the properties of the stress energy bi-tensor and its vacuum expectation value, the noise kernel. See, e.g., [@MV0; @MV1; @MV2; @PH97; @HP0; @PH1; @PH2]. For a broader exploration of ideas and issues based on this theory read the reviews [@stogra; @LivRev]; For a pedagogical introduction with applications, see [@CQGrev; @HVErice].
Thus with the aid of the open system viewpoint it is easy to see that stochastic gravity is a natural extension of the well-established semiclassical gravity theory and a useful framework for the considerations of fluctuations in quantum matter fields and dissipative dynamics of classical spacetimes, including metric fluctuations. Stochastic gravity can address many important issues related to nonequilibrium quantum field processes in curved spacetimes and find applications to many problems in gravitation and cosmology.
Stochastic Gravity in relation to Quantum Gravity
-------------------------------------------------
Before embarking on the discussion of some applications of stochastic gravity, let us illustrate the theory with a simple toy model which minimizes the technical complications. The model will be useful to clarify the role of the noise kernel and illustrate the relationship between the semiclassical, stochastic and quantum descriptions. Let us assume that the gravitational equations are described by a linear field $h(x)$ whose source is a massless scalar field $\phi(x)$ which satisfies the Klein-Gordon equation in flat spacetime $\Box
\phi(x)=0$. The field stress-energy tensor is quadratic in the field, and independent of $h(x)$. The classical gravitational field equations will be given by[^4]$$\Box h(x)=\kappa T(x), \label{2.12a}$$ where $T(x)$ is the (scalar) trace of the stress-energy tensor, $T(x)=\partial_a\phi(x)\partial^a\phi(x)$ and $\kappa \equiv 16
\pi G$, where G is Newton’s constant. Note that this is not a self-consistent theory since $\phi(x)$ does not react to the gravitational field $h(x)$. We should also emphasize that this model is not the standard linearized theory of gravity in which $T$ is also linear in $h(x)$. It captures, however, some of the key features of linearized gravity.
In the Heisenberg representation the quantum field $\hat h(x)$ satisfies $$\Box \hat h(x)=\kappa\hat T(x). \label{2.12}$$ Since $\hat T(x)$ is quadratic in the field operator $\hat\phi
(x)$ some regularization procedure has to be assumed in order for (\[2.12\]) to make sense. Since we work in flat spacetime we may simply use a normal ordering prescription to regularize the operator $\hat T(x)$. The solutions of this equation, i.e. the field operator at the point $x$, $\hat h(x)$, may be written in terms of the retarded propagator $G(x,y)$ as, $$\hat h(x)=\hat h^{(0)}(x) + \kappa \int dx' G(x,x^\prime)\hat
T(x^\prime), \label{2.13}$$ where $\hat h^{(0)}(x)$ is the free field which carries information on the initial conditions and the state of the field. From this solution we may compute, for instance, the symmetric two point quantum correlation function (the anticommutator) $${1\over2}\langle \{\hat h(x),\hat h(y)\}\rangle = {1\over2}
\langle \{\hat h^{(0)}(x),\hat h^{(0)}(y)\}\rangle + {\kappa^2 \over 2}
\int\int dx'dy'G(x,x^\prime)G(y,y^\prime) \langle\{\hat
T(x^\prime),\hat T(y^\prime)\}\rangle, \label{2.14}$$ where the expectation value is taken with respect to the quantum state in which both fields $\phi(x)$ and $h(x)$ are quantized. (We assume for the free field, $\langle \hat h^{(0)}\rangle=0$.)
We can now consider the semiclassical theory for this problem. If we assume that $h(x)$ is classical and the matter field is quantum the semiclassical theory may just be described by substituting into the classical equation (\[2.12a\]) the stress-energy trace by the expectation value of the stress-energy trace operator $\langle \hat T(x)\rangle$, in some quantum state of the field $\hat \phi(x)$. Since in our model $\hat T(x)$ is independent of $h(x)$ we may simply renormalize its expectation value using normal ordering, then for the vacuum state of the field $\hat\phi(x)$, we would simply have $\langle\hat
T(x)\rangle_0=0 $. The semiclassical theory thus reduces to $$\Box h(x)=\kappa \langle \hat T(x)\rangle. \label{2.15a}$$ The two point function $h(x)h(y)$ that one may derive from this equation depends on the two point function $\langle \hat
T(x)\rangle \langle \hat T(y)\rangle $ and clearly cannot reproduce the quantum result (\[2.14\]) which depends on the expectation value of the two point operator $\langle\{\hat T(x),\hat
T(y)\}\rangle$. That is, the semiclassical theory entirely misses the fluctuations of the stress-energy operator $\hat T(x)$.
Let us now see how we can extend the semiclassical theory in order to account for such fluctuations. The first step is to characterize these fluctuations. For this, we introduce the noise kernel as the physical observable that measures the fluctuations of the stress-energy operator $\hat T$. Define $$N(x,y)= \frac{1}{2}\langle\{\hat t(x),\hat t(y)\}\rangle
\label{2.14a}$$ where $\hat t(x)=\hat T(x)-\langle\hat T(x)\rangle$. The bi-scalar $N(x,y)$ is real and positive-semidefinite, a consequence of $\hat t$ being self-adjoint. A simple proof can be given as follows. Let $|\psi\rangle$ be a given quantum state and let $\hat Q$ be a self-adjoint operator, $\hat Q^\dagger=\hat Q$, then one can write $\langle\psi|\hat Q\hat Q|\psi\rangle=
\langle\psi|\hat Q^\dagger Q|\psi\rangle= | \hat
Q|\psi\rangle|^2\geq 0$. Now let $\hat t(x)$ be a self-adjoint operator, then if we define $\hat Q=\int dx f(x) \hat t(x)$ for an arbitrary well behaved function $f(x)$, the previous inequality can be written as $\int dx dy f(x)\langle\psi|\hat t(x)\hat
t(y)|\psi\rangle f(y)\geq 0$, which is the condition for the noise kernel to be positive semi-definite. Note that when considering the inverse kernel $N^{-1}(x,y)$, it is implicitly assumed that one is working in the subspace obtained from the eigenvectors which have strictly positive eigenvalues when the noise kernel is diagonalized.
By the positive semi-definite property of the noise kernel $N(x,y)$ it is possible to introduce a Gaussian stochastic field as follows: $$\langle\xi(x)\rangle_s=0,\quad
\langle\xi(x)\xi(y)\rangle_s=N(x,y). \label{2.14b}$$ where the subscript $s$ means a statistical average. These equations entirely define the stochastic process $\xi(x)$ since we have assumed that it is Gaussian. Of course, higher correlations could also be introduced but we just try to capture the fluctuations to lowest order.
The extension of the semiclassical equation may be simply performed by adding to the right-hand side of the semiclassical equation (\[2.15a\]) this stochastic source $\xi(x)$ which accounts for the fluctuations of $\hat T$ as follows, $$\Box h(x)=\kappa\left( \langle \hat T(x)\rangle+\xi(x)\right).
\label{2.15}$$ This equation is in the form of a Langevin equation: the field $h(x)$ is classical but stochastic and the observables we may obtain from it are correlation functions for $h(x)$. In fact, the solution of this equation may be written in terms of the retarded propagator as, $$h(x)=h^{(0)}(x)+\kappa \int dx^\prime
G(x,x^\prime)\left(\langle\hat T(x^\prime)\rangle
+\xi(x^\prime)\right) , \label{2.16}$$ from where the two point correlation function for the classical field $h(x)$, after using the definition of $\xi(x)$ and that $\langle
h^{(0)}(x)\rangle_s=0$, is given by $$\langle h(x) h(y)\rangle_s = \langle h^{(0)}(x) h^{(0)}(y)\rangle_s
+{\kappa^2 \over 2} \int\int dx^\prime dy^\prime
G(x,x^\prime)G(y,y^\prime) \langle\{\hat T(x^\prime),\hat
T(y^\prime)\}\rangle. \label{2.17}$$ Note that in writing $\left<\dots\right>_s$ here we are assuming a double stochastic average, one is related to the stochastic field $\xi(x)$ and the other is related to the free field $h^{(0)}(x)$ which is assumed also to be stochastic with a distribution function to be specified.
Comparing (\[2.14\]) with (\[2.17\]) we see that the respective second term on the right-hand side are identical provided the expectation values are computed in the same quantum state for the field $\hat \phi(x)$ (recall that we have assumed $T(x)$ does not depend on $h(x)$). The fact that the field $h(x)$ is also quantized in (\[2.14\]) does not change the previous statement. The nature of the first term on the right-hand side of Eqs. (\[2.14\]) and (\[2.17\]) is different: in the first case it is the two point quantum expectation value of the free quantum field $\hat h^{(0)}$ whereas in the second case it is the stochastic average of the two point classical homogeneous field $h^{(0)}$, which depends on the initial conditions. Now we can still make these terms equal to each other if we assume for the homogeneous field $h^{(0)}$ a Gaussian distribution of initial conditions such that $$\langle h^{(0)}(x) h^{(0)}(y)\rangle_s= \frac{1}{2}\langle\{\hat
h^{(0)}(x),\hat h^{(0)}(y)\}\rangle. \label{2.17a}$$ This Gaussian stochastic field $h^{(0)}(x)$ can always be defined due to the positivity of the anti-commutator. Thus, under this assumption on the initial conditions for the field $h(x)$ the two point correlation function of (\[2.17\]) equals the quantum expectation value of (\[2.14\]) exactly. An interesting feature of the stochastic description is that the quantum anticommutator of (\[2.14\]) can be written as the right-hand side of equation (\[2.17\]), where the first term contains all the information on the initial conditions for the stochastic field $h(x)$ and the second term codifies all the information on the quantum correlations of the source. This separation is also seen in the description of some quantum Brownian motion models which are typically used as paradigms of open quantum systems [@CRVopensys; @CRVtunnel].
It is interesting to note that in the standard linearized theory of gravity $T(x)$ depends also on $h(x)$, both explicitly and also implicitly through the coupling of $\phi(x)$ with $h(x)$. The equations are not so simple but it is still true that the corresponding Langevin equation leads to the correct symmetrized two point quantum correlations for the metric perturbations [@MV2; @RouVer03b]. Thus in a linear theory as in the model just described one may just use the statistical description given by (\[2.15\]) to compute the symmetric quantum two point function of equation (\[2.13\]). This does not mean that we can recover all quantum correlation functions with the stochastic description, see Ref. [@CRVopensys] for a general discussion about this point. Note that, for instance, the commutator of the classical stochastic field $h(x)$ is obviously zero, but the commutator of the quantum field $\hat h(x)$ is not zero for timelike separated points; this is the prize we pay for the introduction of the classical field $\xi(x)$ to describe the quantum fluctuations. Furthermore, the statistical description is not able to account for the graviton-graviton effects which go beyond the linear approximation in $\hat h(x)$.
Black Hole Fluctuations and Backreaction
========================================
As the first example we consider the backreaction of Hawking radiation on black holes [@bhbkr] with fluctuations, i.e., how a quantum field and its fluctuations influence the behavior of the background spacetime. We will only sketch the strategy of this program based on stochastic gravity, as detailed calculations are still in progress. The formalism described in Section \[sec:open\] will be useful. Here we study the simpler class of problems of a quasi-static black hole in quasi-equilibrium (a box is required) with its Hawking radiation described by a scalar field. The goal is to obtain a stochastic influence action analogous to (\[Seffxi\]) for this model of a black hole coupled to a scalar field. From it one can derive an Einstein-Langevin equation analogous to (\[langevin\]).
We consider the simplest model of this class described by a perturbed Schwarzschild metric, used by York [@York] to analyze black hole backreaction. We focus on the new aspects of noise and fluctuations, their origin and attributes.
In this model the black hole spacetime is described by a spherically symmetric static nonvacuous metric with line element of the following general form written in advanced time Eddington-Finkelstein coordinates ds\^2 = g\_dx\^dx\^ = -e\^[2]{}(1 - [2mr]{})dv\^2 + 2 e\^[2]{}dvdr + r\^2 d\^2 \[ssmetric\] where $\psi = \psi(r)$ and $m = m(r)$ , $ v
= t + r + 2Mln\left({r\over 2M} -1 \right)$ and $d{\Omega}^2$ is the line element on the two sphere. Hawking radiation is described by a massless, conformally coupled quantum scalar field $\phi$ with the classical action S\_m\[, g\_\] = -d\^n x \[g\^\_\_+ (n) R\^2\] \[phiact\] where $\xi(n) = {(n-2)\over 4(n-1)}$ ($n$ is the dimension of spacetime) and $R$ is the curvature scalar of the spacetime it lives in.
Let us consider linear perturbations $h_{\mu\nu}= g_{\mu\nu} -
g^{(0)}_{\mu\nu}$ off a background Schwarzschild metric $g^{(0)}_{\mu\nu}$ with line element (ds\^2)\^0 = ( 1 - [2Mr]{})dv\^2 + 2dvdr + r\^2d\^2 \[schwarz\] We look for this class of perturbed metrics in the form given by (\[ssmetric\]), (thus restricting our consideration only to spherically symmetric perturbations): e\^1+ (r), m M\[ 1 + (r)\] \[mu\] where ${\epsilon\over \lambda M^2} = {1\over 3}a T_H^4 ; a
={{\pi}^2\over 30} ; \lambda = 90(8^4)\pi^2$. $T_H$ is the Hawking temperature. (This particular parametrization of the perturbation is chosen following York’s [@York] notation.) Thus the only non-zero components of $h_{\mu\nu}$ are h\_[vv]{} = -((1 - [2Mr]{})2(r) + [2M(r)r]{}), h\_[vr]{} = (r) \[hvr\] So this represents a metric with small static and radial perturbations about a Schwarzschild black hole. The initial quantum state of the scalar field is taken to be the Hartle-Hawking vacuum, which is essentially a thermal state at the Hawking temperature as far as static observers are concerned [^5] and it represents a black hole in (unstable) thermal equilibrium with its own Hawking radiation.
The metric perturbation expansion induces a decomposition of the Einstein tensor $G_{\mu\nu} \simeq G^{(0)}_{\mu\nu} + \delta
G_{\mu\nu}$ where $G^{(0)}_{\mu\nu}$ is the Einstein tensor for the background spacetime. The zeroth order solution gives a background metric in empty space, i.e, the Schwarzschild metric. $\delta
G_{\mu\nu}$ is the linear correction to the Einstein tensor in the perturbed metric. The semiclassical Einstein equation in this approximation therefore reduces to $$\delta G_{\mu\nu}(g^{(0)}, h) =
\kappa \langle T_{\mu\nu}\rangle$$ York solved this equation to first order by using the expectation value of the energy momentum tensor for a conformally coupled scalar field in the Hartle-Hawking vacuum in the unperturbed (Schwarzschild) spacetime on the right hand side and $\delta G_{\mu\nu}$ on the left hand side is calculated using (\[hvr\]). This yields the corrections to the background metric induced by the backreaction encoded in the functions $\mu(r)$ and $\rho(r)$, which amounts to the noise-averaged backreaction effects. We are interested in the fluctuations and its effects.
We now derive the CTP effective action for this model, following the treatment of Ref. [@CamHu]. Using the metric (\[schwarz\]) (and neglecting the surface terms that appear in an integration by parts) we have the action for the scalar field written perturbatively as $$S_m[\phi,h_{\mu\nu}]
\ = \ {1\over 2}\int d^nx{\sqrt{-g^{(0)}}}\ \phi
\left[ \Box + V^{(1)} + V^{(2)} + \cdots
\right] \phi,
\label{phipert}$$ where the first and second order perturbative operators $V^{(1)}$ and $V^{(2)}$ are given by $$\begin{aligned}
V^{(1)} & \ \equiv \ & - {1\over \sqrt{-g^{(0)}}} \left\{
[\partial_\mu\left(\sqrt{-g^{(0)}}\bar h^{\mu\nu}(x)\right)]
\partial_\nu
+\bar h^{\mu\nu}(x)\partial_\mu
\partial_\nu
+\xi(n) R^{(1)}(x)
\right\},
\nonumber \\
V^{(2)}
& \ \equiv \ & - {1\over \sqrt{-g^{(0)}}}
\left\{ [\partial_\mu \left(\sqrt{-g^{(0)}} \hat h^{\mu\nu}(x)\right)] \partial_\nu
+\hat h^{\mu\nu}(x)\partial_\mu \partial_\nu
- \xi(n) \left( R^{(2)}(x) +{1\over 2}h(x)R^{(1)}(x) \right) \right\}.\end{aligned}$$ In the above expressions, $R^{(k)}$ is the $k$-order term in the pertubation $h_{\mu\nu}(x)$ of the scalar curvature $R$ and $\bar
h_{\mu\nu}$ and $\hat h_{\mu\nu}$ denote a linear and a quadratic combination of the perturbation, respectively, $$\bar h_{\mu\nu}
\equiv h_{\mu\nu} - {1\over 2} h g^{(0)}_{\mu\nu};
\;\;
\hat h_{\mu\nu}
\equiv h^{\,\, \alpha}_\mu h_{\alpha\nu}
-{1\over 2} h h_{\mu\nu}
+{1\over 8} h^2 g^{(0)}_{\mu\nu}
-{1\over 4} h_{\alpha\beta}h^{\alpha\beta}
g^{(0)}_{\mu\nu}.
\label{eq:def bar h}$$ From quantum field theory in curved spacetime considerations we take the following action for the gravitational field (see [@MV1; @CamHu] for more details) $$S^{(div)}_g[g_{\mu\nu}]
\ = \ {1\over\ell^{n-2}_P}\int d^nx\ \sqrt{-g}R(x)
+{\alpha\bar\mu^{n-4}\over4(n-4)}
\int d^nx\ \sqrt{-g}
\left[ 3R_{\mu\nu\alpha\beta}(x)
R^{\mu\nu\alpha\beta}(x)
\left( 1-360(\xi(n) - {1\over6})^2
\right)R(x)R(x) \right].$$ The first term is the classical Einstein-Hilbert action and the second term is the counterterm in four dimensions used to renormalize the divergent effective action. In this action $\ell^2_P = 16\pi G$, $\alpha = (2880\pi^2)^{-1}$ and $\bar\mu$ is the renormalization mass scale.
We are interested in computing the CTP effective action for the model given by the form (\[phipert\]) for the matter action and when the field $\phi$ is initially in the Hartle- Hawking (HH) vacuum. Since the initial state of the field is described by a thermal density matrix at the HH temperature $T_H$[^6], the finite temperature CTP effective action ($T \equiv 1/\beta$) for this model is given by (for details see [@CamHu]) $$\Gamma^\beta_{CTP}[h^\pm_{\mu\nu}]
\ = \ S^{(div)}_g[h^+_{\mu\nu}]
-S^{(div)}_g[h^-_{\mu\nu}]
-{i\over2}Tr\{ \ln\bar G^\beta_{ab}[h^\pm_{\mu\nu}]\},
\label{eq:eff act two fields}$$ where $\pm$ denote the forward and backward time path of the CTP formalism and $\bar G^\beta_{ab}[h^\pm_{\mu\nu}]$ is the complete $2\times 2$ matrix propagator ($a$ and $b$ take $\pm$ values: $G_{++},G_{+-}$ and $G_{--}$ correspond to the Feynman, Wightman and Schwinger Green’s functions respectively) with thermal boundary conditions for the differential operator $\sqrt{-g^{(0)}}(\Box +
V^{(1)} + V^{(2)} + \cdots)$. The actual form of $\bar G^\beta_{ab}$ cannot be explicitly given. However, it is easy to obtain a perturbative expansion in terms of $V^{(k)}_{ab}$, the $k$-order matrix version of the complete differential operator defined by $V^{(k)}_{\pm\pm} \equiv \pm V^{(k)}_{\pm}$ and $V^{(k)}_{\pm\mp}
\equiv 0$, and $G^\beta_{ab}$, the thermal matrix propagator for a massless scalar field in Schwarzschild spacetime . To second order $\bar G^\beta_{ab}$ reads $$\begin{aligned}
\bar G^\beta_{ab}
\ = \ G^\beta_{ab}
-G^\beta_{ac}V^{(1)}_{cd}G^\beta_{db}
-G^\beta_{ac}V^{(2)}_{cd}G^\beta_{db}
+G^\beta_{ac}V^{(1)}_{cd}G^\beta_{de}
V^{(1)}_{ef}G^\beta_{fb}
+\cdots\end{aligned}$$ Expanding the logarithm and dropping one term independent of the perturbation $h^\pm_{\mu\nu}(x)$, the CTP effective action may be perturbatively written as $$\begin{aligned}
\Gamma^\beta_{CTP}[h^\pm_{\mu\nu}]
& \ = \ & S^{div}_g[h^+_{\mu\nu}] - S^{div}_g[h^-_{\mu\nu}]
+{i\over2}Tr[ V^{(1)}_{+}G^\beta_{++}
-V^{(1)}_{-}G^\beta_{--}
+V^{(2)}_{+}G^\beta_{++}
-V^{(2)}_{-}G^\beta_{--}
]
\nonumber \\
& & -{i\over4}Tr[ V^{(1)}_{+}G^\beta_{++}
V^{(1)}_{+}G^\beta_{++}
+ V^{(1)}_{-}G^\beta_{--}
V^{(1)}_{-}G^\beta_{--}
-2V^{(1)}_{+}G^\beta_{+-}
V^{(1)}_{-}G^\beta_{-+}
].
\label{eq:effective action}\end{aligned}$$ However, unlike the case of Ref.[@CamHu] where $h_{\mu\nu}$ represented a perturbation about flat space and hence one had knowledge of exact “unperturbed" thermal propagators, in this case, since the perturbation is about Schwarzschild spacetime, exact expressions for the corresponding unperturbed propagators $G^\beta_{ab}[h^\pm_{\mu\nu}]$ are not known. Therefore apart from the approximation of computing the CTP effective action to certain order in perturbation theory, an appropriate approximation scheme for the unperturbed Green’s functions is also required. York used the Page approximation [@Page82] for $\langle T_{\mu \nu}\rangle$ in the Schwarzschild metric. The additional complication here is that while to obtain $\langle T_{\mu\nu}\rangle$ the knowledge of only the thermal Feynman Green’s function is required, to calculate the CTP effective action one needs the knowledge of the full matrix propagator, which involves the Feynman, Schwinger and Wightman functions. We can put aside the technical complexity in the calculation of the full thermal matrix propagator $G^\beta_{ab}[h^\pm_{\mu\nu}]$ as our main interest is to identify and analyze the noise term which is the new ingredient in the backreaction problem. We have mentioned that the noise term gives a stochastic contribution $T_{\mu\nu}^{\rm s}$ to the Einstein Langevin equation (\[eq:effective stress tensor\]). We have also stated that this term is related to the variance of fluctuations in $T_{\mu\nu}$, i.e., schematically, to $\langle T^2_{\mu\nu} \rangle$. Since the Influence Functional or CTP formalism itself does not depend on the nature of the approximation, we will attempt to exhibit the general structure and project what is to be expected.
If we denote the difference and the sum of the perturbations $h^\pm_{\mu\nu}$ by $[h_{\mu\nu}] \equiv h^+_{\mu\nu} -
h^-_{\mu\nu}$ and $\{h_{\mu\nu}\} \equiv h^+_{\mu\nu} +
h^-_{\mu\nu}$, respectively, the influence functional form of the thermal CTP effective action may be written to second order in $h_{\mu\nu}$ as [@CamHu] $$\begin{aligned}
\Gamma^\beta_{CTP}[h^\pm_{\mu\nu}]
& \simeq & {1\over2\ell^2_P}\int d^4x\ d^4x'\
[h_{\mu\nu}](x){L}_{(o)}^{\mu\nu,\alpha\beta}(x,x')
\{h_{\alpha\beta}\}(x')
+{1\over2}\int d^4x\
[h_{\mu\nu}](x)T^{\mu\nu}_{(\beta)}(x)
\nonumber \\
& & +{1\over2}\int d^4x\ d^4x'\
[h_{\mu\nu}](x){H}^{\mu\nu,\alpha\beta}(x,x')
\{h_{\alpha\beta}\}(x')
-{1\over2}\int d^4x\ d^4x'\
[h_{\mu\nu}](x)D^{\mu\nu,\alpha\beta}(x,x')
\{h_{\alpha\beta}\}(x')
\nonumber \\
& &+{i\over2}\int d^4x\ d^4x'\
[h_{\mu\nu}](x)N^{\mu\nu,\alpha\beta}(x,x')
[h_{\alpha\beta}](x').
\label{CTPbh}\end{aligned}$$ The first term is the Einstein-Hilbert action to second order in the perturbation $h^\pm_{\mu\nu}(x)$ and ${\mbox{$\rm L$}^{\mu\nu , \alpha\beta}_{(o)}(x)}$ is a symmetric local kernel, [*i.e.*]{} ${\mbox{$\rm L$}^{\mu\nu , \alpha\beta}_{(o)}(x,x')}$ = ${\mbox{$\rm L$}^{\mu\nu , \alpha\beta}_{(o)}(x',x)}$. The second is a local term linear in $h^\pm_{\mu\nu}(x)$. $T_{(\beta)}^{\mu\nu}(x)$ represents the zeroth order contribution to $\langle \hat{T}_{\mu\nu}(x)\rangle$ and far away from the horizon it takes the form of the stress tensor of massless scalar particles at temperature $\beta^{-1}$. The third and fourth terms constitute the remaining quadratic component of the real part of the effective action. The kernels $H^{\mu\nu,\alpha\beta}(x,x')$ and $D^{\mu\nu,\alpha\beta}(x,x')$ are respectively even and odd in $x,x'$. The last term gives the imaginary part of the effective action and the kernel $N^{\mu\nu,\alpha\beta}(x,x')$ is symmetric. This is the general structure of the CTP effective action arising from the calculation of the traces in equation (\[eq:effective action\]). Of course, to write down explicit expressions for the non-local kernels one requires the input of the explicit form of $G^\beta_{ab}[h^\pm_{\mu\nu}]$ , which we have not used. In spite of this limitation we can make some interesting observations from this effective action. Connecting this thermal CTP effective action to the influence functional via equation (\[CTPIFconn\]) we see that the nonlocal imaginary term containing the kernel $N^{\mu\nu,\alpha\beta}(x,x')$ is responsible for the generation of the stochastic noise term in the Einstein-Langevin equation and the real non-local term containing kernel $D^{\mu\nu,\alpha\beta}(x,x')$ is responsible for the non-local dissipation term. The Einstein-Langevin equation can be generated from equation (\[stochem\]) by first constructing the improved semiclassical effective action in accordance with (\[Seffxi\]) and deriving the equation of motion (\[stochem\]) by taking a functional derivative of the above effective action with respect to $[h_{\mu\nu}]$ and equating it to zero. With the identification of the noise and dissipation kernels, one can use them to write down a Fluctuation-Dissipation relation (FDR) analogous to (\[FDR\]) in the context of black holes.
Structure Formation in Inflationary Universes {#sec:strfor}
=============================================
Cosmological structure formation is a key problem in modern cosmology [@Pad93; @Smoot92] and inflation [@infcos; @KolTur90; @Linde90] offers a natural solution to this problem. If an inflationary period is present, the initial seeds for the generation of the primordial inhomogeneities that lead to the large scale structure observed in the present universe have their source in the quantum fluctuations of the inflaton field, the field which is generally responsible for driving inflation. Stochastic gravity provides a sound and natural formalism for the derivation of the cosmological perturbations generated during inflation.
In Ref. [@RouVer03a] it was shown that the correlation functions that follow from the Einstein-Langevin equation which emerges in the framework of stochastic gravity coincide with that obtained with the usual quantization procedures [@MukFelBra92] when both the metric perturbations and the inflaton fluctuations are both linearized. Stochastic gravity, however, can naturally deal with the fluctuations of the inflaton field even beyond the linear approximation.
Here we will illustrate the equivalence with the usual formalism, based on the quantization of the linear cosmological and inflaton perturbations, with one of the simplest chaotic inflationary models in which the background spacetime is a quasi de Sitter universe [@RouVer00; @RouVer03a].
In this chaotic inflationary model [@infcos; @Linde90] the inflaton field $\phi$ of mass $m$ is described by the following Lagrangian density $${\cal
L}(\phi)={1\over 2}g^{ab}\nabla_a\phi \nabla_b\phi + {1\over
2}m^2\phi^2. \label{1.14}$$ The conditions for the existence of an inflationary period, which is characterized by an accelerated cosmological expansion, is that the value of the field over a region with the typical size of the Hubble radius is higher than the Planck mass $m_P$. In order to solve the cosmological horizon and flatness problem more than 60 e-folds of expansion are needed; to achieve this the scalar field should begin with a value higher than $3m_P$. The inflaton mass is small: as we will see, the large scale anisotropies measured in the cosmic background radiation restrict the inflaton mass to be of the order of $10^{-6}m_P$. We will not discuss the naturalness of this inflationary model and we will simply assume that if one such region is found (inside a much larger universe) it will inflate to become our observable universe.
We want to study the metric perturbations produced by the stress-energy tensor fluctuations of the inflaton field on the homogeneous background of a flat Friedmann-Robertson-Walker model, described by the cosmological scale factor $a(\eta)$, where $\eta$ is the conformal time, which is driven by the homogeneous inflaton field $\phi(\eta)=\langle\hat\phi\rangle$. Thus we write the inflaton field in the following form: $\hat\phi=\phi(\eta)+ \hat\varphi (x)$, where $\hat\varphi (x)$ corresponds to a free massive quantum scalar field with zero expectation value on the homogeneous background metric: $\langle\hat\varphi\rangle=0$. We will restrict ourselves to scalar-type metric perturbations because these are the ones that couple to the inflaton fluctuations in the linear theory. We note that this is not so if we were to consider inflaton fluctuations beyond the linear approximation, then tensorial and vectorial metric perturbations would also be driven. The perturbed metric $\tilde g_{ab}=g_{ab}+h_{ab}$ can be written in the longitudinal gauge as, $$ds^2=a^2(\eta)[-(1+2\Phi(x))d\eta^2+(1-2\Psi(x))\delta_{ij}dx^idx^j],
\label{1.16}$$ where the scalar metric perturbations $\Phi(x)$ and $\Psi(x)$ correspond to Bardeen’s gauge invariant variables [@Bar80].
Einstein-Langevin equation for scalar metric perturbations
----------------------------------------------------------
The Einstein-Langevin equation is gauge invariant, thus we can work in a desired gauge and then extract the gauge invariant quantities. It is given by $$G^{(0)}_{ab}-8\pi G\langle\hat T^{(0)}_{ab}\rangle+
G^{(1)}_{ab}(h)-8\pi G\langle\hat T^{(1)}_{ab}(h)\rangle=
8\pi G\xi_{ab},
\label{1.17}$$ where the two first terms cancel, that is $G^{(0)}_{ab}-8\pi
G\langle\hat T^{(0)}_{ab}\rangle=0$, as the background metric satisfies the semiclassical Einstein equations. Here the subscripts $(0)$ and $(1)$ refer to functions in the background metric $g_{ab}$ and linear in the metric perturbation $h_{ab}$, respectively. The stress tensor operator $\hat T_{ab}$ for the minimally coupled inflaton field in the perturbed metric is: $$\hat T_{ab}= \tilde\nabla_{a}\hat\phi
\tilde\nabla_{b}\hat\phi+{1\over2}\tilde g_{ab} (\tilde\nabla_{c}
\hat\phi \tilde\nabla^{c}\hat\phi+ m^2\hat\phi^2). \label{1.18}$$
Using the decomposition of the scalar field into its homogeneous and inhomogeneous part and the metric $\tilde g_{ab}$ into its homogeneous background $g_{ab}$ and its perturbation $h_{ab}$, the renormalized expectation value for the stress-energy tensor operator can be written as $$\langle \hat T^R_{ab}[\tilde g]\rangle= \langle \hat
T_{ab}[\tilde g]\rangle_{\phi\phi}+ \langle \hat T_{ab}[\tilde
g]\rangle_{\phi\varphi}+ \langle \hat T^R_{ab}[\tilde
g]\rangle_{\varphi\varphi}, \label{1.19}$$ where the subindices indicate the degree of dependence on the homogeneous field $\phi$ and its perturbation $\varphi$. The first term in this equation depends only on the homogeneous field and it is given by the classical expression. The second term is proportional to $\langle\hat\varphi[\tilde g]\rangle$ which is not zero because the field dynamics is considered on the perturbed spacetime, [*i.e.*]{}, this term includes the coupling of the field with $h_{ab}$ and may be obtained from the expectation value of the linearized Klein-Gordon equation, $\left( \Box_{g+h}-m^2\right)\hat\varphi =0. \label{1.19a}$ The last term in Eq. (\[1.19\]) corresponds to the expectation value to the stress tensor for a free scalar field on the spacetime of the perturbed metric.
After using the previous decomposition, the noise kernel $N_{abcd}[g;x,y)$ can be written as $$\langle \{\hat t_{ab}[g;x),\hat t_{cd}[g;y)\}\rangle \ = \ \langle
\{\hat t_{ab}[g;x),\hat t_{cd}[g;y)\} \rangle_{(\phi\varphi)^2} +
\langle \{\hat t_{ab}[g;x),\hat t_{cd}[g;y)\}
\rangle_{(\varphi\varphi)^2}, \label{1.20}$$ where we have used the fact that $\langle\hat\varphi\rangle=0
=\langle\hat\varphi\hat\varphi\hat\varphi\rangle$ for Gaussian states on the background geometry. We consider the vacuum state to be the Euclidean vacuum which is preferred in the de Sitter background, and this state is Gaussian. In the above equation the first term is quadratic in $\hat\varphi$ whereas the second one is quartic, both contributions to the noise kernel are separately conserved since both $\phi(\eta)$ and $\hat\varphi$ satisfy the Klein-Gordon field equations on the background spacetime. Consequently, the two terms can be considered separately. On the other hand if one treats $\hat \varphi$ as a small perturbation the second term in (\[1.20\]) is of lower order than the first and may be consistently neglected, this corresponds to neglecting the last term of Eq. (\[1.19\]). The stress tensor fluctuations due to a term of that kind were considered in Ref. [@RouVer00].
We can now write down the Einstein-Langevin equations (\[1.17\]) to linear order in the inflaton fluctuations. It is easy to check [@RouVer03a] that the [*space-space*]{} components coming from the stress tensor expectation value terms and the stochastic tensor are diagonal, i.e. $\langle\hat T_{ij}\rangle=0= \xi_{ij}$ for $i\not= j$. This, in turn, implies that the two functions characterizing the scalar metric perturbations are equal: $\Phi=\Psi$ in agreement with Ref. [@MukFelBra92]. The equation for $\Phi$ can be obtained from the $0i$-component of the Einstein-Langevin equation, which in Fourier space reads $$2ik_i({\cal H}\Phi_k+\Phi'_k)= 8\pi
G(\xi_{0i})_k, \label{1.21}$$ where $k_i$ is the comoving momentum component associated to the comoving coordinate $x^i$, and we have used the definition $\Phi_k(\eta)= \int d^3 x \exp(-i\vec k\cdot\vec x)\Phi(\eta,\vec
x)$. Here primes denote derivatives with respect to the conformal time $\eta$ and ${\cal H}=a'/a$. A nonlocal term of dissipative character which comes from the second term in Eq. (\[1.19\]) should also appear on the left hand side of Eq. (\[1.21\]), but we have neglected it to simplify the forthcoming expressions[^7]. Note, however, that the equivalence of the stochastic approach to linear order in $\hat\varphi$ and the usual linear cosmological perturbations approach is independent of that approximation [@RouVer03a]. To solve Eq. (\[1.21\]), whose left-hand side comes from the linearized Einstein tensor for the perturbed metric [@MukFelBra92], we need the retarded propagator for the gravitational potential $\Phi_k$, $$G_k(\eta,\eta')= -i {4\pi\over k_i m_P^2}\left(
\theta(\eta-\eta') {a(\eta')\over a(\eta)}+f(\eta,\eta')\right),
\label{1.22}$$ where $f$ is a homogeneous solution of Eq. (\[1.21\]) related to the initial conditions chosen and $m_P^2=1/G$. For instance, if we take $f(\eta,\eta')=-\theta(\eta_0-\eta')a(\eta')/a(\eta)$ the solution would correspond to “turning on" the stochastic source at $\eta_0$. With the solution of the Einstein-Langevin equation (\[1.21\]) for the scalar metric perturbations we are in a position to compute the two-point correlation functions for these perturbations.
Correlation functions for scalar metric perturbations
-----------------------------------------------------
The two-point correlation function for the scalar metric perturbations induced by the inflaton fluctuations is thus given by $$\langle\Phi_k(\eta)\Phi_{k'}(\eta')\rangle_s = (2\pi)^2\delta(\vec
k+\vec k') \times\int^\eta \!d\eta_1\int^{\eta'}\!d\eta_2
G_k(\eta,\eta_1) G_{k'}(\eta',\eta_2)
\langle(\xi_{0i})_k(\eta_1)(\xi_{0i})_{k'}(\eta_2)\rangle_s .
\label{1.23}$$ Here the two-point correlation function for the stochastic source, which is connected to the stress-energy tensor fluctuations through the noise kernel is given by, $$\langle (\xi_{0i})_k(\eta_1)(\xi_{0i})_{-k}(\eta_2)\rangle_s \ = \
{1\over2} \langle\{(\hat t_{0i})_k(\eta_1),(\hat
t_{0i})_{-k}(\eta_2)\}\rangle_{\phi\varphi} \ = \
{1\over2}k_ik_i\phi'(\eta_1)\phi'(\eta_2)G_k^{(1)}(\eta_1,\eta_2),
\label{1.24}$$ where $G_k^{(1)}(\eta_1,\eta_2)=\langle\{\hat\varphi_k(\eta_1),
\hat\varphi_{-k}(\eta_2)\}\rangle$ is the $k$-mode Hadamard function for a free minimally coupled scalar field in the appropriate vacuum state on the Friedmann-Robertson-Walker background.
In practice, to make the explicit computation of the Hadamard function we will assume that the field state is in the Euclidean vacuum and the background spacetime is de Sitter. Furthermore, we will compute the Hadamard function for a massless field, and will make a perturbative expansion in terms of the dimensionless parameter $m/m_P$. Thus we consider $$\bar
G_k^{(1)}(\eta_1,\eta_2) = \langle 0|\{\hat y_k(\eta_1),\hat
y_{-k}(\eta_2)\}|0\rangle = 2{\cal
R}\left(u_k(\eta_1)u_k^*(\eta_2)\right),$$ with $\hat y_k(\eta)=
a(\eta)\hat\varphi_k(\eta)= \hat a_k u_k(\eta)+\hat
a_{-k}^\dagger u_{-k}^*(\eta)$ and where $u_k=(2k)^{-1/2}e^{ik\eta}(1-i/\eta)$ are the positive frequency $k$-mode for a massless minimally coupled scalar field on a de Sitter background, which define the Euclidean vacuum state: $\hat
a_k|0\rangle=0$ [@BirDav].
The assumption of a massless field for the computation of the Hadamard function is made because massless modes in de Sitter are much simpler to deal with than massive modes. We can see that this is, however, a reasonable approximation as follows. For a given mode the $m=0$ approximation is reasonable when its wavelength $\lambda$ is shorter that the Compton wavelength, $\lambda_c=1/m$. In our case we have a very small mass $m$ and the horizon size $H^{-1}$, where $H$ is the Hubble constant $H=\dot a/a$ (here $a(t)$ with $t$ the physical time $dt=ad\eta$) satisfies that $H^{-1}<\lambda_c$. Thus, for modes inside the horizon $\lambda<\lambda_c$ and $m=0$ is a reasonable approximation. Outside the horizon massive modes decay in amplitude as $\sim \exp
(-m^2 t/H)$ whereas massless modes remain constant, thus when modes leave the horizon the approximation will eventually break down. However, we only need to ensure that the approximation is still valid after $60$ e-folds, [*i.e.*]{} $Ht\sim 60$, but this is the case since $60\; m^2< H^2$ given that $m\sim 10^{-6}m_P$, and $m\ll H$ as in most inflationary models [@KolTur90; @Pad93].
The background geometry is not exactly that of de Sitter spacetime, for which $a(\eta)=-(H\eta)^{-1}$ with $-\infty <\eta<
0$. One can expand in terms of the “slow-roll" parameters and assume that to first order $\dot\phi(t)\simeq m_P^2(m/m_P)$, where $t$ is the physical time. The correlation function for the metric perturbation (\[1.23\]) can then be easily computed; see Ref. [@RouVer00; @RouVer03a] for details. The final result, however, is very weakly dependent on the initial conditions as one may understand from the fact that the accelerated expansion of de quasi-de Sitter spacetime during inflation erases the information about the initial conditions. Thus one may take the initial time to be $\eta_0=-\infty$ and obtain to lowest order in $m/m_P$ the expression $$\langle\Phi_k(\eta)\Phi_{k'}(\eta')\rangle_s\simeq 8\pi^2 (m/m_P)^2
k^{-3}(2\pi)^3\delta(\vec k+\vec k') \cos k(\eta-\eta'). \label{1.25}$$ From this result two main conclusions are derived. First, the prediction of an almost Harrison-Zel’dovich scale-invariant spectrum for large scales, i.e. small values of $k$. Second, since the correlation function is of order of $(m/m_P)^2$ a severe bound to the mass $m$ is imposed by the gravitational fluctuations derived from the small values of the Cosmic Microwave Background (CMB) anisotropies detected by COBE [@MukFelBra92]. This bound is of the order of $(m/m_P)\sim 10^{-6}$ .
We should now comment on some differences with those works in Ref. [@strforRenegades] which used a self-interacting scalar field or a scalar field interacting nonlinearly with other fields. In those works an important relaxation of the ratio $m/m_p$ was found. The long wavelength modes of the inflaton field were regarded as an open system in an environment made out of the shorter wavelength modes. Then, Langevin type equations were used to compute the correlations of the long wavelength modes driven by the fluctuations of the shorter wavelength modes. In order to get a significant relaxation on the above ratio, however, one had to assume that the correlations of the free long wavelength modes, which correspond to the dispersion of the system initial state, had to be very small. Otherwise they dominate by several orders of magnitude those fluctuations that come from the noise of the environment. This would require a great amount of fine-tuning for the initial quantum state of each mode [@RouVer03a]. We should remark that in the model discussed here there is no environment for the inflaton fluctuations. The inflaton fluctuations, however, are responsible for the noise that induce the metric perturbations.\
[**Acknowledgements**]{} This work is supported in part by NSF grant PHY98-00967, the MICYT Research Project No. FPA-2001-3598 and European project HPRN-CT-2000-00131.
[99]{} B. L. Hu and A. Matacz, “Quantum Noise in Gravitation and Cosmology” in [*Fluctuations and Order*]{}, edited by Marko Millonas (Springer-Verlag, Berlin, 1996). \[astro-ph/9312012\]
B. L. Hu, in [*Proceedings of the Third International Workshop on Thermal Fields and its Applications*]{}, edited by R. Kobes and G. Kunstatter (World Scientific, Singapore, 1994)\[gr-qc/9403061\]
E. Calzetta, B.-L. Hu, F. D. Mazzitelli, Phys. Rep. [**352**]{}, 459-520 (2001).
S. A. Ramsey and B. L. Hu, Phys. Rev. D [**56**]{}, 678 (1997).
S. A. Ramsey, B. L. Hu and A. M. Stylianopoulos, Phys. Rev. D [**57**]{}, 6003 (1998)
A. Raval, B.L. Hu, and J. Anglin, Phys. Rev. D [**53**]{}, 7003-7019 (1996).
A. Raval, B. L. Hu and D. Koks, Phys. Rev. [**D 55**]{}, 4795 (1997).
D. Koks, B. L. Hu, A. Matacz, and A. Raval, Phys. Rev. D [**56**]{}, 4905-4915 (1997).
D. Koks, A. Matacz, and B. L. Hu, Phys. Rev. D [**55**]{}, 5917-5935 (1997)
E. Calzetta and B. L. Hu, Phys. Rev. D [**61**]{}, 025012 (2000).
E. Calzetta and B. L. Hu, in preparation.
G. J. Stephens, E. A. Calzetta, B. L. Hu, and S. A. Ramsey, Phys. Rev. D [**59**]{}, 045009 (1999)
E. Calzetta, A. Roura and E. Verdaguer, Physica A [**319**]{}, 188 (2003).
E. Calzetta, A. Roura and E. Verdaguer, Phys. Rev. D [**64**]{}, 105008 (2001); Phys. Rev. Lett. [**88**]{}, 010403 (2002)
A. Roura and E. Verdaguer, Int. J. Theor. Phys. [**39**]{}, 1831 (2000).
A. Roura and E. Verdaguer, in preparation.
B. L. Hu and K. Shiokawa, Phys. Rev. D [**57**]{}, 3474-3483 (1998)
K. Shiokawa, Phys. Rev. D [**62**]{}, 024002 (2000).
B. L. Hu, A. Raval and S. Sinha, “Notes on Black Hole Fluctuations and Backreaction" in [*Black Holes, Gravitational Radiation and the Universe: Essays in honor of C. V. Vishveshwara*]{} eds. B. Iyer and B. Bhawal (Kluwer Academic Publishers, Dordrecht, 1998) \[gr-qc/9901010\].
S. Sinha, Alpan Raval and B. L. Hu, “Black Hole Fluctuations and Backreaction in Stochastic Gravity”, Found. Phys. 33 (2003) 37-64 \[gr-qc/0210013\].
G. Stephens and B. L. Hu, “Notes on Black Hole Phase Transitions", Int. J. Theor. Phys. 40 (2001) 2183-2200 \[gr-qc/0102052\]
D. J. Gross, M. J. Perry and L. G. Yaffe, Phys. Rev. D [**25**]{}, 330 (1982).
B. F. Whiting and J. W. York, Jr., Phys. Rev. Lett. 61, 1336 (1988)
B. L. Hu, “Stochastic Gravity", Int. J. Theor. Phys. [**38**]{}, 2987 (1999) \[gr-qc/9902064\].
B. L. Hu, E. Verdaguer, “Recent Advances in Stochastic Gravity: Theory and Issues" Erice Lectures May 2001, in [*Advances in the Interplay between Quantum and Gravity Physics*]{} edited by P. Bergmann and V. De Sabbata, (Kluwer, Dortrecht, 2002) \[gr-qc/0110092\].
B. L. Hu and E. Verdaguer, “Stochastic Gravity: A Primer with Applications", Class. Quant. Grav. 20 (2003) R1-R42
B. L. Hu and E. Verdaguer, “Stochastic Gravity: Theory and Applications", [*Living Reviews in Relativity*]{} (2003)
A. Roura and E. Verdaguer, in preparation.
B. L. Hu, “A Kinetic Theory Approach to Quantum Gravity”, Int. J. Theor. Phys. (2002) \[gr-qc/0204069\]
B. S. DeWitt, Phys. Rep. 19, 295(1975).
N. D. Birrell and P. C. W. Davies, [*Quantum Fields in Curved Space*]{} (Cambridge University Press, Cambridge, 1982).
E. Calzetta and B. L. Hu, Phys. Rev. D [**35**]{}, 495 (1987); A. Campos and E. Verdaguer, Phys. Rev. [**D49**]{}, 1861 (1994).
B. L. Hu, “Semiclassical Gravity and Mesoscopic Physics" in [*Quantum Classical Correspondence*]{} eds. D. S. Feng and B. L. Hu (International Press, Boston, 1997) \[gr-qc/9511077\].
B. L. Hu, “General Relativity as Geometro-Hydrodynamics" Invited talk at the Second Sakharov Conference, Moscow, May, 1996 \[ gr-qc/9607070\].
R. Feynman and F. Vernon, Ann. Phys. (NY) [**24**]{}, 118 (1963); R. Feynman and A. Hibbs, [*Quantum Mechanics and Path Integrals*]{}, (McGraw - Hill, New York, 1965).
A. O. Caldeira and A. J. Leggett, Physica [**121A**]{}, 587 (1983); Ann. Phys. (NY) [**149**]{}, 374 (1983). H. Grabert, P. Schramm and G. L. Ingold, Phys. Rep. [**168**]{}, 115 (1988). B. L. Hu, J. P. Paz and Y. Zhang, Phys. Rev. [**D45**]{}, 2843 (1992); [**D47**]{}, 1576 (1993).
B. L. Hu and A. Matacz, Phys. Rev. D49, 6612(1994)
J. Schwinger, J. Math. Phys. [**2**]{} (1961) 407; L. V. Keldysh, Zh. Eksp. Teor. Fiz. [**47** ]{}, 1515 (1964) \[Engl. trans. Sov. Phys. JEPT [**20**]{}, 1018 (1965)\] R. D. Jordan, Phys. Rev. [**D33** ]{}, 444 (1986); E. Calzetta and B. L. Hu, Phys. Rev. D [**35**]{}, 495 (1987);[**37**]{}, 2878 (1988); [**D40**]{}, 656 (1989).
E. Calzetta and B. L. Hu, Phys. Rev. D49, 6636(1994). B. L. Hu and A. Matacz, Phys. Rev. [**D51**]{}, 1577 (1995). B. L. Hu and S. Sinha, Phys. Rev. [**D51**]{}, 1587 (1995). A. Campos and E. Verdaguer, Phys. Rev. D [**53**]{}, 1927 (1996).
L. Parker, Phys. Rev. [**183**]{}, 1057 (1969). S. W. Hawking, Comm. Math. Phys. [**43**]{}, 199 (1975).
A. H. Guth, Phys. Rev. D 23, 347 (1981). A. Albrecht and P. J. Steinhardt, Phys. Rev. Lett. 48, 1220 (1982). A. D. Linde, Phys. Lett. 114B, 431 (1982). Phys. Lett. 162B, 281 (1985)
E.W. Kolb and M. Turner, [*The early Universe*]{} (Addison-Wesley, Reading, Massachusetts, 1990); T. Padmanabhan, [*Structure formation*]{} (Cambridge University Press, Cambridge, England, 1993).
A. Linde, [*Particle physics and inflationary cosmology*]{} (Harwood Academic Publishers, Switzerland, 1990).
B. L. Hu, Physica A158, 399 (1979).
R. Martin and E. Verdaguer, Phys. Lett. B [**465**]{}, 113 (1999).
R. Martin and E. Verdaguer, Phys. Rev. D [**60**]{}, 084008 (1999).
R. Martin and E. Verdaguer, Phys. Rev. D [**61**]{}, 124024 (2000).
N. G. Phillips and B. L. Hu, Phys. Rev. D [**55**]{}, 6132 (1997).
B. L. Hu and N. G. Phillips, Int. J. Theor. Phys. [**39**]{}, 1817 (2000) \[gr-qc/0004006\].
N. G. Phillips and B. L. Hu, Phys. Rev. D [**62**]{}, 084017 (2000).
N. G. Phillips and B. L. Hu, Phys. Rev. D [**63**]{}, 104001 (2001).
C. W. Misner, K. S. Thorne and J. A. Wheeler, [*Gravitation*]{} (Freeman, San Francisco, 1973).
J. M. Bardeen, Phys. Rev. Lett. 46, 382 (1981); P. Hajicek and W. Israel, Phys. Lett. 80A, 9 (1980).
J. W. York, Jr., Phys. Rev. D [**28**]{}, 2929 (1983); D [**31**]{}, 775 (1985); D [**33**]{}, 2092 (1986).
A. Campos and B. L. Hu, Phys. Rev. D58, 125021(1998); Int. J. Theor. Phys. 38, 1253(1999)
D. N. Page, Phys. Rev. [**D25**]{}, 1499 (1982).
G.F. Smoot [*et al.*]{}, Astrophys. J. Lett. [**396**]{}, L1(1992).
T. Padmanabhan, [*Structure Formation*]{} (Cambridge University Press, Cambridge, England, 1993).
V.F. Mukhanov, H.A. Feldman and R.H. Brandenberger, Phys. Rep. [**215**]{}, 203 (1992).
J. M. Bardeen, Phys. Rev. D [**22**]{}, 1882 (1980).
A. Roura and E. Verdaguer, Int. J. Theor. Phys. [**38**]{}, 3123 (1999).
E. Calzetta and B. L. Hu, Phys. Rev. D [**52**]{}, 6770 (1995); A. Matacz, Phys. Rev. D [**55**]{}, [1860]{} (1997); [**56**]{}, 1836 (1997); E. Calzetta and S. Gonorazky, Phys. Rev. D [**55**]{}, [1812]{} (1997)
[^1]: Corresponding author. Electronic address: [email protected]
[^2]: It is for this reason that I think one can view [*general relativity as geometro-hydrodynamics*]{} [@GRhydro], [*semiclassical gravity as a mesoscopic physics*]{} [@meso], and take a [*kinetic theory approach to quantum gravity*]{}[@KinQG]. The only difference is that instead of dealing with the quantum to classical and micro- to macro- transition in the state of matter and fields we are dealing with the corresponding issues for spacetime and geometry.
[^3]: We are labeling the degrees of freedom of the system and the environment by single letters $x$ and $q$ with the understanding that they can represent multiple or even infinite degrees of freedom, e.g. corresponding to a field [@HM2].
[^4]: In this article we use the $(+,+,+)$ sign conventions of Ref. [@MTW], and units in which $c=\hbar=1$.
[^5]: The Hartle-Hawking vacuum is a pure state which is perceived as vacuum by free-falling observers crossing either the black hole or the white hole horizons. However, if one traces out the modes localized in the second asymptotically flat region, one ends up with an incoherent density matrix perceived as a thermal state by static observers in the first asymptotically flat region.
[^6]: As mentioned earlier, this is true for the basis of modes associated to static observers and provided that one is not concerned about quantum correlations (entanglement) between the two asymptotically flat regions, so that the modes localized in the second region can be traced out. Otherwise the full Hartle-Hawking vacuum, which is a pure state, should be considered.
[^7]: Such a term, which leads to an integrodifferential Einstein-Langevin equation, might not be negligible in some situations. Despite the apparent difficulty of dealing with an integrodifferntial equation, in this case the Einstein-Langevin equation can be actually transformed, after suitable manipulation, into an ordinary differential equation, so that the inclusion of the nonlocal term is still tractable [@RouVer03a]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we explore $f(T, \mathcal{T})$, where $T$ and $\mathcal{T}$ denote the torsion scalar and the trace of the energy-momentum tensor respectively. We impose the covariant conservation to the energy-momentum tensor and obtain a cosmological $f(T, \mathcal{T})$ respectively. We impose the covariant conservation to the energy-momentum tensor and obtain a cosmological $f(T, \mathcal{T})$ model. Then, we study the stability of the obtained model for power-law and de Sitter solutions and our result show that the model can be stable for some values of the input parameters, for both power-law and de Sitter solutions.'
author:
- 'M. G. Ganiou'
- 'Ines G. Salako'
- 'M. J. S. Houndjo'
- 'J. Tossa'
title: ' $f(T,\mathcal{T})$ Cosmological Models in Phase Space'
---
10000
Introduction {#sec1}
============
Nowadays the current acceleration of the expansion of the universe is widely confirmed by several independent cosmological observational data as Cosmic Microwave Background Radiation (CMBR) [@Spergel] and the Sloan Digital Sky Survey (SDSS) [@Adelman]. This stage of the universe is explained in the literature through two approaches. The first assumes that the universes if filled by an exotic ith negative pressure, named dark energy known as the responsible of this acceleration of the universe. The second approach, instead of assuming an exotic component, consists to modify the GR by changing the usual Einstein-Hilbert gravitational term, and various theories have been developed in this way and based on the Levi-Civita’s connections, as ( $f(R)$, $f(R,\mathcal{T})$ [@ma1]-[@ma6], $f(G)$) [@mj1]-[@mj5] where $R$ denotes the curvature scalar, $\mathcal{T}$ the trace of the energy-momentum tensor and $G$ the Gauss-Bonnet invariant defined by $G=R^2-4R_{\mu\nu}R^{\mu\nu} + R_{\mu\nu \
lambda\sigma}R^{\mu\nu\lambda\sigma}$. There exists another type theory based the weitzenbock’s connections, equivalent to GR, called Tele-parallel Theory $(TT)$. This theory has been introduced by Ferraro [*et al*]{} [@st1] where they explained the UV modifications to the $TT$ and also the inflation. After this, Ferraro and Bengochea [@st1] have consider the same model to describe the dark energy. other works can be found in [@st2]- [@st41]. In the same, modified versions of this theory have been developed and the one to which we are interested in this paper the $f(T, \mathcal{T})$, $T$ and $\mathcal{T}$ being the torsion scalar and the trace of the energy-momentum tensor, respectively. Specifically, this theory can be view as an homologue to $f(,\mathcal{T})$. Beside several works developed within $f(T,\mathcal{T})$ [@sala1]- [@sala4], we can note the one undertaken by Alvarenga and collaborators [@papierdiego], where they search for the model for which the covariant conservation of the energy-momentum tensor is realized. In that paper they investigate the dynamics of scalar perturbations about the obtained model and focused they attention to the sub-Hubble modes and show that through the quasi-static approximation the result are very different from the ones derived in the frame of concordance $\Lambda CDM$, constraining the validity of this kind of model.
In this paper we are interested to the coincidence cosmic problem and search for the $f(T,\mathcal{T})$ model according to what the tress tensor is conserved. In order to obtain a consistent model, we explore the dynamics and stability about the obtained model. To reach our goal, we assume that it possible to have anti-gravity interactions between the dark energy and the matter because of their unknown nature. We also introduce arbitrarily the terms of interaction between these components because we have not sure of the form of interaction between them. We realize a system of three dynamic equations which take into account the dark energy, the dark matter and the ordinary matter. Consequently, we reconstruct four models and we show that the dynamic equations have two possible attractive solutions namely the phase dominated by the dark matter and that dominated by the dark energy. During this investigation, we have realized that some dynamic systems are unstable; meaning that a model provides that everything disappear in the Universe and leads to an Universe more and more poor in energy. Other models show that the Universe should be filled by dark energy. Another important feature emerging from this work is the stability study of $\Lambda$CDM model under consideration by considering two interesting cosmological solutions i.e the power-law and the de Sitter solutions. We have analyzed the constrains on the input parameters and as results, we have found that the stability is always realized.
This paper is organized as follows: In Sec \[sec2\] we have reconstructed a model by vanishing the covariant derived of energy-momentum tensor. The stability of the obtained critical points of the dynamic systems has been explored in Sec \[sec3\] and the perturbation functions have been determined within the model under consideration in Sec \[sec5\]. The Sec \[sec4\] is devoted to cosmological dynamic study of the considered model. Finally, we have ended our investigation by a conclusion in Sec \[sec6\].
Generality on $f(T,\mathcal{T})$ gravity within FLRW Cosmology {#sec2}
==============================================================
The modified theories of Tele-Parallel gravity are those for which the scalar torsion of Tele-Parallel action is substituted by an arbitrarily function of this latter. As it is done in Tele-Parallel, the modified versions of this theory are also described by the orthonormal tetrads which components are defined on the tangent space of each point of the manifold. The line element is written as $$\begin{aligned}
ds^2=g_{\mu\nu}dx^\mu dx^\nu=\eta_{ij}\theta^i\theta^j\,, \end{aligned}$$ with the following definitions $$\begin{aligned}
d^\mu=e_{i}^{\;\;\mu}\theta^{i}; \,\quad \theta^{i}=e^{i}_{\;\;\mu}dx^{\mu}.\end{aligned}$$ Note that $\eta_{ij}=diag(1,-1,-1,-1)$ is the Minkowskian metric and the $\{e^{i}_{\;\mu}\}$ are the components of the tetrad which satisfy the following identity: $$\begin{aligned}
e^{\;\;\mu}_{i}e^{i}_{\;\;\nu}=\delta^{\mu}_{\nu},\quad e^{\;\;i}_{\mu}e^{\mu}_{\;\;j}=\delta^{i}_{j}.\end{aligned}$$ In General Relativity, one use the following Levi-Civita’s connection which preserves the curvature whereas the torsion vanishes $$\overset{\circ }{\Gamma }{}_{\;\;\mu \nu }^{\rho } =
\frac{1}{2}g^{\rho \sigma }\left(
\partial _{\nu} g_{\sigma \mu}+\partial _{\mu}g_{\sigma \nu}-\partial _{\sigma}g_{\mu \nu}\right)\;,$$ But in the Tele-Parallel theory and its modified version, one keeps the scalar torsion by using Weizenbock’s connection defined as: $$\begin{aligned}
\Gamma^{\lambda}_{\mu\nu}=e^{\;\;\lambda}_{i}\partial_{\mu}e^{i}_{\;\;\nu}=-e^{i}_{\;\;\mu}\partial_\nu e_{i}^{\;\;\lambda}.\end{aligned}$$
From this connection, one obtains the geometric objects. The first is the torsion defined by
$$\begin{aligned}
T^{\lambda}_{\;\;\;\mu\nu}= \Gamma^{\lambda}_{\mu\nu}-\Gamma^{\lambda}_{\nu\mu},\end{aligned}$$
from which we define the contorsion as $$\label{K}
K_{\;\;\mu \nu }^{\lambda} \equiv \widetilde{\Gamma} _{\;\mu \nu }^{\lambda }
-\overset{\circ}{\Gamma }{}_{\;\mu \nu }^{\lambda}=\frac{1}{2}(T_{\mu }{}^{\lambda}{}_{\nu }
+ T_{\nu}{}^{\lambda }{}_{\mu }-T_{\;\;\mu \nu }^{\lambda})\;,$$ Where the expression $\overset{\circ }{\Gamma }{}_{\;\;\mu \nu }^{\lambda}$ designs the above defined connection. Then we can write $$\begin{aligned}
K^{\mu\nu}_{\;\;\;\;\lambda}=-\frac{1}{2}\left(T^{\mu\nu}_{\;\;\;\lambda}-T^{\nu\mu}_{\;\;\;\;\lambda}+T^{\;\;\;\nu\mu}_{\lambda}\right)\,\,.\end{aligned}$$ The two previous geometric objects (the torsion and the contorsion) are used to define another tensor by $$\begin{aligned}
S_{\lambda}^{\;\;\mu\nu}=\frac{1}{2}\left(K^{\mu\nu}_{\;\;\;\;\lambda}+
\delta^{\mu}_{\lambda}T^{\alpha\nu}_{\;\;\;\;\alpha}-\delta^{\nu}_{\lambda}T^{\alpha\mu}_{\;\;\;\;\alpha}\right)\label{S}\end{aligned}$$ From the fact that we are talking about the modified versions of Tele-Parallel gravity, one use a general algebraic function of scalar torsion instead the scalar torsion only as it is done in the initial theory. So, the new action is written as
$$\begin{aligned}
S= \int e \left[\frac{T+f(T,\mathcal{T})}{2\kappa^2} +\mathcal{L}_{m} \right]d^{4}x \label{eq9}\end{aligned}$$
where $\kappa^{2} = 8 \pi G $ is the usual constant coupling to Newton gravitational constant. Varying the action with respect to the tetrad, one obtains the equations of motion as [@sala1]- [@sala4] : $$\begin{aligned}
\label{lagran1}
&&[\partial_\xi(ee^\rho_a
S^{\;\;\sigma\xi}_\rho)-ee^\lambda_a S^{\rho\xi\sigma} T_{\rho\xi\lambda}](1+f_T)
+ e e^\rho_a(\partial_\xi T)S^{\;\;\sigma\xi}_\rho f_{TT} +\frac{1}{4} e e^\sigma_a (T) \nonumber \\
&& =- \frac{1}{4} e e^\sigma_a \Big( f(\mathcal{T})\Big) -e e^\rho_a(\partial_\xi \mathcal{T})S^{\;\;\sigma\xi}_\rho f_{T\mathcal{T}} +
f_{\mathcal{T}}\;\Big(\frac{e\,\Theta^\sigma_{\;\;a}
+ e e^\sigma_a \;p }{2}\Big) + \frac{\kappa^{2}}{2} e\,\Theta^\sigma_{\;\;a} \;,\end{aligned}$$ with $f_{\mathcal{T}} = \partial f/\partial \mathcal{T} $, $f_{T} = \partial f/\partial T$, $ f_{T\mathcal{T}}
= \partial^{2}f/\partial T\partial \mathcal{T}$, $f_{TT} = \partial^{2}f/\partial T^{2}$ et $\Theta^\sigma_{\;\;a}$ is the energy-momentum tensor of matter field. By using some transformations, we can establish the following relations: $$\begin{aligned}
\label{nablaS'}
e^a_\nu e^{-1}\partial_\xi(ee^\rho_a
S^{\;\;\sigma\xi}_\rho)-S^{\rho\xi\sigma}T_{\rho\xi\nu} = -\nabla^\xi S_{\nu\xi}^{\;\;\;\;\sigma}-S^{\xi\rho\sigma}K_{\rho\xi\nu}\;,\end{aligned}$$ $$\label{eqdivs'}
G_{\mu\nu}-\frac{1}{2}\,g_{\mu\nu}\,T
=-\nabla^\rho S_{\nu\rho\mu}-S^{\sigma\rho}_{\;\;\;\;\mu}K_{\rho\sigma\nu}\;,$$ By the end, from the combination of equations Eq. (\[nablaS’\]) and Eq. (\[eqdivs’\], the field equations Eq. (\[lagran1\]) can be written as: $$A_{\mu\nu}(1+ f_T) +\frac{1}{4}g_{\mu\nu}\;T =B_{\mu\nu}^{eff} \label{motion11}$$ where $$\begin{aligned}
\label{motion1add}
&&A_{\mu \nu }=g_{\sigma\mu}e^a_\nu[e^{-1}\partial_\xi(ee^\rho_a
S^{\;\;\sigma\xi}_\rho)-e^\lambda_a S^{\rho\xi\sigma} T_{\rho\xi\lambda}]\\ \nonumber
&&\qquad=-\nabla^\sigma S_{\nu\sigma\mu }-S_{\;\;\;\;\mu }^{\rho\lambda }K_{\lambda \rho \nu }
=G_{\mu \nu }-\frac{1}{2}g_{\mu \nu }T, \; \\ \nonumber
&&B_{\mu\nu}^{eff} = S^{\rho}_{\;\;\;\mu\nu}\; f_{T\mathcal{T}}\; \partial_{\rho} \mathcal{T} -
S^{\rho}_{\;\;\;\mu\nu}\;f_{TT}\; \partial_{\rho} T
- \frac{1}{4} g_{\mu \nu }f +
f_{\mathcal{T}}\;\Big(\frac{\Theta_{\mu \nu }
+ g_{\mu \nu } \;p }{2}\Big) + \frac{\kappa^{2}}{2} \,\Theta_{\mu \nu } \;,\end{aligned}$$ So the relation Eq. (\[motion11\]) can take the following form: $$(1+ f_T)\,G_{\mu\nu}=T_{\mu\nu}^{eff} \label{motion12}$$ where $$\begin{aligned}
\label{motion1add'}
T_{\mu\nu}^{eff} =S^{\rho}_{\;\;\;\mu\nu}\; f_{T\mathcal{T}}\; \partial_{\rho} \mathcal{T} -
S^{\rho}_{\;\;\;\mu\nu}\;f_{TT}\; \partial_{\rho} T - \frac{1}{4} g_{\mu \nu } \Big(T+ f \Big) +\frac{T\,g_{\mu\nu}\,f_T}{2}+
f_{\mathcal{T}}\;\Big(\frac{\Theta_{\mu \nu }
+ g_{\mu \nu } \;p }{2}\Big) + \frac{\kappa^{2}}{2} \,\Theta_{\mu \nu } \;.\end{aligned}$$
Reconstructing of model {#sec3}
=======================
In this section, we are interested to $ T+ Q\,\mathcal{T}^{\mathcal{N}} $ models which can reproduce the different features of $\Lambda CDM$.
In order to point out the expression of the covariant energy-momentum tensor from which one hopes extract a algebraic function, we take the covariant derivative of (\[motion12\]) which leads to: $$\begin{aligned}
\nabla^\mu \Big[(1+ f_T)\, G_{\mu\nu}\Big]& =& \nabla^\mu T_{\mu\nu}^{eff} \cr
&=& \nabla^\mu \Bigg[\frac{T\,g_{\mu\nu}\,f_T}{2}+ S^{\rho}_{\;\;\;\mu\nu}\; f_{T\mathcal{T}}\; \partial_{\rho} \mathcal{T} -
S^{\rho}_{\;\;\;\mu\nu}\;f_{TT}\; \partial_{\rho} T \nonumber\\&-& \frac{1}{4} g_{\mu \nu } \Big(T+ f \Big) +
f_{\mathcal{T}}\;\Big(\frac{\Theta_{\mu \nu }
+ g_{\mu \nu } \;p }{2}\Big) + \frac{\kappa^{2}}{2} \,\Theta_{\mu \nu } \Bigg].\end{aligned}$$ These previous equations lead to the following expression $$\begin{aligned}
&& \nabla^\mu \Theta_{\mu \nu } = \frac{-2}{(f_{\mathcal{T}}+ \kappa^2) }
\Bigg\{ \nabla^\mu \Big[\frac{T\,g_{\mu\nu}\,f_T}{2} + S^{\rho}_{\;\;\;\mu\nu}\; f_{T\mathcal{T}}\;
\partial_{\rho} \mathcal{T} -
S^{\rho}_{\;\;\;\mu\nu}\;f_{TT}\; \partial_{\rho} T - \frac{1}{4} g_{\mu \nu } \Big(T+ f \Big)\Big] + \cr
&& \Big(\frac{\Theta_{\mu \nu }
+ g_{\mu \nu } \;p }{2}\Big)\,
\nabla^\mu f_{\mathcal{T}} + \frac{f_{\mathcal{T}}}{2}\,\nabla^\mu ( g_{\mu \nu } \;p )- G_{\mu\nu} \nabla^\mu (1+ f_T)\Bigg\}.\end{aligned}$$ The $f(T,\mathcal{T})= 0+ f(\mathcal{T})$ gravity field equations namely ($f(T)=0$ or $f_{TT} =f_{T\mathcal{T}} =f_{T}= 0$ ) become $$\begin{aligned}
\nabla_\mu \Theta^\mu_\nu = \frac{1}{2(f_{\mathcal{T}}+ \kappa^2) } \Bigg\{
\frac{1}{4} \delta^\mu_\nu \nabla_\mu f(\mathcal{T}) -
\Big(\frac{\Theta^\mu_\nu
+ \delta^\mu_\nu \;p }{2}\Big)\,
\nabla_\mu f_{\mathcal{T}} - \frac{f_{\mathcal{T}}}{2}\,\delta^\mu_\nu \nabla_\mu \,p \Bigg\},\end{aligned}$$ where we have used the barotropic equation of state $p = \omega\rho$. By fixing $\nu= 0$, one gets: $$\begin{aligned}
\label{afrik}
\dot{\rho}+3H\rho\left(1+\omega\right) = \frac{-\dot{\rho}}{2(f_{\mathcal{T}}+ \kappa^2) }
\Bigg\{\frac{1}{4}\left(1-3\omega\right)\;f_{\mathcal{T}} +
\Big(\frac{1 + \;\omega }{2}\Big)\rho\, \left(1-3\omega\right) f_{\mathcal{T}\mathcal{T}} + \frac{f_{\mathcal{T}}}{2}\,\omega \Bigg\}\end{aligned}$$ To ensure cancellation of the divergence of the energy-momentum tensor, we vanish the second member of the equation Eq.(\[afrik\]), and obtain the following differential equation $$\begin{aligned}
\frac{1}{2} f(\mathcal{T}) + f_{\mathcal{T}}\; \mathcal{T} \;\frac{(1-\omega)}{(1+\omega)}=0, \,\label{s1'},\end{aligned}$$ whose general solution reads $$\begin{aligned}
f(\mathcal{T})= Q\; \mathcal{T}^{\frac{(1+3\omega)}{2(1+\omega)}} , \label{s1}\end{aligned}$$ $$\begin{aligned}
f(T, \mathcal{T})= T -2\Lambda + Q\; \mathcal{T}^{\frac{(1+3\omega)}{2(1+\omega)}} , \label{s1}\end{aligned}$$ We report here that $ Q$ is the integration constant. At the moment, we are pointing out the exact expression of the constant $Q$ by using the wonderful conditions mentioned in [@st28] which stipulates that the algebraic function $ f( \mathcal{T})=\mathcal{T}^{\mathcal{N}} $ must satisfy the following initial conditions $$\begin{aligned}
\left(f\right)_{t=t_i}=T_i,\quad \left(\frac{df}{dt}\right)_{t=t_i}=\left(\frac{dT}{dt}\right)_{t=t_i},\label{arsene1}\end{aligned}$$ with $t_i$ the early time and $T_i$ the initial valor of the scalar torsion associated. By making use of this initial condition (\[arsene1\]) and (\[s1\]), one expresses the constant $Q$ as $$\begin{aligned}
Q=2 \Lambda \; \mathcal{T}_0^{-\frac{(1+3\omega)}{2(1+\omega)}},\label{cvalue}\end{aligned}$$ and the associated algebraic function is $$\begin{aligned}
f(T, \mathcal{T})= T + 2\Lambda \Big[ \Big(\frac{\mathcal{T}}{\mathcal{T}_0}\Big)^{\frac{(1+3\omega)}{2(1+\omega)}}- 1 \Big]
. \label{reconstruit}\end{aligned}$$ We emphasize here the constant $Q$ is positive because of the positivity of $\Lambda$ parameter. Moreover, if it vanishs ($ Q=0 $), we come back to the TT equivalent of RG.
Dynamic study of the systems {#sec4}
=============================
We are working in this section whit the cosmological flat metric of $FLRW$ described by $$\begin{aligned}
ds^{2}= dt^{2} - a^{2}(t)\left(dx^2+dy^2+dz^2\right), \label{metricflat}
\end{aligned}$$ from which we obtein the diagonal matrix for the tetrads as $$\begin{aligned}
\{e^{a}_{\;\; \mu}\}= diag[1,a,a,a]. \label{eq11}\end{aligned}$$ The determinant of the matrix (\[eq11\]) is $e = a^{3}$ and the non zero components of torsion and contorsion are given by $$\begin{aligned}
T^{1}_{\;\;\; 01}= T^{2}_{\;\;\; 02}=T^{3}_{\;\;\; 03}=\frac{\dot{a}}{a},\\
K^{01}_{\;\;\;\;1}=K^{ 02}_{\;\;\;\;2}=K^{ 03}_{\;\;\;\;3}= \frac{\dot{a}}{a}, \label{eq12}\end{aligned}$$ The calculus of components of $S^{\;\;\; \mu\nu}_{\alpha}$ also gives: $$\begin{aligned}
S^{\;\;\; 11}_{0}=S^{\;\;\; 22}_{0}=S^{\;\;\; 33}_{0}=\frac{\dot{a}}{a}. \end{aligned}$$ Therefore, the scalar torsion is expressed as $$\begin{aligned}
T= -6H^{2}, \label{m1}\end{aligned}$$ where $H=\dot{a}/a$ denotes the Hubble parameter. We report also the expression of the trace of energy-momentum tensor related to matter, $ \Theta =\mathcal{T}= (1 - 3 \omega) \rho$. We assume now that the ordinary component of Universe is a perfect fluid with the equation of state $p = \omega \rho$ and $c^2_s= \dot{p}/\dot{\rho}$ so that the energy-momentum is given by $$\begin{aligned}
\Theta_{ \mu \nu} =
diag(1 , -\omega, -\omega,
-\omega ) \rho. \end{aligned}$$ To point out an application of this theory in Cosmology, we insert as needful the flat metric of FLRW (\[metricflat\]) in the field equations ; and obtain consequently the Friedmann modified equations below $$\begin{aligned}
H^{2}&=&\frac{8\pi G}{3} \rho-\frac{1}{6}\left(f+12 H^{2} f_{T}\right)+f_{\mathcal{T}} \left(\frac{\rho+p}{3}\right),\cr
\dot{H}&=&-\frac{4\pi G\left( 1+f_{\mathcal{T}}/8\pi G\right) \left( \rho+p\right) }{1+f_{T}-12H^{2}f_{TT}+H\left( d\rho /dH\right)
\left( 1-3c_{s}^{2}\right)f_{T\mathcal{T}} }.\end{aligned}$$ where $\rho=\rho_m +\tilde{\rho} +\rho_r$, while $\rho_m$, $ \tilde{\rho}$ and $\rho_r$ represent the energy densities of matter, dark energy and the radiation respectively. We also suppose that these three components of the above defined fluid are in interactions. The continuity equations taking into account the different interactions are written as $$\begin{aligned}
\dot{\rho_m}+3H(\rho_m+p_m)&=& E_1, \label{1}\cr
\dot{\tilde{\rho}}+3H(\tilde{\rho}+\tilde{p})&=& E_2, \label{2}\cr
\dot{\rho_r}+3H(\rho_r+p_r) &=& E_3, \label{1'}\end{aligned}$$ where $E_i$, $i=1,2,3$ are the term of interaction between the two fluids. Now, we can define the cosmological density parameters $$\begin{aligned}
y= \frac{\kappa^2 \rho_m}{3H^{2}},\quad x= \frac{\kappa^2\tilde{\rho}}{3H^{2}}.\quad z= \frac{\kappa^2\rho_r}{3H^{2}}
\label{a6}\end{aligned}$$ and $$\begin{aligned}
\frac{\dot{H}}{H^2}
=-\frac{3/2 \left( 1+f_{\mathcal{T}}/8\pi G\right) \left(x+y+z+ x\,\tilde{\omega} +\omega_r\,z \right) }{1+f_{T}-12H^{2}f_{TT}+H\left( d\rho/dH\right)
\left( 1-3c_{s}^{2}\right)f_{T\mathcal{T}} }.\end{aligned}$$ By using the $e$-folding parameter, $ Z= \ln{a}$, $a$ being the scale factor, the continuity equations Eqs. (\[1\]) and (\[2\]) become $$\begin{aligned}
\label{mouss}
\frac{dx}{dZ}&=& 3x \Bigg( \frac{\left( 1+f_{\mathcal{T}}/8\pi G\right) \left(x+y+z+ x\,\tilde{\omega} +\omega_r\,z \right) }{1+f_{T}-12H^{2}f_{TT}+H\left( d\rho/dH\right)
\left( 1-3c_{s}^{2}\right)f_{T\mathcal{T}} } \Bigg) -3x(1 + \tilde{\omega}) + \frac{\kappa^2\,E_1}{3H^3} \label{3}\cr
\frac{dy}{dZ}&=& 3y \Bigg( \frac{\left( 1+f_{\mathcal{T}}/8\pi G\right) \left(x+y+z+ x\,\tilde{\omega} +\omega_r\,z \right) }{1+f_{T}-12H^{2}f_{TT}+H\left( d\rho/dH\right)
\left( 1-3c_{s}^{2}\right)f_{T\mathcal{T}} } \Bigg) -3y + \frac{\kappa^2\,E_2}{3H^3} \label{4}\cr
\frac{dz}{dZ}& =& 3z \Bigg( \frac{\left( 1+f_{\mathcal{T}}/8\pi G\right) \left(x+y+z+ x\,\tilde{\omega} +\omega_r\,z \right) }{1+f_{T}-12H^{2}f_{TT}+H\left( d\rho/dH\right)
\left( 1-3c_{s}^{2}\right)f_{T\mathcal{T}} } \Bigg) -3z(1 + \omega_r) + \frac{\kappa^2\,E_3}{3H^3}, \end{aligned}$$ Where we have used unit of $ \kappa^{2}=1$ and then $Z\equiv N\equiv\ln a$ is used as e-folding parameter. The interacting parameters $E_i$, $i=1,2,3$ are generally functions of the energy densities and the Hubble parameter i.e $E_i= E_i(H,\rho_i)$. We start the analysis of the system of equations in (\[mouss\]) by vanishing the first member of each of these equations in order to extract critical points. Therefore, one perturbs these equations in first order around the critical points and deduce the stability of the system. In our calculation procedure, we force the following parameters $\omega_m=0$, $\omega_r=\frac{1}{3}$ and $\tilde{\omega}$ to be non zero but negative. We are interested to the stable critical points i.e the points for which the eigenvalues of Jacobian matrix associated to the system of equations are negative. Such of points are useful because they represent the attractive solutions of dynamic system.
Analysis of stability in phase space
====================================
In this section, we will erect four models by choosing different forms of coupling parameters $E_i$ and we will analysis the stability of the corresponding dynamic system around the critical points and plot the evolutionary phase diagram associated. To reach this target, we must search for the critical points of (\[mouss\]) and make the system linear around the above points.
Interacting model - I
---------------------
We consider the models with the following interaction terms
$$\label{8a}
E_1=-6bH\tilde{\rho}, \ \ E_2=E_3=3bH\tilde{\rho},$$
where $b$ is a coupling parameter assumed to be positive real in the input parameters. Then, the equation Eq. (\[8a\]) shows that matter and the radiation have energy densities which increase with the time whereas the energy density of dark energy is going to disappears completely. So, the dar energy declines for matter and radiation.
Using (\[reconstruit\]) and (\[8a\]), the system (\[mouss\]) takes the form
$$\begin{aligned}
\frac{dx}{dN}&=&-3x(1+\tilde{\omega})+3x\Big(x+y+z+\tilde{\omega} x+\omega_r z\Big)-6bx,\nonumber\\
\frac{dy}{dN}&=&-3y+3y\Big(x+y+z+\tilde{\omega} x+\omega_r z\Big)+3bx,\label{33}\\
\frac{dz}{dN}&=&-3z(1+\omega_r)+3z\Big(x+y+z+\tilde{\omega} x+\omega_r
z\Big)+3bx.\nonumber\end{aligned}$$
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![ Model I: variation of $x,y,z$ as a function of the $N=\ln(a)$. The initial conditions chosen are $x(0)=0.7,y(0)=0.3,z(0)=0.01$, $\tilde{\omega}=-1.2,\,\,\omega_r=\frac{1}{3}$ and $b=0.5$. []{data-label="fig1"}](model1.pdf "fig:"){width="7cm" height="5cm"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
The critical points are found for this model by vanishing the first member of (\[33\]) and there are the following four points recorded in the below board.
--------------------- -------------------------------------------------------- -------------------------- ------------------------- ----------------------------
Point $(x_c,y_c,z_c)$ $\lambda_1$ $\lambda_2$ $\lambda_2$
\[0.5ex\] $ P_{11}$ $(0,0,0)$ $3(b-1)$ $ -3(b-1)$ $-3(1+2b+\tilde{\omega})$
$ P_{12}$ $(0,(1-b),0)$ $-1 $ $ 3(1-b) $ $ -3(\tilde{\omega}+3b) $
$ P_{13}$ $(0,0,\frac{3}{4}(\frac{4}{3}-b))$ $1$ $ 4-3b $ $ 1-3\tilde{\omega}-9b $
$P_{14}$ $(\frac{(1+\tilde{\omega}+2b)}{1+\tilde{\omega}},0,0)$ $3(1+\tilde{\omega}+2b)$ $ 3(\tilde{\omega}+3b)$ $ -1+3(\tilde{\omega}+3b)$
--------------------- -------------------------------------------------------- -------------------------- ------------------------- ----------------------------
: Critical points and The eigenvalues for the first model[]{data-label="table1"}
The point $P_{11}$ is stable when one the following conditions is satisfied $$\begin{aligned}
\tilde{\omega}<-3,\ \ b<1/18.\\
\tilde{\omega}\geq-3, \ \ \tilde{\omega}<-10/9,\ \ b<1/18.\\
\tilde{\omega}\geq-\frac{10}{9},\ \ \tilde{\omega}<0,\ \ b<-\frac{1}{2}(1+\tilde{\omega}).\end{aligned}$$
$ P_{12}$ is an unstable critical point because even if $\lambda_1<0$ we have $\lambda_2<0$ si $b>1$. $ P_{13}$ is stable if $b>1,\tilde{\omega}>-3b$. In parallel $P_{14}$ is not stable because $\lambda_1>0$.
-------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Model I: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig2"}](1a1.pdf "fig:"){width="8cm" height="6cm"} ![ Model I: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig2"}](1a2.pdf "fig:"){width="8cm" height="6cm"}
![ Model I: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig2"}](1a13.pdf "fig:"){width="8cm" height="6cm"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------
It follows that the matter density dominates for the model [*I*]{} whose parameters stay for the conditions $-1<w_d<-\frac{1}{3},w_u>0,b=0.5$, $w_{tot}>0$. This conclusion is confirmed by the fig.\[fig1\] where the matter density dominates whereas the radiation density is above the dark energy density. The fig \[fig2\] shows the phase diagram of the interaction between dark energy and the both matter and radiation. According to the model [*I*]{}, the dark energy density behaves like quintessence while matter and radiation densities fall with expansion.
Interacting model - II
----------------------
We study another model with the choice of the interaction terms under the following form $$\label{9a}
E_1=-3bH\tilde{\rho},\ \ E_2=3bH(\tilde{\rho}-\rho_m),\ \
E_3=3bH\rho_m.$$
This model shows indeed the situation in which the dark energy looses his density in favor of the matter whereas the radiation density increases because of its interaction with the matter.
$$\begin{aligned}
\label{9}
\frac{dx}{dN}&=&3x\Big(x+y+z+\tilde{\omega} x+\omega_r z \Big)-3bx-3x(1+\tilde{\omega}),\cr
\frac{dy}{dN}&=&3y\Big(x+y+z+\tilde{\omega} x+\omega_r z \Big) + 3b(x-y)-3y,\cr
\frac{dz}{dN}&=&3z\Big(x+y+z+\tilde{\omega} x+\omega_r z\Big) + 3by-3z(1+\omega_r),\end{aligned}$$
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![ Model I: variation of $x,y,z$ as a function of the $N=\ln(a)$. The initial conditions chosen are $x(0)=0.7,y(0)=0.3,z(0)=0.01$, $\tilde{\omega}=-1.2,\,\,\omega_r=\frac{1}{3}$ and $b=0.5$. []{data-label="fig3"}](1a0.pdf "fig:"){width="7cm" height="5cm"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
We have free critical points
[|c|l|p[2cm]{}|m[2.5cm]{}|b[6.5cm]{}|]{} Points & $\lambda_1$&$\lambda_2$ &$\lambda_3$ & $(x_c,y_c,z_c)$\
$ P_{21}$ &$3(b-1)$ & $ -3(b-1)$& $-3(1+2b+\tilde{\omega})$ & $(0,\quad 0,\quad 0)$\
$ P_{22}$ & $-1$ &$3(1-b)$&$ -3(\tilde{\omega}+3b) $ & $(0,\quad0,\quad 1)$\
$ P_{23}$ & $1$& $ 4-3b $ & $ 1-3\tilde{\omega}-9b $ & $\Big(0,\quad (1-3b),\quad 3b \Big)$\
$ P_{24}$ &$3(1+\tilde{\omega}+2b)$& $ 3(\tilde{\omega}+3b)$ & $ -1+3(\tilde{\omega}+3b)$& $ \Bigg( -\,{\frac {3 \left( 2+\tilde{\omega} \right) \left( b-\tilde{\omega}-1 \right)
\left( b-\tilde{\omega}-7/3 \right) }{3\,{\tilde{\omega}}^
{3}+ \left( 16-3\,b \right) {\tilde{\omega}}^{2}+ \left( -6\,b+27 \right)
w_{ {d}}+14+{b}^{2}+b}}, \newline
-\,{\frac {b 3 \left( b-\tilde{\omega}-7/3 \right)
\left( b-\tilde{\omega}-1 \right) }{3\,{\tilde{\omega}}^{3}+ \left( 16-3\,b
\right) {\tilde{\omega}}^{2}+ \left( -6\,b+27 \right) \tilde{\omega}+14+{b}^{2}+b}}, \newline
\,{\frac { 3 \left( b-\tilde{\omega}-1 \right) {b}^{2
}}{14+16\,{\tilde{\omega}}^{2}-6\,\tilde{\omega}b+{b}^{2}-3\,{\tilde{\omega}}^{2}b+3\,{w_{{
d}}}^{3}+27\,\tilde{\omega}+b}}
\Bigg)$\
------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Model II: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig4"}](2phase22.pdf "fig:"){width="8cm" height="6cm"} ![ Model II: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig4"}](2phase33.pdf "fig:"){width="8cm" height="6cm"}
![ Model II: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig4"}](2phase44.pdf "fig:"){width="8cm" height="6cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------
$ P_{21}$, $ P_{24}$ are conditionally stable if $b>1+\tilde{\omega}$ (for $ P_{21}$) and $\tilde{\omega}>-2$ and then
$b<1+\tilde{\omega}$ ( for $ P_{24}$ ). But $ P_{22}$ and $ P_{23}$ are unstable because $\lambda_1>0$0.
The figures \[fig3\] and \[fig4\] show the dynamic of the model [*II*]{}. We notice for this model that there is a great domination of dark energy while the energy densities of the radiation and the matter have declined considerably. This situation is well compatible with the recent observational data which show that the dark energy is the very important responsible of the expansion of Universe. We also point out from these figures that if $N\sim2$, the radiation declines and goes towards zero. The figure \[fig4\] is related to the phase diagram of radiation and dark energy interaction. For the model in study, the behavior of the dark energy is similar to quintessence while the matter and radiation tumble during the expansion.
Interacting Model - III
-----------------------
Let us take the following interaction terms [@jamil8]
$$\label{12}
E_1=-6b\kappa^2 H^{-1}\tilde{\rho}\rho_r, \ \
E_2=E_3=3b\kappa^2 H^{-1}\tilde{\rho}\rho_r.$$
The system in (\[mouss\]) becomes $$\begin{aligned}
\frac{dx}{dN}&=&3x\Big(x+y+z+\tilde{\omega}x+\omega_r z \Big)
-3x-3\tilde{\omega}x-18bxz,\nonumber\\
\frac{dy}{dN}&=&3y\Big(x+y+z+\tilde{\omega}x+w_ rz\Big)-3y+9bxz,\\
\frac{dz}{dN}&=&3z\Big(x+y+z+\tilde{\omega}x+\omega_r
z\Big)-3z-3\omega_rz+9bxz.\nonumber\end{aligned}$$ .
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![ Model III: variation of $x,y,z$ as a function of the $N=\ln(a)$. The initial conditions chosen are $x(0)=0.7,y(0)=0.3,z(0)=0.01$, $\tilde{\omega}=-1.2,\,\,\omega_r=\frac{1}{3}$ and $b=0.5$.[]{data-label="fig5"}](4.pdf "fig:"){width="8cm" height="6cm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
[|c|l|p[3cm]{}|m[3cm]{}|b[6cm]{}|]{} Points & $\lambda_1$&$\lambda_2$ &$\lambda_3$ & $(x_c,y_c,z_c)$\
$ P_{31}$ &$-3$ & $ -4$& $-3(1+\tilde{\omega})$ & $\ (0,\quad 0,\quad0)$\
$ P_{32}$ & $3$ &$-1$&$ -3\tilde{\omega} $ & $\ (0,\quad 1,\quad0)$\
$ P_{33}$ & $ 3\tilde{\omega} $& $ 3(1+\tilde{\omega}) $ & $ +9\,b-1+3\,\tilde{\omega} $ & $\ (0,\quad 0,\quad 1) $\
$ P_{34}$ & $4$& $ 1-9b $ & $ -9b+1-3\tilde{\omega} $ & $ \ (1,\quad0,\quad0)$\
$ P_{35}$ & -& - & - &\
$ P_{35}$ &-& -& -& $\Bigg(
{\frac { \left( \tilde{\omega}+6\,b-\frac{1}{3} \right) }{
b \left(+6\,b-\frac{1}{3}+2\,\tilde{\omega} \right)
}},\newline
{\frac {\,-3\,b+18\,{b}^{2}+{\tilde{\omega}}^{2}-2\,\tilde{\omega} \frac{1}{3}+9\,\tilde{\omega}b+\frac{1}{9}}{3b \left(6\,b-\frac{1}{3}+ 2\,\tilde{\omega} \right) }},\newline
-\,{\frac
{\tilde{\omega} \left(+3\,b-\frac{1}{3}+\tilde{\omega} \right) } {3b
\left(6\,b-\frac{1}{3}+2\,\tilde{\omega} \right) }}
\Bigg)$\
-------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Model III: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig6"}](3phase11.pdf "fig:"){width="8cm" height="6cm"} ![ Model III: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig6"}](3phase22.pdf "fig:"){width="8cm" height="6cm"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------
$P_{31}$ is stable for $w_d>-1$. $P_{32}$ , $P_{34}$ are unstable. $P_{33}$ is systematically stable when $w_d<-1$, $b<\frac{1-3w_d}{9(1+\alpha)}$.
We present here the dynamic of model [*III*]{} through the figures \[fig5\] and \[fig6\]. Here we note a gradual increase for the dark energy whereas the energy densities of the radiation and the matter are tending to zero. These facts are compatible with the recent observational data showing that Universe is accelerated expansion because of the strong presence of dark energy in Universe. This analysis shows also that becomes nonexistent because of the strong domination of dark energy. The phase diagram of interaction between dark energy and the both matter and radiation is plotted in figure \[fig6\]. In parallel with the previous models, this model is also one of those where the behavior the energy density of dark energy is that of quintessence while the densities of radiation and matter are going to vanish during the the expansion.
Interacting Model - IV
----------------------
Let’s search for new model generalized by the new following interaction terms: $$\begin{aligned}
\label{11}
E_1&=&-3b\kappa^2 H^{-1}\tilde{\rho}\rho_r,\nonumber\\
E_2&=&3b\kappa^2 H^{-1}(\tilde{\rho}\rho_r-\rho_m\rho_r),\nonumber\\
E_3&=&3b\kappa^2 H^{-1}\rho_m\rho_r.\end{aligned}$$
The system in (\[mouss\]) takes the form
$$\begin{aligned}
\frac{dx}{dN}&=&3x\Big(x+y+z+\tilde{\omega}x+\omega_r z\Big)-3x-3\tilde{\omega}x-9bxz,\nonumber\\
\frac{dy}{dN}&=&3y\Big(x+y+z+\tilde{\omega}x+\omega_r z\Big)-3y+9b(xz-yz),\\
\frac{dz}{dN}&=&3z\Big(x+y+z+\tilde{\omega}x+\omega_r
z\Big)-3z-3\omega_rz+9byz.\nonumber\end{aligned}$$
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![ Model IV: variation of $x,y,z$ as a function of the $N=\ln(a)$. The initial conditions chosen are $x(0)=0.7,y(0)=0.3,z(0)=0.01$, $\tilde{\omega}=-1.2,\,\,\omega_r=\frac{1}{3}$ and $b=0.5$.[]{data-label="fig7"}](model4.pdf "fig:"){width="8cm" height="6cm"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
One obtains seven critical points
[|c|l|p[4cm]{}|m[4cm]{}|b[4cm]{}|]{} Points & $\lambda_1$&$\lambda_2$ &$\lambda_3$ & $(x_c,y_c,z_c)$\
$ P_{41}$ &$-3$ & $ -4$& $-3(1+\tilde{\omega})$ & $\ (0,\quad 0,\quad0)$\
$ P_{42}$ & $3$ &$-1$&$ -3\tilde{\omega} $ & $\ (1,\quad 0,\quad0)$\
$ P_{43}$ & $ 3\tilde{\omega}$& $ 3\tilde{\omega}-1 $ & $ 3\tilde{\omega}-1$ & $\ (0,\quad 1,\quad 0) $\
$ P_{44}$ & $3(1+\tilde{\omega})$& $ -9b+1$ & $ -3\tilde{\omega}-9b+1 $ & $ \ (0,\quad0,\quad1)$\
$ P_{45}$ & -& -& - & $ \ \Big(\frac{1}{3}\frac{\frac{4}{3}\tilde{\omega}}{b(1+\tilde{\omega})},\frac{4}{9b},
-\frac{1}{3}\frac{1+\tilde{\omega}}{b}\Big)$\
$ P_{46}$ & -& - & -& $
\Big( \,{\frac {\tilde{\omega}+3\,b+3\,-1/3}{3b}}, \newline
-\,{\frac { \left( -\frac{1}{3}+\tilde{\omega} \right)
\left(\tilde{\omega}+3\,b+-\frac{1}{3} \right) }{3b \left(
3\,b-\frac{1}{3} \right) } } ,\newline
{\frac
{\tilde{\omega} \left( -\frac{1}{3}+\tilde{\omega} \right) }{3b \left( 3\,b-\frac{1}{3} \right) }}\Big)$\
$ P_47$ & $-3\tilde{\omega}$& $ \frac {\sqrt {3}\sqrt {b
\left(-4/9+3\,b+\,b \right) }}{b }$ & $ -{\frac {\sqrt {3}\sqrt {b \left(-4/9+3\,b+\,b \right) }}{b }} $ & $ \ (0, \frac{4}{9b}, -\frac{1}{3b})$\
------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Model IV: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig8"}](4phase11.pdf "fig:"){width="8cm" height="6cm"} ![ Model IV: Phase space for $\tilde{\omega}=(-1,-1.2,-1.5),\, b=0.5,\,\omega_r=\frac{1}{3}$[]{data-label="fig8"}](4phase22.pdf "fig:"){width="8cm" height="6cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------
We remark for this model that $ P_{41}$ is stable for $\tilde{\omega}>-1$ , $ P_{42}$ is also stable for $\tilde{\omega}<-1$ while $ P_{43}$ and $ P_{47}$ are unstable. $ P_{44}$ is stable for $\tilde{\omega}<-1$ and $b>\frac{1}{9}$. It is also possible to determine the stability of point $ P_{45}$. The stabilities of the point $P_{47}$ and also $ P_{35}$ from the model III are not easy to be studied because the matrix of Jacobi in these cases is not diagonal and the behaviors of its eigenvalues are not trivial. This means that we can not theoretically know if these points are stable or not. Consequently these cases are analytically and numerically impossible to be studied. Therefore, the dynamic behaviors of the model IV behavior have been plotted in the figures \[fig7\] and \[fig8\] which show that the density of the dark energy quickly rise up to $N \sim 0,5$ and after decreases sharply (same behavior with quintessence) whereas the energy densities of radiation and matter decrease and tend to zero when $N \sim 1,8$.
Stability of $ T+ \mathcal{T}^{\mathcal{N}} $ model {#sec5}
====================================================
This section is devoted to the study of the stability of model $ f(T, \mathcal{T})= T+ \mathcal{T}^{\mathcal{N}} $ by using the power-law and the de Sitter solutions.
We are interested here to the perturbation of both geometric parts and matter of the generalized equations of motion. To do so, we have focused our attention on the Hubble parameter for geometric perturbation and energy density for ordinary primordial matter perturbation and we have followed the same way as it is done in [@antoniode; @diegoalvaro].
$$\begin{aligned}
H(t) = H_{b}(t)(1+\delta(t)),\quad \rho(t)= \rho_{b}(t) (1+\delta_{m}(t))\label{7}.\end{aligned}$$
$H_b(t)$ and $\rho_b(t)$ denote the Hubble parameter and the energy density of the ordinary matter of the background respectively. Taking into consideration the interaction term, the continuity equation of the ordinary matter becomes the following differential equation
$$\begin{aligned}
\dot{\rho_{b}}(t)+3H_{b}(t)\rho_{b}(t)(1+\omega+q)= 0\,, \label{8}\end{aligned}$$
whose resolution leads to: $$\begin{aligned}
\rho_{b}(t)=\rho_{0}\ e^{-3(1+\omega+q)\int {H_{b}(t)}dt},\label{9}\end{aligned}$$ where $\rho_{0}$ is an integration constant. In order to study the linear perturbation about $ H(t)$ and $ \rho(t)$, we develop $\mathcal{T}^{\mathcal{N}} $ in a series of $\mathcal{T}_{b} = \rho_{b}(1-3\omega)$ as: $$\begin{aligned}
f(\mathcal{T}) = f^{b}+f_{\mathcal{T}}^{b}(\mathcal{T}-\mathcal{T}_b) + O^{2}\label{10}, \end{aligned}$$ The function $\mathcal{T}^{\mathcal{N}} $ and its derivatives are computed at $\mathcal{T}=\mathcal{T}_{h}$. According to the Einstein-Hilbert term, the strangeness here is the effect coming from $ \mathcal{T}^{\mathcal{N}} $. By putting (\[7\]) into (\[10\]) in the the first generalized Friedmann equation; one gets $$3H^{2}= \rho-\frac{1}{2}\left(f+12 H^{2} f_{T}\right)+ 3f_{\mathcal{T}} \left(\frac{\rho + p}{3}\right),
\label{Friedmann1}$$ which gives after simplification $$\begin{aligned}
6H_{b}^{2}(t)\delta(t)= \big[\rho_{b}+\rho_{b}f_{\mathcal{T}}^{b}(\frac{3-\omega}{2})+
\rho_{b}^{2}(1-2\omega-3\omega^{2})f_{\mathcal{T}\mathcal{T}}^{b}\big]\delta_{m}(t).
\label{11}\end{aligned}$$ Considering that the ordinary matter is essentially the dust, we obtain the simple expression $$\begin{aligned}
6H_{b}^{2}(t)\delta(t)= \big[\rho_{b}+3\rho_{b}f_{\mathcal{T}}^{b}+
2\rho_{b}^{2}f_{\mathcal{T}\mathcal{T}}^{b}\big]\delta_{m}(t),\label{12}.\end{aligned}$$ For matter pertubation function, we have the following differential equation $$\begin{aligned}
\dot{\delta_{m}}(t)+3(1+\omega+q)H_{b}(t)\delta(t) = 0. \label{13}\end{aligned}$$ Eliminating $\delta(t)$ between (\[11\]) and (\[13\]), we obtain also the differential equation $$\begin{aligned}
2H_{b}\dot{\delta}_{m}(t)+(1+\omega+q)
\big[\rho_{b}+\rho_{b}f_{\mathcal{T}}^{b}\left(\frac{3-\omega}{2}\right)+
\rho_{b}^{2}(1-2\omega-3\omega^{2})f_{\mathcal{T}\mathcal{T}}^{b}\big]\delta_{m}(t)=0.\label{14}\end{aligned}$$ The direct resolution of this differential equation gives $$\begin{aligned}
\delta_{m}(t)= C_{0}\exp\left\{-\left(\frac{1+\omega+q}{2}\right)\int
\frac{\rho_{b}}{H_{b}}\left[1+f_{\mathcal{T}}^{b}\left(\frac{3-\omega}{2}\right)+\rho_{b}
(1-2\omega-3\omega^{2})f_{\mathcal{T}\mathcal{T}}^{b}\right]dt \right\} ,
\label{15} \end{aligned}$$ where $C_{0}$ is an integration constant. From Eq. (\[13\]) one can extract $$\begin{aligned}
\delta(t) = \frac{C_{0} C_{\mathcal{T}}}{6H_{b}}\exp\left\{ -\left(\frac{1+\omega+q}{2}\right)\int C_{\mathcal{T}}dt \right\} ,
\label{16}\end{aligned}$$ with $$\begin{aligned}
C_{\mathcal{T}}=\frac{\rho_{b}}{H_{b}}\left[1+f_{\mathcal{T}}^{b}
\left(\frac{3-\omega}{2}\right)+\rho_{b}\left(1-2\omega-3\omega^{2}\right)f_{\mathcal{T}\mathcal{T}}^{b}\right].\end{aligned}$$
Stability of de Sitter solutions {#sec5.1}
--------------------------------
In this case, the Hubble parameter is written as $$\begin{aligned}
H_{b}(t) = H_{0} \rightarrow a(t) = a_{0}e^{H_{0}t}.\end{aligned}$$ The expression (\[9\]) becomes, $$\begin{aligned}
\rho_{b}(t)=\rho_{0} e^{-3(1+\omega+q)H_{0}t}.
\label{17}\end{aligned}$$ From relation $ d{\rho_{b}}=-3(1+\omega+q)H_{0}\rho_{b}dt$ and within an elementary transformation, we get $$\begin{aligned}
\int C_{\mathcal{T}}dt &=&-\frac{1}{3H_{0}(1+\omega+q)}\int\frac{1}{\rho_{b}}C_{\mathcal{T}}d{\rho_{b}} \nonumber\\
&=&-\frac{1}{3H_{0}^{2}(1+q+\omega)}\Bigg\{ \rho_{b}
+Q\; \frac{(3-\omega)}{2} \rho_{b}^{\mathcal{N}}(1-3\omega)^{\mathcal{N}-1}
\nonumber\\
&+& Q\;(1-2\omega-3\omega^{2}) (\mathcal{N}-1)(1-3\omega)^{\mathcal{N}-2}\rho_{b}^{\mathcal{N}} \Bigg\}.\end{aligned}$$ By replacing this expression in \[15\], we obtains $$\begin{aligned}
\delta_{m}(t)= C_{0}\exp \left\{ \frac{1}{6H_{0}^{2}}\left[\rho_{b}+Q\; \frac{(3-\omega)}{2} \rho_{b}^{\mathcal{N}}(1-3\omega)^{\mathcal{N}-1}
+ Q\;(1-2\omega-3\omega^{2}) (\mathcal{\mathcal{N}}-1)(1-3\omega)^{\mathcal{N}-2}\rho_{b}^{\mathcal{N}}\right]\right\}.
\label{18}\end{aligned}$$ Therefore the perturbation function about the geometry can be obtained and given by $$\begin{aligned}
\delta(t) &=& \frac{C_{0} C_{\mathcal{T}}}{6H_{0}}\exp \left\{ \frac{1}{6H_{0}^{2}}\left[\rho_{b}+Q\; \frac{(3-\omega)}{2}
\rho_{b}^{\mathcal{N}}(1-3\omega)^{\mathcal{N}-1} +
Q\;(1-2\omega-3\omega^{2}) (\mathcal{N}-1)(1-3\omega)^{\mathcal{N}-2}\rho_{b}^{\mathcal{N}}\right]\right\},
\label{19}\end{aligned}$$ with $$\begin{aligned}
C_{\mathcal{T}}= \frac{1}{H_{0}}\left\{\rho_{b}+Q\;\mathcal{N} \frac{(3-\omega)}{2} \rho_{b}^{\mathcal{N}}(1-3\omega)^{\mathcal{N}-1} + Q\;\mathcal{N} (\mathcal{N}-1)(1-2\omega-3\omega^{2})(1-3\omega)^{\mathcal{N}-2}\rho_{b}^{\mathcal{N}}\right\}\end{aligned}$$ and
$$\begin{aligned}
f(\mathcal{T}) = Q\; \mathcal{T}^{\mathcal{N}}. \end{aligned}$$
with $Q$ the one defined in (\[cvalue\]) and $\mathcal{N} =-(1-3\omega)/((1+\omega))$. For some suitable values of the input parameters consistent with cosmological observational data, we plot the curve characterizing the behavior of the perturbation function at the left side in Fig \[fig2\]. We see that as the universe expands, i.e., increasing $Z$, the matter and geometric perturbations functions, $\delta_m$ and $\delta$ respectively, goes towards positive values more less than $0.1$ when the time evolves.
Stability of Power-Law solutions {#sec5.2}
--------------------------------
Here, the scale factor is written as $$\begin{aligned}
a(t)\propto t^{n} \quad \rightarrow H_{b}(t) = \frac{n}{t}\,,\end{aligned}$$ and the ordinary energy density (\[9\]) becomes $$\begin{aligned}
\rho_{b}=\rho_{0}t^{-3n(1+\omega+q)}\label{20}\end{aligned}$$ By making the substitution of $\rho_{b}$ in (\[14\]), one gets after resolution, the following expression $$\begin{aligned}
\delta_{m}(t)= C_{1}\exp\left\{ -A\left(\frac{A_{1}}{2+B}t^{2+B} +\frac{A_{2}}{2+\mathcal{N} B}t^{2+\mathcal{N} B}\right)\right\},\end{aligned}$$ with $C_{1}$ an integration constant, and $$\begin{aligned}
A = \frac{(1+\omega+q)}{2n}, \quad A_{1}= \rho_{0},\quad B = -3n(1+\omega+q)\nonumber\\ A_{2}=Q\;\mathcal{N} \rho_{0}^{\mathcal{N}}\left\{ \frac{(18\omega^{3}+9\omega^{2}-14\omega+3)}
{2(1-3\omega)^{2-\mathcal{N}}}+\frac{\mathcal{N}}{{(1-3\omega)}^{(1-\mathcal{N})}}\right\}.\end{aligned}$$ The use of the relation (\[13\] leads to $$\begin{aligned}
\delta(t)=\frac{C_{1}}{6n^{2}}\left( A_{1}t^{2+\mathcal{N}}+A_{2} t^{2+B \mathcal{N}}\right) \exp\left\{ -A\left(\frac{A_{1}}{2+B}t^{2+B} +\frac{A_{2}}{2+\mathcal{N} B}t^{2+\mathcal{N} B}\right)\right\}.
\end{aligned}$$ As we have done in the previous section, we present here the evolution of the perturbation functions in Fig \[fig21\] for suitable values of input parameters.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ The graph at the left side of the figure presents the evolution of the perturbation functions $\delta_m$ and $\delta$ within the de Sitter solutions, while the one at the right side shows the evolution of the perturbation functions within the power-law solutions. The graph are plotted for $n=2/3$, $\Lambda=1.7\times 10^{-121}$, $\rho_0=0.1\times 10^{-121}$, $\omega=0$ and $C_1=1$ .[]{data-label="fig21"}](sitter.pdf "fig:"){height="5cm" width="8.5cm"} \[e3\] ![ The graph at the left side of the figure presents the evolution of the perturbation functions $\delta_m$ and $\delta$ within the de Sitter solutions, while the one at the right side shows the evolution of the perturbation functions within the power-law solutions. The graph are plotted for $n=2/3$, $\Lambda=1.7\times 10^{-121}$, $\rho_0=0.1\times 10^{-121}$, $\omega=0$ and $C_1=1$ .[]{data-label="fig21"}](Power.pdf "fig:"){height="5cm" width="9cm"} \[e4\]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Conclusion {#sec6}
==========
We undertook in this work cosmological analysis about a model in the framework of the so-called $f(T, \mathcal{T})$ theory. In order to obtain a viable $f(T, \mathcal{T})$ model, we first impose the covariant conservation of the energy-momentum, from which, we get a model of the type $T+f(\mathcal{T})$, being a sort of trace depending function correction to the TT. The obtained model includes parameters depending on the cosmological constant $\Lambda$ and the parameter $\omega$ of the ordinary equation of state. These parameters play a main role in the whole study developed in this manuscript. By the way, we study the dynamics of the cosmological system, analyzing the stability about the critical points. We solve the equations and it appears that for some specific expressions of the interaction term one can obtain attractor solutions. We numerically integrate the equations and show that the evolution of the dark energy density mimics three diffract behaviors: phantom, quintessence and cosmological constant in some interactive forms. We argue that this interaction is purely phenomenological and is consistent with the observational data. Our result shows that for both de Sitter and power-law solutions, the perturbations functions converge traducing the stability of the model.
Moreover, the stability of the model is checked within the de Sitter and power-law solutions by performing linear perturbation about the physical critical points. We see that for the both considered solutions, the model presents stability through the convergence of the geometric and matter perturbation functions $\delta$ and $\delta_m$.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank IMSP for hospitality during the elaboration of this work.
[9]{}
C. W. Misner, K. S. Thorne, and J. H. Wheeler, [*Gravitation*]{}, W. H. Freeman and Company, 1973. Szekeres P: The Gravitational Compass, J. Maths. Phys. 6 (1965), 1387. J. L. Synge. On the Deviation of Geodesics and Null Geodesics, Particularly in Relation to the Properties of Spaces of Constant Curvature and Indefinite Line Element. Ann. Math. 35:705 (1934). F. A. E. Pirani. On the Physical Significance of the Riemann Tensor. Acta Phys. Polon. 15:389 (1956). G. F. R. Ellis and H. Van Elst. Deviation of geodesics in FLRW spacetime geometries (1997). Preprint in \[arXiv:gr-qc/9709060v1\]. S.L. Shapiro and S.A. Teukolsky, Black Holes, White Dwarfs and Neutron Satrs (Wile-Interscience, New York 1983). K.S. Thorne, in S. Hawking and W. Israel, eds, 300 Year of Gravitation (Cambridge University, Cambridge 1987) p. 330. Raychaudhuri A K: Relativistic Cosmology, Phys. Rev. 98 (1955), 1123. Mattig W: Uber den Zusammenhang zwischen Rotverschiebung und scheinbarer Helligkeit, Astr. Nach. 284 (1958), 109. Pirani F A E: On the Physical Significance of the Riemann Tensor, Acta Phys. Polon. 15 (1956),389. D. N. Spergel, et. al., Astrophys. J. Suppl. [**170**]{}, (2007) 377, arXiv:astro-ph/0603449. J. K. Adelman-McCarthy, et. al., Astrophys. J. Suppl. [**175**]{}, (2008) 297, arXiv:0707.3413\[astro-ph\]. R.Aldrovandi and J.G.Pereira,TELEPARALLEL GRAVITY,\
in http://www.ift.unesp.br/users/jpereira/tele.pdf. ; S. Nojiri and S. D. Odintsov, ECONF C [**0602061**]{}, 06 (2006); Int. J. Geom. Meth. Mod. Phys. [**4**]{}, 115-146 (2007) \[arXiv:hep-th/0601213\]; Phys. Rept. [**505**]{}, 59-144 (2011) \[arXiv:1011.0544\]. T. Harko, F. S. N. Lobo, S. Nojiri and S. D. Odintsov, ?f(R, T ) gravity,? Phys. Rev. D [**84**]{} (2011) 024020. \[arXiv:1104.2669 \[gr-qc\]\]. M. J. S. Houndjo, Int. J. Mod. Phys. D. [**21**]{}, 1250003 (2012). arXiv: 1107.3887 \[astro-ph.CO\]. M. J. S. Houndjo and O. F. Piattella, Int. J. Mod. Phys. D. [**21**]{}, 1250024 (2012). arXiv: 1111.4275 \[gr.qc\]. D. Momeni, M. Jamil and R. Myrzakulov, Euro. Phys. J. C [**72**]{}, arXiv: 1107.5807\[physics.gen-ph\]. M. J. S. Houndjo, C. E. M. Batista, J. P. Campos and O. F. Piattella,?? \[arXiv:1203.6084 \[gr-qc\]\]. F. G. Alvarenga, M. J. S. Houndjo, A. V. Monwanou and Jean. B. Chabi-Orou,?? arXiv: 1205.4678 \[gr-qc\]. S. ’i. Nojiri and S. D. Odintsov, “Modified Gauss-Bonnet theory as gravitational alternative for dark energy,” Phys. Lett. B [**631**]{}, 1 (2005) \[hep-th/0508049\]; S. Nojiri, S. D. Odintsov, A. Toporensky, P. Tretyakov, arXiv:0912.2488. K. Bamba, S. D. Odintsov, L. Sebastiani, S. Zerbini, arXiv:0911.4390. K. Bamba, C.-Q. Geng, S. Nojiri, S. D. Odintsov, arXiv:0909.4397. M.E. Rodrigues, M.J.S. Houndjo, D. Momeni, R. Myrzakulov, arXiv:1212.4488. M. J. S. Houndjo, M. E. Rodrigues, D. Momeni, R. Myrzakulov . arXiv:1301.4642 \[gr-qc\]. J. Amorós, J. de Haro and S. D. Odintsov, “Bouncing Loop Quantum Cosmology from $f(T)$ gravity,” Physical Review D 87, [**104037**]{} (2013) \[arXiv:1305.2344 \[gr-qc\]\]; K. Bamba, J. de Haro and S. D. Odintsov, “Future Singularities and Teleparallelism in Loop Quantum Cosmology,” JCAP [**1302**]{} (2013) 008 \[arXiv:1211.2968 \[gr-qc\]\]; K. Bamba, S. ’i. Nojiri and S. D. Odintsov, “Effective $f(T)$ gravity from the higher-dimensional Kaluza-Klein and Randall-Sundrum theories,” arXiv:1304.6191 \[gr-qc\]; G. R. Bengochea, R. Ferraro and , Phys. Rev. D [**79**]{}, 124019 (2009) \[arXiv:0812.1205 \[astro-ph\]\]. E. V. Linder, Phys.Rev. D [**81**]{}, 127301 (2010) \[Erratum-ibid. D 82, 109902 (2010)\] \[arXiv:1005.3039 \[astro-ph.CO\]\]. M. Jamil, D. Momeni and R. Myrzakulov, Eur. Phys. J. C [**72**]{} (2012) 2267 \[arXiv:1212.6017 \[gr-qc\]\]. R. Myrzakulov, Entropy [**14**]{} (2012) 1627\[arXiv:1212.2155 \[gr-qc\]\]. M. R. Setare and N. Mohammadipour, JCAP [**1211**]{} (2012) 030 \[arXiv:1211.1375 \[gr-qc\]\]. M.R. Setare, N. Mohammadipour, JCAP [**01**]{} (2013) 015 \[arXiv: 1301.4891\]. M. Jamil, D. Momeni, R. Myrzakulov and P. Rudra, J. Phys. Soc. Jap. [**81**]{} (2012) 114004 \[arXiv:1211.0018 \[physics.gen-ph\]\]. M. E. Rodrigues, M. J. S. Houndjo, D. Saez-Gomez and F. Rahaman, Phys. Rev. D [**86**]{} (2012) 104059 \[arXiv:1209.4859 \[gr-qc\]\]. M. Jamil, D. Momeni and R. Myrzakulov, Eur. Phys. J. C [**72**]{} (2012) 2122 \[arXiv:1209.1298 \[gr-qc\]\]. R. Myrzakulov, Eur. Phys. J. C [**72**]{} (2012)2203 \[arXiv:1207.1039 \[gr-qc\]\]. M. J. S. Houndjo, D. Momeni and R. Myrzakulov, Int. J. Mod. Phys. D [**21**]{} (2012) 1250093 \[arXiv:1206.3938 \[physics.gen-ph\]\]. M. E. Rodrigues, M. H. Daouda and M. J. S. Houndjo, arXiv:1205.0565 \[gr-qc\]. M. R. Setare and M. J. S. Houndjo, arXiv:1203.1315 \[gr-qc\]. K. Bamba, M. Jamil, D. Momeni and R. Myrzakulov, arXiv:1202.6114 \[physics.gen-ph\]. K. Bamba, R. Myrzakulov, S. ’i. Nojiri and S. D. Odintsov, Phys. Rev. D [**85**]{} (2012)104036 \[arXiv:1202.4057 \[gr-qc\]\]. M. Jamil, D. Momeni and R. Myrzakulov, Eur. Phys. J. C [**72**]{} (2012) 2267 \[arXiv:1212.6017\[gr-qc\]\]. M. Jamil, D. Momeni and R. Myrzakulov, Gen. Rel. Grav. [**45**]{} (2013) 263 \[arXiv:1211.3740 \[physics.gen-ph\]\]. M. Jamil, D. Momeni and R. Myrzakulov, Eur. Phys. J. C [**72**]{} (2012) 2122 \[arXiv:1209.1298 \[gr-qc\]\]. M. Jamil, D. Momeni and R. Myrzakulov, Eur. Phys. J. C [**72**]{} (2012) 2075 \[arXiv:1208.0025 \[gr-qc\]\]. M. Jamil, K. Yesmakhanova, D. Momeni and R. Myrzakulov, Central Eur. J. Phys. [**10**]{} (2012) 1065 \[arXiv:1207.2735 \[gr-qc\]\]. M. J. S. Houndjo, D. Momeni and R. Myrzakulov, Int. J. Mod. Phys. D [**21**]{} (2012) 1250093 \[arXiv:1206.3938 \[physics.gen-ph\]\]. M. Jamil, D. Momeni and R. Myrzakulov, Eur. Phys. J. C [**72**]{} (2012) 1959 \[arXiv:1202.4926 \[physics.gen-ph\]\]. M. H. Daouda, M. E. Rodrigues and M. J. S. Houndjo, Phys. Lett. B [**715**]{} (2012) 241 \[arXiv:1202.1147 \[gr-qc\]\]. M. Jamil, S. Ali, D. Momeni, R. Myrzakulov and Eur. Phys. J. C [**72**]{}, 1998 (2012) \[arXiv:1201.0895 \[physics.gen-ph\]\]. M. Jamil, D. Momeni, N. S. Serikbayev, R. Myrzakulov and , Astrophys. Space Sci.[ **339**]{}, 37 (2012) \[arXiv:1112.4472 \[physics.gen-ph\]\]. M. Jamil, D. Momeni, M. A. Rashid and , Eur. Phys. J. C [**71**]{}, 1711 (2011) \[arXiv:1107.1558 \[physics.gen-ph\]\]. M. Hamani Daouda, M. E. Rodrigues and M. J. S. Houndjo, Eur. Phys. J. C [**72**]{} (2012) 1893 \[arXiv:1111.6575 \[gr-qc\]\]. M. Hamani Daouda, M. E. Rodrigues and M. J. S. Houndjo, Eur. Phys. J. C [**72**]{} (2012) 1890 \[arXiv:1109.0528 \[physics.gen-ph\]\]. R. Myrzakulov, Gen. Rel. Grav. [**44**]{} (2012) 3059 \[arXiv:1008.4486 \[physics.gen-ph\]\]. K. K. Yerzhanov, S. .R. Myrzakul, I. I. Kulnazarov and R. Myrzakulov, arXiv:1006.3879 \[gr-qc\]. R. Myrzakulov, Eur. Phys. J. C [**71**]{} (2011) 1752 \[arXiv:1006.1120 \[gr-qc\]\]. M. E. Rodrigues, M. J. S. Houndjo, D. Momeni, R. Myrzakulov and , arXiv:1302.4372 \[physics.gen-ph\]. J. M. Bardeen, B. Carter, S. W. Hawking , Commun. Math. Phys. [**31**]{} (1973) 161-170. N. Tamanini and C. G. Boehmer, Phys. Rev. D [**86**]{}, 044009 (2012), arXiv:1204.4593 \[gr-qc\]. Baojiu Li, T. P. Sotiriou and J. D. Barrow, Phys. Rev. D [**83**]{}, 064035 (2011); Phys. Rev. D [**83**]{}, 104030 (2011). M. J. S. Houndjo, D. Momeni, R. Myrzakulov and M. E. Rodrigues,arXiv:1304.1147. C. Deliduman and B. Yapiskan, arXiv:1103.2225v3 \[gr-qc\]. M. Hamani Daouda, M. E. Rodrigues and M. J. S. Houndjo, Eur. Phys. J. C [**71**]{} (2011) 1817 \[arXiv:1108.2920 \[astro-ph.CO\]\]. I.G.Salako, M.E.Rodrigues, A.V.Kpadonou, M. J.S.Houndjo and J.Tossa JCAP [ **060**]{}, 1475-7516 (2013) M. E. Rodrigues, I. G. Salako, M. J. S. Houndjo, J. Tossa Int. J. Mod. Phys. D [ **23**]{}, 1450004 (2014) Davood Momeni, Ratbay Myrzakulov.: arXiv:1405.5863 \[gr-qc\] Tiberiu Harko, Francisco S. N. Lobo, G. Otalora,and Emmanuel N. Saridakis http://arxiv.org/abs/1405.0519v1 Ines G. Salako, Abdul Jawad, Surajit Chattopadhyay, S. B. Nassur, M. J. S. Houndjo, A. V. Kpadonou, M. E. Rodrigues, J. Tossa, http://arxiv.org/abs/1506.09161 F. G. Alvarenga, A. de la Cruz-Dombriz, M. J. S. Houndjo, M. E. Rodrigues, D. Sáez-Gómez, Phys. Rev. D [**87**]{}, 103526 (2013). arXiv:1302.1866 \[gr-qc\]. M. Jamil, D. Momeni, M. A. Rashid, Eur. Phys. J. C 71, 1711 (2011); S. Chen, J. Jing, Class. Quan. Grav. 26, 155006 (2009). A. De Felice and S. Tsujikawa, Phys. Lett. B [**675**]{}, 1-8 (2009); arXiv: 0810.5712 \[hep-th\]. A. de la Cruz-Dombriz and D. Sáez-Gómez, Class. Quantum Grav [**29**]{}, 245014 (2012), arXiv: 1112.4481 \[gr-qc\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The ability to detect and count certain substructures in graphs is important for solving many tasks on graph-structured data, especially in the contexts of computational chemistry and biology as well as social network analysis. Inspired by this, we propose to study the expressive power of graph neural networks (GNNs) via their ability to count attributed graph substructures, extending recent works that examine their power in graph isomorphism testing and function approximation. We distinguish between two types of substructure counting: matching-count and containment-count, and establish both positive and negative answers for popular GNN architectures. Specifically, we prove that Message Passing Neural Networks (MPNNs), $2$-Weisfeiler-Lehman ($2$-WL) and $2$-Invariant Graph Networks ($2$-IGNs) cannot perform matching-count of substructures consisting of 3 or more nodes, while they can perform containment-count of star-shaped substructures. We also prove positive results for $k$-WL and $k$-IGNs as well as negative results for $k$-WL with limited number of iterations. We then conduct experiments that support the theoretical results for MPNNs and $2$-IGNs, and demonstrate that local relational pooling strategies inspired by @murphy2019relational are more effective for substructure counting. In addition, as an intermediary step, we prove that $2$-WL and $2$-IGNs are equivalent in distinguishing non-isomorphic graphs, partly answering an open problem raised in @maron2019open.'
author:
- Zhengdao Chen
- Lei Chen
- Soledad Villar
- Joan Bruna
bibliography:
- 'ref.bib'
title: 'Can graph neural networks count substructures?'
---
Introduction
============
In recent years, graph neural networks (GNNs) have achieved empirical success on processing data from various fields such as social networks, quantum chemistry, particle physics, knowledge graphs and combinatorial optimization [@scarselli2008graph; @bruna2013spectral; @duvenaud2015convolutional; @kipf2016semi; @defferrard2016convolutional; @bronstein2017geometric; @dai2017combopt; @nowak2017note; @ying2018diffpool; @zhou2018graph; @choma2018graph; @zhang2018link; @you2018gcpn; @you2018graphrnn; @you2019g2sat; @yao2019experimental; @ding2019cognitive; @stokes2020deep]. Thanks to such progress, there have been growings interest in studying the expressive power of GNNs. One line of work does so by studying their ability to distinguish non-isomorphic graphs. In this regard, @xu2018powerful and @morris2019higher show that GNNs based on neighborhood-aggregation schemes are at most as powerful as the classical Weisfeiler-Lehman (WL) test [@weisfeiler1968reduction] and propose GNN architectures that can achieve such level of power. While graph isomorphism testing is very interesting from a theoretical viewpoint, one may naturally wonder how relevant it is to real-world tasks on graph-structured data. Moreover, WL is powerful enough to distinguish almost all pairs of non-isomorphic graphs except for rare counterexamples [@babai1980random]. Hence, from the viewpoint of graph isomorphism testing, existing GNNs are in some sense already not far from being maximally powerful, which could make the pursuit of more powerful GNNs appear unnecessary.\
Another perspective is the ability of GNNs to approximate permutation-invariant functions on graphs. For instance, @maron2019universality and @keriven2019universal propose architectures that achieve universal approximation of permutation-invariant functions on graphs, though such models involve tensors with order growing in the size of the graph and are therefore impractical. Importantly, @chen2019equivalence establishes an equivalence between the ability to distinguish any pair of non-isomorphic graphs and the ability to approximate arbitrary permutation-invariant functions on graphs. Nonetheless, for GNNs used in practice, which are not universally approximating, more efforts are needed to characterize what they can and cannot do. For example, @loukas2019graph shows that GNNs under assumptions are Turing universal but loses power when its depth and width are limited, though the arguments rely on the nodes all having distinct features and the focus is on the asymptotic depth-width tradeoff. Concurrently to our work, @garg2020generalization provide impossibility results of several classes of GNNs to decide graph properties including girth, circumference, diameter, radius, conjoint cycle, total number of cycles, and $k$-cliques. Despite these interesting results, we still need a perspective for understanding the expressive power of different classes of GNNs in a way that is intuitive, relevant to goals in practice, and potentially helpful in guiding the search for more powerful architectures.\
Inspired by the relevance of detecting and counting *graph substructures* in applications, we propose to understand the power of GNN architectures via the substructures that they can and cannot count. Also referred to by various names including *graphlets*, *motifs*, *subgraphs* and *graph fragments*, graph substructures are well-studied and relevant for graph-related tasks in computational chemistry [@deshpande2002automated; @murray2009rise; @duvenaud2015convolutional; @jin2018junction; @jin2019hierarchical; @jin2020composing], computational biology [@koyuturk2004biological] and social network studies [@jiang2010finding]. In organic chemistry, for example, certain patterns of atoms called functional groups are usually considered indicative of the molecules’ properties [@lemke2003review; @pope2018discovering]. In the literature of molecular chemistry, substructure counts have been used to generate molecular fingerprints [@morgan1965generation; @oboyle2016comparing] and compute similarities between molecules [@alon2008biomolecular; @rahman2009small]. In addition, for general graphs, substructure counts have been used to create graph kernels [@shervashidze2009efficient] and compute spectral information [@preciado2010local]. The connection between GNNs and graph substructures is explored empirically by @ying2019gnn as a way to interpret the predictions made by GNNs. Thus, the ability of different GNN architectures to count graph substructures not only serves as an intuitive theoretical measure of their expressive power but also is highly relevant to real-world scenarios. While people have proposed variants of GNNs that take advantage of substructure information [@monti2018motifnet; @liu2018ngram; @liu2019neuralsubgraph], often they rely on handcrafting rather than learning such information. More importantly, there is a lack of a systematic theoretical study of the ability of existing GNNs to count substructures.\
In this work, we first build a theoretical framework for studying the ability of GNNs to count *attributed* substructures based on both function approximation and graph discrimination. In particular, we distinguish between *containment-count* and *matching-count*, each corresponding to having *subgraphs* and *induced subgraphs* isomorphic to a given pattern, respectively. Next, we look at classical GNN architectures and prove the following results.
1. Focusing on matching-count, we establish that neither Message Passing Neural Networks (MPNNs) [@gilmer2017neural] nor $2$nd-order Invariant Graph Networks ($2$-IGNs) [@maron2019universality] can count any connected substructure of $3$ or more nodes. For any such pattern, we prove this by constructing a pair of graphs that provably cannot be distinguished by any MPNN or $2$-IGN but with different matching-counts of the given pattern. This result points at an important class of simple-looking tasks that are provably hard for classical GNN architectures.
2. We show positive results for containment-count of star-shaped patterns by MPNNs and $2$-IGNs, generalizing results in @arvind2018weisfeiler, as well as for both matching- and containment-count of size-$k$ patterns by $k$-WL and $k$-IGNs. The latter result hints at a hierarchy of the increasing power of $k$-WL’s in terms of counting substructures, which would be more intuitive than the hierarchy in terms distinguishing non-isomorphic graphs as shown in @cai1992optimal, and therefore concretely motivates the search for GNNs with higher expressive power than $2$-WL or MPNN.
3. While a tight negative result for general $k$-WL is difficult to obtain, we show that $T$ iterations of $k$-WL is unable to perform matching-count for the path pattern of $(k+1)2^T$ or more nodes. It is relevant since real-life GNNs are often shallow, and also demonstrates an interplay between $k$ and depth.
We complement these theoretical results with synthetic experiments of counting triangles and stars in random graphs. In addition, while our negative theoretical results are worst-case in nature, the experiments illustrate an average-case difficulty for classical GNNs to count even the simplest graph substructures such as triangles. On the other hand, instead of performing iterative equivariant aggregations of information as is done in MPNNs and IGNs, we propose a type of locally powerful models based on the observation that substructures present themselves in local neighborhoods known as *egonets*. One idea is to apply the Relational Pooling approach [@murphy2019relational] to egonets, resulting in a model we call *Local Relational Pooling*. We demonstrate that it can perform both matching- and containment-count in the experiments.
Framework
=========
Attributed graphs, (induced) subgraphs and two types of counting
----------------------------------------------------------------
An *unattributed graph* $G$ with $n$ nodes is usually denoted by $G = (V, E)$, where typically $V = [n] := \{1, ..., n\}$ is the vertex set and $E \subset V^2 := V \times V$ is the edge set. We define an *attributed graph* or *weighted graph* as $G = (V, E, x, e)$, where in addition to $V$ and $E$, we let $x_i \in \mathcal{X}$ represent the node feature (or node attribute) of node $i$, and $e_{i, j} \in \mathcal{Y}$ represent the edge feature of edge $(i, j)$ if $(i, j) \in E$. For simplicity, we only consider undirected graphs (i.e. if $(i, j) \in E$ then $(j, i) \in E$ and $e_{i, j} = e_{j, i}$), and we do not allow self-connections (i.e., $(i, i) \notin E$) or multi-edges (so that $E$ is a well-defined set). Note that an unattributed graph can be viewed as an attributed graph with identical node and edge features. If a graph has only node features and no edge features, we can also represent it as $G = (V, E, x)$.\
Unlike the node and edge features, the indices of the nodes are not inherent properties of the graph. Rather, different ways of ordering the nodes result in different representations of the same underlying graph. This is characterized by the definition of *graph isomorphism*: Two attributed graphs $G^{[\mathtt{1}]}=(V^{[\mathtt{1}]}, E^{[\mathtt{1}]}, x^{[\mathtt{1}]}, e^{[\mathtt{1}]})$ and $G^{[\mathtt{2}]}=(V^{[\mathtt{2}]},E^{[\mathtt{2}]},x^{[\mathtt{2}]}, e^{[\mathtt{2}]})$ are *isomorphic* if there exists a bijection $\pi:V^{[\mathtt{1}]}\to V^{[\mathtt{2}]}$ such that (1) $(i,j)\in E^{[\mathtt{1}]}$ if and only if $(\pi(i), \pi(j))\in E^{[\mathtt{2}]}$, (2) $x^{[\mathtt{1}]}_i=x^{[\mathtt{2}]}_{\pi(i)}$ for all $i$ in $V^{[\mathtt{1}]}$, and (3) $e^{[\mathtt{1}]}_{i, j}=e^{[\mathtt{2}]}_{\pi(i), \pi(j)}$ for all $(i,j)\in E^{[\mathtt{1}]}$.\
Before defining substructure counting, we first need to define *subgraphs* and *induced subgraphs*. For $G = (V, E, x, e)$, a *subgraph* of $G$ is any graph $G^{[\mathtt{S}]} = (V^{[\mathtt{S}]}, E^{[\mathtt{S}]}, x, e)$ with $V^{[\mathtt{S}]} \subseteq V$ and $E^{[\mathtt{S}]} \subseteq E$. An *induced subgraphs* of $G$ is any graph $G^{[\mathtt{S'}]} = (V^{[\mathtt{S'}]}, E^{[\mathtt{S'}]}, x, e)$ with $V^{[\mathtt{S'}]} \subseteq V$ and $E^{[\mathtt{S'}]} = E \cap (V^{[\mathtt{S'}]})^2$. In words, the edge set of an induced subgraph needs to include all edges in $E$ that have both end points belonging to $V^{[\mathtt{S'}]}$. Thus, an induced subgraph of $G$ is also its subgraph, but the converse is not true.\
We now define two types of counting attributed substructures: *matching* and *containment*, illustrated in Figure \[fig:motif\]. Let $G^{[\mathtt{P}]} = (V^{[\mathtt{P}]}, E^{[\mathtt{P}]}, x^{[\mathtt{P}]}, e^{[\mathtt{P}]})$ be a (typically smaller) graph that we refer to as a *pattern* or *substructure*. We define $\mathsf{C}(G, G^{[\mathtt{P}]})$, called the *containment-count* of $G^{[\mathtt{P}]}$ in $G$, to be the number of *subgraphs* of $G$ that are isomorphic to $G^{[\mathtt{P}]}$. We define $\mathsf{M}(G; G^{[\mathtt{P}]})$, called the *matching-count* of $G^{[\mathtt{P}]}$ in $G$, to be the number of *induced subgraphs* of $G$ that are isomorphic to $G^{[\mathtt{P}]}$. Since all induced subgraphs are subgraphs, we always have $\mathsf{M}(G; G^{[\mathtt{P}]})\leq \mathsf{C}(G; G^{[\mathtt{P}]})$.\
Moreover, on a space of graphs $\mathcal{G}$, we call $\mathsf{M}(\cdot; G^{[\mathtt{P}]})$ the *matching-count function* of the pattern $G^{[\mathtt{P}]}$, and $\mathsf{C}(\cdot; G^{[\mathtt{P}]})$ the *containment-count function* of $G^{[\mathtt{P}]}$. To formalize the probe into whether certain GNN architectures can count different substructures, a natural question to study is whether they are able to approximate the matching-count and the containment-count functions arbitrarily well. Formally, given a target function $g: \mathcal{G} \to \mathbb{R}$, and family of functions, $\mathcal{F}$, which in our case is typically the family of functions that a GNN architecture can represent, we say $\mathcal{F}$ is able to approximate $g$ on $\mathcal{G}$ if for all $ \epsilon > 0$ there exists $f \in \mathcal{F}$ such that $|g(G) - f(G)| < \epsilon$, for all $G \in \mathcal{G}$.\
However, such criterion based on function approximation is hard to work with directly when we look at concrete examples later on. For this reason, below we will look for an alternative and equivalent definition from the perspective of graph discrimination.
![Illustration of the two types of counts of the pattern $G^{[\mathtt{P}]}$ in graphs $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$. The edge and node features are represented by colors. For $G^{[\mathtt{1}]}$, the matching-count $\mathsf{M}(G^{[\mathtt{1}]}; G^{[\mathtt{P}]})=0$ but the containment-count $\mathsf{C}(G^{[\mathtt{1}]}; G^{[\mathtt{P}]})=1$. For $G^{[\mathtt{2}]}$, $\mathsf{M}(G^{[\mathtt{2}]}; G^{[\mathtt{P}]})=\mathsf{C}(G^{[\mathtt{2}]}; G^{[\mathtt{P}]})=0$ since the edge features do not match. []{data-label="fig:motif"}](motif "fig:"){width="25.00000%"} ![Illustration of the two types of counts of the pattern $G^{[\mathtt{P}]}$ in graphs $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$. The edge and node features are represented by colors. For $G^{[\mathtt{1}]}$, the matching-count $\mathsf{M}(G^{[\mathtt{1}]}; G^{[\mathtt{P}]})=0$ but the containment-count $\mathsf{C}(G^{[\mathtt{1}]}; G^{[\mathtt{P}]})=1$. For $G^{[\mathtt{2}]}$, $\mathsf{M}(G^{[\mathtt{2}]}; G^{[\mathtt{P}]})=\mathsf{C}(G^{[\mathtt{2}]}; G^{[\mathtt{P}]})=0$ since the edge features do not match. []{data-label="fig:motif"}](motif3.pdf "fig:"){width="25.00000%"} ![Illustration of the two types of counts of the pattern $G^{[\mathtt{P}]}$ in graphs $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$. The edge and node features are represented by colors. For $G^{[\mathtt{1}]}$, the matching-count $\mathsf{M}(G^{[\mathtt{1}]}; G^{[\mathtt{P}]})=0$ but the containment-count $\mathsf{C}(G^{[\mathtt{1}]}; G^{[\mathtt{P}]})=1$. For $G^{[\mathtt{2}]}$, $\mathsf{M}(G^{[\mathtt{2}]}; G^{[\mathtt{P}]})=\mathsf{C}(G^{[\mathtt{2}]}; G^{[\mathtt{P}]})=0$ since the edge features do not match. []{data-label="fig:motif"}](motif2.pdf "fig:"){width="25.00000%"}
$G^{[\mathtt{P}]}$ $G^{[\mathtt{1}]}$ $G^{[\mathtt{2}]}$
From function approximation to graph discrimination
---------------------------------------------------
Say $\mathcal{G}$ is a space of graphs, and $\mathcal{F}$ is a family of functions from $\mathcal{G}$ to $\mathbb{R}$. Given two graphs $G^{[\mathtt{1}]}, G^{[\mathtt{2}]} \in \mathcal{G}$, we say $\mathcal{F}$ is able to distinguish them if there exists $ f \in \mathcal{F}$ such that $f(G^{[\mathtt{1}]}) \neq f(G^{[\mathtt{2}]})$. Such a perspective has been explored in @chen2019equivalence, for instance, to build an equivalence between function approximation and graph isomorphism testing by GNNs. In the context of substructure counting, it is clear that the ability to approximate the count functions entails the ability to distinguish graphs in the following sense:
If $\mathcal{F}$ is able to approximate the matching-count (or containment-count) function of a pattern $G^{[\mathtt{P}]}$ on the space $\mathcal{G}$, then for all $G^{[\mathtt{1}]}, G^{[\mathtt{2}]} \in \mathcal{G}$ such that $\mathsf{M}(G^{[\mathtt{1}]}, G^{[\mathtt{P}]}) \neq \mathsf{M}(G^{[\mathtt{2}]}, G^{[\mathtt{P}]})$ (or $\mathsf{C}(G^{[\mathtt{1}]}, G^{[\mathtt{P}]}) \neq \mathsf{C}(G^{[\mathtt{2}]}, G^{[\mathtt{P}]})$), they can be distinguished by $\mathcal{F}$.
What about the converse? When the space $\mathcal{G}$ is finite, such as if the graphs have bounded numbers of nodes and the node as well as edge features belong to finite alphabets, we can show a slightly weaker statement than the exact converse. Following @chen2019equivalence, we define an augmentation of families of functions using feed-forward neural networks as follows:
Given $\mathcal{F}$, a family of functions from a space $\mathcal{X}$ to $\mathbb{R}$, we consider an augmented family of functions also from $\mathcal{X}$ to $\mathbb{R}$ consisting of all functions of the following form $$x \mapsto h_{\mathcal{NN}}([f_1(x), ..., f_d(x)]),$$ where $d \in \mathbb{N}$, $h_1, ..., h_d \in \mathcal{F}$, and $h_\mathcal{NN}$ is a feed-forward neural network / multi-layer perceptron. When $\mathcal{NN}$ is restricted to have $L$ layers at most, we denote this augmented family by $\mathcal{F}^{+L}$.
Suppose $\mathcal{X}$ is a finite space, $g$ is a finite function on $\mathcal{X}$, and $\mathcal{F}$ is a family of functions on $\mathcal{X}$. Then, $\mathcal{F}^{+1}$ is able to approximate $f$ on $\mathcal{G}$ if $\forall x_1, x_2 \in \mathcal{X}$ with $g(x_1) \neq g(x_2)$, $\exists f \in \mathcal{F}$ such that $f(x_1) \neq f(x_2)$.
Since $\mathcal{X}$ is a finite space, for some large enough integer $d$, $\exists$ a collection of $d$ functions, $f_1, ..., f_d \in \mathcal{F}$ such that, if we define the function $\mathbf{f}(x) = (f_1(x), ..., f_d(x)) \in \mathbb{R}^d$, then it holds that $\forall x_1, x_2 \in \mathcal{X}, \mathbf{f}(x_1) = \mathbf{f}(x_2) \Rightarrow g(x_1) = g(x_2)$. (In fact, we can choose $d \leq \frac{|\mathcal{X}| \cdot (|\mathcal{X}|-1)}{2}$, since in the worst case we need one $f_i$ per pair of $x_1, x_2 \in \mathcal{X}$ with $x_1 \neq x_2$.) Then, $\exists$ a well-defined function $h$ from $\mathbb{R}^d$ to $\mathbb{R}$ such that $\forall x \in \mathcal{X}, g(x) = h(\mathbf{f}(x))$. By the universal approximation power of neural networks, $h$ can then be approximated arbitrarily well by some neural network $h_{\mathcal{NN}}$.
Thus, in the context of substructure counting, we have the following observation.
Suppose $\mathcal{G}$ is a finite space. If $\forall G^{[\mathtt{1}]}, G^{[\mathtt{2}]} \in \mathcal{G}$ with $\mathsf{M}(G^{[\mathtt{1}]}, G^{[\mathtt{P}]}) \neq \mathsf{M}(G^{[\mathtt{2}]}, G^{[\mathtt{P}]})$ (or $\mathsf{C}(G^{[\mathtt{1}]}, G^{[\mathtt{P}]}) \neq \mathsf{C}(G^{[\mathtt{2}]}, G^{[\mathtt{P}]})$), $\mathcal{F}$ is able to distinguish $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$, then $\mathcal{F}^{+1}$ is able to approximate the matching-count (or containment-count) function of the pattern $G^{[\mathtt{P}]}$ on $\mathcal{G}$.
For many GNN families, $\mathcal{F}^{+1}$ in fact has the same expressive power as $\mathcal{F}$. For example, consider $\mathcal{F}_{\textsc{MPNN}}$, the family of all Message Passing Neural Networks on $\mathcal{G}$. $\mathcal{F}_{\textsc{MPNN}}^{+1}$ consists of functions that run several MPNNs on the input graph in parallel and stack their outputs to pass through an MLP. However, running several MPNNs in parallel is equivalent to running one MPNN with larger dimensions of hidden states and messages, and moreover the additional MLP at the end can be merged into the readout function. Similar holds for the family of all $k$-Invariant Graph Functions ($k$-IGNs). Hence, for such GNN families, we have an exact equivalence on finite graph spaces $\mathcal{G}$.\
Therefore, we define substructure counting alternatively as follows, which are equivalent thanks to the results above and easier to work with when we study particular GNN architectures:
We say $\mathcal{F}$ is able to *perform matching-count (or containment-count)* of a pattern $G^{[\mathtt{P}]}$ on $\mathcal{G}$ if $\forall G^{[\mathtt{1}]}, G^{[\mathtt{2}]} \in \mathcal{G}$ such that $\mathsf{M}(G^{[\mathtt{1}]}, G^{[\mathtt{P}]}) \neq \mathsf{M}(G^{[\mathtt{2}]}, G^{[\mathtt{P}]})$ (or $\mathsf{C}(G^{[\mathtt{1}]}, G^{[\mathtt{P}]}) \neq \mathsf{C}(G^{[\mathtt{2}]}, G^{[\mathtt{P}]})$), $\mathcal{F}$ is able to distinguish $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$.
Another benefit of this definition is that it naturally allows us to also define the ability of graph isomorphism tests to count substructures. A graph isomorphism test, such as the Weisfeiler-Lehman (WL) test, takes as input a pair of graphs and returns whether or not they are believed to be isomorphic. Typically, the test will return true if the two graphs are indeed isomorphic but does not necessarily return false for every pair of non-isomorphic graphs. Given such a graph isomorphism test, we say it is able to perform matching-count (or containment-count) of a pattern $G^{[\mathtt{P}]}$ on $\mathcal{G}$ if $\forall G^{[\mathtt{1}]}, G^{[\mathtt{2}]} \in \mathcal{G}$ such that $\mathsf{M}(G^{[\mathtt{1}]}, G^{[\mathtt{P}]}) \neq \mathsf{M}(G^{[\mathtt{2}]}, G^{[\mathtt{P}]})$ (or $\mathsf{C}(G^{[\mathtt{1}]}, G^{[\mathtt{P}]}) \neq \mathsf{C}(G^{[\mathtt{2}]}, G^{[\mathtt{P}]})$), the test can tell these two graphs apart.\
Additional notations used in the proofs are given in Appendix \[app.notations\].
Message Passing Neural Networks and $k$-Weisfeiler-Lehman tests
===============================================================
Message Passing Neural Network (MPNN) is a generic model that incorporates many popular architectures, and it is based on learning local aggregations of information in the graph [@gilmer2017neural]. When applied to an undirected graph $G = (V, E, x, e)$, an MPNN with $T$ layers is defined iteratively as follows. For $t < T$, to compute the message $m_i^{(t+1)}$ and the hidden state $h_i^{(t+1)}$ for each node $i \in V$ at the $(t+1)$th layer, we apply the following update rule: $$\begin{split}
m_i^{(t+1)} &= \sum_{\mathcal{N}(i)} M_t(h_i^{(t)}, h_j^{(t)}, e_{i, j}) \\
h_i^{(t+1)} &= U_t(h_i^{(t)}, m_i^{(t+1)})
\end{split}$$ where $\mathcal{N}(i)$ is the neighborhood of node $i$ in $G$, $M_t$ is the message function at layer $t$ and $U_t$ is the vertex update function at layer $t$. Finally, a graph-level prediction is computed as $$\hat{y} = R(\{ h_i^{(T)}: i \in V \}),$$ where $R$ is the readout function. Typically, the hidden states at the first layer are set as $h_i^{(0)} = x_i$. Learnable parameters can appear in the functions $M_t$, $U_t$ (for all $t \leq T$) and $R$.\
@xu2018powerful and @morris2019higher show that, when the graphs’ edges are unweighted, such models are at most as powerful as the Weisfeiler-Lehman (WL) test in distinguishing non-isomorphic graphs. We will prove an extension of this result that incorporates edge features, which MPNNs naturally accommodate, so that by examining the ability of $2$-WL to count substructures, we can draw conclusions for MPNNs. Before that, we will first introduce the hierarchy of $k$-WL tests.
The hierarchy of $k$-Weisfeiler-Lehman ($k$-WL) tests {#sec.kwl}
-----------------------------------------------------
We will introduce the general $k$-WL test for $k \in \mathbb{N}^*$ applied to a pair of graphs, $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$. Assume that the two graphs have the same number of vertices, since otherwise they can be told apart easily. Without loss of generality, we assume that they share the same set of vertex indices, $V$ (but can differ in $E$, $x$ or $e$). For each of the graphs, at iteration $0$, the test assigns an initial color in some color space to every $k$-tuple in $V^k$ according to its isomorphism type[^1], and then updates the coloring in every iteration. For any $k$-tuple $s = (i_1, ..., i_k) \in V^k$, we let $\cb_k^{(t)}(s)$ denote the color of $s$ in $G^{[\mathtt{1}]}$ assigned at $t$th iteration, and let ${\cb'}_k^{(t)}(s)$ denote the color it receives in $G^{[\mathtt{2}]}$. $\cb_k^{(t)}(s)$ and ${\cb'}_k^{(t)}(s)$ are updated iteratively as follows. For each $w \in [k]$, define the neighborhood $$N_w(s) = \{ (i_1, ..., i_{w-1}, j, i_{j+1}, ..., i_k): j \in V \}$$ Given $\cb_k^{(t-1)}$ and ${\cb'}_k^{(t-1)}$, define $$\begin{split}
C_w^{(t)}(s) &= \textsc{Hash}_{t, 1} \Big( \boldsymbol{\{} \cb_k^{(t-1)}(\tilde{s}): \tilde{s} \in N_w(s) \boldsymbol{\}} \Big) \\
{C'}_w^{(t)}(s) &= \textsc{Hash}_{t, 1} \Big( \boldsymbol{\{} {\cb'}_k^{(t-1)}(\tilde{s}): \tilde{s} \in N_w(s) \boldsymbol{\}} \Big)
\end{split}$$ with “$\boldsymbol{\{} \boldsymbol{\}}$” representing a multiset, and $\textsc{Hash}_{t, 1}$ being some hash function that maps injectively from the space of multisets of colors to some intermediate space. Then let $$\begin{split}
\cb_k^{(t)}(s) &= \textsc{Hash}_{t, 2} \bigg( \bigg(\cb_k^{(t-1)}(s), \Big( C_1^{(t)}(s), ..., C_k^{(t)}(s) \Big) \bigg) \bigg) \\
{\cb'}_k^{(t)}(s) &= \textsc{Hash}_{t, 2} \bigg( \bigg({\cb'}_k^{(t-1)}(s), \Big( {C'}_1^{(t)}(s), ..., {C'}_k^{(t)}(s) \Big) \bigg) \bigg)
\end{split}$$
where $\textsc{Hash}_{t, 2}$ maps injectively from its input space to the space of colors. The test will terminate and return the result that the two graphs are not isomorphic if at some iteration $t$, the following two multisets differ: $$\boldsymbol{\{} \cb_k^{(t)}(s): s \in V^k \boldsymbol{\}} \neq \boldsymbol{\{} {\cb'}_k^{(t)}(s): s \in V^k \boldsymbol{\}}$$
#### Some properties of $k$-WL
For graphs with unweighted edges, $1$-WL and $2$-WL are known to have the same discriminative power [@maron2019provably]. For $k \geq 2$, it is known that $(k+1)$-WL is strictly more powerful than $k$-WL, in the sense that there exist pairs of graph distinguishable by the former but not the latter [@cai1992optimal]. Thus, with growing $k$, the set of $k$-WL tests forms a hierarchy with increasing discriminative power. Note that there has been an different definition of WL in the literature, sometimes known as *Folklore Weisfeiler-Lehman* (FWL), with different properties [@maron2019provably; @morris2019higher]. When people use the term “Weisfeiler-Lehman test” without specifying “$k$”, it usually refers to $1$-WL, $2$-WL or $1$-FWL.\
Extending the aforementioned results by @xu2018powerful [@morris2019higher] in a nontrivial way to incorporate edge features, we present the following theorem, to be proved in Appendix \[app.mpnn\_2wl\].
\[thm.mpnn\_2wl\] Say two graphs $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$ cannot be distinguished by $2$-WL. Then there is no MPNN that can distinguish them.
*Proof intuition:* If $2$-WL cannot distinguish the two graphs, then at any iteration $t$, $\boldsymbol{\{} \cb_k^{(t)}(s): s \in V^2 \boldsymbol{\}} = \boldsymbol{\{} {\cb'}_k^{(t)}(s): s \in V^2 \boldsymbol{\}}$. This guarantees the existence of a bijective map from pairs of nodes in $G^{[\mathtt{1}]}$ to pairs of nodes in $G^{[\mathtt{2}]}$ that preserve the coloring. Through examining the update rules of $2$-WL and MPNNs, we will show by induction that for any MPNN, at the $t$th layer, such a map will also preserve the hidden states of the nodes involved in the pair as well as the edge feature. This implies that any MPNN with $t$ layers will return identical outputs when applied to the two graphs.\
This result motivates us to study what patterns $2$-WL can and cannot count in the next subsection.
![Illustration of the construction in the proof of Theorem \[thm.wl\_mc\] for the pattern from Figure \[fig:motif\] (left). Note that $\mathsf{M}(G^{[\mathtt{1}]}; G^{[\mathtt{P}]})=0$ whereas $\mathsf{M}(G^{[\mathtt{2}]}; G^{[\mathtt{P}]}) = 2$. The graphs $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$ are not distinguishable by MPNNs, 2-WL, or 2-IGNs.[]{data-label="fig:my_label"}](motif_c2.pdf "fig:"){width="27.00000%"} ![Illustration of the construction in the proof of Theorem \[thm.wl\_mc\] for the pattern from Figure \[fig:motif\] (left). Note that $\mathsf{M}(G^{[\mathtt{1}]}; G^{[\mathtt{P}]})=0$ whereas $\mathsf{M}(G^{[\mathtt{2}]}; G^{[\mathtt{P}]}) = 2$. The graphs $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$ are not distinguishable by MPNNs, 2-WL, or 2-IGNs.[]{data-label="fig:my_label"}](motif_c1.pdf "fig:"){width="27.00000%"}\
$G^{[\mathtt{1}]}$ $G^{[\mathtt{2}]}$
Substructure counting by $2$-WL and MPNNs
-----------------------------------------
Whether or not $2$-WL can perform matching-count of a pattern is completely characterized by the number of nodes in the pattern. Any connected pattern with $1$ or $2$ nodes (i.e., representing a node or an edge) can be easily counted by an MPNN with $0$ and $1$ layer of message-passing, respectively, or by $2$-WL with $0$ iteration[^2]. In contrast, for all other patterns, we provide the following negative result, to be proved in Appendix \[app.2wl\_mc\_neg\].
\[thm.wl\_mc\] $2$-WL cannot perform matching-count of any connected pattern with 3 or more nodes.
*Proof Intuition.* Given any connected pattern of at least $3$ nodes, we can construct a pair of graphs that have different matching-counts of the pattern but cannot be distinguished from each other by $2$-WL. For instance, if we run $2$-WL on the pair of graphs in Figure \[fig:my\_label\], then there will be $\cb_2^{(t)}((1, 3)) = {\cb'}_2^{(t)}((1, 3))$, $\cb_2^{(t)}((1, 2)) = {\cb'}_2^{(t)}((1, 6))$, $\cb_2^{(t)}((1, 6)) = {\cb'}_2^{(t)}((1, 2))$, and so on. We can in fact show that $\boldsymbol{\{} \cb_2^{(t)}(s): s \in V^2 \boldsymbol{\}} = \boldsymbol{\{} {\cb'}_2^{(t)}(s): s \in V^2 \boldsymbol{\}}, \forall t$, which implies that $2$-WL cannot distinguish the two graphs.\
Thus, together with Theorem \[thm.mpnn\_2wl\], we have
\[cor.mp\_mc\] MPNNs cannot perform matching-count of any connected pattern with 3 or more nodes.
For containment-count, if both nodes and edges are unweighted, @arvind2018weisfeiler show that the only patterns $1$-WL (and equivalently $2$-WL) can count are either star-shaped patterns and pairs of disjoint edges. We prove the positive result that MPNNs can count star-shaped patterns even when node and edge features are allowed, utilizing a result in @xu2018powerful that the message functions are able to approximate any function on multisets.
\[thm.mpnn\_cc\] MPNNs can perform containment-count of star-shaped patterns.
By Theorem \[thm.mpnn\_2wl\], this implies that
\[cor.wl\_cc\] $2$-WL can perform containment-count of star-shaped patterns.
Substructure counting by $k$-WL
-------------------------------
There have been efforts to extend the power of GNNs by going after $k$-WL for higher $k$, such as @morris2019higher. Thus, it is also interesting to study the patterns that $k$-WL can and cannot count. Firstly, since $k$-tuples are assigned initial colors based on their isomorphism types, the following is easily seen, and we provide a proof in Appendix \[app.kwl\_knodes\].
\[thm.kwl\_pos\] $k$-WL, at initialization, is able to perform both matching-count and containment-count of patterns consisting of at most $k$ nodes.
This establishes a potential hierarchy of increasing power in terms of substructure counting by $k$-WL. However, tighter results can be much harder to achieve. For example, to show that $2$-FWL (and therefore $3$-WL) cannot count cycles of length $8$, @furer2017combinatorial has to rely on performing computer counting on the classical Cai-Fürer-Immerman counterexamples to $k$-WL [@cai1992optimal]. We leave the pursuit of general and tighter characterizations of $k$-WL’s substructure counting power for future research, but we are nevertheless able to provide a partial negative result concerning finite iterations of $k$-WL.
A path pattern of size $m$, denoted by $H_m$, is an unattributed graph, $H_m = (V^{[\mathtt{H_m}]}, E^{[\mathtt{H_m}]})$, where $V^{[\mathtt{H_m}]} = [m]$, and $E^{[\mathtt{H_m}]} = \{ (i, i+1) : 1 \leq i < m \} \cup \{ (i+1, i) : 1 \leq i < m \}$.
\[thm.kwl\_path\] Running $T$ iterations of $k$-WL cannot perform matching-count of any path pattern of $(k+1)2^{T}$ or more nodes.
The proof is given in Appendix \[app.kwl\_neg\]. This bound grows quickly when $T$ becomes large. However, since in practice, many if not most GNN models are designed to be shallow [@zhou2018graph; @wu2019comprehensive], we believe this result is still relevant for studying finite-depth GNNs that are based on $k$-WL.
Invariant Graph Networks {#2ign}
========================
Recently, diverging from the strategy of local aggregation of information as adopted by MPNNs and $k$-WLs, an alternative family of GNN models called *Invariant Graph Networks (IGNs)* was introduced in @maron2018invariant [@maron2019universality; @maron2019provably]. Here we restate its definition.
A $k$th-order Invariant Graph Network ($k$-IGN) is a function $F: \mathbb{R}^{n^{k} \times d_{0}} \to \mathbb{R}$ that can be decomposed in the following way: $$F = m \circ h \circ L^{(T)} \circ \sigma \circ \dots \circ \sigma \circ L^{(1)},$$ where each $L^{(t)}$ is a linear equivariant layer from $\mathbb{R}^{n^{k} \times d_{t-1}}$ to $\mathbb{R}^{n^{k} \times d_{t}}$, $\sigma$ is a pointwise activation function, $h$ is a linear invariant layer from $\mathbb{R}^{n^{k} \times d_{T}}$ to $\mathbb{R}$, and $m$ is an MLP.
@maron2019universality show that if $k$ is allowed to grow as a function of the size of the graphs, then $k$-IGNs can achieve universal approximation of permutation-invariant functions on graphs. Nonetheless, due to the quick growth of computational complexity and implementation difficulty as $k$ increases, in practice it is hard to have $k>2$, while if $k=2$, it is proven to lose the universal approximation power [@chen2019equivalence]. However, it remains interesting to study what are the things that $2$-IGNs are capable of doing, especially from the perspective of substructure counting. Note that the a $2$-IGN takes as input a third-order tensor, $\Bb^{(0)}$, defined for a given graph $G = (V=[n], E, x, e)$ in the following way. Supposing without loss of generality that the node and edge features both have dimension $d$, we have $\Bb^{(0)} \in \mathbb{R}^{n \times n \times (d+1)}$, such that: $\forall i \in [n], \Bb_{i, i, 2:(d+1)}^{(0)} = x_i$; $\forall i, j \in [n]$ with $i \neq j$, $\Bb_{i, j, 1}^{(0)} = A_{i, j}$ and $\Bb_{i, j, 2:(d+1)}^{(0)} = e_{i, j}$. If we use $\Bb^{(t)}$ to denote the output of the $t$th layer of the $2$-IGN, then they are obtained iteratively by $$\label{eq:IGN_layer}
\Bb^{(t+1)} = \sigma(L^{(t)}(\Bb^{(t)}))$$
$2$-IGNs equivalent to $2$-WL
-----------------------------
Before studying how well can $2$-IGNs count substructures, we first relate it to $2$-WL. It is known that $2$-IGNs are at least as powerful as $2$-WL, while the other direction remains an open problem [@maron2019universality; @maron2019open]. Here we answer the question by proving the converse, that $2$-IGNs are no more powerful than $2$-WL. The full argument can be found in Appendix \[app.2ign\_2wl\].
\[thm.2ign\_leq\_wl\] If two graphs $G^{[\mathtt{1}]}$ and $G^{[\mathtt{2}]}$ cannot be distinguished by the $2$-WL test, then there is no $2$-IGN that can distinguish them either.
*Proof intuition:* Given two nodes $i, j \in V$ with $i \neq j$, we can partition $V^2$ as the union of nine disjoint subsets: $\mathcal{A}_{1} = \{ (i, j)\}$, $\mathcal{A}_{2} = \{(i, i)\}$, $\mathcal{A}_{3} = \{ (j, j) \}$, $\mathcal{A}_{4} = \{(i, k): k \neq i \text{ or } j\}$, $\mathcal{A}_{5} = \{(k, i): k \neq i \text{ or } j\}$, $\mathcal{A}_{6} = \{(j, k): k \neq i \text{ or } j\}$, $\mathcal{A}_{7} = \{(k, j): k \neq i \text{ or } j\}$, $\mathcal{A}_{8} = \{ (k, l): k \neq l \text{ and } \{k, l \} \cap \{ i, j \} = \emptyset \}$, and $\mathcal{A}_{9} = \{ (k, k): k \notin \{ i, j \} \}$. If $2$-WL cannot distinguish the two graphs in $t$ iterations, then there exists not only a color-preserving bijective map from pairs of nodes in $G^{[\mathtt{1}]}$ to pairs of nodes in $G^{[\mathtt{2}]}$, mapping $(i, j)$ to some $(i', j')$, but also a color-preserving bijective map from $\mathcal{A}_{w}$ to $\mathcal{A}'_{w}$ for each $w \in [9]$, where $\mathcal{A}'_{w}$ is the corresponding subset of $V^2$ associated with $(i', j')$. By the update rule of $2$-IGNs, this allows us to show that $\Bb^{(t)}_{i, j} = {\Bb'}^{(t)}_{i', j'}$, and hence a $t$-layer $2$-IGN cannot return distinct outputs when applied to the two graphs.
$2$-IGNs are exactly as powerful as $2$-WL.
Substructure counting by $2$-IGNs
---------------------------------
Thanks to the equivalence shown above, the two following theorems are direct consequences of Theorem \[thm.wl\_mc\] and Corollary \[cor.wl\_cc\], though we also provide a direct proof of Corollary \[cor.2ign\_mc\] in Appendix \[app.2ign\_mc\_neg\].
\[cor.2ign\_mc\] $2$-IGNs cannot perform matching-count of any connected pattern with 3 or more nodes.
2-IGNs can perform containment-count of star-shaped patterns.
Substructure counting by $k$-IGNs
---------------------------------
Since $k$-IGNs are no less powerful than $k$-WL [@maron2019provably], we have as a corollary of Theorem \[thm.kwl\_pos\] that
$k$-IGNs can perform both matching-count and containment-count of patterns consisting of at most $k$ nodes.
Local Relational Pooling
========================
Though MPNNs and $2$-IGNs are able to aggregate information from multi-hop neighborhoods, we have seen above that they are unable to preserve information such as the matching-counts of nontrivial patterns. To bypass such limitations, we suggest going beyond the strategy of iteratively aggregating information in an equivariant way, which underlies both MPNNs and IGNs. One helpful observation is that, if a pattern is present in the graph, it can always be found in a sufficiently large local neighborhood, or *egonet*, of some node in the graph [@preciado2012structural]. An egonet of depth $l$ centered at a node $i$ is the induced subgraph consisting of $i$ and all nodes within distance $l$ from it. Note that any pattern with radius $r$ is a subgraph of some egonet of depth $l=r$. Hence, by applying a powerful local model to each egonet separately and then aggregating the outputs, we could potentially obtain a model capable of counting patterns.\
For such a local model, we adopt the Relational Pooling (RP) idea from @murphy2019relational. In summary, it creates a powerful permutation-invariant model by symmetrizing a powerful model that is not necessarily permutation-invariant, where the symmetrization is performed by averaging or summing over all permutations of the nodes’ ordering. Formally, if $\Bb \in \mathbb{R}^{n \times n \times d}$ is a node-ordering-dependent representation of the graph $G$, such as the adjacency matrix or the $\Bb^{(0)}$ defined above for $2$-IGNs, then define $$f_{\textrm{RP}}(G) = \sum_{\pi \in S_n} \Bar{f} (\pi \circ \Bb),$$ where $\bar{f}$ can be some non-permutation-invariant function, $S_n$ is the set of permutations on $n$ nodes, and $\pi \circ \Bb$ is $\Bb$ transformed by permuting its first two dimensions according to $\pi$. Such $f$’s are shown to be an universal approximators of permutation-invariant functions [@murphy2019relational]. The summation quickly becomes intractable once $n$ is large, and hence approximation methods have been introduced. In our case, however, since we apply this model to egonets that are usually smaller than the entire graph, the tractability issue is greatly alleviated. Moreover, since egonets are rooted graphs, we can reduce the symmetrization over all permutations in $S_n$ to the subset $S_n^{\textrm{BFS}} \subseteq S_n$ of permutations compatible with breath-first-search (BFS) to further reduce the complexity, as suggested in @murphy2019relational.\
Concretely, we define $G^{[\mathtt{ego}]}_{i, l}$ as the egonet centered at node $i$ of depth $l$, $\Bb^{[\mathtt{ego}]}_{i, l}$ as the corresponding representation and $n_{i, l}$ as the number of nodes in $G^{[\mathtt{ego}]}_{i, l}$. For computational efficiency, every tensor representation of egonet $\Bb$ is cropped into a fixed-sized subtensor $C_k(\Bb)=\Bb_{[k], [k], :} \in \mathbb{R}^{k\times k\times d}$. Then our model over the entire graph $G$ is expressed as $$f_{\textrm{LRP}}^{l,k}(G) = \sum_{i \in V} \sum_{\pi \in S_{n_{i, l}}^{\textrm{BFS}}} \Bar{f} \left(C_k(\pi \circ \Bb^{[\mathtt{ego}]}_{i, l})\right)$$ We call it depth-$l$ size-$k$ *Local Relational Pooling (LRP-$l$-$k$)*. If node degrees are upper-bounded by $D$, the time complexity is $O(n \cdot (D!)^{D^{l}} \cdot k^2)$, and hence linear in $n$ if $D, k$ and $l$ are fixed. In the experiments below, we implement a variant of LRP-$1$-$4$ designed as, with bias terms ignored, $$\tilde{f}_{\textrm{LRP}}^{1,4}(G) = \mathbf{W}_1\sum_{i \in V}\sigma\left[\frac{\textsc{Mlp}(D_i)}{|S_{n_{i, 1}}^{\textrm{BFS}}|} \odot \sum_{\pi \in S_{n_{i, 1}}^{\textrm{BFS}}} f_* (\pi \circ \Bb^{[\mathtt{ego}]}_{i, 1})\right],$$ where $D_i$ is the degree of node $i$, $\sigma$ is ReLU, $\textsc{Mlp}$ maps from $\mathbb{R}$ to $\mathbb{R}^{H}$, where $H$ is the hidden dimension, $\mathbf{W}_1 \in \mathbb{R}^{1 \times H}$ and $\forall j\in[H]$, $(f_*(\mathbf{X}))_j=\text{tanh}(\sum\mathbf{W}_{2,j}\odot C_4(\mathbf{X}))\in\mathbb{R}$ with $\mathbf{W}_{2, j} \in \mathbb{R}^{4 \times 4 \times d}$. The motivation of $\textsc{Mlp}(D_i)$ is to adaptively learn an invariant function over permutation, such as summing and averaging.
Conclusions
===========
We propose a theoretical framework to study the expressive power of classes of GNNs based on their ability to count substructures. We distinguish two kinds of counting: containment-count (counting subgraphs) and matching-count (counting induced subgraphs). We prove that neither MPNNs nor $2$-IGNs can matching-count any connected structure with $3$ or more nodes; $k$-IGNs and $k$-WL can containment-count and matching-count any pattern of size $k$. We also provide an upper bound on the size of “path-shaped” substructures that finite iterations of $k$-WL can matching-count. To establish these results, we prove an equivalence between approximating graph functions and discriminating graphs. Also, as intermediary results, we prove that MPNNs are no more powerful than $2$-WL on attributed graphs, and that $2$-IGNs are equivalent to $2$-WL in distinguishing non-isomorphic graphs, which partly answers an open problem raised in @maron2019open. In addition, we perform numerical experiments that support our theoretical results and show that the Local Relational Pooling approach inspired by @murphy2019relational can successfully count certain substructures. In summary, we build the foundation for using substructure counting as an intuitive and relevant measure of the expressive power of GNNs, and our concrete results for existing GNNs motivate the search for more powerful designs of GNNs.\
One limitation of our theory is that it only pertains to the expressive power of GNNs and does not speak about optimization or generalization. In addition, our theoretical results are worse-case in nature and cannot predict average-case performance, which is interesting to study as well. Nonetheless, even within this new framework, many interesting questions remain, including better characterizing the ability to count substructures of general $k$-WL and $k$-IGNs as well as other architectures such as spectral GNNs [@chen2019cdsbm] and polynomial IGNs [@maron2019open]. Another interesting future direction is to study the relevance of substructure counting in empirical tasks, following the work of @ying2019gnn. Finally, we hope our framework can help guide the search for more powerful GNNs by having substructure counting as a criterion.
#### Acknowledgements
We would like to thank Haggai Maron and Jiaxuan You for nice conversations. This work is partially supported by the Alfred P. Sloan Foundation, NSF RI-1816753, NSF CAREER CIF 1845360, NSF CHS-1901091, Samsung Electronics, and the Institute for Advanced Study. SV is partly supported by NSF DMS 1913134, EOARD FA9550-18-1-7007 and the Simons Algorithms and Geometry (A&G) Think Tank.
Additional notations {#app.notations}
====================
For two positive integers $a$ and $b$, we define $\textsc{Mod}_a(b)$ to be $a$ if $a$ divides $b$ and the number $c$ such that $b \equiv c \pmod{a}$ otherwise. Hence the value ranges from $1$ to $a$ as we vary $b \in \mathbb{N}^*$.\
For a positive integer $c$, let $[c]$ denote the set $\{1, ..., c \}$.\
Two $k$-typles, $(i_i, ..., i_k), (j_1, ..., j_k) \in V^k$ are said to be in the same *equivalent class* if $\exists$ a permutation $\pi$ on $V$ such that $(\pi(i_i), ..., \pi(i_k)) = (j_1, ..., j_k)$. Note that belonging to the same equivalence class is a weaker condition than having the same isomorphism type, as will be defined in Appendix \[app.isotypes\], which has to do with what the graphs look like.\
For any $k$-tuple, $s = (i_1, ..., i_k)$, and for $w \in [k]$, use $\mathtt{I}_w(s)$ to denote the $w$th entry of $s$, $i_w$.
Isomorphism types of $k$-tuples in $k$-WL for attributed graphs {#app.isotypes}
===============================================================
Say $G^{[\mathtt{1}]}=(V^{[\mathtt{1}]}, E^{[\mathtt{1}]}, x^{[\mathtt{1}]}, e^{[\mathtt{1}]})$, $G^{[\mathtt{2}]}=(V^{[\mathtt{2}]}, E^{[\mathtt{2}]}, x^{[\mathtt{2}]}, e^{[\mathtt{2}]})$.\
a) $\forall s = (i_1, ..., i_k), s' = (i'_1, ..., i'_k) \in (V^{[\mathtt{1}]})^k$, $s$ and $s'$ are said to have the same isomorphism type if
1. $\forall \alpha, \beta \in [k], i_\alpha = i_\beta \Leftrightarrow i'_\alpha = i'_\beta$
2. $\forall \alpha \in [k], x^{[\mathtt{1}]}_{i_\alpha} = x^{[\mathtt{1}]}_{i'_\alpha}$
3. $\forall \alpha, \beta \in [k], (i_\alpha, i_\beta) \in E^{[\mathtt{1}]} \Leftrightarrow (i'_\alpha, i'_\beta) \in E^{[\mathtt{1}]}$, and moreover, if either side is true, then $e^{[\mathtt{1}]}_{i_\alpha, i_\beta} = e^{[\mathtt{1}]}_{i'_\alpha, i'_\beta}$
b\) Similar if both $s, s' \in (V^{[\mathtt{2}]})^k$.\
c) $\forall s = (i_1, ..., i_k) \in (V^{[\mathtt{1}]})^k, s' = (i'_1, ..., i'_k) \in (V^{[\mathtt{2}]})^k$, $s$ and $s'$ are said to have the same isomorphism type if
1. $\forall \alpha, \beta \in [k], i_\alpha = i_\beta \Leftrightarrow i'_\alpha = i'_\beta$
2. $\forall \alpha \in [k], x^{[\mathtt{1}]}_{i_\alpha} = x^{[\mathtt{2}]}_{i'_\alpha}$
3. $\forall \alpha, \beta \in [k], (i_\alpha, i_\beta) \in E^{[\mathtt{1}]} \Leftrightarrow (i'_\alpha, i'_\beta) \in E^{[\mathtt{2}]}$, and moreover, if either side is true, then $e^{[\mathtt{1}]}_{i_\alpha, i_\beta} = e^{[\mathtt{2}]}_{i'_\alpha, i'_\beta}$
In $k$-WL tests, two $k$-tuples $s$ and $s'$ in either $(V^{[\mathtt{1}]})^k$ or $(V^{[\mathtt{2}]})^k$ are assigned the same color at iteration $0$ if and only if they have the same isomorphism type.\
For a reference, see @maron2019provably.
Proof of Theorem \[thm.mpnn\_2wl\] (MPNNs are no more powerful than $2$-WL) {#app.mpnn_2wl}
===========================================================================
Proof of Theorem \[thm.wl\_mc\] ($2$-WL is unable to matching-count patterns of $3$ or more nodes) {#app.2wl_mc_neg}
==================================================================================================
Proof of Theorem \[thm.mpnn\_cc\] (MPNNs are able to containment-count star-shaped patterns) {#app.mpnn_cc_star}
============================================================================================
Proof of Theorem \[thm.kwl\_pos\] ($k$-WL is able to count patterns of $k$ or fewer nodes) {#app.kwl_knodes}
==========================================================================================
Proof of Theorem \[thm.kwl\_path\] ($T$ iterations of $k$-WL cannot matching-count path patterns of size $(k+1)2^T$ or more) {#app.kwl_neg}
============================================================================================================================
Proof of Theorem \[thm.2ign\_leq\_wl\] ($2$-IGNs are no more powerful than $2$-WL) {#app.2ign_2wl}
==================================================================================
Direct proof of Corollary \[cor.2ign\_mc\] ($2$-IGNs are unable to matching-count patterns of $3$ or more nodes) {#app.2ign_mc_neg}
================================================================================================================
Specific GNN architectures {#appdx_gnn_archi}
==========================
Experiment results {#app.all_exp_res}
==================
[^1]: We define isomorphism types rigorously in Appendix \[app.isotypes\].
[^2]: Rigorously, this is a special case of Theorem \[thm.kwl\_pos\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop a framework for obtaining linear programming bounds for spherical codes whose inner products belong to a prescribed subinterval $[\ell,s]$ of $[-1,1)$. An intricate relationship between Levenshtein-type upper bounds on cardinality of codes with inner products in $[\ell,s]$ and lower bounds on the potential energy (for absolutely monotone interactions) for codes with inner products in $[\ell,1)$ (when the cardinality of the code is kept fixed) is revealed and explained. Thereby, we obtain a new extension of Levenshtein bounds for such codes. The universality of our bounds is exhibited by a unified derivation and their validity for a wide range of codes and potential functions.'
address:
- |
Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, 8 G Bonchev Str., 1113 Sofia, Bulgaria\
and Faculty of Mathematics and Natural Sciences, South-Western University, Blagoevgrad, Bulgaria.
- 'Department of Mathematical Sciences, Purdue University Fort Wayne, IN 46805, USA '
- 'Center for Constructive Approximation, Department of Mathematics, Vanderbilt University, Nashville, TN 37240, USA '
- 'Faculty of Mathematics and Informatics, Sofia University, 5 James Bourchier Blvd., 1164 Sofia, Bulgaria'
author:
- 'P. G. Boyvalenkov $^\dagger$'
- 'P. D. Dragnev $^{\dagger \dagger}$'
- 'D. P. Hardin$^*$'
- 'E. B. Saff$^*$'
- 'M. M. Stoyanova$^\dagger$'
title: '**On spherical codes with inner products in a prescribed interval**'
---
[**Keywords.**]{} Spherical codes, Linear programming, Bounds for codes, $h$-energy of a code
[**MSC Codes.**]{} 94B65, 52A40, 74G65,
Introduction
============
In the seminal paper of Cohn and Kumar [@CK], many classical maximal spherical codes with applications to communications, such as the Korkin-Zolotarev kissing number configuration on $\mathbb{S}^7$, the Leech lattice configuration in 24 dimensions, the $600$-cell, etc., were shown to be universally optimal in the sense that they have minimal potential energy for a large class of potential interactions. The notion of universal optimality was further developed for Hamming spaces in [@CZ].
As important as these particular configurations are, it is of significant interest to study bounds for codes of general cardinality. The theory of universal bounds for codes and designs in polynomial metric spaces was laid out by Levenshtein in [@Lev]. The interplay between Levenshtein’s framework and universal lower bounds (ULB) on potential energy of codes was established recently by the authors for Euclidean spaces in [@BDHSS-CA] and for Hamming spaces in [@BDHSS-DCC]. In this paper we further that interplay to codes with inner products in a prescribed subinterval $[\ell,s]$ of $[-1,1)$ and as a result derive an extension of Levenshtein’s framework to this setting.
Let $\mathbb{S}^{n-1}\subset \mathbb{R}^n$ denote the $(n-1)$-dimensional unit sphere. A nonempty finite set $C \subset \mathbb{S}^{n-1}$ is called a [*spherical code*]{}. For $-1 \leq \ell <s < 1$ denote by $$\mathcal{C}(\ell,s):=\{C \subset \mathbb{S}^{n-1}: \ \ell \leq \langle x,y \rangle \leq s , \ x,y \in C, x \neq y\},$$ the set of spherical codes with prescribed maximum diameter and minimum pairwise distance, where $\langle x,y \rangle$ denotes the inner product of $x$ and $y$. We establish upper bounds on the quantity $$\mathcal{A}(n;[\ell,s]):=\max\{|C|:C \in \mathcal{C}(\ell,s)\},$$ which is a classical problem in coding theory.
Given a (potential) function $h(t):[-1,1] \to [0,+\infty]$ and a code $C \subset \mathbb{S}^{n-1}$, we define the [*potential energy*]{} (also referred to as [*$h$-energy*]{}) of $C$ as $$E(C;h):=\sum_{x, y \in C, x \neq y} h(\langle x,y \rangle).$$ In what follows we shall consider potential functions $h$ that are absolutely monotone, namely $h^{(k)}(t) \geq 0$ for every $k \geq 0$ and $t\in [-1,1)$. For such potentials we establish ULB for the quantity $$\mathcal{E}(n,M,\ell;h):=\inf\{E(C;h):C \in \mathcal{C}(\ell,1), |C|=M \}.$$ As in [@BDHSS-CA], the use of linear programming reveals a strong connection between our ULB on $\mathcal{E}(n,M,\ell;h)$ and our Levenshtein-type upper bounds on $\mathcal{A}(n;[\ell,s])$.
Throughout, $P_k^{(n)}(t)$, $k=0,1,\dots$, will denote the Gegenbauer polynomials [@Sze] normalized with $P_k (1)=1$. We consider functions $f(t):[-1,1]\to \mathbb{R}$, $$f(t)=\sum_{k=0}^\infty f_k P_k^{(n)}(t), \ \ {\rm where} \ \
f(1)=\sum_{k=0}^\infty f_k <\infty.$$ The function $f$ is called [*positive definite*]{} (*strictly positive definite*) if all coefficients $f_k$ are non-negative (positive). Following Levenshtein’s notation we denote the class of all positive definite (strictly positive definite) functions by the symbol ${\mathcal F}_{\geq}$ (${\mathcal F}_{>}$). When $f$ is a polynomial, the definition of ${\mathcal F}_{>}$ does not include $f_k$ for $k>\deg(f)$ (since $f_k=0$ for such $k$).
The Kabatiansky-Levenshtein [@KL] approach (see also [@DGS]) is based on the inequality $$\label{MaxCodesLP} \mathcal{A}(n;[\ell,s]) \leq \min_{f\in \mathcal{F}_{n,\ell,s}} f(1)/f_0,$$ where $$\mathcal{F}_{n,\ell,s}:=\{ f \in {\mathcal F}_{\geq} \ | \ f(t) \leq 0, \, t \in [\ell,s], \, f_0>0\}.$$
Similarly, the Delsarte-Yudin approach (see [@Y]) uses the inequality $$\label{EnergyLP} \mathcal{E}(n,M,\ell;h)\ge \max_{g\in \mathcal{G}_{n,\ell;h}} M(Mg_0-g(1)),$$ where $$\mathcal{G}_{n,\ell;h}:=\{ g\in {\mathcal F}_{\geq} \ | \ g(t) \leq h(t), \, t \in [\ell,1), \, g_0>0\}.$$ The determination of the right-hand sides of the bounds and over the respective classes defines two infinite linear programs. To determine his universal bounds on $\mathcal{A}(n,s):=\mathcal{A}(n;[-1,s])$ Levenshtein [@Lev] found explicitly the solution of the linear program posed by when restricted to $\mathcal{F}_{n,-1,s}\cap \mathcal{P}_m$, where $\mathcal{P}_m$ denotes the class of real polynomials of degree at most $m$.
In [@BDHSS-CA] the authors considered the linear program in over $\mathcal{G}_{n,-1;h} \cap \mathcal{P}_m$ and found its solution as the Hermite interpolation polynomial of $h(t)$ at the zeros of the Levenshtein polynomial. This implies the ULB on $\mathcal{E}(n,M;h):=\mathcal{E}(n,M,-1;h)$. The interplay between the two optimal solutions is that the zeros of the Levenshtein polynomials serve also as nodes of an important Radau or Lobato quadrature formulae. In this paper we further develop the intricate connection between the maximum cardinality and minimum energy problems, which is described in our main result Theorem \[main-th-subintervals\]. For this purpose a central role is played by an $\ell$-modification of the so-called ‘strengthened Krein condition’ introduced by Levenshtein (see Section 4).
The outline of the paper is as follows. In Section 2, we introduce certain signed measures and establish their positive definiteness up to an appropriate degree. Properties of their associated orthogonal polynomials are also discussed. In Section 3, Levenshtein-type polynomials $f_{2k}^{(n,\ell,s)}(t)$ are constructed and corresponding quadrature formulas are derived. These formulas are used in Section 4, together with linear programming techniques, to derive the Levenshtein-type bounds on the cardinality of maximal codes $\mathcal{A}(n;[\ell,s])$ and ULB-type (in the sense of [@BDHSS-CA]) energy bounds on $\mathcal{E}(n,M,\ell;h)$. In the last section some special examples and numerical evidence of an $\ell$-strengthened Krein property are presented.
Positive definite signed measures and associated orthogonal polynomials
=======================================================================
In this section we establish the positive definiteness up to certain degrees of the signed measures that are used in the proof of our main result, Theorem \[main-th-subintervals\].
We shall denote the measure of orthogonality of Gegenbauer polynomials as $$\label{dmu} d\mu(t):=\gamma_n (1-t^2)^{\frac{n-3}{2}}\, dt, \quad t\in [-1,1], \quad
\gamma_n := \frac{\Gamma(\frac{n}{2})}{\sqrt{\pi}\Gamma(\frac{n-1}{2})},$$ where $ \gamma_n$ is a normalizing constant that makes $\mu$ a probability measure.
Levenshtein used the [*adjacent*]{} (to Gegenbauer) [*polynomials*]{} $$\label{Adjacent}
P_k^{1,0} (t): =P_k^{(\frac{n-1}{2},\frac{n-3}{2})}(t)/P_k^{(\frac{n-1}{2},\frac{n-3}{2})}(1)=
\eta_k^{1,0} t^k + \cdots ,\quad \eta_k^{1,0}>0,$$ where $P_k^{(\alpha,\beta)}(t)$ denotes the classical Jacobi polynomial (the normalization is again chosen so that $P_k^{1,0} (1)=1$). The polynomials are orthogonal with respect to the probability measure $$\label{AdjacentMeas} d\chi(t) := (1-t) d\mu(t).$$ They also satisfy the following three-term recurrence relation $$(t-a_i^{1,0})P_i^{1,0} (t)=b_i^{1,0} P_{i+1}^{1,0} (t)+c_i^{1,0} P_{i-1}^{1,0} (t), \quad i=1,2, \dots ,$$ where $$P_0^{1,0} (t) = 1, \ P_1^{1,0} (t) = \frac{nt+1}{n+1},$$ $$b_i^{1,0} =\frac{\eta_i^{1,0}}{\eta_{i+1}^{1,0}}>0, \
c_i^{1,0} =\frac{r_{i-1}^{1,0} b_{i-1}^{1,0}}{r_i^{1,0}}, \ a_i^{1,0}=1-b_i^{1,0}-c_i^{1,0},$$ $$r_i^{1,0}:=\left( \int_{-1}^1 \left[ P_i^{1,0} (t)\right]^2 d\chi(t) \right)^{-1} = \left(\frac{n+2i-1}{n-1}\right)^2\binom{n+i-2}{i}.$$
Let $t_{i,1}^{1,0}< t_{i,2}^{1,0} < \dots < t_{i,i}^{1,0}$ be the zeros of the polynomial $P_i^{1,0} (t)$, which are known to interlace with the zeros of $P_{i-1}^{1,0} (t)$.
We next recall the definition of positive definite signed measures up to degree $m$ (see [@CK Definition 3.4]).
A signed Borel measure $\nu$ on $\mathbb{R}$ for which all polynomials are integrable is called [**positive definite up to degree $\mathbf {m}$**]{} if for all real polynomials $p \not\equiv 0$ of degree at most $m$ we have $\int p(t)^2 d \nu(t) > 0$.
Given $\ell$ and $s$ such that $\ell < t_{k,1}^{1,0}<t_{k,k}^{1,0}<s$, we define the signed measures on $[-1,1]$ (see and ) $$\begin{aligned}
\label{SignedMeasures1}
d\nu_\ell(t) &:=& (t-\ell)d\chi(t),\\
\label{SignedMeasures2}
d\nu_s(t) &:=& (s-t)d\chi(t), \\
\label{SignedMeasures3}
d\nu_{\ell,s}(t) &:=& (t-\ell)(s-t)d\chi(t), \\
\label{SignedMeasures4}
d\mu_\ell (t) &:=& (t-\ell )d\mu (t).\end{aligned}$$ The following lemma establishes the positive definiteness of these signed measures up to certain degrees, which in turn allows us to define orthogonal polynomials with respect to these signed measures. This equips us with the essential ingredients for modifying Levenshtein’s framework.
\[lem\_pos\_def\] For given $k>1$, let $s$ and $\ell$ satisfy $\ell < t_{k,1}^{1,0}<t_{k,k}^{1,0}<s$. Then the measures $d\nu_\ell(t)$, $d\nu_s (t)$, and $d\mu_\ell (t)$ are positive definite up to degree $k-1$ and the measure $d\nu_{\ell,s} (t)$ is positive definite up to degree $k-2$.
[*Proof.*]{} We first note that the system of $k+1$ nodes $$M_{k+1}:=\{ t_{k,1}^{1,0}< t_{k,2}^{1,0}< \dots< t_{k,k}^{1,0}<1:=t_{k,k+1}^{1,0} \}$$ defines a positive Radau quadrature with respect to the measure $\mu$ that is exact for all polynomials of degree at most $2k$ (see e.g. [@DR pp. 102-105], [@BCV Theorem 2.4]), namely the quadrature formula $$\label{w_quadrature}
f_0:= \int_{-1}^1 f(t) d\mu(t)= w_{k+1} f(1)+\sum_{i=1}^k w_i f(t_{k,i}^{1,0} )$$ holds for all polynomials $f$ of degree at most $2k$, and the weights $w_i$, $i=1,
\dots, k+~1$ are positive. Let now $q(t)$ be an arbitrary polynomial of degree at most $k-1$. From we have that $$\begin{aligned}
\int_{-1}^1 q^2(t) d\nu_\ell (t) &=& \int_{-1}^1 q^2(t)(1-t)(t-\ell) d\mu(t) \\
&=& \sum_{i=1}^k w_i q^2(t_{k,i}^{1,0} )(1-t_{k,i}^{1,0} )(t_{k,i}^{1,0} -\ell) \geq 0,\end{aligned}$$ where equality may hold only if $q(t_{k,i}^{1,0} )=0$ for all $i=1,\dots,k$, which would imply that $q(t) \equiv 0$. Therefore the measure $d\nu_\ell (t)$ is positive definite up to degree $k-1$ as asserted.
Similarly, for the measure $d\nu_s (t)$ and ${\rm deg}\ q \leq k-1$ we have $$\begin{aligned}
\int_{-1}^1 q^2(t) d\nu_s (t) &=& \int_{-1}^1 q^2(t)(1-t)(s-t) d\mu(t) \\
&=& \sum_{i=1}^k w_i q^2(t_{k,i}^{1,0} )(1-t_{k,i}^{1,0} )(s-t_{k,i}^{1,0} ) \geq 0,\end{aligned}$$ where again equality holds only for $q(t) \equiv 0$.
Next, if $q(t) \not\equiv 0$ is of degree at most $k-2$, then we utilize again to derive that $$\int_{-1}^1 q^2(t) d\nu_{\ell,s} (t) = \sum_{i=1}^k w_i q^2(t_{k,i}^{1,0} )(1-t_{k,i}^{1,0} )(t_{k,i}^{1,0} -\ell)(s-t_{k,i}^{1,0} ) > 0.$$ Hence, $d\nu_{\ell,s} (t)$ is positive definite up to degree $k-2$.
To verify the assertion about the measure $d\mu_\ell (t)$ we employ a similar argument but with a quadrature rule defined on the collection of $k$ nodes $$\widetilde{M}_k:=\{ t_{k,1}< t_{k,2}< \cdots < t_{k,k} \},$$ where $t_{k,i}$ are the zeros of the regular Gegenbauer polynomials $P_k^{(n)}(t)$. We note that from [@Lev Lemma 5.29, Eq. (72)] we have $t_{k,1}^{1,0}<t_{k,1}$. Using the associated Lagrange basis polynomials $\widetilde{L}_i$, $i=1,2,\dots, k$, we define the weights $v_i:= \int_{-1}^1 \widetilde{L}_i (t) d\mu(t)$, $i=1,2, \dots, k$. Then, as in the proof of Gaussian quadrature, one shows that the formula $$f_0:= \int_{-1}^1 f(t) d\mu(t)= \sum_{i=1}^k v_i f(t_{k,i} )$$ is exact for polynomials of degree up to $2k-1$. Thus, for any polynomial $q(t)$ of degree less than or equal to $k-1$, we have $$\begin{aligned}
\int_{-1}^1 q^2(t) d\mu_\ell (t) &=& \int_{-1}^1 q^2(t)(t-\ell) d\mu(t) \\
&=& \sum_{i=1}^k v_i q^2(t_{k,i} )(t_{k,i} -\ell) \geq 0,\end{aligned}$$ with equality if and only if $q(t)\equiv 0$. This concludes the proof of the lemma. $\Box$
Applying Gram-Schmidt orthogonalization (see, for example, [@CK Lemma 3.5]) one derives the existence and uniqueness (for the so-chosen normalization) of the following classes of orthogonal polynomials with respect to the signed measures -.
\[cor\_ortho\] Let $\ell < t_{1,k}^{1,0}<t_{k,k}^{1,0}<s$. The following classes of orthogonal polynomials are well-defined: $$\{ P_j^{0,\ell} (t) \}_{j=0}^k, \ {\rm orthogonal\ w.r.t.} \ d\mu_\ell (t), \ P_j^{0,\ell} (1)=1;$$ $$\{ P_j^{1,\ell} (t) \}_{j=0}^k, \ {\rm orthogonal\ w.r.t.} \ d\nu_\ell (t), \ P_j^{1,\ell} (1)=1;$$ $$\{ P_j^{1,s} (t) \}_{j=0}^k, \ {\rm orthogonal\ w.r.t.} \ d\nu_s (t), \ P_j^{1,s} (1)=1;$$ $$\{ P_j^{1,\ell,s} (t) \}_{j=0}^{k-1}, \ {\rm orthogonal\ w.r.t.}\ d\nu_{\ell,s} (t), \ P_j^{1,\ell,s} (1)=1.$$ The polynomials in each class satisfy a three-term recurrence relation and their zeros interlace.
We note that if $t_{k+1,1}^{1,0}<\ell<t_{k,1}^{1,0}$ is such that $P_{k+1}^{1,0}(\ell)/P_k^{1,0}(\ell)=1$, then $P_k^{1,\ell}(1)=0$ and the normalization above fails. However, for our purposes we shall restrict to values of $\ell$ such that $P_{k+1}^{1,0}(\ell)/P_k^{1,0}(\ell)<1$.
Utilizing the Christoffel-Darboux formula (see, for example [@Sze Th. 3.2.2], [@Lev Eq. (5.65)]) we are able to construct these polynomials explicitly. Let $$\label{kernelT}
\begin{split}
T_i^{1,0} (x,y) &:= \sum_{j=0}^i r_j^{1,0} P_j^{1,0}(x) P_j^{1,0}(y) \\
&= r_i ^{1,0}b_i^{1,0} \frac{P_{i+1}^{1,0}(x) P_i^{1,0}(y)-P_{i+1}^{1,0}(y) P_i^{1,0}(x)}{x-y}.
\end{split}$$ Note that in the limiting case $x=y$ we use appropriate derivatives.
Levenshtein [@Lev] uses the Christoffel-Darboux formula to prove the interlacing properties $t_{j+1,i}<t_{j,i}^{1,0}<t_{j,i}$, $i=1,2,\dots,j$, of the zeros of $P_j^{1,0}$ and the Gegenbauer polynomials. Similarly, from the representation $$P_j^{1,\ell} (t)=\frac{(1-\ell)\left( P_{j+1}^{1,0} (t)-P_j^{1,0} (t)P_{j+1}^{1,0} (\ell)/P_{j}^{1,0} (\ell) \right)}{(t-\ell)\left(1-P_{j+1}^{1,0} (\ell)/P_{j}^{1,0} (\ell)\right)},$$ which is verified in the next theorem, we derive interlacing properties of the zeros of $P_j^{1,\ell}$ with respect to the zeros of $P_i^{1,0}$.
\[rootsofP1l\] Let $\ell$ and $k$ be such that $t_{k+1,1}^{1,0}<\ell<t_{k,1}^{1,0}$ and $P_{k+1}^{1,0}(\ell)/P_{k}^{1,0}(\ell)<1$. Then all zeros $\{ t_{i,j}^{1,\ell}\}_{j=1}^i$ of $P_i^{1,\ell}(t)$ are in the interval $[\ell,1]$ and we have $$\label{Pi1l'}
P_i^{1,\ell}(t) = \frac{T_i^{1,0} (t,\ell)}{T_i^{1,0}(1,\ell)}=\eta_i^{1,\ell} t^i + \cdots,\quad i=0,1,\dots, k,$$ with all leading coefficients $\eta_i^{1,\ell}>0$ and $t_{k,k}^{1,\ell}<1$. Finally, the interlacing rules $$\label{Interlacing}\begin{split}
t_{i,j}^{1,\ell} &\in (t_{i,j}^{1,0}, t_{i+1,j+1}^{1,0}), \ i=1,\dots, k-1, j=1,\dots, i ;\\
t_{k,j}^{1,\ell} &\in (t_{k+1,j+1}^{1,0}, t_{k,j+1}^{1,0}), \ j=1,\dots,k-1,
\end{split}$$ hold.
As the proof below shows the condition $P_{k+1}^{1,0}(\ell)/P_{k}^{1,0}(\ell)<1$ is equivalent with $t_{k,k}^{1,\ell}<1$. In general, the orthogonal polynomial $P_k^{1,\ell} (t)$ is well defined for all $t_{k+1,1}^{1,0}<\ell<t_{k,1}^{1,0}$, but its largest root leaves the interval $[-1,1]$ and the leading coefficient becomes negative.
[*Proof.*]{} For any polynomial $p(t)$ of degree less than $i$ we have $$\begin{aligned}
&& \int_{-1}^1 T_i^{1,0} (t,\ell) p(t) \, d\nu_\ell(t) \\ &=& r_i^{1,0} b_i ^{1,0}\int_{-1}^1
\left( P_{i+1}^{1,0}(t) P_i^{1,0}(\ell)-P_{i+1}^{1,0}(\ell) P_i^{1,0}(t) \right)p(t) \, d\chi(t)=0,\end{aligned}$$ and follows from the positive definiteness of the measure $d\nu_\ell(t)$ and the uniqueness of the Gram-Schmidt orthogonalization process.
We next focus on the location of the zeros of $P_i^{1,\ell}(t)$. From and they are solutions of the equation $$\label{FracEq}
\frac{P_{i+1}^{1,0}(t)}{P_i^{1,0}(t)} = \frac{P_{i+1}^{1,0}(\ell)}{P_i^{1,0}(\ell)}.$$ For all $i<k$ the zeros of $P_{i+1}^{1,0}(t)$ and $P_i^{1,0}(t)$ are interlaced and contained in $[t_{k,1}^{1,0},t_{k,k}^{1,0}]$. Observe that ${\rm sign}\,P_i^{1,0}(\ell)=(-1)^i$, so $P_{i+1}^{1,0}(\ell)/P_i^{1,0}(\ell)<0$. The function $P_{i+1}^{1,0}(t)/P_i^{1,0}(t)$ has simple poles at $t_{i, j}^{1,0}$, $j=1,\dots, i$, and simple zeros at $t_{i+1,j}^{1,0}$, $j=1,\dots, i+1$; therefore, there is at least one solution $t_{i,j}^{1,\ell}$ of on every subinterval $(t_{i,j}^{1,0}, t_{i+1,j+1}^{1,0})$, $j=1,\dots,i$, which accounts for all zeros of $P_i^{1,\ell}(t)$.
When $i=k$ we note first that $P_{k+1}^{1,0}(\ell)/P_k^{1,0}(\ell)>0$. Moreover, $\ell$ is contained in the interval $(t_{k+1,1}^{1,0},t_k^{1,0})$, so we can account similarly for only the first $k-1$ solutions of , namely $$t_{k,j}^{1,\ell} \in (t_{k+1,j+1}^{1,0}, t_{k,j+1}^{1,0}), \ j=1,\dots,k-1.$$ This establishes the interlacing properties . To account for the last zero of $P_{k}^{1,\ell}(t)$ we utilize the fact that $P_{k+1}^{1,0}(t)/P_k^{1,0}(t)>0$ for $t\in (t_{k+1,k+1}^{1,0},\infty)$. As $\lim_{t\to \infty}P_{k+1}^{1,0}(t)/P_k^{1,0}(t) = \infty$, we have one more solution $t_{k,k}^{1,0}$ of .
Since $P_{k+1}^{1,0}(\ell)/P_{k}^{1,0}(\ell)<1$, we conclude that $t_{k,k}^{1,0}<1$ because $P_{k+1}^{1,0}(1)/P_k^{1,0}(1)=1$. Comparison of coefficients in yields $\eta_k^{1,\ell}>0$. $\Box$
Construction of the Levenshtein-type polynomials
================================================
Given some $\ell>-1$, we choose $k=k(\ell)$ to be the largest $k$ such that the condition $\ell<t_{1,k}^{1,0}$ is satisfied.
We first construct the polynomials $P_i^{1,\ell,s} (t)$ utilizing the system $\{ P_i^{1,\ell} (t) \}_{i=0}^k$ from the previous section. The positive definiteness of the measure $d\nu_\ell(t)$ implies that $$r_i^{1,\ell}:=\left( \int_{-1}^1 \left( P_i^{1,\ell} (t)\right)^2 \,d\nu_\ell (t) \right)^{-1} >0, \quad i=0,1,\dots,k-1.$$ The three-term recurrence relation from Corollary \[cor\_ortho\] can be written as $$(t-a_i^{1,\ell}) P_i^{1,\ell} (t)=b_i^{1,\ell} P_{i+1}^{1,\ell} (t)+c_i^{1,\ell} P_{i-1}^{1,\ell} (t), \quad i=1,2, \dots ,k-1,$$ where $$P_0^{1,\ell}(t)=1, \ P_1^{1,\ell}(t) = \frac{(n\ell +1)t+\ell +1}{(n+1)\ell +2},$$ $$b_i^{1,\ell}=\frac{\eta_{i+1}^{1,\ell}}{\eta_i^{1,\ell}}>0, \ c_i^{1,\ell}=\frac{r_{i-1}^{1,\ell} b_{i-1}^{1,\ell}}{b_i^{1,\ell}}>0, \ a_i^{1,\ell}=1-b_i^{1,\ell}-c_i^{1,\ell}.$$
By Corollary \[cor\_ortho\] we have that the zeros of $\{ P_i^{1,\ell} (t) \}$ interlace; i.e. $$t_{j,i}^{1,\ell}<t_{j-1,i}^{1,\ell}<t_{j,i+1}^{1,\ell}, \ \ i=1,2,\dots, j-1.$$
We next consider the Christoffel-Darboux kernel (depending on $\ell$) associated with the polynomials $\{ P_j^{1,\ell} \}_{j=0}^k$: $$\begin{aligned}
\label{Chr_kernel_q}
R_i (x,y;\ell) &:=& \sum_{j=0}^i r_j^{1,\ell} P_j^{1,\ell}(x) P_j^{1,\ell} (y) \\
&=& r_i^{1,\ell} b_i^{1,\ell} \frac{P_{i+1}^{1,\ell} (x) P_i^{1,\ell} (y) - P_{i+1}^{1,\ell} (y) P_i^{1,\ell} (x)}{x-y},
\ \ 0\leq i \leq k-1 .\end{aligned}$$ Given $t_{k,k}^{1,0} \leq s \leq \min \{1, t_{k,k}^{1,\ell}\}$, we define $$\begin{aligned}
\label{poly_sub}
P_{k-1}^{1,\ell,s}(t) &:=& \frac{R_{k-1} (t,s;\ell)}{R_{k-1} (1,s;\ell)}\nonumber \\
&=& \frac{1-s}{1- P_k^{1,\ell} (s)/P_{k-1}^{1,\ell} (s)}
\frac{P_k^{1,\ell} (t) - P_{k-1}^{1,\ell} (t) P_k^{1,\ell} (s)/P_{k-1}^{1,\ell} (s) }{t-s}.
\end{aligned}$$
We now define the Levenshtein-type polynomial $$\label{f_{2k}}
f_{2k}^{(n,\ell,s)}(t):= (t-\ell)(t-s)\left( P_{k-1}^{1,\ell,s} (t)\right)^2,$$ and proceed with an investigation its properties.
\[rootsofP1ls\] Let $n$, $\ell$, $s$ and $k$ be such that $t_{k+1,1}^{1,0}<\ell<t_{k,1}^{1,0}$, $P_{k+1}^{1,0}(\ell)/P_{k}^{1,0}(\ell)<1$, $t_{k,k}^{1,0} \leq s \leq t_{k,k}^{1,\ell}$, and $P_k^{1,\ell} (s)/P_{k-1}^{1,\ell} (s)>P_k^{1,\ell}(\ell)/P_{k-1}^{1,\ell}(\ell)$. Then the polynomial $P_{k-1}^{1,\ell,s} (t)$ has $k-1$ simple zeros $\beta_1<\beta_2<\cdots<\beta_{k-1}$ such that $\beta_1 \in (\ell,t_{k-1,1}^{1,\ell})$ and $\beta_{i+1} \in (t_{k-1,i}^{1,\ell},t_{k-1,i+1}^{1,\ell})$, $i=1,2,\ldots,k-2$.
[*Proof.*]{} The proof is similar to that of Theorem \[rootsofP1l\]. It follows from that the roots of the equation $$\frac{P_k^{1,\ell} (t)}{P_{k-1}^{1,\ell} (t)}=\frac{P_k^{1,\ell} (s)}{P_{k-1}^{1,\ell} (s)}$$ are $s$ and the zeros of $P_{k-1}^{1,\ell,s} (t)$, say $\beta_1<\beta_2<\cdots<\beta_{k-1}$.
The function $P_k^{1,\ell} (t)/P_{k-1}^{1,\ell} (t)$ has $k-1$ simple poles at the zeros $t_{k-1,i}^{1,\ell}$, $i=1,2,\ldots,k-1$, of $P_{k-1}^{1,\ell}(t)$. Therefore, there is a zero of $P_{k-1}^{1,\ell,s} (t)$ in each interval $(t_{k-1,i}^{1,\ell},t_{k-1,i+1}^{1,\ell})$, $i=1,2,\ldots,k-2$, which accounts for $k-2$ zeros.
Since $P_k^{1,\ell}(s)/P_{k-1}^{1,\ell} (s)<0$ and the function $P_k^{1,\ell} (t)/P_{k-1}^{1,\ell} (t)$ increases from $-\infty$ to 1 for $t \in (t_{k-1,k-1}^{1,\ell},1]$, we have the root $s$ in this interval. Finally, in the interval $[-\infty,t_{k-1,1}^{1,\ell})$, the function $\frac{P_k^{1,\ell} (t)}{P_{k-1}^{1,\ell} (t)}$ increases from $-\infty$ to $+\infty$ and the condition $P_k^{1,\ell} (s)/P_{k-1}^{1,\ell} (s)>P_k^{1,\ell}(\ell)/P_{k-1}^{1,\ell}(\ell)$ implies that the smallest zero $\beta_1$ of $P_{k-1}^{1,\ell,s} (t)$ lies in the interval $(\ell,t_{k-1,1}^{1,\ell})$. $\Box$
The next theorem is an analog of Theorem 5.39 from [@Lev]. It involves the zeros of $f_{2k}^{(n,\ell,s)} (t)$ to form a right end-point Radau quadrature formula with positive weights.
\[QFtheorem\] Let $\beta_1<\beta_2<\dots<\beta_{k-1}$ be the zeros of the polynomial $P_{k-1}^{1,\ell,s}(t)$. Then the Radau quadrature formula $$\label{QF}
\begin{split}
f_0 &= \int_{-1}^1 f(t) (1-t^2)^{\frac{n-3}{2}}\, dt \\
&=\rho_0 f( \ell )+\sum_{i=1}^{k-1} \rho_i f(\beta_i) +\rho_k f(s)+\rho_{k+1} f(1)=:QF(f)
\end{split}$$ is exact for all polynomials of degree at most $2k$ and has positive weights $\rho_i>0$, $i=0,1,\dots, k+1$.
[*Proof.*]{} Let us denote with $L_i (t)$, $i=0,1,\dots,k+1$, the Lagrange basic polynomials generated by the nodes $\beta_0:=\ell<\beta_1<\dots<\beta_{k-1}<\beta_k:=s<1=:\beta_{k+1}$. Defining $\rho_i:=\int_{-1}^1 L_i(t) d\mu(t)$, $i=0,1,\dots,k+1$, we observe that is exact for the Lagrange basis and hence for all polynomials of degree $k+1$.
We write any polynomial $f(t)$ of degree at most $2k$ as $$f(t)=q(t)(t-\ell)(t-s)(1-t)P_{k-1}^{1,\ell,s} (t)+g(t),$$ where $q(t)$ is of degree at most $k-2$ and $g(t)$ is of degree at most $k+1$. Then the orthogonality of $P_{k-1}^{1,\ell,s}(t)$ to all polynomials of degree at most $k-2$ with respect to the measure $d\nu_{\ell,s}(t)=(t-\ell)(s-t)d\chi(t)$ and the fact that $QF(f)=QF(g)$ show the exactness of the quadratic formula for polynomials up to degree $2k$, namely $$\int_{-1}^1 f(t) d\mu(t) = \int_{-1}^1 g(t) d\mu(t) =QF(g)=QF(f).$$
We next show the positivity of the weights $\rho_i$, $i=0,\dots, k$. Substituting in the polynomial $f(t)=(s-t)(1-t){\left( {P_{k-1}^{1,\ell,s}}(t)\right)}^2$ of degree $2k$, we obtain $$\begin{aligned}
\rho_0 (s-\ell)(1-\ell)\left( P_{k-1}^{1,\ell,s} (\ell)\right)^2=f_0&=&\int_{-1}^1 (s-t)(1-t)\left( P_{k-1}^{1,\ell,s} (t)\right)^2 \, d\mu(t)\\
&=&\int_{-1}^1 \left( P_{k-1}^{1,\ell,s} (t)\right)^2 \, d\nu_s (t)>0 ,\end{aligned}$$ from which we derive $\rho_0 >0$.
To derive that $\rho_i>0$ for $i=1,2,\dots,k-1$, we substitute $$f(t)=(1-t)(t-\ell)(s-t) u_{k-1,i}^2 (t)$$ in , where $u_{k-1,i}(t)=P_{k-1}^{1,\ell,s}(t)/(t-\beta_i)$. Then clearly $$\rho_i (1-\beta_i)(s-\beta_i)(\beta_i-\ell)u_{k-1,i}^2 (\beta_i)=f_0=\int_{-1}^1 u_{k-1,i}^2 (t) \, d\nu_{\ell,s}(t)>0 .$$
Similarly, utilizing the polynomial $f(t)=(1-t)(t-\ell)\left(P_{k-1}^{1,\ell,s}(t)\right)^2$ of degree $2k$ and the positive definiteness of the measure $d\nu_\ell (t)$ up to degree $k-1$ we show that $\rho_k >0$.
Finally, we compute the weight $\rho_{k+1}$ and show that it is positive. In this case we use $f(t)=f_{2k}^{(n,\ell,s)}(t)$ in and easily find that $$\rho_{k+1}=\frac{f_0}{f(1)}=\frac{f_{2k,0}^{(n,\ell,s)}}{f_{2k}^{(n,\ell,s)}(1)}=\frac{f_{2k,0}^{(n,\ell,s)}}{(1-s)(1-\ell)}.$$
Computing $f_0=f_{2k,0}^{(n,\ell,s)}$ using (recall that $P_{k-1}^{1,\ell,s}(1)=1$) we get $$\label{f0}
\begin{split}
f_0&=
\int_{-1}^1 (t-\ell)(s-t)(1-t)P_{k-1}^{1,\ell,s} (t) \frac{P_{k-1}^{1,\ell,s}(t)-P_{k-1}^{1,\ell,s} (1)}{t-1} d\mu(t) \\
& \quad \quad + \int_{-1}^1 (t-\ell)(t-s)P_{k-1}^{1,\ell,s} (t) d\mu(t) \\
&= \frac{1-s}{1-P_k^{1,\ell} (s)/P_{k-1}^{1,\ell} (s)} \int_{-1}^1 (t-\ell)\left( P_k^{1,\ell} (t)-\frac{P_k^{1,\ell} (s)}{P_{k-1}^{1,\ell} (s)} P_{k-1}^{1,\ell} (t) \right) \, d\mu(t).
\end{split}$$ By Lemma \[Pos1\] and the fact that $-P_k^{1,\ell} (s)/P_{k-1}^{1,\ell} (s) >0$ we have that the integrand in is positive definite and in particular its zero-th coefficient (which is the integral in ) is positive. This proves the theorem. $\Box$
For any fixed $-1 < \ell < t_{1,k}^{1,0}$ and $t_{k,k}^{1,0} < s < t_{k,k}^{1,\ell}$ the Levenshtein-type bound is defined to be $$L_{2k}(n;[\ell,s]):=\frac{1}{\rho_{k+1}}=\frac{f_{2k}^{n,\ell,s}(1)}{f_0}=\frac{(1-\ell)(1-s)}{f_0}.$$
Bounding cardinalities and energies
====================================
In the proof of the positive definiteness of his polynomials Levenshtein uses what he called the strengthened Krein condition $$(t+1)P_i^{1,1}(t)P_j^{1,1}(t) \in {\mathcal F}_{>}$$ (see [@Lev (3.88) and (3.92)]). We need a following modification.
We say that the polynomials $\{P_i^{1,\ell}(t)\}_{i=0}^k$ satisfy $\ell$-strengthened Krein condition if $$\label{LSK}
(t-\ell)P_i^{1,\ell}(t)P_j^{1,\ell}(t) \in {\mathcal F}_{>}$$ for every $i,j \in \{0,1,\ldots,k\}$ except for $i=j=k$.
The strengthened Krein condition holds true for every $i$ and $j$ by a classical result of Gasper [@Gas]. However, the $\ell$-strengthened Krein condition is not true for every $\ell$, and for fixed $\ell$, is not true for every $k$.
For fixed $n$ and $k$, denote $$\ell(n,k):=\sup\{ \ell \in [-1,0] : \mbox{$\ell$-strengthened Krein condition holds true} \}.$$
Our computations ensure strong evidence that the following conjecture is true.
\[conj-lsk\] For fixed $n$ and $k$ the condition holds true for every $\ell \in [-1,\ell(n,k)]$.
The Christoffel-Darboux formula $$(t-\ell)P_k^{1,\ell} (t)=\frac{(1-\ell)\left( P_{k+1}^{1,0} (t)-P_k^{1,0} (t)P_{k+1}^{1,0} (\ell)/P_{k}^{1,0} (\ell) \right)}{1-P_{k+1}^{1,0} (\ell)/P_{k}^{1,0} (\ell)}$$ yields easily that the inequality $ P_{k+1}^{1,0} (\ell)/P_{k}^{1,0} (\ell)<1$ is a necessary condition for the $\ell$-strengthened condition to hold. Therefore, we assume from now on that it holds (see the hypothesis of Theorems \[rootsofP1l\] and \[rootsofP1ls\]).
Our computations suggest also that $\ell(n,k)$ is always less (but not much less!) than $t_{k,1}^{1,0}$ and the smallest root of the equation $P_{k+1}^{1,0}(t)/P_{k}^{1,0}(t)=1$. Hence the $\ell$-strengthened Krein condition is stronger than the conditions imposed so far. This corresponds to the Levenshtein’s theory, where the strengthened Krein condition appears to be the most significant obstacle.
The following Lemma demonstrates the reasonableness of Conjecture \[conj-lsk\].
\[Pos1\] The polynomials $\{ (t-\ell) P_i^{1,\ell} (t)\}_{i=0}^{k-1}$ are strictly positive definite provided that $-1\leq \ell <t_{k,1}^{1,0}$.
[*Proof.*]{} From the definition of the kernels $T_i^{1,0} (x,y)$ and we have that $$\begin{aligned}
(t-\ell)P_i^{1,\ell} (t)&=& \frac{(1-\ell)(P_{i+1}^{1,0}(t)P_i^{1,0}(\ell)-P_i^{1,0}(t)P_{i+1}^{1,0}(\ell))}{P_{i+1}^{1,0}(1) P_i^{1,0}(\ell)-P_i^{1,0}(1) P_{i+1}^{1,0}(\ell) }\\
&=&
\frac{1-\ell}{ 1 -P_{i+1}^{1,0}(\ell)/P_i^{1,0}(\ell)}
\left(P_{i+1}^{1,0}(t) -\frac{P_{i+1}^{1,0}(\ell) }{P_i^{1,0}(\ell)} P_i^{1,0}(t) \right)\end{aligned}$$ Since the zeros of $\{ P_i^{1,0} (t) \}$ interlace we have that for all $i \leq k$ the zeros of $P_i^{1,0} (t) $ lie in the interval $[t_{k,1}^{1,0},t_{k,k}^{1,0} ]$ we have that $P_{i+1}^{1,0}(\ell)/P_i^{1,0}(\ell)<0$ for all $i\leq k-1$. Indeed, the numerator and denominator polynomials have different signs on $(-\infty,t_{k,1}^{1,0})$. Since $P_i^{1,0} (t) $ are strictly positive definite (see [@Lev Eq. (3.91)]), we conclude the proof of the Lemma. $\Box$
\[Pos2\] The polynomials $(t-\ell)(t-s)P_{k-1}^{1,\ell,s} (t)$ and $(t-\ell)P_{k-1}^{1,\ell,s} (t)$ are strictly positive definite.
[*Proof.*]{} The interlacing property of $\{P_i^{1,\ell}\}$ implies that $ \frac{P_k^{1,\ell} (s)}{P_{k-1}^{1,\ell} (s)}<0$. Applying Lemma \[Pos1\] one concludes that $(t-\ell)(t-s)P_{k-1}^{1,\ell,s}(t)$ is positive definite. Furthermore, as $P_j^{1,\ell}(s)>0$ and $(t-\ell)P_j^{1,\ell} (t)$ is strictly positive definite for $j=0,1,\dots,k-1$, we derive using that $(t-\ell)P_{k-1}^{1,\ell,s} (t)$ is also strictly positive definite. $\Box$
In our main result we use the $\ell$-strengthened Krein condition relying on the following observation. For fixed $n$, $k$, and $\ell$, we check numerically whether is satisfied for every pair $(i,j)$, $i,j \in \{0,1,\ldots,k\}$, except for $i=j=k$. This is done for every $\ell=-1+m\varepsilon$, $\varepsilon=10^{-3}$ (of course, doing the step $\varepsilon$ smaller is only a matter of computations), $m=1,2,\ldots$, until holds true. In practice, when one needs to compute bounds in the class $C(\ell,s)$, he can consider instead $C(\ell_0,s)$, where $\ell_0<\ell$ is the largest for which the $\ell$-strengthened Krein condition holds true.
The next assertion is the analog of Theorem 5.42 of [@Lev]. It uses a seemingly weaker[^1] version of the $\ell$-strengthened Krein condition.
\[PD\_Lev\] Let $n$, $k$, and $\ell$ be such that the polynomials $(t-\ell)P_i^{1,\ell}(t)P_j^{1,\ell}(t)$ are positive definite for $i \in \{k,k-1\}$ and every $j \leq k-1$. Let $t_{k,k}^{1,0} \leq s \leq t_{k,k}^{1,\ell}$ and $P_k^{1,\ell} (s)/P_{k-1}^{1,\ell} (s)>P_k^{1,\ell}(\ell)/P_{k-1}^{1,\ell}(\ell)$. Then the Levenshtein-type polynomial $f_{2k}^{(n,\ell,s)}(t)$ is positive definite.
[*Proof.*]{} It follows from the definition that the Levenshtein-type polynomial can be represented as follows $$\label{f_lev_pd1}
f_{2k}^{(n,\ell,s)}(t)= c(t-\ell)\left(P_k^{1,\ell} (t)+c_1 P_{k-1}^{1,\ell} (t)\right)
\sum_{i=0}^{k-1} r_i^{1,\ell} P_i^{1,\ell} (t)P_i^{1,\ell}(s),$$ where $c=(1-s)/(1+c_1)R_{k-1}(1,\ell,s)>0$ and $c_1=-P_k^{1,\ell} (s)/P_{k-1}^{1,\ell} (s)>0$ under the assumptions for $\ell$ and $s$. Since $P_i^{1,\ell} (s)>0$ for $0 \leq i \leq k-1$, the polynomial $f_{2k}^{(n,\ell,s)}(t)$ becomes positive linear combination of terms like $(t-\ell)P_i^{1,\ell}(t)P_j^{1,\ell}(t)$, where $i \in \{k,k-1\}$ and $j \leq k-1$. $\Box$
The main result in this paper is the following.
\[main-th-subintervals\] Assume that $\ell \in [-1,t_{k,1}^{1,0})$ and $s \in (t_{k,k}^{1,0},t_{k,k}^{1,\ell})$ and that the $\ell$-strengthened Krein condition holds true. Then $$\label{L-like-bound}
\mathcal{A}(n;[\ell,s]) \leq \frac{f_{2k}^{(n,\ell,s)}(1)}{f_0}=\frac{1}{\rho_{k+1}}.$$ Furthermore, for $h$ being an absolutely monotone function, and for $M$ determined by $f_{2k}^{(n,\ell,s)}(1)=Mf_0$, the Hermite interpolant[^2] $$g(t)=H((t-s)f_{2k}^{(n,\ell,s)}(t);h)$$ belongs to $\mathcal{G}_{n,\ell;h}$, and, therefore, $$\label{ULB-like-bound}
\mathcal{E}(n,M,\ell;h) \geq M(Mg_0-g(1))=M^2\sum_{i=0}^{k} \rho_i h(\beta_i).$$
[*Proof.*]{} We first verify the positive definiteness of the polynomials $f_{2k}^{(n,\ell,s)}(t)$ and $g(t)$. We have $f_{2k}^{(n,\ell,s)}(t) \in \mathcal{F}_>$ by Theorem \[PD\_Lev\].
Denote by $t_1\leq t_2\leq \cdots \leq t_{2k}$ the zeros of $f_{2k}^{(n,\ell,s)}(t)$ counting multiplicity. Observe, that $t_1=\ell$, $t_{2i}=t_{2i+1}=\beta_i$, $i=1,\dots,k-1$, and $t_{2k}=s$. It follows from [@CW Lemma 10] that the polynomial $$g(t):=H((t-s)f_{2k}^{(n,\ell,s)}(t);h)$$ is a linear combination with nonnegative coefficients of the partial products $$\prod_{j=1}^{m} (t-t_j), \ \ m=1, 2,\ldots,2k.$$
Since $t_{2i}$, $i=1,\ldots,k$, are the roots of $P_k^{1,\ell}(t)+\alpha P_{k-1}^{1,\ell}(t)$ (see ) it follows from [@CK Theorem 3.1] that the partial products $\prod_{j=1}^{m} (t-t_{2j})$, $m=1,\ldots,k-1$, have positive coefficients when expanded in terms of the polynomials $P_{i}^{1,\ell}(t)$. Then $g(t)$ is a linear combination with positive coefficients of terms $(t-\ell)P_i^{1,\ell}(t)P_j^{1,\ell}(t)$ and the last partial product which is in fact $f_{2k}^{(n,\ell,s)}(t)$. Now the positive definiteness of $g(t)$ follows from the validity of the $\ell$-strengthened Krein condition and Theorem \[PD\_Lev\].
The bounds and now hold true since $f_{2k}^{(n,\ell,s)}(t) \leq 0 $ for every $t \in [\ell,s]$ (see ) and $g(t) \leq h(t)$ for every $t \in [\ell,1)$ by [@CW Lemma 9].
The expressions of the bounds via the weights $\rho_i$ and the nodes $\beta_i$ follow from Theorem \[QFtheorem\]. $\Box$
Examples and numerical results
==============================
On the $\ell$-strengthened Krein condition
------------------------------------------
In the table below we present our computations of the value of $\ell(n,k)$, the maximum $\ell$ for fixed $n$ and $k$, such that the $\ell$-strengthened Krein condition is true.
[**Table.**]{} Conjectured values of $\ell(n,k)$ for $3 \leq n,k \leq 10$. The rows after the corresponding $\ell(n,k)$ show the value of the smallest root of the equation $P_{k+1}^{1,0}(t)/P_{k}^{1,0}(t)=1$. The real numbers are truncated after the third digit.
\[tab:1\]
$k$ 10 9 8 7 6 5 4 3 2 1
-------------- --------- --------- --------- --------- --------- --------- --------- --------- --------- ---------
$\ell(3,k)$ $-.979$ $-.974$ $-.969$ $-.962$ $-.951$ $-.936$ $-.912$ $-.870$ $-.787$ $-.577$
$-.978$ $-.973$ $-.967$ $-.959$ $-.948$ $-.930$ $-.902$ $-.854$ $-.754$ $-.500$
$\ell(4,k)$ $-.967$ $-.961$ $-.953$ $-.942$ $-.927$ $-.906$ $-.874$ $-.821$ $-.723$ $-.499$
$-.965$ $-.958$ $-.950$ $-.938$ $-.922$ $-.897$ $-.860$ $-.796$ $-.676$ $-.400$
$\ell(5,k)$ $-.955$ $-.947$ $-.936$ $-.923$ $-.905$ $-.879$ $-.840$ $-.779$ $-.672$ $-.447$
$-.952$ $-.944$ $-.932$ $-.917$ $-.896$ $-.866$ $-.821$ $-.748$ $-.615$ $-.333$
$\ell(6,k)$ $-.942$ $-.933$ $-.921$ $-.904$ $-.883$ $-.853$ $-.810$ $-.744$ $-.631$ $-.408$
$-.939$ $-.929$ $-.915$ $-.897$ $-.872$ $-.838$ $-.787$ $-.706$ $-.566$ $-.285$
$\ell(7,k)$ $-.930$ $-.919$ $-.905$ $-.887$ $-.863$ $-.830$ $-.783$ $-.712$ $-.597$ $-.377$
$-.926$ $-.914$ $-.898$ $-.878$ $-.850$ $-.811$ $-.755$ $-.670$ $-.526$ $-.250$
$\ell(8,k)$ $-.918$ $-.906$ $-.890$ $-.870$ $-.843$ $-.808$ $-.758$ $-.685$ $-.568$ $-.353$
$-.914$ $-.900$ $-.882$ $-.859$ $-.828$ $-.787$ $-.727$ $-.638$ $-.492$ $-.222$
$\ell(9,k)$ $-.907$ $-.893$ $-.876$ $-.854$ $-.825$ $-.788$ $-.736$ $-.660$ $-.543$ $-.333$
$-.901$ $-.886$ $-.866$ $-.841$ $-.808$ $-.764$ $-.702$ $-.610$ $-.463$ $-.200$
$\ell(10,k)$ $-.895$ $-.880$ $-.862$ $-.838$ $-.808$ $-.769$ $-.715$ $-.638$ $-.520$ $-.316$
$-.889$ $-.872$ $-.851$ $-.824$ $-.789$ $-.743$ $-.678$ $-.585$ $-.439$ $-.181$
System of bounds for fixed $n$ and $\ell$
-----------------------------------------
We present here as example the system of bounds for $\mathcal{A}(n;[\ell,s])$, where $n=4$ and $\ell=-0.95$ are fixed and $s$ is varying. According to the above table, the $\ell$-strengthened Krein conditions holds true for $k \leq 7$ and corresponding bounds $$\mathcal{A}(4;[-0.95,s]) \leq L_{2k}(4;[-0.95,s])=1/\rho_{k+1}, \ k=1,2,\ldots,7,$$ hold true.
On the figure below we show the first four bounds $$L_{2k}(4;[-0.95,s]), \ k=1,2,3,4,$$ together with the Levenshtein odd degree bounds $L_{2u-1}(4,s)$, $u=1,2,3,4$. The subscripts are missed for short. The behaviour of the bounds is as follows. For $s \in [-0.95,t_1^{1,0}]$, $t_1^{1,0}=-1/4$, the Levenshtein bound $L_1(4,s)$ is better, then for $s \in [t_1^{1,0},0.0175]$ our bound $L_{2}(4;[-0.95,s])=1/\rho_{2}$ is better, for $s \in [-0.0175,t_2^{1,0}]$, $t_2^{1,0} \approx 0.27429$, the Levenshtein bound $L_3(4,s)$ is better, then for $s \in [t_2^{1,0},0.4195]$ our bound $L_{4}(4;[-0.95,s])=1/\rho_{4}$ is better, etc. This is the typical situation for all reasonable values of $n$ and $\ell$ we have checked.
{width="100.00000%"}
Bounds for $\mathcal{E}(n,M,\ell;h)$
------------------------------------
We use the system of bounds from Section 5.2 to derive our ULB-like bounds for $\mathcal{E}(n,M,\ell;h)$. Given $n$, $\ell$, and $M$ we consecutively construct the polynomials $f_{2k}^{n,\ell,s}(t)$ and their bounds as above until we reach the maximum $k$ such that $f_{2k}^{n,\ell,s}(t) \in \mathcal{F}_>$ and the equality $$f_{2k}^{n,\ell,s}(1)=Mf_0$$ holds. Then, as Theorem \[main-th-subintervals\] states, we construct the interpolant $$g(t)=H((t-s)f_{2k}^{(n,\ell,s)}(t);h)$$ and compute the bound $$\mathcal{E}(n,M,\ell;h) \geq M(Mg_0-g(1))=M^2\sum_{i=0}^{k} \rho_i h(\beta_i).$$
System of bounds for fixed $\ell$ and $s$
-----------------------------------------
For the case $k=2$, the Levenshtein polynomial is given by $$f_{4}^{(n,\ell,s)}(t):= (t-\ell)(t-s)\left( P_{1}^{1,\ell,s} (t)\right)^2,$$ where the zero of $P_{1}^{1,\ell,s} (t)$ is $\alpha=-\frac{3+(n+2)(\ell s+\ell+s)}{(n+2)(n\ell s+\ell+s+1)}$. Thus, from Theorem \[main-th-subintervals\] we obtain $$\label{U4}
\mathcal{A}(n;[\ell,s]) \leq \frac{f_{4}^{(n,\ell,s)}(1)}{f_0}=
\frac{n(1-\ell)(1-s)[3+(n+2)(n\ell s+\ell s+2\ell+2s+1)]}{(n+2)[n\ell^2s^2-(\ell-s)^2]-6\ell s+3}$$ subject to $$\ell+s+2\alpha \leq 0,$$ $$\alpha^2+2(\ell+s)\alpha+\ell s+\frac{6}{n+4} \geq 0,$$ $$(\ell+s)\alpha^2+2\alpha(\ell s+\frac{3}{n+2})+\frac{3(\ell+s)}{n+2} \leq 0.$$ The bound is attained by codes of parameters $$(n,M,s)=\left(3m^2-5,\frac{m^4(3m^2-5)}{2},\frac{1}{m+1}\right),$$ known only for $m=2$ (here $\ell=-1$) and 3 (here $\ell=-1/4$). Such codes are derived from corresponding tight spherical 7-designs in dimensions $3m^2-4$ (see [@DGS]).
Let $n$, $M$, and $\ell$ be such that $k=2$ be the maximal value of $k$ such that $f_{2k}^{n,\ell,s}(1)=Mf_0$ holds true and the above $f_{4}^{(n,\ell,s)}(t)$ is positive definite (this fixes $s$ as well). Then, according to Theorem \[main-th-subintervals\], the $h$-energy (for any absolutely monotone $h$) bound is given by the polynomial $g_4(t) \in \mathcal{G}_{n,\ell;h}$ of degree 4 which interpolates $h$ by $$g_4(\ell)=h(\ell), \ g_4(\alpha)=h(\alpha), \ g_4^\prime(\alpha)=h^\prime(\alpha), \
g_4(s)=h(s), \ g_4^\prime(s)=h^\prime(s).$$
In right ranges for $\ell$ and $s$ both bounds are optimal in the sense that they can not be improved by using linear programming with polynomials of degree at most 4.
[**Acknowledgement.**]{} The authors thank Konstantin Delchev, Tom Hanson, and Nikola Sekulov for their computational work that independently verified Conjecture \[conj-lsk\].
[99]{}
A. Bultheel, R. Cruz-Barroso and M. Van Barel, On Gauss-type quadrature formulas with prescribed nodes anywhere on the real line, [*Calcolo*]{} 47 (2010), 21-48.
B.Beckermann, J.Bustamante, R. Martinez-Cruz, and J. Quesada, Gaussian, Lobatto and Radau positive quadrature rules with a prescribed abscissa, [*Calcolo*]{} 51, (2014), 319-328.
S.Borodachov, D.Hardin, and E.Saff, *Minimal Discrete Energy on Rectifiable Sets*, Springer, 2018 (to appear).
P.Boyvalenkov, P.Dragnev, D.Hardin, E.Saff, and M.Stoyanova. Universal lower bounds for potential energy of spherical codes, [*Constr. Approx.*]{} 44, 2016, 385-415.
P.Boyvalenkov, P.Dragnev, D.Hardin, E.Saff, and M.Stoyanova, Energy bounds for codes and designs in Hamming spaces, [*Designs, Codes and Cryptography*]{}, 82(1), (2017), 411-433 (arxiv:1510.03406).
H.Cohn and A.Kumar, Universally optimal distribution of points on spheres, *J. Amer. Math. Soc.* 20, (2007), 99-148.
H.Cohn and J.Woo, Three point bounds for energy minimization, *J. Amer. Math. Soc.* 25, (2012), 929-958.
H.Cohn and Y.Zhao, Energy-minimizing error-correcting codes, *IEEE Trans. Inform. Theory* 60, (2014), 7442-7450 (arXiv:1212.1913).
P. J. Davis and P. Rabinowitz, *Methods of Numerical Integration*, 2nd ed. Academic Press, New York (1984).
P.Delsarte, J.-M.Goethals, and J. J.Seidel, Spherical codes and designs, *Geom. Dedicata* 6, (1977), 363-388.
G. Gasper, Linearization of the product of Jacobi polynomials, II, *Canad. J. Math.* 22, (1970), 582-593.
G. A. Kabatyanskii and V. I. Levenshtein, Bounds for packings on a sphere and in space, [*Probl. Inform. Transm.*]{} 14, (1989), 1-17.
V.I.Levenshtein, Designs as maximum codes in polynomial metric spaces, *Acta Appl. Math.* 25, (1992), 1-82.
V.I.Levenshtein, Universal bounds for codes and designs, *Handbook of Coding Theory*, V.S. Pless and W.C. Huffman, Eds., Elsevier, 1998, Ch. 6, 499–648.
, On extremal polynomials used to estimate the size of codes, [*Probl. Inform. Transm.*]{} 16 (1980), 174-186.
G.Szegő, *Orthogonal polynomials*, Amer. Math. Soc. Col. Publ., [**23**]{}, Providence, RI, 1939.
V.A.Yudin, Minimal potential energy of a point system of charges, *Discret. Mat.* 4, (1992), 115-121 (in Russian); English translation: *Discr. Math. Appl.* 3, (1993), 75-81.
[^1]: In fact, we suspect that both conditions are equivalent.
[^2]: The notation $g=H(f;h)$ is taken from [@CK]; it signifies that $g$ is the Hermite interpolant to the function $h$ at the zeros (taken with their multiplicity) of $f$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The muon decay-at-rest ($\mu$-DAR) facility provides us with an ideal platform to probe purely muonic charged-current non-standard neutrino interactions (NSIs). We probe this class of NSI effects using antineutrinos from a muon decay-at-rest ($\mu$-DAR) source. Even though muonic NSI are absent in neutrino production at the future Tokai to Kamioka superbeam experiment with megaton Hyper Kamiokande detector (T2HK), we show that our proposed hybrid set-up comprising of [$\mu$-DAR]{} and T2HK helps in alleviating the parameter degeneracies that are present in data. Analytic considerations reveal that oscillation probability is most sensitive to the NSI parameter in the $\mu$-e sector. For this parameter, we show, that the $\mu$-DAR setup can improve on the existing bounds down to around 0.01, especially when the data is combined with neutrino data from T2HK experiment due to the lifting of parameter degeneracies. The high precision with which $\mu$-DAR can measure $\delta_{\rm{CP}}$ is shown to be robust even in the presence of the considered NSIs. Finally, we show that the combination of $\mu$-DAR along with T2HK can also be used to put mild constraints on the NSI phase in the vicinity of the maximal CP-violating value.'
author:
- 'Soumya C.'
- Monojit Ghosh
- 'Sushant K. Raut'
- Nita Sinha
- Poonam Mehta
bibliography:
- 'references.bib'
title: 'Probing muonic charged current non-standard interactions at Muon Decay at Rest facility in conjunction with T2HK'
---
Introduction {#sec:intro}
============
Neutrino physics has entered a precision era with a smooth transition (over the past few decades) of the goals, from measuring precisely the parameters of the neutrino mixing matrix to looking for signals of physics beyond the Standard Model (SM). While most of oscillation parameters in the standard three flavour paradigm have been measured successfully to varying degrees of precision (for latest global fit to neutrino data, see [@Esteban:2016qun]), one needs to measure the value of $\delta_{CP}$, figure out the correct octant of $\theta_{23}$ and address the issue of neutrino mass hierarchy. In our quest for physics beyond the SM, some of the currently ongoing and future neutrino experiments are expected to play pivotal role in constraining the new physics parameter space. While effects due to new physics are more intensively looked for at the neutrino experiments, collider experiments allow for a complementary probe of new physics [@Davidson:2011kr].
Many directions have been explored to study new physics in the neutrino sector such as sterile neutrinos, non-standard interactions (NSIs), large extra dimensions, non-unitarity, neutrino decays, violation of CPT and Lorentz symmetry, quantum decoherence etc (see [@Dev:2019anc] and references therein). Of these, NSIs of neutrinos is one of the most widely studied new physics scenarios, an idea which originated in an important paper by Wolfenstein [@Wolfenstein:1977ue] (for reviews, see Refs. [@Miranda:2015dra; @Ohlsson:2012kf; @Farzan:2017xzy; @Dev:2019anc]). Some of the other motivations for NSIs include electroweak leptogenesis [@Pilaftsis:2005rv], neutrino magnetic moment [@BARBIERI1990251; @PhysRevLett.63.228], neutrino condensate as dark energy [@CHOUDHURY1990113] and direct detection of dark matter [@Gonzalez-Garcia:2018dep]. Typically, in the general effective field theory description, the effective Lagrangian can be of the form of charged current (CC) or neutral current (NC) interaction which can alter production, detection and propagation of neutrinos and lead to rich phenomenology. Detailed analyses pertaining to the constraints on these NSI parameters have been summarized in [@Davidson:2003ha; @Biggio:2009nt; @Farzan:2017xzy]. Propagation NSI has been the topic of much recent interest at long baseline accelerator experiments such as Deep Underground Neutrino Experiment (DUNE) (eg: [@Masud:2015xva; @Coloma:2015kiu; @deGouvea:2015ndi; @Blennow:2016etl; @Deepthi:2017gxg; @Masud:2018pig]) as well as reactor (eg: [@Kopp:2007ne]), solar (eg: [@Das:2010sd]) and atmospheric (eg: [@Chatterjee:2014gxa; @Choubey:2015xha]) experiments. While the treatment of propagation NSI does not depend upon the details of particular source or detector type, it should be noted that descripition of NSI at source or detector is governed by the details of the production or detection mechanism at play.
In terrestrial experiments, neutrino production occurs either via beta decay or pion decay processes. Typically, the interactions of neutrinos at source and detector are CC interactions involving both leptons and quarks. The impact of these NSI on neutrino oscillation phenomenology has been studied extensively, for example, in the context of ESS$\nu$SB [@Blennow:2015nxa], DUNE [@Blennow:2016etl; @Bakhti:2016gic], astrophysical neutrinos [@Blennow:2009rp], neutrino factory [@Meloni:2009cg], explaining data from MiniBoone and LSND [@Akhmedov:2010vy] as well as solar and reactor neutrino experiments [@Bolanos:2008km; @Leitner:2011aa; @Khan:2013hva; @Girardi:2014kca; @Khan:2014zwa; @Agarwalla:2014bsa; @Khan:2017oxw]. In addition, neutrinos may be also produced via muon-decay process and it is worthwhile to investigate impact of NSIs on processes involving muons especially in light of the various anomalies (involving muons) such as muon $g-2$ or results involving $B$-meson decays [@Bennett:2006fi; @Albrecht:2018vsa].
Clean measurement of the CP violating phase ${\delta_{\rm CP}}$ is a formidable task especially in the context of superbeam experiments [@Rout:2017udo] and it was proposed that experiments with a Muon Decay at Rest ([$\mu$-DAR]{}) source could help accomplish this [@Alonso:2010fs; @Evslin:2015pya; @Agarwalla:2010nn]. Another proposed experiment for clean measurement of the CP phase is called the MuOn-decay MEdium baseline NeuTrino beam facility (MOMENT) which has neutrino beam of around 300 MeV and high flux [@Cao:2014bea]. In a typical [$\mu$-DAR]{} experiment, ${\ensuremath{\bar{\nu}_\mu}}$ produced from muon decay (with energy around a few $10$s of MeV) oscillates into $\bar \nu_{e}$ and is detected through the inverse beta-decay (IBD) process. It should be noted that in this set-up, not only the flux is very well-known but also the IBD detection cross-section of these neutrinos is large and well-measured. Also, the much smaller systematic uncertainties coupled with large detection efficiency and fewer backgrounds allow for a clean measurement of ${\delta_{\rm CP}}$. In a recent study, it is proposed that the [$\mu$-DAR]{} set-up is useful in addressing the question of neutrino mass hierarchy [@Agarwalla:2017nld].
New physics in the context of [$\mu$-DAR]{} has been studied in [@Ge:2016xya; @Ge:2016dlx; @Ge:2018uhz]. New physics effects have been studied in the context of the MOMENT in [@Tang:2017qen]. In the present work, we attempt to constrain the “CC muonic NSI parameters" at the neutrino production stage using the [$\mu$-DAR]{} set-up. We also demonstrate the impact of combining data from the future Tokai to Kamioka with megaton Hyper Kamiokande detector (T2HK) superbeam experiment [@Abe:2016ero] which aids in lifting parameter degeneracies [@Ghosh:2015ena] even if CC muonic NSI parameters do not play any role in case of neutrino production in T2HK. The CC NSIs at the detection stage involving quarks have been neglected in the present work. In general, one must also consider NC NSIs that affect the propagation of neutrinos through matter. However, such effects are energy dependent, similar to the MSW matter effect. Therefore, at low energies, one may assume them to be negligible.
The outline of this article is as follows. In Sec. \[sec:mudar\], we describe in detail the specifications of the experiments, [$\mu$-DAR]{} and T2HK considered in the present work. Section \[sec:formalism\] contains the formalism for CC NSIs for the case of muon decay with a brief discussion on the existing constraints on relevant parameters. In Sec. \[sec:results\], we present our results and discuss the outcome. Finally, we end with concluding remarks in Sec. \[sec:conc\].
Experimental details : [$\mu$-DAR]{} facility and T2HK {#sec:mudar}
======================================================
For the [$\mu$-DAR]{} facility, we use the configuration described in Ref. [@Evslin:2015pya]. The [$\mu$-DAR]{} source collides low energy protons into a large enough target, thereby, creating charged pions, $\pi^+$ and $\pi^-$ which stop in the target. Of these, the $\pi^-$ may either get absorbed or decay into $\nu_\mu$ and $\mu^-$ that are absorbed in a reasonably high-$Z$ target. While, the $\pi^+$ decay at rest into $\mu^+$ and $\nu_\mu$. The $\mu^+$ thus produced stops and decays at rest, producing $e^+$, $\bar{\nu}_\mu$ and $\nu_e$. The spectra of neutrinos obtained via the process $$\begin{aligned}
\pi^+ &\to& \mu^+ + \nu_\mu
\nonumber\\
&&
{\LARGE{\hookrightarrow}} ~~e^+ + \bar \nu_\mu +\nu_e
\end{aligned}$$ is depicted in Fig. \[fig0\].
{width="50.00000%"}
The [$\mu$-DAR]{} spectrum is known precisely. The [$\mu$-DAR]{} experiment is well-suited to study $\bar{\nu}_\mu \rightarrow \bar{\nu}_e$ oscillations. As the $\bar \nu_e$ have energies around $30 - 50$ MeV, these can be detected via the IBD process with precisely known cross-section. The proposal is to place the [$\mu$-DAR]{} accelerator complex in the southern hills of Toyama about 15 km from the site of the Super-Kamiokande (SK) detector and about 23 km from the Hyper-Kamiokande (HK) detector. The SK detector is a 22.5 kt water Cerenkov detector and the proposed HK detector will consist of two water Cerenkov detectors, each having mass of 187 kt.
The experimental project comprising of neutrinos produced at the J-PARC accelerator and detected by the SK and HK detectors is known as the T2HK project [@Abe:2016ero]. In T2HK, both the SK and HK detectors will be located at a distance of 295 km from the J-PARC accelerator site.
As mentioned in the introduction, our main goal is to constrain the relevant source NSI parameters at [$\mu$-DAR]{} facility. It should be noted that simultaneous operation of [$\mu$-DAR]{} and T2HK is very much feasible since the low energy $\bar \nu_e$ can be distinguished at SK and HK from pulsed higher energy neutrinos from J-PARC. We also explore the impact of adding the data from T2HK on our results. In Table \[tab:specs\], we have given the details of the [$\mu$-DAR]{} set-up [@Evslin:2015pya] and T2HK [@Abe:2016ero].
Experiment Detector mass (kt) Baseline (km) Total p.o.t. Systematic errors
---------------- -------------------- ----------------- ---------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------
[$\mu$-DAR]{} 22.5(SK) + 374(HK) 15(SK) + 23(HK) $1.1 \times 10^{25}$ ($\bar{\nu}$) 5% for both sg and bg (same for both SK and HK).
T2HK 374 295 $27 \times 10^{21}$ ($\nu$) OR $6.75 \times 10^{21}$ ($\nu$) + $20.25 \times 10^{21}$ ($\overline{\nu}$) $\nu$: 3.3% for app and disap, $\bar{\nu}$: 6.2% (4.5%) for app (disap). Errors are same for sg and bg.
NSI in the context of [$\mu$-DAR]{} {#sec:formalism}
===================================
In the SM, the CC interactions of neutrinos are flavour-diagonal by definition. The inclusion of NSI effects can alter this, so that the neutrino produced in association with the charged lepton $\ell_\alpha$ need not be purely $\nu_\alpha$ (as in the case of SI) but can also have an admixture of other flavour $\nu_\beta$ proportional to the (subdominant) NSI term $\varepsilon^{\mu e}_{\alpha\beta}$. Thus, up to a normalization factor, we have $$\nu^s_\alpha = \sum_\beta (\delta_{\alpha\beta} + \varepsilon^{\mu e}_{\alpha\beta}) \nu_\beta ~.$$ The superscript $s$ stands for production NSIs at the source, while the label $\mu e$ on the NSI parameters indicates that these neutrinos are produced in muon-decay in conjunction with a muon and an electron. The effective operator responsible for muon decay is $\sim G_F \ \bar{e} \gamma_\rho P_L \nu_e^s \ \bar{\nu}_\mu^s \gamma^\rho P_L \mu$, which modifies the muon-decay neutrino flux as [@Antusch:2006vwa] $$\Phi^{NSI} = \Phi^{SM} \left( 1+2{\textrm{Re}({\varepsilon^{\mu e}_{ee}})} \right) \left( 1+2{\textrm{Re}({\varepsilon^{\mu e}_{\mu\mu}})} \right) ~.$$ Two neutrinos of unknown flavour are produced, of which the antineutrino $\bar{\nu}_\mu^s$ is detected after oscillation while the neutrino $\nu_e^s$ escapes undetected. Therefore the probability of $\bar{\nu}^s_\mu \to \bar{\nu}_\rho$ is $$\begin{aligned}
P(\bar{\nu}^s_\mu \to \bar{\nu}_\rho) & = & \sum_X P(\bar{\nu}^s_\mu \nu^s_e \to \bar{\nu}_\rho \nu_X) \nonumber \\
& = & \sum_X \left| \sum_{\beta,\alpha} (\delta_{\mu\beta} + \varepsilon^{\mu e *}_{\mu\beta}) \bar{\mathcal{A}}_{\beta \rho} (\delta_{e\alpha} + \varepsilon^{\mu e}_{e\alpha}) \mathcal{A}_{\alpha X} \right|^2 \nonumber \\
& = & \sum_X \left[ \sum_{\beta,\alpha} (\delta_{\mu\beta} + \varepsilon^{\mu e *}_{\mu\beta}) \bar{\mathcal{A}}_{\beta \rho} (\delta_{e\alpha} + \varepsilon^{\mu e}_{e\alpha}) \mathcal{A}_{\alpha X} \right] \left[ \sum_{\delta,\gamma} (\delta_{\mu\delta} + \varepsilon^{\mu e}_{\mu\delta}) \bar{\mathcal{A}}^*_{\delta \rho} (\delta_{e\gamma} + \varepsilon^{\mu e *}_{e\gamma}) \mathcal{A}^*_{\gamma X} \right] ~, \nonumber
\label{eq:pmeformal}\end{aligned}$$ where $\mathcal{A}_{\alpha\beta}$ (without/with bar) denotes the standard amplitude for the process $\nu_\alpha \to \nu_\beta$ (for neutrinos/antineutrinos). From the completeness of neutrino states, we have $ \sum_X \mathcal{A}_{\alpha X} \mathcal{A}^*_{\gamma X} = \delta_{\alpha\gamma}$, and hence $$\begin{aligned}
P(\bar{\nu}^s_\mu \to \bar{\nu}_\rho) & = & \left( 1+2{\textrm{Re}({\varepsilon^{\mu e}_{ee}})} \right) \left| \sum_\beta (\delta_{\mu\beta} + \varepsilon^{\mu e *}_{\mu\beta}) \bar{\mathcal{A}}_{\beta \rho} \right|^2 \nonumber \\
& \approx & (1+2{\textrm{Re}({\varepsilon^{\mu e}_{ee}})}) \left[ (1+2{\textrm{Re}({\varepsilon^{\mu e}_{\mu\mu}})}) \bar{P}_{\mu \rho} + 2{\textrm{Re}({\varepsilon^{\mu e}_{\mu e}}\bar{\mathcal{A}}_{e\rho} \bar{\mathcal{A}}_{\mu\rho})} + 2{\textrm{Re}({\varepsilon^{\mu e}_{\mu\tau}}\bar{\mathcal{A}}_{\tau\rho} \bar{\mathcal{A}}_{\mu\rho} )} \right] ~.\end{aligned}$$ It is easy to check that $P(\bar{\nu}^s_\mu \to \bar{\nu}_e) + P(\bar{\nu}^s_\mu \to \bar{\nu}_\mu) + P(\bar{\nu}^s_\mu \to \bar{\nu}_\tau) = (1+2{\textrm{Re}({\varepsilon^{\mu e}_{ee}})})(1+2{\textrm{Re}({\varepsilon^{\mu e}_{\mu\mu}})})$, therefore the neutrino states in presence of NSIs (and hence the probability) should be appropriately normalized. These are the same factors $(1+2{\textrm{Re}({\varepsilon^{\mu e}_{ee}})})(1+2{\textrm{Re}({\varepsilon^{\mu e}_{\mu\mu}})})$ that appear in the modified neutrino flux. Instead of normalizing the probability [and]{} multiplying the same factor to the SM flux, we choose to do neither with the understanding that these factors cancel out in the calculation of event rates [@Antusch:2006vwa]. Notice that only the parameters ${\varepsilon^{\mu e}_{\mu e}}$, ${\varepsilon^{\mu e}_{\mu\mu}}$ and ${\varepsilon^{\mu e}_{\mu\tau}}$ are relevant in the normalized probability formula.
The formalism described above is completely general and is applicable to any muon-decay neutrino source where one of the neutrinos goes undetected. In the specific case of [$\mu$-DAR]{} facility, the only relevant oscillation channel is $P(\bar{\nu}^s_\mu \to \bar{\nu}_e)$, since the low energy neutrino cannot produce a $\mu$ or $\tau$ lepton in the detector. Therefore, the relevant formula for the probability (before normalization) is $$P(\bar{\nu}^s_\mu \to \bar{\nu}_e) = (1+2{\textrm{Re}({\varepsilon^{\mu e}_{ee}})}) \left[ (1+2{\textrm{Re}({\varepsilon^{\mu e}_{\mu\mu}})}) \bar{P}_{\mu e} + 2{\textrm{Re}({\varepsilon^{\mu e}_{\mu e}}\bar{\mathcal{A}}_{ee} \bar{\mathcal{A}}_{\mu e})} + 2{\textrm{Re}({\varepsilon^{\mu e}_{\mu\tau}}\bar{\mathcal{A}}_{\tau e} \bar{\mathcal{A}}_{\mu e} )} \right] ~.$$
An analytical formula for the oscillation probability in the presence of NSIs can be calculated perturbatively in the small parameters $\alpha={\Delta m^2_{21}}/{\Delta m^2_{31}}$, $\sin\theta_{13}$ and $\varepsilon^{\mu e}_{\alpha\beta}$ [@Kopp:2007ne]. Every $\varepsilon^{\mu e}_{\alpha\beta}$ can be complex and therefore introduces two new real parameters – a magnitude $|\varepsilon^{\mu e}_{\alpha\beta}|$ and the corresponding phase $\phi^{\mu e}_{\alpha\beta}$. In general, the formula also depends on the Mikheyev-Smirnov-Wolfenstein (MSW) matter potential, $A$. However, since we will restrict our discussion to experiments with short baselines and low energies, it suffices to consider the vacuum oscillation formula and study the role of NSI parameters. Up to first order in the NSI parameters, we have $$\begin{aligned}
P^{NSI} \equiv P(\nu^s_\mu \to \nu_e) & = & 4 \sin^2\theta_{13} \sin^2\theta_{23} \sin^2\Delta \nonumber \\
& + & 2\alpha\sin\theta_{13} \sin 2\theta_{12} \sin 2\theta_{23} \cos(\Delta+{\delta_{\rm CP}}) \Delta \sin\Delta \nonumber \\
& - & 4 |{\varepsilon^{\mu e}_{\mu e}}| \sin\theta_{13} \sin\theta_{23} \sin(\Delta+{\delta_{\rm CP}}+{\phi^{\mu e}_{\mu e}}) \sin\Delta \nonumber \\
& - & 2 |{\varepsilon^{\mu e}_{\mu e}}| \alpha \sin 2\theta_{12} \cos\theta_{23} \sin{\phi^{\mu e}_{\mu e}}\Delta ~,
\label{eq:pme}\end{aligned}$$ where $\Delta = {\Delta m^2_{31}}L/4E$. For antineutrinos (which are relevant for our study), we make the replacements ${\delta_{\rm CP}}\to-{\delta_{\rm CP}}$ and ${\phi^{\mu e}_{\mu e}}\to-{\phi^{\mu e}_{\mu e}}$. The last two terms of the formula show the effect of NSIs. Up to lowest order, the formula involves ${\varepsilon^{\mu e}_{\mu e}}$ but the parameters ${\varepsilon^{\mu e}_{\mu\mu}}$ and ${\varepsilon^{\mu e}_{\mu\tau}}$ do not appear. Therefore, we expect the sensitivity of this channel to be better for ${\varepsilon^{\mu e}_{\mu e}}$ than the other two parameters. Apart from modulating the amplitude of the probability spectrum, the NSI also introduces an additional source of CP violation through the phase. We will refer back to this formula to explain features of the experimental sensitivity. However, it should be noted that all the numerical results presented in the present work are exact.
The $\varepsilon^{\mu e}_{\alpha \beta}$ can be constrained from (a) tests of lepton universality, i.e., by comparing the Fermi constant from muonic and electronic processes, and (b) the non-observation of zero-distance flavour conversion at experiments like KARMEN [@Eitel:2000by] and NOMAD [@Astier:2003gs]. This analysis was carried out in Ref. [@Biggio:2009nt] (assuming one non-zero NSI parameter at a time) and the following $90\%$ C.L. bounds were obtained on the magnitudes of the NSIs: $$|\varepsilon^{\mu e}_{\alpha \beta}| < \left(
\begin{array}{ccc}
0.025 & 0.030 & 0.030 \\
0.025 & 0.030 & 0.030 \\
0.025 & 0.030 & 0.030
\end{array}
\right)$$ In our study, we will constrain the parameters (${\varepsilon^{\mu e}_{\mu e}}$, ${\varepsilon^{\mu e}_{\mu\mu}}$, and ${\varepsilon^{\mu e}_{\mu\tau}}$) using the [$\mu$-DAR]{} set-up and compare them to their existing bounds (0.025, 0.030, 0.030, respectively).
---------------------------------------------- ------------------------ -----------------------
Oscillation Parameter True value Marginalization range
$\theta_{12}$ \[$^\circ$\] $33.5$ -
$\theta_{13}$ \[$^\circ$\] $8.5$ -
$\theta_{23}$ \[$^\circ$\] $45$ $40 - 50$
${\Delta m^2_{21}}$ \[$\textrm{eV}^2$\] $7.5 \times 10^{-5}$ -
${\Delta m^2_{31}}$ (NH) \[$\textrm{eV}^2$\] $+2.45 \times 10^{-3}$ -
${\Delta m^2_{31}}$ (IH) \[$\textrm{eV}^2$\] $-2.46 \times 10^{-3}$
${\delta_{\rm CP}}$ \[$^\circ$\] $-90$ $-180 - 180$
---------------------------------------------- ------------------------ -----------------------
: Oscillation parameters in the standard three flavour paradigm used in our study [@Esteban:2016qun]. []{data-label="tab:parameters"}
Results {#sec:results}
=======
{width="98.00000%"}
![The bounds on source NSIs at [$\mu$-DAR]{} facility (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. The neutrino mass hierarchy is assumed to be NH (IH) in the upper (lower) panel.[]{data-label="fig:newPnue"}](DAR_mu_e-ME.pdf "fig:"){width="5.4cm" height="5.5cm"} ![The bounds on source NSIs at [$\mu$-DAR]{} facility (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. The neutrino mass hierarchy is assumed to be NH (IH) in the upper (lower) panel.[]{data-label="fig:newPnue"}](DAR_mu_e-MM.pdf "fig:"){width="5.4cm" height="5.5cm"} ![The bounds on source NSIs at [$\mu$-DAR]{} facility (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. The neutrino mass hierarchy is assumed to be NH (IH) in the upper (lower) panel.[]{data-label="fig:newPnue"}](DAR_mu_e-MT.pdf "fig:"){width="5.4cm" height="5.5cm"}\
![The bounds on source NSIs at [$\mu$-DAR]{} facility (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. The neutrino mass hierarchy is assumed to be NH (IH) in the upper (lower) panel.[]{data-label="fig:newPnue"}](IH-DAR_mu_e-ME.pdf "fig:"){width="5.4cm" height="5.5cm"} ![The bounds on source NSIs at [$\mu$-DAR]{} facility (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. The neutrino mass hierarchy is assumed to be NH (IH) in the upper (lower) panel.[]{data-label="fig:newPnue"}](IH-DAR_mu_e-MM.pdf "fig:"){width="5.4cm" height="5.5cm"} ![The bounds on source NSIs at [$\mu$-DAR]{} facility (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. The neutrino mass hierarchy is assumed to be NH (IH) in the upper (lower) panel.[]{data-label="fig:newPnue"}](IH-DAR_mu_e-MT.pdf "fig:"){width="5.4cm" height="5.5cm"}\
In the present study, we study the impact of NSIs in the context of a [$\mu$-DAR]{} facility, which was conceived for clean and precise measurement of the parameters (especially ${\delta_{\rm CP}}$) in standard neutrino oscillation scenario [@Alonso:2010fs; @Evslin:2015pya; @Agarwalla:2010nn]. In particular, we address the following questions which are relevant -
1. How well we can measure or constrain the relevant NSI parameters at such a facility ?
2. What is the impact of NSIs on the measurement of the standard CP phase, ${\delta_{\rm CP}}$ ?
Let us begin with a discussion of (a) which is about the constraints that can be placed on the relevant NSI parameters at a [$\mu$-DAR]{} facility. We end this section with a discussion of the second question in (b) which pertains to testing the robustness of the precision with which ${\delta_{\rm CP}}$ can be measured at such a facility.
The numerical computation is carried out using the General Long Baseline Experiment Simulator (GLoBES) [@Huber:2004ka] package along with the new physics plugin for sterile neutrinos and NSIs [@Kopp:2006wp]. The neutrino oscillation parameters used in our analysis are given in Table \[tab:parameters\][^1].
In order to comprehend our results better, let us first attempt to understand the expected sensitivity at the probability level to the NSI parameters at the [$\mu$-DAR]{} set-up. In Fig. \[fig1\], we plot in the plane of $|\varepsilon^{\mu e}_{\mu \beta}| - \phi^{\mu e}_{\mu \beta}$, the relative absolute difference of the probabilities $({| P_{SI} - P_{NSI} |}/{P_{SI}})$ where $P_{SI}$ ($P_{NSI}$) and the oscillation probability with SI (NSI) effects. The top (bottom) row depicts the effect at SK (HK) which corresponds to a baseline of 15 (23) km and energy of 35 (45) MeV. The three columns correspond to the different source NSI parameters, ${\varepsilon^{\mu e}_{\mu e}}$, ${\varepsilon^{\mu e}_{\mu\mu}}$ and ${\varepsilon^{\mu e}_{\mu\tau}}$ respectively.
The first observation is that the [$\mu$-DAR]{} is most sensitive to the source NSI parameter, $|{\varepsilon^{\mu e}_{\mu e}}|$. The sinusoidal nature of the curve in the first column can be easily understood from the $\sin{\phi^{\mu e}_{\mu e}}$ dependence in Eq. \[eq:pme\] at the oscillation maximum (when $\Delta=\pi/2$) and for the currently favoured value of ${\delta_{\rm CP}}=-90^\circ$. From Eq. \[eq:pmeformal\], we see that the probability depends only on the real part of ${\varepsilon^{\mu e}_{\mu\mu}}$, hence the second column shows a $\cos{\phi^{\mu e}_{\mu\mu}}$ dependence, with maximum at ${\phi^{\mu e}_{\mu\mu}}=0,180^\circ$. This same feature is also seen for ${\varepsilon^{\mu e}_{\mu\tau}}$ in the last column. We will see later that these characteristic shapes are also reflected in the $\chi^2$ analysis.
Next, we discuss the bounds on the NSI parameters. We present our results for both normal hierarchy (NH) and inverted hierarchy (IH), and assume in each case that the hierarchy is known from other experiments, i.e., we do not marginalize over hierarchy. To obtain a bound on any of the NSI parameters using [$\mu$-DAR]{} facility and either SK or HK, we generate the true event spectrum with the standard oscillation framework using the parameters listed in Table \[tab:parameters\] and compare it with a test event spectrum which is simulated with NSI scenario with only one of the NSI parameters taken non-zero at a time. We show the 90% C.L. exclusion regions in the plane of $|\varepsilon_{\alpha\beta}^{\mu e}|-\phi_{\alpha\beta}^{\mu e}$ for each of the three sectors ($\mu - e $, $\mu- \mu$ and $\mu -\tau$ respectively) as shown in Fig. \[fig:newPnue\]. The total $\chi^2$ is obtained by adding the individual contributions from [$\mu$-DAR]{} at SK and [$\mu$-DAR]{} at HK for each point in the test parameter space, $$\chi^2_{total}=\chi^2_{\textrm{{$\mu$-DAR}@}SK}+ \chi^2_{\textrm{{$\mu$-DAR}@}HK} \nonumber$$ [We add a Gaussian prior for $\sin^2\theta_{23}$ with $\sigma(\sin^2\theta_{23}=0.015)$ and marginalize over ${\delta_{\rm CP}}$ and $\sin^2\theta_{23}$ in their allowed in Table \[tab:parameters\]]{}[^2]. The bounds are presented for four different true values of ${\delta_{\rm CP}}$ i.e., $-180^\circ$, $-90^\circ$, $0^\circ$ and $90^\circ$. The left, middle and the right panels correspond to the NSI parameters ${\varepsilon^{\mu e}_{\mu e}}$, ${\varepsilon^{\mu e}_{\mu\mu}}$ and ${\varepsilon^{\mu e}_{\mu\tau}}$ respectively. In all the three panels, the black dotted vertical line indicates the present 90% C.L. upper bound of the parameter.
From Fig. \[fig:newPnue\], we can make the following observations. The value of the NSI phase plays an important role as the sensitivity to NSI in a given sector is very much dependent on the value of the NSI phase.
- $\mu -e$ sector : The left panel depicts the bounds on $|{\varepsilon^{\mu e}_{\mu e}}|$ as a function of ${\phi^{\mu e}_{\mu e}}$. Marginalization over the unknown test value of ${\delta_{\rm CP}}$ results in degeneracy between ${\delta_{\rm CP}}$ and ${\phi^{\mu e}_{\mu e}}$ at higher values of $|{\varepsilon^{\mu e}_{\mu e}}|$, resulting in the features seen in the plot. Consequently, values of ${\phi^{\mu e}_{\mu e}}$ in the upper half plane are more constraining for this sector, especially for ${\delta_{\rm CP}}=\pm90^\circ$. We obtain an upper bound, $|{\varepsilon^{\mu e}_{\mu e}}| < 0.012$ for ${\delta_{\rm CP}}=\pm 90^\circ$. This is clearly an improvement over the existing bounds.
It is worth mentioning that if the value of ${\delta_{\rm CP}}$ is assumed to be known and fixed at its true value, the bound oscillates mildly [[between 0.01 and 0.02]{}]{} for all values of ${\phi^{\mu e}_{\mu e}}$. A similar effect will be seen later in Fig. \[fig:combined\], where addition of T2HK data leads to shrinking of contours due to lifting of the parameter degeneracies, thereby leading to results that are fixed parameter like.
- $\mu -\mu$ sector : From the middle panel, we note that for ${\varepsilon^{\mu e}_{\mu\mu}}$ the sensitivity is better at CP conserving values, ${\phi^{\mu e}_{\mu\mu}}= 0^\circ$, $180^\circ$, but the bounds obtained from [$\mu$-DAR]{} facility are weaker than the existing ones.
- $\mu -\tau$ sector : The conclusions are similar to those for the $\mu - \mu $ sector as mentioned above.
- The bounds for IH as seen in the lower panels are qualitatively similar to the ones for NH, with variations depending on the choice of ${\delta_{\rm CP}}$ and NSI phase.
The results of the analysis can be improved by constraining the values of the other oscillation parameters (in particular, ${\delta_{\rm CP}}$) that affect the probability in the multi-dimensional parameter space. It should be noted that during the running of [$\mu$-DAR]{} facility, the T2HK experiment will be also taking data. Therefore, we discuss the possibility of adding the T2HK data to the [$\mu$-DAR]{} data to improve the sensitivity to considered NSI parameters. At T2HK, neutrinos are produced from pion decay and are hence unaffected by the $\varepsilon^{\mu e}_{\alpha\beta}$ parameters which are specific to muon decay production mechanism. However, its ability to discriminate between true and spurious degenerate regions in the standard oscillation parameter space leads to a synergy with [$\mu$-DAR]{} facility and can improve the sensitivity to the NSI parameters. In the case of T2HK experiment, we simulate both true and test event spectra without new physics and obtain the $\chi^2$ by comparing these event spectra. The total $\chi^2$ is obtained by adding the individual contributions from [$\mu$-DAR]{} at SK, [$\mu$-DAR]{} at HK and T2HK for each point in the test parameter space, $$\chi^2_{total}=\chi^2_{\textrm{{$\mu$-DAR}@}SK}+ \chi^2_{\textrm{{$\mu$-DAR}@}HK}+ \chi^2_{T2HK} \nonumber$$
![The bounds on source NSIs at T2HK + [$\mu$-DAR]{} (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. Top row corresponds to T2HK running fully in neutrino mode bottom row corresponds to when the neutrino to antineutrino running ratio of T2HK is 1:3. []{data-label="fig:combined"}](T2HK_NU-DAR_mu_e-ME.pdf "fig:"){width="5.5cm" height="5.5cm"} ![The bounds on source NSIs at T2HK + [$\mu$-DAR]{} (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. Top row corresponds to T2HK running fully in neutrino mode bottom row corresponds to when the neutrino to antineutrino running ratio of T2HK is 1:3. []{data-label="fig:combined"}](T2HK_NU-DAR_mu_e-MM.pdf "fig:"){width="5.5cm" height="5.5cm"} ![The bounds on source NSIs at T2HK + [$\mu$-DAR]{} (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. Top row corresponds to T2HK running fully in neutrino mode bottom row corresponds to when the neutrino to antineutrino running ratio of T2HK is 1:3. []{data-label="fig:combined"}](T2HK_NU-DAR_mu_e-MT.pdf "fig:"){width="5.5cm" height="5.5cm"}\
![The bounds on source NSIs at T2HK + [$\mu$-DAR]{} (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. Top row corresponds to T2HK running fully in neutrino mode bottom row corresponds to when the neutrino to antineutrino running ratio of T2HK is 1:3. []{data-label="fig:combined"}](T2HK-DAR_mu_e-ME.pdf "fig:"){width="5.5cm" height="5.5cm"} ![The bounds on source NSIs at T2HK + [$\mu$-DAR]{} (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. Top row corresponds to T2HK running fully in neutrino mode bottom row corresponds to when the neutrino to antineutrino running ratio of T2HK is 1:3. []{data-label="fig:combined"}](T2HK-DAR_mu_e-MM.pdf "fig:"){width="5.5cm" height="5.5cm"} ![The bounds on source NSIs at T2HK + [$\mu$-DAR]{} (SK and HK) in $\mu-e$, $\mu- \mu$, and $\mu-\tau$ sectors are respectively given in left, middle and right panels. Top row corresponds to T2HK running fully in neutrino mode bottom row corresponds to when the neutrino to antineutrino running ratio of T2HK is 1:3. []{data-label="fig:combined"}](T2HK-DAR_mu_e-MT.pdf "fig:"){width="5.5cm" height="5.5cm"}\
We present our results in Fig. \[fig:combined\]. In the top row, we given the results with T2HK running fully in the neutrino mode with the addition of [$\mu$-DAR]{} and in the bottom row we have given the results when the neutrino to antineutrino runtime ratio at T2HK is 1:3 with the addition of [$\mu$-DAR]{}. In each row the left, middle, and right panel corresponds to NSI parameter ${\varepsilon^{\mu e}_{\mu e}}$, ${\varepsilon^{\mu e}_{\mu\mu}}$, and ${\varepsilon^{\mu e}_{\mu\tau}}$ respectively. From the figure, we observe that the addition of T2HK data significantly improves the bounds on the NSI parameter $|{\varepsilon^{\mu e}_{\mu e}}|$. This is achieved due to the synergy between the [$\mu$-DAR]{} facility and T2HK experiment in constraining ${\delta_{\rm CP}}$. The resulting bounds lie between 0.01 and 0.02 depending on the values of the two phases, which is better than the existing bound over most of the parameter space. There is also an improvement in the bounds on $|{\varepsilon^{\mu e}_{\mu\mu}}|$ and $|{\varepsilon^{\mu e}_{\mu\tau}}|$, significantly for values of ${\delta_{\rm CP}}$ in the lower half-plane. However, for these two parameters the bounds do not improve the existing ones. This is expected from the fact that these play a sub-leading role, as discussed in Sec. \[sec:formalism\]. The results for ${\varepsilon^{\mu e}_{\mu\mu}}$ and ${\varepsilon^{\mu e}_{\mu\tau}}$ can be improved slightly by changing the neutrino to antineutrino ratio, as seen in the lower panels. [The results for T2HK neutrino to antineutrino ratio of 1:3 is better than T2HK pure neutrino run because of the fact that the antineutrino run helps to resolve the octant degeneracy and hence improving the ${\delta_{\rm CP}}$ measurement [@Ghosh:2014zea].]{}
![Range of allowed values of ${\delta_{\rm CP}}$ as a function of its true value. The true values of NSI parameters are ${\varepsilon^{\mu e}_{\mu e}}=0.01$ and ${\phi^{\mu e}_{\mu e}}=0$. Left panel: fit against oscillations in the presence of NSI; Right panel: fit against standard oscillations.[]{data-label="fig:dcpprec"}](nh-NSI-t2hk+mdr_me-phi-0.pdf "fig:"){width="45.00000%"} ![Range of allowed values of ${\delta_{\rm CP}}$ as a function of its true value. The true values of NSI parameters are ${\varepsilon^{\mu e}_{\mu e}}=0.01$ and ${\phi^{\mu e}_{\mu e}}=0$. Left panel: fit against oscillations in the presence of NSI; Right panel: fit against standard oscillations.[]{data-label="fig:dcpprec"}](nh-NSI-SI-t2hk+mdr_me-phi-0.pdf "fig:"){width="45.00000%"}
We now address the question pertaining to the robustness of ${\delta_{\rm CP}}$-precision, i.e., what is the impact of inclusion of additional NSI parameters on the measurement of standard CP violation at the combined set-up T2HK(1:3)+[$\mu$-DAR]{} facility. The range of allowed values of ${\delta_{\rm CP}}$ is shown as a function of its true value in Fig. \[fig:dcpprec\]. The true values of NSI parameters are chosen to be ${\varepsilon^{\mu e}_{\mu e}}=0.01$ and ${\phi^{\mu e}_{\mu e}}=0$ for illustrative purposes. From the left panel, we see the result of fitting the simulated data against the test hypothesis including the NSI. The allowed values of ${\delta_{\rm CP}}$ are very closely correlated with its true value, signalling the robustness of the ${\delta_{\rm CP}}$-measurement at this set-up. However, when data from this set-up is first analysed, it is likely to be fitted against standard oscillations without NSIs. The result of such a fit is shown in the right panel. Once again, we find allowed regions around the true value of ${\delta_{\rm CP}}$. Any strong degenerate effects in the parameter space would show up as allowed regions away from the $y=x$ line in these panels. Their absence indicates the robustness of ${\delta_{\rm CP}}$-measurement. In addition, the mild tension in the data (absence of $1\sigma$ allowed region close to CP-conserving values of ${\delta_{\rm CP}}$) is an indicator of new physics. Similar results can be generated for various chosen values of ${\phi^{\mu e}_{\mu e}}$. In Table \[tab:prec\], we list the precision in ${\delta_{\rm CP}}$ for various choices of the true values of ${\delta_{\rm CP}}$ and ${\phi^{\mu e}_{\mu e}}$.
-- ----- ----------------------- --------------- --------------------
Standard oscillations
T2HK+[$\mu$-DAR]{} [$\mu$-DAR]{} T2HK+[$\mu$-DAR]{}
-90 103 65
0 77 57
90 59 49
180 65 52
-90 88 19
0 60 30
90 46 27
180 59 30
-90 108 61
0 83 55
90 66 50
180 71 52
-90 92 32
0 65 30
90 47 27
180 59 30
-- ----- ----------------------- --------------- --------------------
![The correlation between $\delta_{{\rm CP}}$ and $\phi_{\mu e}$ in the test parameter plane. The true value of $\delta_{{\rm CP}}$ is assumed to be -90$^\circ$ in all cases. Whereas, the true value of $\phi_{\mu e}$ is assumed to be 90$^\circ$, 180$^\circ$ respectively in top left and right panels and -90$^\circ$, 0$^\circ$ respectively in bottom left and right panels. The neutrino mass hierarchy is assumed to be normal.[]{data-label="fig:correlation"}](T2HK-DAR-ME-DCP--90-PHI-90.pdf "fig:"){width="40.00000%"} ![The correlation between $\delta_{{\rm CP}}$ and $\phi_{\mu e}$ in the test parameter plane. The true value of $\delta_{{\rm CP}}$ is assumed to be -90$^\circ$ in all cases. Whereas, the true value of $\phi_{\mu e}$ is assumed to be 90$^\circ$, 180$^\circ$ respectively in top left and right panels and -90$^\circ$, 0$^\circ$ respectively in bottom left and right panels. The neutrino mass hierarchy is assumed to be normal.[]{data-label="fig:correlation"}](T2HK-DAR-ME-DCP--90-PHI-180.pdf "fig:"){width="40.00000%"}\
![The correlation between $\delta_{{\rm CP}}$ and $\phi_{\mu e}$ in the test parameter plane. The true value of $\delta_{{\rm CP}}$ is assumed to be -90$^\circ$ in all cases. Whereas, the true value of $\phi_{\mu e}$ is assumed to be 90$^\circ$, 180$^\circ$ respectively in top left and right panels and -90$^\circ$, 0$^\circ$ respectively in bottom left and right panels. The neutrino mass hierarchy is assumed to be normal.[]{data-label="fig:correlation"}](T2HK-DAR-ME-DCP--90-PHI--90.pdf "fig:"){width="40.00000%"} ![The correlation between $\delta_{{\rm CP}}$ and $\phi_{\mu e}$ in the test parameter plane. The true value of $\delta_{{\rm CP}}$ is assumed to be -90$^\circ$ in all cases. Whereas, the true value of $\phi_{\mu e}$ is assumed to be 90$^\circ$, 180$^\circ$ respectively in top left and right panels and -90$^\circ$, 0$^\circ$ respectively in bottom left and right panels. The neutrino mass hierarchy is assumed to be normal.[]{data-label="fig:correlation"}](T2HK-DAR-ME-DCP--90-PHI-0.pdf "fig:"){width="40.00000%"}
Finally, we address the question of measurement of the NSI phase itself. We do so by finding the allowed region in the plane of ${\delta_{\rm CP}}$ and ${\phi^{\mu e}_{\mu e}}$ as shown in Fig \[fig:correlation\] for $|{\varepsilon^{\mu e}_{\mu e}}|=0.01$. While the true value of ${\delta_{\rm CP}}$ is always chosen to be $-90^\circ$, we choose four representative values for ${\phi^{\mu e}_{\mu e}}=-90^\circ,0^\circ,90^\circ,180^\circ$ in the four panels. The 90% C.L. contours are shown for [$\mu$-DAR]{} alone, as well as in conjunction with T2HK. The synergy between [$\mu$-DAR]{} and T2HK data helps to constrain the parameter space when ${\phi^{\mu e}_{\mu e}}=-90^\circ$ although T2HK itself is insensitive to muonic NSIs. The light band in the background represents the allowed range of ${\delta_{\rm CP}}$ from T2HK+[$\mu$-DAR]{} in the absence of NSIs. We find that for a large part of the parameter space, the precision in ${\delta_{\rm CP}}$ which is given by the thickness of the allowed region does not get altered appreciably in the presence of NSIs, thereby pointing towards the robustness of the CP-measurement. The ${\phi^{\mu e}_{\mu e}}$-precision of our set-up is better for the CP-violating values of ${\phi^{\mu e}_{\mu e}}=\pm 90^\circ$ than at the CP-conserving values, ${\phi^{\mu e}_{\mu e}}= 0,180^\circ$. This is in contrast to the standard CP measurement which is better at ${\delta_{\rm CP}}= 0,180^\circ$. This fact can this be explained on the basis of the top panels in Fig. \[fig1\]. Since the difference between probabilities in the two scenarios (SI and NSI) is highest around the maximally CP-violating values ${\phi^{\mu e}_{\mu e}}=\pm90^\circ$, the precision with which this parameter can be resolved is better.
Summary and Conclusions {#sec:conc}
=======================
Beyond SM scenarios that arise at a high energy scale can manifest themselves in low energy neutrino phenomena as NSIs. Thus, the search for NSIs in the neutrino sector is complementary to collider searches for new physics. At MeV scale, the NC NSIs in the neutrino propagation can be safely ignored, allowing us to focus on the CC NSIs that affect production/detection processes. In this work, we concentrate specifically on the muonic NSI parameters i.e., $\varepsilon^{\mu e}_{\alpha\beta}$ which uniquely impact the production of neutrinos at a [$\mu$-DAR]{} source.
We compute the bounds that a [$\mu$-DAR]{} set-up can place on the magnitude of the relevant NSI parameters $\varepsilon^{\mu e}_{\mu\beta}$. For the parameter ${\varepsilon^{\mu e}_{\mu e}}$, the 90% C.L. bounds are competitive with existing bounds from tests of lepton universality for around half the range of the phase ${\phi^{\mu e}_{\mu e}}$. Addition of data from the superbeam experiment, T2HK substantially improves the bounds to better than the current ones for most of the parameter space. This is surprising, given that neutrinos produced at T2HK are not at all expected to be affected by the muonic NSIs. This result can only be explained by invoking a synergy between T2HK and [$\mu$-DAR]{} in constraining the standard oscillation parameters, thus lifting parameter degeneracies that hinder the measurement of NSIs.
Next, we show that the precise measurement of ${\delta_{\rm CP}}$ at the [$\mu$-DAR]{} set-up is robust even in the presence of NSIs. The precision in this parameter does not worsen considerably because of any parameter degeneracies even for the most unfavourable combinations of the two CP-violating phases - the standard Dirac phase ${\delta_{\rm CP}}$ and the non-standard phase, ${\phi^{\mu e}_{\mu e}}$.
Finally, we discuss correlations between ${\delta_{\rm CP}}$ and ${\phi^{\mu e}_{\mu e}}$. We find that the precision in ${\phi^{\mu e}_{\mu e}}$ is limited, and a substantial range of values of this phase can be excluded when it is close to $\pm 90^\circ$. In conclusion, we find that the [$\mu$-DAR]{} set-up is well-suited to measure ${\delta_{\rm CP}}$ not only in the standard oscillation scenario but also in the presence of CC muonic NSIs at source. In addition, this set-up can also constrain the magnitude and phase of the NSI parameters depending upon its true value.
This work was initiated during the working group discussions at Workshop on High Energy Physics Phenomenology (WHEPP) held at IIT Kanpur during December 2015 and the authors would like to thank the organisers for the vibrant atmosphere and warm hospitality during the meeting. SR acknowledges support by IBS (Project Code IBS-R018-D1), Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2019R1A6A1A10073887) and NRF Strategic Research Program (NRF2017R1E1A1A01072736). The work of PM is supported by the Indian funding from University Grants Commission under the second phase of University with Potential of Excellence (UPE II) at JNU and Department of Science and Technology under DST-PURSE at JNU. PM would like to acknowledge partial funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skodowska-Curie grant agreement No 690575 and 674896.
[^1]: We have checked that the oscillation parameters that significantly impact the results are $\theta_{23}$ and ${\delta_{\rm CP}}$. Marginalization over ${\Delta m^2_{31}}$ does not affect the results much and therefore we keep it fixed.
[^2]: Since at a [$\mu$-DAR]{} facility, it is not possible to study $\nu_\mu \to \nu_\mu$ oscillations, we have added a prior on $\sin^2\theta_{23}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present two data-driven methods for estimating reachable sets with probabilistic guarantees. Both methods make use of a probabilistic formulation allowing for a formal definition of a data-driven reachable set approximation that is correct in a probabilistic sense. The first method recasts the reachability problem as a binary classification problem, using a Gaussian process classifier to represent the reachable set. The quantified uncertainty of the Gaussian process model allows for an adaptive approach to the selection of new sample points. The second method uses a Monte Carlo sampling approach to compute an interval-based approximation of the reachable set. This method comes with a guarantee of probabilistic correctness, and an explicit bound on the number of sample points needed to achieve a desired accuracy and confidence. Each method is illustrated with a numerical example.'
author:
-
bibliography:
- 'refs.bib'
title: '**Data-Driven Reachable Set Computation using Adaptive Gaussian Process Classification and Monte Carlo Methods** '
---
Introduction
============
Reachable sets characterize the states to which a system may evolve using the knowledge of where it starts, what inputs may affect the system, and how long the system may evolve. Computing reachable sets is a critical step in the solution to control problems involving objectives such as safety, recurrence, and more complicated requirements expressed as automata or temporal logic specifications. However, accurate reachable sets are generally very expensive to compute, and common practice is to use a tractable relaxation, such as an overapproximation that is guaranteed to contain the true reachable set.
In relaxing the problem, the analyst must make a trade-off between computational tractability and accuracy of the overapproximation. There are many reachable set overapproximation methods that lie at different points of the tractability-accuracy spectrum. At one extreme, reachability methods based on the Hamilton-Jacobi-Bellman equations [@MitchellTomlinHJProjections; @TomlinHJGames] and dynamic programming [@BertsekasMinimax], such as those used in the Level Set Toolbox [@mitchell2005toolbox], yield reachable set approximations that are very accurate but slow to compute. Zonotope-based methods [@althoff2008verification], such as those used in the CORA toolbox[@althoff2015introduction], are faster to compute at the cost of some accuracy. At the opposite extreme, interval reachability methods [@meyer2019tira; @meyer2017hierarchical; @moor2002abstraction] give overapproximations that require a minimum of resources to compute and store, but due to their strict geometry they are generally conservative.
In this paper we introduce a *data-driven* approach that allows for improvements in both tractability and accuracy, at the cost of a relaxed guarantee of correctness. The essence of this relaxation is to place a suitable probability measure over the initial set and the controls, and to define reachable sets as events on the induced probability space. Then, a sample of simulated system trajectories can be used to make probabilistic estimates of the true reachable set. To achieve the lowest computational complexity possible, we minimize the number of sample trajectories while maintaining a probabilistic guarantee of a given accuracy.
Probabilistic methods have been used to analyze the reachability of stochastic systems [@SastryStochasticHybridSystems; @althoff2008stochastic; @margellos2014road; @yang2016multi] and as an exploratory tool to guide deterministic reachability analysis [@LygerosATCMC]. Here, we investigate the probabilistic approach as a rigorous method in its own right to analyze the reachability of deterministic systems. Data-driven methods have also been used as a tool for robustness analysis of uncertain control systems [@tempo2012randomized], which allow for probabilistic verification of robustness against various types of uncertainty. This paper provides a similar approach to the problem of reachable set computation.
We present two data-driven methods for computing reachable set approximations that make use of the probabilistic relaxation. The first method uses a Gaussian process classifier (GPC) to construct a probabilistic reachable set of arbitrary accuracy. The prediction uncertainty of the GPC allows us to employ an *active learning* method [@settles2009active], where we sequentially select samples in order to maximize information gain. The second method uses a Monte Carlo sampling approach to construct an interval overapproximation of the probabilistic reachable set. Although less accurate, this method comes with a provable probabilistic guarantee. The two methods are complementary: the GPC method allows for approximations of higher accuracy (since it is not restricted to interval approximations), while the Monte Carlo method can make faster approximations. When probabilistic guarantees are acceptable for the problem at hand, the formalism and methods described in this paper can offer a significant computational speedup. An additional advantage of the data-driven approach is that it may be used in a model-free way: we need only to be able to sample system trajectories, so the system itself is allowed to be a black box or otherwise inaccessible. Indeed, many high-fidelity models are either available only in black-box form, or are too complex to analyze with standard reachability tools.
Reachable Sets
==============
Suppose we have a dynamical system with state transition function $\Phi(t; t_0, x_0, u)$ that maps an initial state $x_0\in{\ensuremath{\mathbb{R}}}^n$ at time $t_0$ to a unique final state at time $t_1$, under the influence of an input $u\in C^{[t_0,t_1]}$ and the system dynamics. For example, if the system is defined as a vector ordinary differential equation $$\dot{x}(t)=f(x(t),u(t),t) \label{eq:vector_ode}$$ whose solutions are well-defined and unique on the interval $[t_0,t]$, then $\Phi(t; t_0, x_0, u)$ is the solution to (\[eq:vector\_ode\]) satisfying the initial condition $\Phi(t_0; t_0, x_0, u)=x_0$.
Now, suppose we have an *initial set* ${\ensuremath{\mathcal{X}_0} }\subset{\ensuremath{\mathbb{R}}}^n$, and an *input set* ${\ensuremath{\mathcal{U}} }\subset C^{[t_0,t_1]}$. We would like to know all of the states to which the system may evolve between times $[t_0, t_1]$ starting in the initial set, and subjected to any allowable input. The set of all such states is the *forward reachable set*, which we write as $$R_{[t_0,t_1]}=\{x | x=\Phi(t_1;t_0, x_0, u)\text{ for some } x_0\in{\ensuremath{\mathcal{X}_0} }, u\in{\ensuremath{\mathcal{U}} }\}.$$ When the state transition function is invertible, we also consider the inverse of this problem. Suppose we have a *final set* set ${\ensuremath{\mathcal{X}_1} }\subset{\ensuremath{\mathbb{R}}}^n$, and we would like to know all of the states that can reach [$\mathcal{X}_1$ ]{}in the time $[t_0,t_1]$. The set of all such states is called the *backward reachable set*, that is $$B_{[t_0,t_1]}=\{x | \Phi(t_1;t_0, x, u)\in{\ensuremath{\mathcal{X}_1} }\text{ for some } u\in{\ensuremath{\mathcal{U}} }\}.$$ We may also be interested in finding the set [$\mathcal{X}_e$ ]{}of all initial states for which some *event*, characterized by $h(x, t, u)=0$, occurs at some time $t_e\ge t_0$. The set of all such states is called the *event set*, that is $$E_{t_0}=\{x | h(\Phi(t_e;t_0, x, u), t_e, u(t_e))=0 \mbox{ for some }t_e \ge t_0, u\in{\ensuremath{\mathcal{U}} }\}.$$ This is similar to the backwards reachable set problem, except that $t_e$ is not known *a priori*. Further, $t_e$ will in general not be the same for each state that leads to the event.
Probabilistic Reachable Sets
============================
To frame the data-driven approach, we consider a *probabilistic relaxation* of the reachable set problems described above. The methods in this paper consist of sampling initial states and inputs, evaluating the transition function at these sample points, and using the results to estimate the reachable set. The state transition function may be available directly through numerical integration of (\[eq:vector\_ode\]), through more advanced computer simulations, or even through physical experiments.
A reachable set computed using a sample-based method can be at best only *probabilistically accurate*, so we would like a way to represent this notion as well. To formalize the notion of sampling from [$\mathcal{X}_0$ ]{}, we define a random variable $X_0 \sim p_0$ over the initial set. The probability distribution $p_0: {\ensuremath{\mathcal{X}_0} }\to [0,1]$ is called the *initial distribution*, and may be any distribution whose support is [$\mathcal{X}_0$ ]{}. Similarly, we will define a random variable $U\sim p_u$ over the input set, with *input distribution* $p_u:{\ensuremath{\mathcal{U}} }\to[0,1]$.
These two random variables, together with the state transition function, define the family of *successor random variables* $X_t=\Phi(t;t_0, X_0, U)\sim p_t$ for $t\ge t_0$. In general, the distribution $p_t$ will be unknown, since the state transition function is not known. The successor distribution can be used to define a probability space whose sample space is the state space ${\ensuremath{\mathbb{R}}}^n$, whose events are the Borel sets of ${\ensuremath{\mathbb{R}}}^n$, and whose probability measure is $p_t$. In this probability space, the probability of an event $\omega$ corresponds to the probability that the successor of a random initial state and input is an element of $\omega$. This means that the true forward reachable set $R_{[t_0,t_1]}$ corresponds to the smallest event of probability 1. With that in mind, we define the *$\epsilon$-accurate reachable sets*, denoted $R_{[t_0,t_1],\epsilon}$ as the smallest events with probability $1-\epsilon$. A set $R\subset{\ensuremath{\mathbb{R}}}^n$ such that $p_t(R)\ge 1-\epsilon$ is an *overapproximation* of an $\epsilon$-accurate reachable set, since it must contain an $\epsilon$-accurate reachable set. The relationship between the deterministic and probabilistic cases for forward reachable sets for a two-state system is illustrated in Figure \[fig:reachability-cartoon\].
![A diagram of a forward reachable set (upper graph), and its relaxation to an $\epsilon$-accurate probabilistic forward reachable set (lower graph), and an overapproximation of the probabilistic reachable set. []{data-label="fig:reachability-cartoon"}](frs "fig:"){width="45.00000%"} ![A diagram of a forward reachable set (upper graph), and its relaxation to an $\epsilon$-accurate probabilistic forward reachable set (lower graph), and an overapproximation of the probabilistic reachable set. []{data-label="fig:reachability-cartoon"}](frs-prob "fig:"){width="45.00000%"}
We define a similar probabilistic formulation for backward reachable sets. The only difference is that we will choose a final random variable $X_1$ and $U$, and let $X_0=\Phi^{-1}(t;t_0, X_1, U)$, where $$\Phi^{-1}(t;t_0,x_1,u)=\{x | x_1=\Phi(t;t_0,x,u),u\in{\ensuremath{\mathcal{U}} }\}.$$
For event sets, we are not interested specifically in the probabilistic behavior of $\Phi$, but instead in the likelihood that a given sample in an initial set will lead to the event. Essentially, we would like to use samples to inform our belief about the location of the event set, so it is sensible to adopt a Bayesian formulation for the probabilistic event set. We employ a distribution over the initial set, $p_0(x)$, which represents our belief that $x$ is in the event set prior to seeing any samples. Then the posterior distribution conditioned on the sample trajectories represents an updated belief that the point $x$ belongs to the event set that takes information from the sample trajectories into account. We call this posterior distribution the *event distribution*, $p_e$.
Gaussian Process Classification (GPC) With Adaptive Sampling
============================================================
The first method we present uses a Gaussian stochastic process to construct a binary classifier that identifies an estimate of the reachable set. A point in the state space is either in the reachable set or out of it, so determining the set of points in the reachable set has a natural representation as a binary classification problem.
A Gaussian process $g$ is a random variable defined over a space of functions with the property that the joint distribution of any finite selection of point evaluations of the function is distributed as a joint Gaussian random variable [@gpml].The covariance between any two point evaluations $g(x_1)$ and $g(x_2)$ is $k(x_1,x_2)$, where $k$ is the *kernel function* of the process.
Suppose we have a set of $m$ sample points $x^{(i)}\in{\ensuremath{\mathbb{R}}}^n$ and their associated labels $y^{(i)}\in\{a,b\}$, where $a,b\in{\ensuremath{\mathbb{R}}}$, such that $y^{(i)}=a$ if $x^{(i)}$ is in the reachable set, and $y^{(i)}=b$ otherwise. For example, a suitable choice of labels would be $a=1, b=0$. We use a Gaussian process to construct a classifier that minimizes the *regularized least-squares classification risk*, that is a function $g:{\ensuremath{\mathbb{R}}}^n\to{\ensuremath{\mathbb{R}}}$ that minimizes $$\sum_{i=1}^{m} \left(g(x^{(i)})-y^{(i)}\right)^2 + ||g||_{k}$$ where $||\cdot||_{k}$ is a norm that depends on the kernel function. A classifier that minimizes this risk is called a *least-squares classifier*. Least-squares classification is attractive here because the mean $\mu_{\hat{g}}$ and variance $\sigma_{\hat{g}}$ of the Gaussian process $\hat{g}$ that minimizes this risk have analytic expressions that can be computed quickly.
To make predictions using this classifier, we select a *threshold* $\gamma\in (a,b)$, and declare that a point $x$ is predicted to be in the reachable set if $\hat{g}(x)\ge\gamma$, and not in the reachable set otherwise. For example, in the $a=1,b=0$ case, $\gamma=0.5$ is suitable. With a threshold chosen, the reachable set estimate produced by this method is the sublevel set $$\hat{R}=\{ x | \mu_{\hat{g}}(x) < \gamma\}.$$ To construct a data set, we select a set of sample points $x^{(i)}$, and use the state transition function to assign a label $y^{(i)}$ to each of the sample points based on whether or not it is in the reachable set.
In principle, we may select the sample points in any way we like, e.g. uniform sampling over the region of interest, or using Latin hypercube sampling. However, since we wish to minimize the number of transition function evaluations, we use the GPC model of the reachable set to inform our choice of future sample points. This kind of sampling is called *adaptive sampling*, since our selection method adapts according to the incoming data, and is an *active learning* method. The use of adaptive sampling to guide the construction of a Gaussian process model is motivated by a method from optimal experiment design known as *Adaptive Kriging* [@ak-mcs; @schobi-pck], in which a Gaussian process regression is used to form a *surrogate model* for an expensive computational model.
To have the GPC inform our selection of sample points, we use the prediction uncertainty of the GPC to sequentially select sample points that maximize some measure of information gain. The most relevant measure of information gain for this problem is the *probability of misclassification* [@bect2012sequential], that is the probability that a state in the reachable set is identified as being outside of it, or vice versa. With classifier threshold $\gamma$, the probability of misclassification is $$P_{misclass}(x) = \Phi\left( -\frac{|\mu_{\hat{g}}(x) - \gamma|}{\sigma_{\hat{g}}(x)} \right)
\label{eq:p_misclass}$$ where $\Phi$ in (\[eq:p\_misclass\]) is the cumulative distribution function of the standard normal distribution.
When selecting a new sample point, ideally we would like to find the point in the state space with the highest probability of misclassification. However, this is a nonconvex and potentially high-dimensional optimization problem. Instead of searching the entire state space for a new sample, we use a stochastic optimization approach proposed in [@ak-mcs] and search over a large pool of randomly-selected *candidate samples*. We calculate the probability of misclassification for each candidate, and select the one with the highest probability of misclassification to be the next sample. The sample pool is selected using a Latin Hypercube, so that the candidate samples will be evenly distributed over a compact region of the state space. Note that this is distinct from selecting samples directly by a Latin hypercube: after we have selected the candidate pool, only a small number of candidate points will be selected as sample points, and the distribution of the selected points will be guided by the probability of misclassification.
Example: Safe Set Estimation for Adaptive Cruise Control {#example-safe-set-estimation-for-adaptive-cruise-control .unnumbered}
--------------------------------------------------------
![Estimated safe set boundaries (dashed lines) for the ACC model computed with the GPC method compared with the true safe set boundary (solid lines), which is calculated analytically. The sample locations are also shown: an ‘o’ indicates that a collision occurred, and an ‘x’ indicates that it did not. The model parameters are set at $a=4.9$, $b=1$, and the initial follower velocity is fixed at $v_F(0)=5$. Top row: $m=50$ sample points. Bottom row: $m=200$ sample points. For adaptive sampling, a candidate pool of $m_{candidate}=1000$ samples was used in both cases. For the two sample sizes shown, adaptive sampling is able to make the most accurate approximation of the event set out of the three sampling methods used.[]{data-label="fig:safeset_gpc"}](safeset-m50.png "fig:"){width="\textwidth"} ![Estimated safe set boundaries (dashed lines) for the ACC model computed with the GPC method compared with the true safe set boundary (solid lines), which is calculated analytically. The sample locations are also shown: an ‘o’ indicates that a collision occurred, and an ‘x’ indicates that it did not. The model parameters are set at $a=4.9$, $b=1$, and the initial follower velocity is fixed at $v_F(0)=5$. Top row: $m=50$ sample points. Bottom row: $m=200$ sample points. For adaptive sampling, a candidate pool of $m_{candidate}=1000$ samples was used in both cases. For the two sample sizes shown, adaptive sampling is able to make the most accurate approximation of the event set out of the three sampling methods used.[]{data-label="fig:safeset_gpc"}](safeset-m200.png "fig:"){width="\textwidth"}
Consider the Adaptive Cruise Control (ACC) scenario depicted in Figure \[fig:acc\_diagram\]. In this scenario, a car being operated by ACC (the *follower*) is driving behind another car (the *leader*). The follower and leader are initially traveling with positive velocities $v_F(0)$ and $v_L(0)$ respectively. At $t=0$, the leader begins to brake and eventually comes to a halt. If the distance between the leader and follower becomes zero at any $t>0$, then the two cars have collided. To prevent this, we determine what initial states (that is, velocities and relative positions at $t=0$) give the follower enough time to prevent a collision. We call the set of all such initial states a “safe set”.
We use the following point-mass model for the dynamics of the two vehicles: $$\begin{aligned}
\dot{h}(t) &= v_{L}(t)-v_{F}(t) \\
\dot{v}_{L}(t) &= -a - b v_{L}(t)^2\\
\dot{v}_{F}(t) &= -a - b v_{F}(t)^2 \end{aligned}$$ where $v_L(t)$ and $v_F(t)$ are the velocities of the leader and follower, respectively, and $h(t)$ is the distance between the two cars. The acceleration of each car has a constant term from the brakes, as both cars are applying the brakes fully, and a quadratic term from drag force.
![Diagram of the leader and follower, and the associated state variables, in the ACC braking model. If $h(t_e)=0$ for some $t_e\ge t_0$, the two cars have collided.[]{data-label="fig:acc_diagram"}](ACCsafety4CDCpaper.png){width="45.00000%"}
This problem is an event set estimation problem because the safe set we wish to determine corresponds to the complement of the set of initial conditions $x(0)= \begin{bmatrix} h(0) & v_{L}(0) & v_{F}(0) \end{bmatrix}^T$ for which the event $h(x)=h=0$ occurs.
This model can be solved analytically, and the true event set is shown in the Appendix to be
$$E = \{ (h,v_L,v_F)|
h
+ \frac{1}{2b}\log\left(1 + \frac{b}{a}v_L^2 \right)
- \frac{1}{2b}\log\left(1 + \frac{b}{a}v_F^2 \right)
\ge 0
\}.$$
Since we know the true event set, we can directly observe how well the GPC method approximates the true event set under different conditions.
For convenience of visualization, we hold the initial velocity of the follower constant at $v_{F}(0)=0.5$. We restrict our attention to a compact region of the state space, specifically $$\begin{aligned}
0 \le h &\le 2\\
0 \le v_{L} &\le 5.\end{aligned}$$
Using sample points from this region, we construct a least-squares GPC using a *squared-exponential* kernel, that is we take $$k(x_1,x_2)=\sigma \exp(-(x_2-x_1)^T \Lambda (x_2-x_1)),$$ where $\sigma$ and the diagonal matrix $\Lambda=\text{diag}(\ell_1,\ell_2,\ell_3)$ are *hyperparameters* that are selected using maximum likelihood.
To demonstrate the effectiveness of the adaptive sampling method, we compare it to two other non-adaptive sampling strategies: sampling uniformly at random over the region, and sampling with a Latin hypercube over the region. To demonstrate how the number samples affects the quality of the predicted event set, we form two sample sets for each of the sampling methods, with $m=50$ and $m=200$ each.
For adaptive sample selection, we begin by selecting a pool of $m_{candidate}=1000$ candidate samples from the region of interest using a Latin hypercube. Three samples are selected at random to serve as the initial set for the GPC model, and the remaining $m-3$ are selected by sequentially minimizing the probability of misclassification.
The GPC-estimated event sets are shown in Figure \[fig:safeset\_gpc\]. The true reachable set is also shown, to confirm that the estimated reachable sets are converging to the ground truth. For both sample sizes, the adaptive sampling method makes the most accurate event set estimate out of each of the three methods. By maximizing the probability of misclassification with each new sample, the adaptive method will either select a new sample with high prediction variance, which will be far away from the other samples, or one whose prediction mean is close to the threshold; that is, one close to the border. After minimizing the variance over most of the state space region of interest, the adaptive method begins to select samples that are likely to be near the border of the event set. Using this strategy, the adaptive GPC reduces the number of samples required to identify points near the boundary of the event set. By contrast, the uniform and Latin hypercube methods can only select states at random, so the likelihood of selecting a sample point near the boundary of the event set never increases.
[ We used a point-mass model here so that the true event set could be derived, as has been done in the Appendix. However, this same analysis could be carried out with a high-fidelity model with no change to the methodology. ]{}
Monte Carlo Interval Overapproximation
======================================
We now present a Monte Carlo Sampling (MCS) approach to produce *interval* overapproximations of epsilon-accurate reachable sets, that is overapproximations of the form $$\hat{R} = [\underline{x}, \overline{x}]= \{x | \underline{x} \le x \le \overline{x}, \underline{x}\in{\ensuremath{\mathbb{R}}}^n,\overline{x}\in{\ensuremath{\mathbb{R}}}^n\}$$ where $\le$ is the vector inequality corresponding to the positive orthant cone of ${\ensuremath{\mathbb{R}}}^n$. Geometrically, the set $[\underline{x}, \overline{x}]$ is an axis-aligned hyperrectangle of dimension $n$ whose least point is $\underline{x}$ and whose greatest point is $\overline{x}$. In general, even the tightest interval overapproximation of the reachable set will be inaccurate, since intervals have such restricted geometry. Despite the inaccurate nature of their approximation of the true reachable set, interval approximations can be an appropriate design choice when quick computation and low memory requirements are preferable to reachable set accuracy.
An important example of when interval approximation is a suitable design choice is *symbolic control*, where controller synthesis is carried out on a finite-state machine *abstraction* that simulates the continuous-state dynamical system [@lunze1994qualitative; @alur2000discrete; @moor2002abstraction; @tabuada; @gazit11; @belta; @7519063]. The states of the abstraction represent the cells of a partition of ${\ensuremath{\mathbb{R}}}^n$, and the transitions are derived from the intersection of the forward reachable sets of each cell with the other cells. For high-dimensional state spaces, the number of reachable sets that must be computed and stored grows rapidly, so it is necessary to use a reachable set overapproximation that can be computed quickly and is memory-efficient. A simple method to calculate the interval approximation is a Monte Carlo approach. For the forward reachable set case, this would consist of the following steps:
1. take a set of $m$ samples each from the initial distribution and input distribution, $\{x_0^{(i)}\}_{i=1}^{m}$ and $\{u^{(i)}\}_{i=1}^{m}$;
2. Evaluate the sample successor states $x_1^{(i)}=\Phi(t;t_0,x_0^i,u^i)$;
3. Take $\hat{R}^{(m)}$ as the smallest interval containing all of the $x_1^{(i)}$.
Despite its simplicity, the Monte Carlo Sampling (MCS) method described above is provably effective at overapproximating $\epsilon$-accurate reachable sets with intervals. In particular, the inequality (\[eq:mc\_sample\_bound\]), adapted from an example in [@Vidyasagar] serves as a lower bound on the number of sample points required to ensure that the method described above produces an overapproximation of a desired accuracy and confidence. Although similar in form, the sample complexity bound (\[eq:mc\_sample\_bound\]) is distinct from the Chernoff bounds commonly used in the analysis of Monte Carlo methods, in ways that will be discussed after the proof.
Let $\epsilon$, $\delta\in (0,1)$. If $$m \ge \frac{2n}{\epsilon}\log\left(\frac{2n}{\delta}\right),
\label{eq:mc_sample_bound}$$ then $\hat{R}^{(m)}$ overapproximates an $\epsilon$-accurate reachable set with confidence $\delta$, i.e. $P(R_{[t_0,t_1],\epsilon} \subset \hat{R}^{(m)}) \ge 1-\delta$.
First, suppose that we have $m$ samples of the successor $\{x_1^{(i)}\}_{i=1}^m$, and that the resulting $\hat{R}^{(m)}$ overapproximates the $\epsilon$-accurate reachable set.
An $n$-dimensional interval has $2n$ “faces”, which are all axis-aligned. To each “face” $f_i$ of the interval we associate the half space $\mathcal{H}_i$ whose boundary hyperplane coincides with that face, and faces “away from” the interval.
Now, let $\mathcal{P}_i$ be a half space parallel to $\mathcal{H}_i$ such that $p_{t_1}(\mathcal{P}_i)\ge\frac{\epsilon}{2n}$. Furthermore, let $\mathcal{P}_i$ be the *smallest* such half space, in the sense that any other half space $\mathcal{P}$ parallel to $\mathcal{H}_i$ such that $p_{t_1}(\mathcal{P}_i)\ge\frac{\epsilon}{2n}$ contains $\mathcal{P}_i$. By the right-continuity of $p_1$, this half space is unique. In the case that $p_1$ is continuous, this is just the unique hyperplane parallel to $\mathcal{H}_i$ such that $p_{t_1}(\mathcal{P}_i)=\frac{\epsilon}{2n}$.
The probability that none of the $m$ samples is in the half space $\mathcal{P}_i$ is $(1-\frac{\epsilon}{2n})^m$. This implies that the probability that at least one of the $\mathcal{P}_i$ contains no samples less than $2n(1-\frac{\epsilon}{2n})^m$, and that the probability that each $\mathcal{P}_i$ contains at least one sample is greater than $1-2n(1-\frac{\epsilon}{2n})^m$.
If one of the samples is in $\mathcal{P}_i$, then $\mathcal{H}_i\subset\mathcal{P}_i$, meaning that $p_{t_1}(\mathcal{H}_i) \le p_{t_1}(\mathcal{P}_i)\le\frac{\epsilon}{2n}$. If this is true for each $\mathcal{P}_i$, then $p_{t_1}(\bigcup_i \mathcal{H}_i)=p_{t_1}(\hat{R}^{(n)\complement})\le\epsilon$, and so $p_{t_1}(\hat{R}^{(i)})\ge 1-\epsilon$. This implies that $$P(p_{t_1}(\hat{R}^{(i)})\ge 1-\epsilon) \ge 1-2n(1-\frac{\epsilon}{2n})^m.$$ To ensure that $p_{t_1}(\hat{R}^{(i)})\ge 1-\epsilon)$, i.e. that $\hat{R}^{(m)}$ overapproximates contains an $\epsilon$-accurate reachable set, with probability $\ge 1-\delta$, it suffices to ensure $$\delta \ge 2n\left(1-\frac{\epsilon}{2n}\right)^m.
\label{eq:mc_implicit_bound}$$ Using the identity $\log(1-x) \le -x$, valid for $0\le x \le 1$, we can conclude that (\[eq:mc\_sample\_bound\]) guarantees (\[eq:mc\_implicit\_bound\]).
For standard Monte Carlo methods, such as Monte Carlo Integration, probability inequalities such as *Chernoff bounds* can be used to determine the number of samples needed for a desired accuracy. However, Chernoff bounds are not applicable in this case. Chernoff bounds apply to the tail probability of a sum of iid random variables (e.g. a Monte Carlo integral) exceeding an arbitrary parameter, and are derived by applying Markov’s inequality to the moment generating function of the random sum. In our case, however, we are interested in bounding the tail probability of the *maximum* of an iid sequence of random variables exceeding a specific observed value (the empirical maximum). The strategy for deriving a Chernoff bound does not work in this case, which is why a different strategy was needed to prove Theorem 1.
Example: Robustness Analysis For a Powered Lower Limb Orthosis Through Forward Reachable Sets {#example-robustness-analysis-for-a-powered-lower-limb-orthosis-through-forward-reachable-sets .unnumbered}
---------------------------------------------------------------------------------------------
![Diagram of the orthosis and its user, from [@na]. State variables are labeled in red and green. Parameters are labeled in blue and black.[]{data-label="fig:ortho_diagram"}](ortho-fig.png){width="30.00000%"}
We now present an application of the MCS method to a reachable-set problem posed in [@na] for robustness analysis of a sit-to-stand motion controller for powered lower limb orthoses.
The Powered Lower Limb Orthosis and its user are modeled in [@na] as a three-link planar robot with three joints. The model has six states (angles and angular velocities of the three joints, labeled in red and green in Figure \[fig:ortho\_diagram\]) and twelve parameters (the lengths, masses, and moments of inertia of the three links, and the distances between each joint and the center of moment of its corresponding link, labeled in blue and black in Figure \[fig:ortho\_diagram\]). The twelve parameters are *uncertain*, since they all depend on the weight of the user. The sit-to-stand reference trajectory is defined in terms of the position in the $x-$ and $y-$ (horizontal and vertical in Figure \[fig:ortho\_diagram\]) directions of the center of mass (CoM) of the user and orthosis, which can be computed from the state variables and parameters.
The authors of [@na] designed a finite time horizon LQR controller to track a reference trajectory that brings the user and orthosis from a sitting position to a standing position. To analyze the robustness of the motion to parameter variations, the authors used sensitivity-based reachability methods to compute an interval overapproximation of the CoM trajectories subject under parameter changes induced by a 5% variation in user body weight. In this example, we approach this robustness analysis problem by using the MCS method to compute an interval overapproximation of the CoM trajectory with accuracy $\epsilon=0.05$ and confidence $\delta=0.001$.
To recast the problem from a robustness analysis problem into a reachability problem, we add a new state variable $p_i$ for each parameter with constant dynamics, i.e. $\dot{p_i}=0$. The parameters may be selected by choosing the initial conditions of the $p_i$. This way, we can perform the robustness verification by computing reachable set overapproximations of the parameter-augmented 18-state system.
The parameter intervals are derived from human biometric data, and the initial state is assumed to be fixed. We choose an initial distribution to be the uniform distribution over the initial set. In this case, the initial distribution has a physical significance, namely that the parameters are subject to change from user to user. If we had more detailed statistical information about the parameters (e.g. that they are Gaussian-distributed with some mean and variance), that could be used to define the initial set instead.
Using (\[eq:mc\_sample\_bound\]) with values $\epsilon=0.05$, $\delta=0.001$, and $n=18$, we know that $m=7554$ sample trajectories will suffice to compute an $\hat{R}^{(m)}$ that has at the desired levels of accuracy and confidence.
![Reachable set overapproximations (black boxes) of the center of mass trajectories at three time instants over sit-to-stand movement, calculated using the Monte Carlo interval overapproximation method with $\epsilon=0.05$, $\delta=0.001$. $m=7554$ samples trajectories (also shown) were required to ensure the specified accuracy and confidence over the parameter-augmented 18-dimensional state space.[]{data-label="fig:ortho_result"}](ortho-result-xyonly.png){width="50.00000%"}
The resulting interval overapproximations of the position of the CoM and is shown in Figure \[fig:ortho\_result\]. Specifically, we show the overapproximation for three points in the sit-to-stand movement; at the beginning ($t=0$), in the middle ($t=1.75$), and at the end ($t=3.5$) of the movement. In addition, the volumes of the intervals in Figure \[fig:ortho\_result\] are reported in Table \[tab:ortho\_volumes\], and compared to the volumes of the intervals computed using sensitivity-based methods in [@na]. The table also reports the volumes of the interval overapproximations to the velocity of the CoM, which come from the same calculation as the position overapproximations. At each time, the volumes of the MCS rectangles is lower than those of the sensitivity method volumes, meaning that the MCS method gives less conservative estimates of the reachable sets.
interval Volume from [@na] Volume from MCS
-------------------------------------------- --------------------- ---------------------
$(x_{CoM}, y_{CoM})$, $t=0$ $2.4\times 10^{-3}$ $1.7\times 10^{-3}$
$(x_{CoM}, y_{CoM})$, $t=1.75$ $4.0\times 10^{-3}$ $2.5\times 10^{-3}$
$(x_{CoM}, y_{CoM})$, $t=3.5$ $3.2\times 10^{-3}$ $5.0\times 10^{-4}$
$(\dot{x}_{CoM}, \dot{y}_{CoM})$, $t=0$ $0$ $0$
$(\dot{x}_{CoM}, \dot{y}_{CoM})$, $t=1.75$ $1.0\times 10^{-3}$ $7.4\times 10^{-4}$
$(\dot{x}_{CoM}, \dot{y}_{CoM})$, $t=3.5$ $4.8\times 10^{-6}$ $1.1\times 10^{-6}$
: Comparison of Interval Overpproximation Volumes[]{data-label="tab:ortho_volumes"}
Conclusion
==========
In this paper, we showed the utility of a probabilistic formulation of the problem of reachable set approximation by developing data-driven reachable set computation methods, and by verifying them with numerical examples. Between the two methods, the analyst can make a reachable set approximation to any degree of accuracy. Sets computed with the MCS method also come with a rigorous probabilistic guarantee of accuracy and confidence. Since the problem formulation makes few assumptions, these two methods can be used on a wide range of systems, many of which are too complex for standard reachability methods.
Both of the methods we present are useful in their present form, as shown by the examples. However, they can both be improved and extended with future work. For the MCS interval overapproximation method, the proof of (\[eq:mc\_sample\_bound\]) makes no assumptions on the successor distribution $p_1$, other than that $p_1(x)$ be well-defined for all $x\in{\ensuremath{\mathcal{X}_0} }$. On one hand, having a distribution-free result is useful, since we don’t in general have any knowledge of $p_1$. On the other hand, it suggests that if we did have some knowledge about $p_1$, the bound could perhaps be improved by using importance sampling or a related technique. Additionally, the proof as given depends on the geometry of hyperrectangles, and does not generalize to other classes of sets that we may want to use for interval reachable set overapproximation. For the GPC method, an useful extension would be a probabilistic guarantee, of a similar kind to the guarantee that the MCS method enjoys. It may also be useful to consider other classification methods than least-squares, which may be more accurate and amenable to analysis.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported in part by the grants ONR N00014-18-1-2209, AFOSR FA9550-18-1-0253, NSF ECCS-1906164.
Appendix: Derivation of the True ACC Safe Set {#appendix-derivation-of-the-true-acc-safe-set .unnumbered}
=============================================
For the system $$\begin{aligned}
\dot{x}&=&v, \\
\quad \dot{v}&=&-bv^2-a, \label{vdot}\end{aligned}$$ the distance traveled before coming to a full stop is (from the derivation below): $$\label{main}
\frac{1}{2b}\ln\left(1+\frac{b}{a}v(0)^2\right).$$ Since $h$ is a monotonic function of $t$ (see below), to check for a collision it suffices to check the sign of $h$ after both cars have come to a stop. if $h\ge0$ when both cars have stopped, then $h(t)\ge0$ for all prior $t$, and no collision occurred. On the other hand, if $h<0$ after the cars stopped, at some point $h$ changed sign, at which time there was a collision. With this in mind, from (\[main\]) the safe set in the $(h(0),v_L(0))$ space is: $$\begin{gathered}
E(v_F(0))=\bigg\{(h,v): h+\frac{1}{2b}\ln\left(1+\frac{b}{a}v^2\right)\\
-\frac{1}{2b}\ln\left(1+\frac{b}{a}v_F(0)^2\right)\ge 0\bigg\}.\end{gathered}$$ Derivation of (\[main\]): the solution of (\[vdot\]) is $$\label{v}
v(t)=\sqrt{\frac{a}{b}}\tan\left( \tan^{-1}\left(\sqrt{\frac{b}{a}}v(0) \right)-\sqrt{ab}\,t\right)$$ and integration gives $$\label{x}
x(t)-x(0)=\frac{1}{b}\ln \left( \frac{\cos\left(\tan^{-1}\left( \sqrt{\frac{b}{a}}v(0)\right)-\sqrt{ab}\,t\right)}{\cos\left(\tan^{-1}\left( \sqrt{\frac{b}{a}}v(0)\right)\right)}\right).$$ Note that $v(t)=0$ when $t$ is such that the argument of the tangent term in (\[v\]) is zero. Since the cosine in the numerator of (\[x\]) has the same argument, the numerator equals one when $v(t)=0$. When, in addition, the denominator is simplified as $$\cos\left(\tan^{-1}\left( \sqrt{\frac{b}{a}}v(0)\right)\right)=\left( 1+\frac{b}{a}v(0)^2\right)^{-1/2},$$ the right-hand side of (\[x\]) becomes (\[main\]).
Monotonicity of $h(t)$: introducing some abbreviations, we can write (\[x\]) as $$x_i(t)-x_i(0) = \frac{1}{b} \ln \left(\frac{\cos(\alpha_i-\beta t)}{\cos(\alpha_i)}\right)
\label{x simplified}$$ where $i=L\text{ or }F$. This This expression gives us $\dot{h}$ as $$\dot{h}=\dot{x}_L-\dot{x}_F=\frac{\beta}{b} \left(\tan(\alpha_L-\beta t)-\tan(\alpha_F-\beta t) \right).$$ Now we can consider two cases: $\alpha_L>\alpha_F$, and $\alpha_L\le\alpha_F$. In the first case, we will have $\alpha_L-\beta t > \alpha_F-\beta t$, and $\tan(\alpha_L-\beta t) > \tan(\alpha_F-\beta t)$. This makes $\dot{h}(t)>0$ for all $t$, meaning that $h(t)$ is monotonically increasing. By analogous reasoning, in the second case $h(t)$ is monotonically decreasing.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Video representation is an important and challenging task in the computer vision community. In this paper, we assume that image frames of a moving scene can be modeled as a Linear Dynamical System. We propose a sparse coding framework, named adaptive video dictionary learning (AVDL), to model a video adaptively. The developed framework is able to capture the dynamics of a moving scene by exploring both sparse properties and the temporal correlations of consecutive video frames. The proposed method is compared with state of the art video processing methods on several benchmark data sequences, which exhibit appearance changes and heavy occlusions.'
address: |
Department of Electrical Engineering and Information Technology\
Technische Universit[ä]{}t M[ü]{}nchen, Arcisstr. 21, 80333 Munich, Germany\
{xian.wei, hao.shen, kleinsteuber}@tum.de\
title: |
An Adaptive Dictionary Learning Approach for\
Modeling Dynamical Textures
---
Dynamic textures modeling, sparse representation, dictionary learning, linear dynamical systems.
Introduction {#sec:01}
============
Temporal or dynamic textures (DT) are image sequences that exhibit spatially repetitive and certain stationarity properties in time. This kind of sequences are typically videos of processes, such as moving water, smoke, swaying trees, moving clouds, or a flag blowing in the wind. Study and analysis of DT is important in several applications such as video segmentation [@chan2009layered], video recognition [@saisan2001dynamic], and DT synthesizing [@doretto2003dynamic].
One classical approach is to model dynamic scenes via the optical flow [@horn1981determining]. However, such methods require a certain degree of motion smoothness and parametric motion models [@chan2009layered]. Non-smoothness, discontinuities, and noise inherence to rapidly varying, non-stationary DTs (e.g. fire) pose a challenge to develop optical flow based algorithms. Another technique, called particle filter [@djuric2003particle], models the dynamical course of DTs as a Markov process. A reasonable assumption in DT modeling is that each observation is correlated to an underlying latent variable, or “state”, and then derive the parameter transition operator between these states.
Some approaches directly view each observation as a state, and then focus on transitions between the observations in the time domain. For instance, the work in [@schodl2000video] treats this transition as an associated probability problem, and other methods construct a spatio-temporal autoregressive model (STAR) or position affine operator for this transition [@szummer1996temporal; @kwatra2003graphcut].
Differently, feature-based models capture the intrinsic law and underlying structures of the data by projecting the original data onto a low-dimensional feature space via feature extracted techniques, such as principle component analysis (PCA). G. Doretto et al. [@saisan2001dynamic; @doretto2003dynamic] model the evolution of the dynamic textured scenes as a linear dynamical system (LDS) under a Gaussian noise assumption. As a popular method in dynamic textures, LDS and its derivative algorithms have been successfully used for various dynamic texture applications [@doretto2003dynamic; @saisan2001dynamic]. However, constraints are imposed on the types of motion and noise that can be modeled in LDS. For instance, it is sensitive to input variations due to various noise. Especially, it is vulnerable to non-Gaussian noise, such as missing data or occlusion of the dynamic scenes. Moreover, stability is also a challenging problem for LDS [@boots2007constraint]. To tackle these challenges, the approach taken here is to explore an alternative method to model the DTs by appealing to the principle of sparsity. Instead of using the Principle Components (PCs) as the transition “states” in LDS, sparse coefficients over a learned dictionary are imposed as the underlying “states”. In this way, the dynamical process of DTs exhibits a transition course of corresponding sparse events. These sparse events can be obtained via a recent technique on linear decomposition of data, called dictionary learning [@elad2006image; @hawe2013separable]. Formally, these sparse representations $x\in \mathbb{R}^k$ to a signal $y\in \mathbb{R}^m$, can be written as $$y = Dx$$ where $D\in \mathbb{R}^{m\times k}$ is a dictionary, and $x$ is sparse, i.e. most of its entries are zero or small in magnitude. That is, the signal $y$ can be sparsely represented only using a few elements from some dictionary $D$. In this work, we start with a brief review of the dynamic texture model from the viewpoint of convex $\ell_2$ optimization, and then deduce a combined regression associated with several regularizations for a joint process—“states extraction” and “states transition”. Then we treat the solution of the above combined regression as an adaptive dictionary learning problem, which can achieve two distinct yet tightly coupled tasks— efficiently reducing the dimensionality via sparse representation and robustly modeling the dynamical process. Finally, we cast this dictionary learning problem as the optimization of a smooth non-convex objective function, which is efficiently resolved via a gradient descent method.
Adaptive Video Dictionary Learning {#sec:03}
==================================
In this section, we start with a brief introduction to the linear dynamical systems (LDS) model and develop an adaptive dictionary learning framework for sparse coding.
Linear Dynamical Systems {#sec:02}
------------------------
Let us denote a given sequence of $(n+1)$ frames by $Y := [y_{0}, \ldots, y_{n}]
\in \mathbb{R}^{m\times (n+1)}$, where the time is indexed by $i = 0,1, \ldots, n$. The evolution of a LDS is often described by the following two equations $$\label{eq:lds}
\left\{\!\!
\begin{array}{rl}
x_{i+1} &\!\!\!\! = A x_{i} + w_{i} \\
y_{i} &\!\!\!\! = D x_{i} + v_{i},
\end{array}
\right.$$ where $y_{i} \in \mathbb{R}^{m}$, $x_{i} \in \mathbb{R}^{k}$, $w_{i} \in \mathbb{R}^{k}$ and $v_{i} \in \mathbb{R}^{m}$ denote the observation, its hidden state or feature, state noise, and observation noise, respectively. The system is described by the dynamics matrix $A \in \mathbb{R}^{k\times k}$, and the modeling matrix $D \in
\mathbb{R}^{m \times k}$. Here we are interested in estimating the system parameters $A$ and $D$, together with the hidden states, given the sequence of observations $Y$.
The problem of learning the LDS can be considered as a coupled linear regression problem [@boots2007constraint]. Let us denote $X = [x_{0}, \ldots, x_{n}] \in \mathbb{R}^{k \times (n+1)}$, $X_{0} = [x_{0}, \ldots, x_{n-1}] \in \mathbb{R}^{k \times n}$, and $X_{1} = [x_{1}, \ldots, x_{n}] \in \mathbb{R}^{k \times n}$. The system dynamics and modeling matrix are expected to be caught by solving the following minimization problem, $$\label{FeatureBased1}
\min_{A,D,X} \big\| X_{1} - A X_{0} \big\|_{F}^2 \quad s.t. \;
\big\| Y - D X \big\|_{F}^2\leq \varepsilon,
$$ where $\varepsilon$ is a small positive constant. In our approach, we assume that all observations $y_{i}$ admit a sparse representation with respect to an unknown dictionary $D \in \mathbb{R}^{m \times k}$, i.e. $$y_{i} = D x_{i}, \qquad\text{for all}~i=0,1,\ldots,n,$$ where $x_{i} \in \mathbb{R}^{k}$ is sparse. Without loss of generality, we further assume that all columns of the dictionary $D$ have unit norm. We then define the set $$\mathcal{S}(m,k) := \{ D \in \mathbb{R}^{m \times k} | \operatorname{ddiag}(D^{\top} D) = I_{k} \},$$ where $\operatorname{ddiag}(Z)$ is the diagonal matrix whose entries on the diagonal are those of $Z$, $I_{k}$ denotes the identity matrix. The set $\mathcal{S}(m,k)$ is the product of $k$ unit spheres, and is hence a $k(m-1)$ dimensional smooth manifold. Finally, by adopting the common sparse coding framework to problem , we have the following minimization problem $$\label{FeatureBased2}
\min_{A,D,X} \big\| X_{1} - A X_{0} \big\|_{F}^2 + \mu_{1}
\big\| Y - D X \big\|_{F}^2 + \mu_{2} \|X\|_{1},
$$ where $D\in \mathcal{S}(m,k)$, $\|\cdot\|_{F}$ denotes the Frobenius norm of matrices, and $\|\cdot\|_{1}$ is the $\ell_{1}$ norm, which measures the overall sparsity of a matrix. The parameter $\mu_{2} > 0$ weighs the sparsity measurement against the residual errors.
A Dictionary Learning Model for Dynamical Scene
-----------------------------------------------
Solving the minimization problem as stated in Eq. is a very challenging task. In this work, we employ an idea similar to *subspace identification methods* [@boots2007constraint], which treat the state as a function of $(A,D)$. Here, we confine ourselves to the sparse solution of an elastic-net problem, which is proposed in [@zou2005regularization], as $$\label{eq:elastic-net}
x^{*} := \operatorname*{argmin}_{x \in \mathbb{R}^{k}}
\tfrac{1}{2} \| y - D x \|_{2}^{2} + \lambda_{1} \|x\|_{1}
+ \tfrac{\lambda_{2}}{2} \|x\|_{2}^{2},$$ where $\lambda_1 > 0$ and $\lambda_2 > 0$ are regularization parameters, which play an important role in ensuring stability and uniqueness of the solutions. Let us define the set of indices of the non-zero entries of the solution $x^{*} = [
x_{1}^{*}, \ldots, x_{k}^{*}]^{\top} \in \mathbb{R}^{k}$ as $$\Lambda := \{i\in\{1,\ldots,k\} | x^{*}_{i} \neq 0\}.$$ Then the solution $x^{*}$ has a closed-form expression as $$\label{eq:enet}
x_{y}^{*}(D) := \left( D_{\Lambda}^{\top} D_{\Lambda} - \lambda_{2}
I_{m} \right)^{-1} \left( D_{\Lambda}^{\top} y - \lambda_{1}
s_{\Lambda} \right),$$ where $s_{\Lambda} \in \{ \pm 1\}^{|\Lambda|}$ carries the signs of $x_{\Lambda}^{*}$, $D_{\Lambda}$ is the subset of $D$ in which the index of atoms (rows) fall into support $\Lambda$. Furthermore, it is known that the solution $x_{y}^{*}(D)$ as given in is a locally twice differentiable function at $D$. By an abuse of notation, we define $$\begin{split}
X_{0} \colon \mathcal{S}(m,k) \to &\, \mathbb{R}^{k \times n} \\
D \mapsto &\, [x_{y_{0}\!}^{*}(D), \ldots, x_{y_{n-1}\!}^{*}(D)].
\end{split}$$ In a similar way, $X_{1} \colon \mathcal{S}(m,k) \to \mathbb{R}^{k \times n}$ is defined. Thus, the cost function reads as $$\label{main_AVDL_unstable}
\begin{split}
f \colon \mathbb{R}^{k \times k} \times \mathcal{S}(m,k) \to &\, \mathbb{R} \\
(A, D) \mapsto &\, \tfrac{1}{2} \left\| X_{1}(D) - A X_{0}(D) \right\|_{F}^{2}.
\end{split}$$
It is known that an LDS with the dynamic matrix $A$ is said to be stable, if the largest eigenvalue of $A$ is bounded by $1$ [@boots2007constraint]. Let $\sigma$ be the largest eigenvalue of $A$, then $|\sigma|\leq \|A\|_F.$ Thus, we enforce the small $\sigma$ via imposing a penalty $\|A\|_F^2$ on , and then end up with the cost function as $$\label{main_AVDL}
\begin{split}
\widetilde{f} \colon \mathbb{R}^{k \times k} \times \mathcal{S}(m,k) \to &\, \mathbb{R} \\
(A, D) \mapsto &\, f(A,D) + \tfrac{\gamma}{2}\|A\|_F^2,
\end{split}$$
Development of the Algorithm
----------------------------
In this section, we firstly derive a gradient descent algorithm to minimize and then discuss some details of the choice of the parameters in the final implementation.
We start with the computation of the first derivative of the sparse solution of the elastic-net problem $x_{y}^{*}(D)$ as given in . Given the tangent space of $\mathcal{S}(m,k)$ at $D$ as $$T_{D}\mathcal{S}(m,k) := \{ X \in \mathbb{R}^{m \times k} | \operatorname{ddiag}(X^{\top} D) = 0 \},$$ the orthogonal projection of a matrix $H \in \mathbb{R}^{m \times k}$ onto the tangent space $T_{D}\mathcal{S}(p,n)$ with respect to the inner product $\langle X,Y \rangle = \operatorname{tr}(X^{\top}Y)$ is given by $$\Pi_{D}(H) := H - D \operatorname{ddiag}(D^{\top} H).$$ Let us denote $K := D_{\Lambda}^{\top} D_{\Lambda} - \lambda_{2} I_{k}$. The first derivative of $x_{y}^{*}$ in the direction $H \in T_{D}\mathcal{S}(m,k)$ is $$\begin{split}
\operatorname{D}x_{y}^{*}(D)H = &\,
K^{-1} H_{\Lambda}^{\top} y - K^{-1}\!(D_{\Lambda}^{\top}
H_{\Lambda} \\
& + H_{\Lambda}^{\top} D_{\Lambda} )
\cdot K^{-1} \left( D_{\Lambda}^{\top} y - \lambda_{1}
s_{\Lambda} \right).
\end{split}
\label{eq:diriv_direction2}$$
By the product structure of $\mathbb{R}^{k \times k} \times \mathcal{S}(m,k)$, the Riemannian gradient of the function $\widetilde{f}$ is $$\label{eq:main_diriv}
\operatorname{grad} \widetilde{f}(A,D) = \left(\nabla_{\widetilde{f}}(A),
\Pi_{D}\big(\nabla_{\widetilde{f}}(D)\big) \right).$$ Here, the Euclidean gradient $\nabla_{\widetilde{f}}(A)$ of $\widetilde{f}$ with respect to $A$ is computed as $$\label{eq:diffA}
\nabla_{\widetilde{f}}(A) = \left(A X_{0}(D) - X_{1}(D) \right)X_{0}(D)
+ \gamma A,$$ with $e_{i}$ being the $i$-th standard basis vector of $\mathbb{R}^{n}$. Using the shorthand notation, $r_{i} := D_{\Lambda_{i}}^{\top} y_{i} - \lambda_{1}s_{\Lambda_{i}}$, $\Delta x_{i} := x_{y_{i}}^{*}(D)-A_{\Lambda_i}x_{y_{i-1}}^{*}(D)$, and $q_i := r_i\Delta x_{i}^{\top}$, the Euclidean gradient $\nabla_{\widetilde{f}}(D)$ of $
\widetilde{f}$ with respect to $D$ is $$\begin{split}
\label{eq:diff_D}
& \nabla_{\widetilde{f}}(D) = \sum\limits_{i=1}^{n} y_{i} (\Delta x_{i})^{\top}
K_{i}^{-1} - D_{\Lambda_{i}} K_{i}^{-1} (q_i+ q_i^{\top}) \\
&\cdot K_{i}^{-1}- y_{i-1} (\Delta x_{i})^{\top} \!A_{\Lambda_i} \!(K_{i-1})^{-1} + D_{\Lambda_{i-1}} \\
& \cdot (K_{i-1})^{-1} (A_{\Lambda_{i-1}}q_{i-1} + q_{i-1}^{\top}A_{\Lambda_{i-1}}^{\top}) (K_{i-1})^{-1}.
\end{split}$$
For a gradient search iteration on manifolds, we employ the following smooth curve on $\mathcal{S}(m,k)$ through $D \in \mathcal{S}(m,k)$ in direction $H \in T_{D}\mathcal{S}(m,k)$ $$\begin{split}
\tau \colon &(-\lambda,\lambda) \to \mathcal{S}(m,k) \\
t \mapsto & (D + t H) \big(\operatorname{ddiag}((D + t H)^{\top}
(D + t H))\big)^{-\tfrac{1}{2}}
\end{split}$$ with $\lambda >0$. It essentially normalizes all columns of $D + t H$. For a detailed overview on optimization on matrix manifold, refer to [@absil2009optimization].
Training data $Y$ Initialize the parameters $\lambda_1$,$\lambda_2$,$\gamma$, initial dictionary $D$, and initial transition matrix $A$. *Sparse Coding Stage*\
$\quad$ Use Lasso algorithm to compute $x$ via\
$\quad$ $x\leftarrow\underset{x}{\min}\frac{1}{2}\|y-Dx\|_2^2 + \lambda_1\|x\|_1 + \frac{\lambda_2}{2} \|x\|_2^2$\
$\quad$ Compute the active set $\Lambda$ for each $x$.\
Compute the gradient of $\widetilde{f}(A,D)$ according to and . Update the parameters $A$ and $D$ $$\begin{split}
\label{eq:diriv_direction}
&A_i\leftarrow A_{i-1}-\rho_i \nabla_{\widetilde{f}}(A_{i-1}), \\
&D_i\leftarrow D_{i-1}-\rho_i \nabla_{\widetilde{f}}(D_{i-1}).
\end{split}$$
Until now, we have computed the gradient of $\widetilde{f}$ as defined in with respect to its two arguments $D$ and $A$. An iterative scheme (such as the gradient descent method or conjugate gradient method) can be used to find the optimal $D$ and $A$, using the gradient expression above. The procedure displayed in Algorithm is the version of AVDL based on gradient descent procedure. The learning rate $\rho_i$ can be computed via the well-known backtracking line search method, similar to [@hawe2013separable]. Here, considering the high coherence among the temporal frames, we prefer non-redundant dictionary, that is, $k\ll m$ for the dictionary $D\in \mathbb{R}^{m\times k}$. For parameters $(\lambda_1,\lambda_2)$ in the elastic net, we put an emphasis on sparse solutions and choose $\lambda_2 \in (0,\frac{\lambda_1}{10})$, as proposed in [@zou2005regularization].
---------------------- ------------------- ------------------- ------------------- ------------------- -------- --------- --------- --------
64 128 256 1 50 100 200 400
Compression rate (%) 6.25 12.50 25.00 1.02 3.29 3.41 3.50 3.55
$\sigma$ 0.9802 0.9833 0.9849 1.78 1.06 0.9992 0.9994 0.9994
$e_y$ $1.35\times 10^5$ $1.35\times 10^5$ $1.35\times 10^5$ $1.36\times 10^3$ 60.29 $58.82$ $55.97$ 71.27
$e_x$ $101.58$ $135.88$ $168.95$ $3.75\times 10^4$ 171.99 75.52 61.96 46.18
---------------------- ------------------- ------------------- ------------------- ------------------- -------- --------- --------- --------
Numerical Experiments {#sec:04}
=====================
We carry out a few experiments on natural image sequences data, and demonstrate the practicality of the proposed algorithm. Our test dataset comprises of videos from DynTex++ [@ghanem2010maximum], and data from internet sources (for instance, YouTube). Firstly, we show the performance on reconstruction and synthesizing with a grayscale video of burning candle, which is corrupted by Gaussian noise or occlusion. This video has 1024 frames with size of $32\times 32$, see figure \[fig\_candle\_sys\_rec\]. The initial dictionary is $1024\times 512$. After the acquisition of the dictionary $D$ and the transition $A$, the synthesized data can be generated easily by $x_{i+1} = Ax_ix_i^{\top}x_i(x_i^\top x_i)^{-1}$, or more precisely, using a convex formulation $$\min_{x_{i+1}}\frac{1}{2}\|x_{i+1}-Ax_i\|_2^2+\lambda\|x_{i+1}\|_1.$$ Table \[candle\_sys\] shows the performance of synthesizing on burning candle with Gaussian noise. The error pairs $(e_x, e_y)$ are defined as $e_{y}=\sum_{i}\|y_i-Dx_i\|$, $e_{x}=\sum_i\|x_{i+1}-Ax_i\|$, and the largest eigenvalue of $A$ is denoted by $\sigma$. The compression rate for AVDL is sparsity of $x$ to $m\times(n+1)$, and for LDS is number of PCs to $m$. Table \[candle\_sys\] shows AVDL can obtain the stable dynamic matrix $A$ $(\sigma\leq1)$, smaller compression rate and smaller error $(e_x, e_y)$ of cost function , by increasing the numbers of main loops in Algorithm 1.
Figure \[fig\_candle\_sys\_rec\] $(a\sim c)$ is the visual comparison between LDS and AVDL. AVDL performs well on denoising against corruption by Gaussian noise. In the case of occlusion in figure \[fig\_candle\_sys\_rec\] (d), random 50 frames of the 1024 burning candle video are corrupted by a $(6\times7)$ rectangle. The length of both synthesizing data is 1024, based on first frame of the burning candle. $87.01\%$ of the synthesizing data from LDS are corrupted by this rectangle, but $9.47\%$ for AVDL.
[lllll]{} Occlusion rate (%) &0 &5 &15 &30\
LDS-NN (128PCs) &69.72 &45.00 &25.14 &14.17\
AVDL-SRC &70.28 &64.72 &44.44 &22.36\
The second experiment is about scenes classification on DynTex++, which contains DTs from 36 classes. Each class has 100 subsequences of length 50 frames with $50\times 50$ pixels. 20 videos are randomly chosen in each class and total 720 videos are used for our experiments. Classification for LDS is performed using the Martin distance with a nearest-neighbor classifier on its parameters pair $(A,D)$ [@saisan2001dynamic]. Another classifier is AVDL associated with the sparse representation-based classifier (SRC) [@wright2009robust; @ghanem2010sparse], in which the class of a test sequence is determined by the smallest reconstruction error $e_{y}$ and transition error $e_{x}$. Table \[table:recognition\] provides the recognition results with increasing occlusion rates for test data. Compared to LDS with nearest-neighbor classifier (LDS-NN), Table \[table:recognition\] shows the proposed AVDL with SRC (AVDL-SRC) performs better while the test videos are corrupted by increasing occlusion.
Conclusions {#sec:05}
===========
This paper proposes an alternative method, called AVDL, to model the dynamic process of DTs. In AVDL, the sparse events over a dictionary are imposed as transition states. The proposed method show a robust performance for synthesizing, reconstruction and recognition on DTs corrupted by Gaussian noise. Especially, AVDL exhibits more powerful in the case of test data with non-Gaussian noise, such as occlusion. One possible future extension is to learn a dictionary for large scale DT sequences based on AVDL.
[10]{}
Antoni B Chan and Nuno Vasconcelos, “Layered dynamic textures,” , vol. 31, no. 10, pp. 1862–1879, 2009.
Payam Saisan, Gianfranco Doretto, Ying Nian Wu, and Stefano Soatto, “Dynamic texture recognition,” in [*Computer Vision and Pattern Recognition. IEEE Computer Society Conference on*]{}. IEEE, 2001, vol. 2, pp. II–58.
Gianfranco Doretto, Alessandro Chiuso, Ying Nian Wu, and Stefano Soatto, “Dynamic textures,” , vol. 51, no. 2, pp. 91–109, 2003.
Berthold KP Horn and Brian G Schunck, “Determining optical flow,” , vol. 17, no. 1, pp. 185–203, 1981.
Petar M Djuric, Jayesh H Kotecha, Jianqui Zhang, Yufei Huang, Tadesse Ghirmai, M[ó]{}nica F Bugallo, and Joaquin Miguez, “Particle filtering,” , vol. 20, no. 5, pp. 19–38, 2003.
Arno Sch[ö]{}dl, Richard Szeliski, David H Salesin, and Irfan Essa, “Video textures,” in [*Proceedings of the 27th annual conference on Computer graphics and interactive techniques*]{}. ACM Press/Addison-Wesley Publishing Co., 2000, pp. 489–498.
Martin Szummer and Rosalind W Picard, “Temporal texture modeling,” in [*International Conference on Image Processing.*]{} IEEE, 1996, vol. 3, pp. 823–826.
Vivek Kwatra, Arno Sch[ö]{}dl, Irfan Essa, Greg Turk, and Aaron Bobick, “Graphcut textures: image and video synthesis using graph cuts,” in [*Graphics (TOG), ACM Transactions on*]{}. ACM, 2003, vol. 22, pp. 277–286.
Byron Boots, Geoffrey J Gordon, and Sajid M Siddiqi, “A constraint generation approach to learning stable linear dynamical systems,” in [*Advances in Neural Information Processing Systems*]{}, 2007, pp. 1329–1336.
Michael Elad and Michal Aharon, “Image denoising via sparse and redundant representations over learned dictionaries,” , vol. 15, no. 12, pp. 3736–3745, 2006.
Simon Hawe, Matthias Seibert, and Martin Kleinsteuber, “Separable dictionary learning,” in [*Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on*]{}, June 2013, pp. 438–445.
Hui Zou and Trevor Hastie, “Regularization and variable selection via the elastic net,” , vol. 67, no. 2, pp. 301–320, 2005.
P-A Absil, Robert Mahony, and Rodolphe Sepulchre, , Princeton University Press, 2009.
Bernard Ghanem and Narendra Ahuja, “Maximum margin distance learning for dynamic texture recognition,” in [*European Conference on Computer Vision*]{}, pp. 223–236. Springer, 2010.
J. Wright, A.Y. Yang, A. Ganesh, S.S. Sastry, and Y. Ma, “Robust face recognition via sparse representation,” , vol. 31, no. 2, pp. 210–227, 2009.
Bernard Ghanem and Narendra Ahuja, “Sparse coding of linear dynamical systems with an application to dynamic texture recognition,” in [*Pattern Recognition (ICPR), 2010 20th International Conference on*]{}. IEEE, 2010, pp. 987–990.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[We derive an exact, analytic expression for the fourth virial coefficient of a system of polydisperse spheres under the constraint that the smallest sphere has a radius smaller than a given function of the radii of the three remaining particles.]{}'
address: |
FOM Institute for Atomic and Molecular Physics, Kruislaan 407,\
1098 SJ Amsterdam, The Netherlands.
author:
- Ronald Blaak
title: Exact analytic expression for a subset of fourth virial coefficients of polydisperse hard sphere mixtures
---
Introduction {#sec:intro}
============
It is surprisingly difficult to calculate analytic expressions for virial coefficients. Even the exact calculation of the second virial coefficient is in general extremely difficult. If particles have no other interaction than hard core repulsion, things are slightly easier. In the isotropic phase, the second virial coefficient $B_2$ of two arbitrary, convex particles, $A$ and $B$, can be determined by [@Isihara:50]
$$\label{eq:b2-iso}
B_2(A,B) = \frac{1}{8 \pi} \left( V(A) G(B) + S(A) M(B) + M(A) S(B) +
G(A) V(B) \right)$$
where only geometrical quantities of both particles are used: the volume $V$, the surface area $S$, and the integrals $M$ and $G$ over the mean respectively Gaussian curvature. For higher virial coefficients only approximations of a similar type are known [@Boublik:86].
The second and third virial coefficients of identical spheres are known analytically and are obtained after a straightforward calculation. The fourth virial coefficient, for which the analytic expression due to Boltzmann is also known [@Boltzmann:1899], is much more difficult. Beyond the fourth virial coefficient only numerical data are available [@Kratky:77]. More recently, data of the fourth and fifth virial coefficients of binary mixtures have been published by Saija [*et al*]{} [@Saija:96-B4; @Saija:96-B5; @Saija:96-B4er].
In this paper we will focus on hard spheres with different radii, for which until now, no analytic results are known beyond the third virial coefficient. In section \[sec:2-3\] we show the results for the second and third virial coefficients. In section \[sec:fourth\] we derive the known results for the fourth virial coefficient. The main problem, the calculation of the complete star diagram, is derived in section \[sec:star\] together with an inequality, which has to be satisfied, in order for the given expression to be valid. In section \[sec:discussion\] we finish by discussing our result and some concluding remarks.
Second and third virial coefficient {#sec:2-3}
===================================
The general expression for the $n$th virial coefficient $B_n$ of a gas with pairwise, additive interaction $\phi_{ij}$ between particles $i$ and $j$ is given in terms of Mayer $f_{ij}$ functions by $$\label{eq:Bn}
B_n = \frac{1-n}{n!} \lim_{V \rightarrow \infty} V^{-1} \int \cdots \int V_n
d{\bf r}_1 \cdots d{\bf r}_n,$$ where ${\bf r}_i$ are the spatial coordinates of particle $i$, and $V$ is the volume accessible for the particles. $V_n$ is the sum over all labeled stars with $n$ points given by $$\label{eq:sum}
V_n \equiv \sum_{\{S_n \}} \prod_{i>j}^n f_{ij}$$ and expressed in the Mayer $f$ functions, which are related to the interaction potential by $$\label{eq:Mayer}
f_{ij} \equiv \exp(-\beta \phi_{ij}) - 1.$$ For hard particles the Mayer $f$ function reduces to $$\label{eq:Mayer-hard}
f = \left\{ \begin{array}{cl} 0 & \mbox{if no overlap} \\ -1 &
\mbox{if overlap}
\end{array} \right.$$ because the interaction potential is either zero or infinity in the case that particles do overlap, respectively do not overlap.
The second virial coefficient $B_2(A,B)$ for two particles $A$ and $B$, has only one contributing diagram, and is given by $$B_2(A,B) =
\frac{1}{2}\raisebox{-18pt}{\psfig{figure=figures/V2.eps,width=.6in}} =
\frac{1}{2}\raisebox{-18pt}{\psfig{figure=figures/V2AB.eps,width=.6in}},$$ where we have explicitly labeled the diagram. We use diagrams to represent the different integrals. Integrations are over all possible positions of the different particles with one particle fixed. Each solid line represents an overlap between the connected particles, and ,hence, some constraint on the position vectors. The signs, which are determined by the Mayer functions, are, for convenience, immediately put in front of the diagrams. The diagram in $B_2$ is the volume from which the second particle is excluded, in order not to cause an overlap. For spheres with radii $A$ and $B$ this is simply the volume of a sphere with radius $A+B$, leading to the well known result $$B_2(A,B) = \frac{2}{3} \pi (A+B)^3.$$ Also the third virial coefficient $B_3(A,B,C)$ of three particles $A$, $B$ and $C$ has only one contributing diagram
$$B_3(A,B,C) =
\frac{2}{3!}\raisebox{-18pt}{\psfig{figure=figures/V3.eps,width=.6in}} =
\frac{1}{3} \raisebox{-18pt}{\psfig{figure=figures/V3ABC.eps,width=.6in}}.$$
Only in the case of simple objects, e.g. spheres and discs, are analytic expressions for the third virial coefficient known.
In order to proceed we will assume that the radii of the three spheres $A$, $B$ and $C$ satisfy the inequality $A \geq B \geq C$. We position the biggest particle at the origin and particle $B$ on the positive $z$-axis. The integral corresponding to the diagram consists of two contributions. The first contribution comes from the case where $B$ is completely inside particle $A$, in which case particle $C$ is only required to overlap with particle $B$, leading automatically to an overlap with particle $A$. The second contribution comes from the case where $B$ is overlapping with $A$, but not enclosed. Particle $C$ now has to overlap with both particles, $A$ and $B$, and hence its center has to be placed in the overlapping volume of the spheres placed at $A$ with radius $A+C$ and the sphere placed at $B$ with radius $B+C$. This volume $Z(z_{AB},A+C,B+C)$ is the sum of two segments of different spheres and depends on the distance $z_{AB}$ between $A$ and $B$. The formula $Z(r,R_1,R_2)$ can easily be evaluated if we assume $R_1 \geq R_2$ $$Z(r,R_1,R_2) = \left\{ \begin{array}{l}
\frac{4 \pi}{3} R_2^3 \mbox{\hspace{1cm} $r < R_1 - R_2$}\\
\frac{\pi}{12 r} (R_1 + R_2 - r)^2 (r^2 - 3(R_1 -
R_2)^2 + 2 r (R_1 + R_2)) \\
0 \mbox{\hspace{1.75cm} $r > R_1 + R_2$} \end{array} \right.$$ if $r < R_1 - R_2$ the smaller sphere is completely enclosed by the larger sphere, and, hence, the overlap is equal to the volume of the smaller sphere. If $r > R_1 + R_2$ the spheres are not overlapping at all. The diagram for the third virial coefficient is therefore given by $$\raisebox{-18pt}{\psfig{figure=figures/V3ABC.eps,width=.6in}} =
\int\limits_0^{A+B} d z_{AB} 4 \pi z_{AB}^2 Z(z_{AB},A+C,B+C)$$ which results for the third virial coefficient $B_3(A,B,C)$ in $$\begin{aligned}
B_3(A,B,C) = \frac{16 \pi^2}{27} ( A^3 B^3 + B^3 C^3 + A^3 C^3
+ \nonumber \\
3 A B C [ A + B + C ] [A B + B C + C A] ) .\end{aligned}$$ Although we have assumed that $A \geq B \geq C$, the resulting formula is fully symmetric in the radii.
Fourth virial coefficient {#sec:fourth}
=========================
In order to obtain the fourth virial coefficient $B_4$ for particles $A$, $B$, $C$ and $D$, we need to evaluate three different diagrams, although this could be reduced to only two modified star diagrams (see ref. [@Ree:64]) $$\label{eq:B4}
B_4(A,B,C,D) = -\frac{3}{4!} \left(
3 \raisebox{-18pt}{\psfig{figure=figures/V40.eps,width=.6in}} -
6 \raisebox{-18pt}{\psfig{figure=figures/V41.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/V42.eps,width=.6in}}
\right).$$ The fourth virial coefficient is here expressed in unlabeled diagrams. However, labeling now becomes important because we want to evaluate those diagrams for the case of non-identical particles. This leads to the following expressions $$\label{eq:V4} \begin{array}{lll}
3 & \raisebox{-18pt}{\psfig{figure=figures/V40.eps,width=.6in}} = &
\raisebox{-18pt}{\psfig{figure=figures/V40ABCD.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/V40ACDB.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/V40ADBC.eps,width=.6in}}
\\
6 & \raisebox{-18pt}{\psfig{figure=figures/V41.eps,width=.6in}} = &
\raisebox{-18pt}{\psfig{figure=figures/V41ABCD.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/V41ACDB.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/V41ADBC.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/V41BACD.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/V41BCDA.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/V41CADB.eps,width=.6in}}
\\
& \raisebox{-18pt}{\psfig{figure=figures/V42.eps,width=.6in}} = &
\raisebox{-18pt}{\psfig{figure=figures/V42ABCD.eps,width=.6in}}
\end{array}$$
In order to proceed, we need to assume that the radii of the spheres satisfy $A \geq B \geq C \geq D$. The labeled diagrams in the first line of (\[eq:V4\]) are given by expressions of the form $$\label{eq:B40}
\raisebox{-18pt}{\psfig{figure=figures/V40ABCD.eps,width=.6in}} =
\int\limits_0^{A+C+2 D} d z_{AC} 4 \pi z_{AC}^2 Z(z_{AC},A+B,C+B)
Z(z_{AC},A+D,C+D).$$ The limits of integration are a consequence of the differences in size and in this case determined by the fact that $B>D$. The diagrams in the second line of (\[eq:V4\]) are similar, and only differ in the integration limits because of the extra overlap. They are of the form $$\label{eq:B41}
\raisebox{-18pt}{\psfig{figure=figures/V41ABCD.eps,width=.6in}} =
\int\limits_0^{A+C} d z_{AC} 4 \pi z_{AC}^2 Z(z_{AC},A+B,C+B)
Z(z_{AC},A+D,C+D).$$ The other diagrams are obtained by permutation symmetry. The resulting integrals of (\[eq:B40\]) and (\[eq:B41\]) can easily be evaluated, but lead to lengthy expressions. The summations over the different labelings do not lead to symmetric expressions in terms of the radii. If we however take the combination of the diagrams according to the definition of the fourth virial coefficient (\[eq:B4\]), we obtain
$$\begin{aligned}
\label{eq:simple}
3 \raisebox{-18pt}{\psfig{figure=figures/V40.eps,width=.6in}} -
6 \raisebox{-18pt}{\psfig{figure=figures/V41.eps,width=.6in}} =
- \frac{64 \pi^3}{9} (
A^3 B^3 C^3 + A^3 B^3 D^3 + A^3 C^3 D^3 + B^3 C^3 D^3 +
\nonumber \\
3 A B C D [A B + A C + B C + A D + B D + C D][A B C + A B D
+ A C D + B C D] ),\end{aligned}$$
which is symmetric in the different radii of the spheres.
The complete star diagram {#sec:star}
=========================
The problem of calculating the fourth virial coefficient lies in the remaining diagram, the complete star diagram in which all pairs of the four particles overlap. An analytic expression for this diagram in general is not known. What is known is the special case of four identical radii, which is due to Boltzmann [@Boltzmann:1899; @Happel:6]. $$\raisebox{-18pt}{\psfig{figure=figures/V42RRRR.eps,width=.6in}} =
\frac{256 \pi^2 R^9}{945} \left( 3419 \pi - 438 \sqrt{2} - 8262
\arccos(1/\sqrt{3}) \right),$$ where we used $R$ to denote the radius of the particles.
Some limiting cases for different radii are easily obtained. In the limit that the radius of the smallest sphere goes to zero the three remaining particles have to overlap with a point $$\label{eq:lim1}
\lim_{D \to 0}
\raisebox{-18pt}{\psfig{figure=figures/V42ABCD.eps,width=.6in}} =
\left( \frac{4 \pi}{3}\right)^3 A^3 B^3 C^3.$$ In the limit where the radius of the biggest sphere goes to infinity, the other three particles give rise to the diagram related to $B_3$ $$\label{eq:lim2}
\lim_{A \to \infty}
\raisebox{-18pt}{\psfig{figure=figures/V42ABCD.eps,width=.6in}}
\approx \frac{4 \pi}{3} A^3
\raisebox{-18pt}{\psfig{figure=figures/V3BCD.eps,width=.6in}}.$$ In the limit where both $A$ and $B$ go to infinity $$\label{eq:lim3}
\lim_{A,B \to \infty}
\raisebox{-18pt}{\psfig{figure=figures/V42ABCD.eps,width=.6in}}
\approx \lim_{A,B \to \infty}
\raisebox{-18pt}{\psfig{figure=figures/V3ABC.eps,width=.6in}}
\frac{4 \pi}{3} (C+D)^3 \approx \frac{64 \pi^3}{27} A^3 B^3 (C+D)^3.$$ In order to explore the star diagram for non-limiting cases we must look in more detail to the diagrams and introduce the $\tilde f$-bonds as defined by Ree and Hoover [@Ree:64] which we denote by a dashed line. Such a connection refers to non-overlapping particles, solid connections to overlapping particles and no connection allows for both, overlap and no overlap.
We will now consider diagrams in which the particles $A$, $B$ and $C$ all overlap with the smallest particle $D$. The diagram without other constraints can be written as the following summation of diagrams: $$\begin{aligned}
\raisebox{-18pt}{\psfig{figure=figures/SA.eps,width=.6in}} = &
\raisebox{-18pt}{\psfig{figure=figures/S0.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S1.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S2.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S3.eps,width=.6in}} +
\nonumber \\ &
\raisebox{-18pt}{\psfig{figure=figures/S4.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S5.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S6.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S7.eps,width=.6in}},\end{aligned}$$ where we have denoted at the right hand side explicitly whether pairs of particles overlap or not. The diagram at the left hand side can, however, be evaluated immediately
$$\raisebox{-18pt}{\psfig{figure=figures/SA.eps,width=.6in}} =
\raisebox{-18pt}{\psfig{figure=figures/V2AD.eps,width=.6in}} \times
\raisebox{-18pt}{\psfig{figure=figures/V2BD.eps,width=.6in}} \times
\raisebox{-18pt}{\psfig{figure=figures/V2CD.eps,width=.6in}}.$$
The same is true for combinations of the form: $$\raisebox{-18pt}{\psfig{figure=figures/S4567.eps,width=.6in}} =
\raisebox{-18pt}{\psfig{figure=figures/S4.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S5.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S6.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S7.eps,width=.6in}} =
\raisebox{-18pt}{\psfig{figure=figures/V2CD.eps,width=.6in}} \times
\raisebox{-18pt}{\psfig{figure=figures/V3ABD.eps,width=.6in}}.$$ As well, some combinations have already been evaluated before (\[eq:B41\]), e.g. $$\raisebox{-18pt}{\psfig{figure=figures/S67.eps,width=.6in}} =
\raisebox{-18pt}{\psfig{figure=figures/S6.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S7.eps,width=.6in}} =
\raisebox{-18pt}{\psfig{figure=figures/V41ACDB.eps,width=.6in}}.$$ This leads to the following expression for the complete star diagram $$\begin{aligned}
\label{eq:star}
\raisebox{-18pt}{\psfig{figure=figures/S7.eps,width=.6in}} = &
\raisebox{-18pt}{\psfig{figure=figures/SA.eps,width=.6in}} -
\raisebox{-18pt}{\psfig{figure=figures/S4567.eps,width=.6in}} -
\raisebox{-18pt}{\psfig{figure=figures/S2367.eps,width=.6in}} -
\raisebox{-18pt}{\psfig{figure=figures/S1357.eps,width=.6in}} +
\nonumber \\ &
\raisebox{-18pt}{\psfig{figure=figures/S67.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S57.eps,width=.6in}} +
\raisebox{-18pt}{\psfig{figure=figures/S37.eps,width=.6in}} -
\raisebox{-18pt}{\psfig{figure=figures/S0.eps,width=.6in}}.\end{aligned}$$ For all diagrams on the right hand side we have an analytic expression with only one exception - the last diagram. We demonstrate below that this last diagram is identical to zero if the radii of the spheres satisfy a simple inequality.
In order for this diagram to be non-zero the spheres with radius $A$, $B$ and $C$ should all overlap with the smallest sphere with radius $D$, but have no overlaps with each other. If the spheres $A$, $B$ and $C$ are mutually non-overlapping as shown in figure \[fig:Apollonian\], there is a smallest sphere $D$ that can be constructed that will touch with all three spheres. The centers of mass must all lie in the same plane, and therefore this problem reduces to the one of touching circles, known as Apollonian circles since the construction of this fourth circle was first solved by the Greek mathematician Apollonius of Perga (200 BC) [@Coxeter:68]. If the radius $D$ were smaller, at least one of the spheres $A$, $B$ or $C$ would have no overlap with this circle.
If this were the case for all possible configurations of non-overlapping spheres $A$, $B$ and $C$ there would be no contribution to the last diagram of (\[eq:star\]), and hence we would know the fourth virial coefficient. The smallest radius $D$ which can contribute is found in the case that all four spheres touch as depicted in figure \[fig:Apollonian\_Dec\]. For this special case of the Apollonian problem, a simple relation between the radii of the circles was derived by the French philosopher and mathematician Descartes [@Coxeter:68], known as the Descartes circle theorem: $$\label{eq:Decartes}
\left(\frac{1}{A}+\frac{1}{B}+\frac{1}{C}+\frac{1}{D}\right)^2 =
2 \left(\frac{1}{A^2}+\frac{1}{B^2}+\frac{1}{C^2}+\frac{1}{D^2}\right).$$ Solving this equation gives the upper limit for radius $D$ $$\label{eq:maxD}
D \leq \frac{A B C}{A B + A C + B C + 2 \sqrt{A B C (A + B + C)}},$$ which will lead to a zero-valued diagram, and hence enables us to give an analytic expression for the complete star diagram. Therefore we obtain an analytic expression for the fourth virial coefficient of polydisperse spheres under the constraint (\[eq:maxD\]) $$\begin{array}{l}
B_4(A,B,C,D) = \left( \frac{16 \pi^3}{27} \right) \times (
A^3 B^3 C^3 + A^3 B^3 D^3 + A^3 C^3 D^3 + B^3 C^3 D^3 + \\
~~~ 3 A B C D [ A B + A C + B C + A D + B D + C D][A B C + A B D
+ A C D + B C D]) - \\
~~~\frac{16 \pi^3}{3} D^3 A^2 B^2 C^2 +
\frac{ 8 \pi^3}{3} D^4 A B C (A B + A C + B C) -\\
~~~\frac{ 8 \pi^3}{15} D^5 (A^2 B^2 - 2 A^2 B C - 2 A B^2 C + A^2 C^2 -
2 A B C^2 + B^2 C^2) - \\
~~~\frac{ 8 \pi^3}{5} D^6 (A + B) (A + C) (B + C) -
\frac{32 \pi^3}{35} D^7 (A^2 + 3 A B + B^2 + 3 A C + 3 B C + C^2) - \\
~~~\frac{ 8 \pi^3}{7} D^8 (A + B + C) -
\frac{ 8 \pi^3}{21} D^9
\end{array}$$ The first part is symmetric in the four different radii and proportional to (\[eq:simple\]), the last part, however, is only symmetric in the radii $A$, $B$ and $C$. Note that the limiting cases (\[eq:lim1\]), (\[eq:lim2\]) and (\[eq:lim3\]) are in agreement with this result.
In the case that $A=B=C=1$, and, since according to (\[eq:maxD\]) $D
\leq (2\sqrt{3} - 3)/3 \approx 0.1547$, this formula reduces to $$\begin{aligned}
B_4(1,1,1,D) = & \left( \frac{8 \pi^3}{945} \right) \left( 70 + 630 D +
2520 D^2 + 1470 D^3 + 945 D^4 + \right. \nonumber \\ & \left.
189 D^5 - 1512 D^6 - 1296 D^7 - 405 D^8 - 45 D^9 \right).\end{aligned}$$ This result can be compared with the numerical data for the binary mixture [@Saija:96-B4er]. Only the first three numbers in the first column satisfy the inequality (\[eq:maxD\]) $$\begin{array}{llll}
D_{1112}({\cal R}=\frac{1}{20}) & = \frac{11058144323491
\pi^3}{6193152000000000} & = 0.0553630659 & ~~~[0.05539(2)]\\
D_{1112}({\cal R}=\frac{2}{20}) & = \frac{31952948861
\pi^3}{12096000000000} & = 0.0819065785 & ~~~[0.08189(2)] \\
D_{1112}({\cal R}=\frac{3}{20}) & = \frac{23207369313073
\pi^3}{6193152000000000} & = 0.1161886732 & ~~~[0.11617(3)]
\end{array}$$ The numerical values obtained by Saija [*et al.*]{} [@Saija:96-B4er] are given in square brackets.
Discussion {#sec:discussion}
==========
To our knowledge this is the first time that an exact and analytic solution of the fourth virial coefficient of a hard sphere mixture is found. One should note, however, that in order for the expression to be valid, the radius of the smallest sphere is at most 0.1547 times that of the largest sphere. Our result agrees perfectly with the relevant, numerical results of Saija [*et al*]{} [@Saija:96-B4er]. However, in the description of a polydisperse mixture of spheres, the most interesting case is that of a few large and many small particles. Unfortunately, this also means that the inequality (\[eq:maxD\]) is, in general, not satisfied. The inequality requires that the smallest radius is strictly smaller than the other three, which, for instance in a binary mixture, allows only one of the three mixed virial coefficients to be evaluated.
The method described in this article can almost certainly be used to evaluate the fourth virial coefficient of other particles, e.g. discs, and can possibly be used to simplify calculations of higher virial coefficients of asymmetric mixtures.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Bela Mulder, José Cuesta and James Polson for a critical reading of the manuscript. The work of the FOM Institute is part of the research program of FOM and is made possible by financial support from the Netherlands Organization for Scientific Research (NWO).
[10]{}
, 1950, [*J. Chem. Phys.*]{}, [**18**]{}, 1446.
, and [Nezbeda, I.]{}, 1986, [*[Czech. chem. Commun. ]{}*]{}, [**51**]{}, 2301.
, 1899, [*[Verslag Gewone Vergader. Afdel. Natuurk. Koninkl. Ned. Akad. Wetenschap]{}*]{}, [**7**]{}, 484.
, 1977, [*Physica A*]{}, [**87**]{}, 584.
, [Fiumara, G.]{}, and [Giaquinta, P. V.]{}, 1996, [*Mol. Phys.*]{}, [**87**]{}, 991.
, [Fiumara, G.]{}, and [Giaquinta, P. V.]{}, 1996, [*Mol. Phys.*]{}, [**89**]{}, 1181.
, [Fiumara, G.]{}, and [Giaquinta, P. V.]{}, 1997, [*Mol. Phys.*]{}, [**92**]{}, 1089.
, and [Hoover, W. G.]{}, 1964, [*J. Chem. Phys.*]{}, [ **40**]{}, 939.
, 1906, [*[Ann. Physik]{}*]{}, [**21**]{}, 342.
, 1968, [*Amer. Math. Monthly*]{}, [**75**]{}, 5.
[**FIGURE CAPTIONS**]{}
1. The Apollonian problem of constructing a circle $D$ which is tangent to three arbitrary spheres $A$, $B$ and $C$.
2. A special case of the Apollonian problem in which all spheres are tangent to each other, and for which case a simple relation (\[eq:Decartes\]) between the radii was derived by Descartes.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove a Torelli theorem for the moduli space of semistable parabolic Higgs bundles over a smooth complex projective algebraic curve under the assumption that the parabolic weight system is generic. When the genus is at least two, using this result we also prove a Torelli theorem for the moduli space of semistable parabolic bundles of rank at least two with generic parabolic weights. The key input in the proofs is a method of [@Hu].'
address:
- 'School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Bombay 400005, India'
- 'Instituto de Ciencias Matemáticas (CSIC-UAM-UC3M-UCM), C/ Nicolas Cabrera 15, 28049 Madrid, Spain'
- 'Instituto de Ciencias Matemáticas (CSIC-UAM-UC3M-UCM), C/ Nicolas Cabrera 15, 28049 Madrid, Spain'
author:
- Indranil Biswas
- 'Tomás L. Gómez'
- Marina Logares
title: Integrable systems and Torelli theorems for the moduli spaces of parabolic bundles and parabolic Higgs bundles
---
\[section\] \[thm\][Proposition]{} \[thm\][Lemma]{} \[thm\][Corollary]{}
\[thm\][Definition]{} \[thm\][Example]{}
\[thm\][Remark]{}
Introduction
============
The classical theorem by R. Torelli [@crs] says that a smooth complex algebraic curve is determined by the isomorphism class of its polarized Jacobian up to isomorphism. Similar theorems in many contexts have been worked out, e.g., for moduli spaces of stable vector bundles [@T; @NR; @MN] and moduli spaces of stable Higgs bundles [@BG]. As far as moduli spaces of parabolic or parabolic Higgs bundles with fixed determinant (see definition below) are concerned, a set of Torelli theorems were proved [@BBB; @BHK; @sebastian; @GL]. Here we deal with the non-fixed determinant situation.
In [@Hu], Hurtubise investigated algebraically completely integrable systems satisfying certain conditions. His main result is to extract an algebraic surface out of an integrable system. We observe that a moduli space of parabolic Higgs bundles is an example of the model of completely integrable systems studied in [@Hu].
The above mentioned assumption that the determinant is not fixed stems from the fact that in the set-up of [@Hu] the Lagrangians in the fibers are required to be Jacobians, while fixing the determinant amounts to making the fibers Prym varieties. To consider the moduli spaces with fixed determinant with our techniques, we would need an analogue of our main tool, namely Theorem 1.11 of [@Hu], but for an integrable system in which the fibers are Prym varieties instead of Jacobians. This is planned for future work.
We will prove the following theorems.
\[theoremPH\] Let $X$ and $X'$ be smooth projective curves with genus $g$ and parabolic points $D$ and $D'$ respectively. Let ${\mathcal{M}}_{X}(d,r,\alpha)$ (respectively, ${\mathcal{M}}_{X'}(d,r,\alpha)$) be the moduli space of stable parabolic Higgs bundles over $X$ (respectively, $X'$) endowed with the usual $\mathbb{C}^*$ action (cf. ) and the determinant line bundle ${\mathcal{L}}$ (respectively, ${\mathcal{L}}'$) (cf. ). If there is a $\mathbb{C}^*$-equivariant isomorphism between ${\mathcal{M}}_X(d,r,\alpha)$ and ${\mathcal{M}}_{X'}(d,r,\alpha)$, such the the pullback of the Néron-Severi class $NS({\mathcal{L}}')$ is $NS({\mathcal{L}})$, then there exists an isomorphism between $X$ and $X'$ inducing a bijection between the parabolic points $D$ and $D'$, whenever the following conditions on the genus and the rank are satisfied,
- if $g=2$ then $r\ge 5$,
- if $g=3$ then $r\ge 3$,
- if $g\ge 4$ then $r\ge 2$.
Since the moduli space of stable parabolic bundles sits inside the moduli space of stable parabolic Higgs bundles, in all cases where its codimension is greater than two, we get the following extension of the Torelli theorem for the moduli space of stable parabolic bundles given in [@BBB].
\[theoremP\] Let $X$ and $X'$ be smooth projective curves with genus $g$ and parabolic points $D$ and $D'$ respectively. Let $M_X(d,r,\alpha)$ be the moduli space of stable parabolic bundles over $X$ (respectively, $M_{X'}(d,r,\alpha)$), and let ${\mathcal{L}}$ (respectively, ${\mathcal{L}}'$) be the determinant line bundle (cf. ). If there is an isomorphism between $M_X(d,r,\alpha)$ and $M_{X'}(d,r,\alpha)$ such that the pullback of $NS({\mathcal{L}}')$ is $NS({\mathcal{L}})$, then there exists an isomorphism between $X$ and $X'$ inducing a bijection between the parabolic points $D$ and $D'$, whenever the following conditions on the genus and the rank are satisfied,
- if $g=2$ then $r\ge 5$,
- if $g=3$ then $r\ge 3$,
- if $g\ge 4$ then $r\ge 2$.
Preliminaries
=============
Let $X$ be an irreducible smooth projective algebraic curve over $\mathbb C$. The holomorphic cotangent bundle of $X$ will be denoted by $K$. Let $\{p_1\, ,\cdots\, , p_n\}$ be a set of distinct *parabolic points* in $X$ and let $D
\,=\, p_1+\ldots +p_n$ the corresponding reduced effective divisor. A *parabolic bundle* on $X$ with parabolic structure over $D$ consists of a holomorphic vector bundle $E$ equipped with a weighted flag over each *parabolic point* $p\in D$, that is a filtration of subspaces $$E\vert_p\,=\,E_{p,1}\,\supset\, \cdots \,\supset\, E_{p,r(p)}\,\supset\,
E_{p,r(p)+1}\,=\, 0$$ together with a system of *parabolic weights* $$0\,\leq\, \alpha_{1}(p)\,<\,\cdots \,<\, \alpha_{r(p)}(p)\,<\,1\, .$$ The *parabolic degree* and *parabolic slope* of $E$ are defined as follows: $$\operatorname{pardeg \,}(E) \,:=\,\deg (E) + \sum_{p \in D}\sum_{i=1}^{r(p)} \alpha_i(p)\cdot m_i (p)
\quad\quad \operatorname{par\mu \,}(E) \,:=\, \frac{\operatorname{pardeg \,}(E)}{\operatorname{rk \,}(E)}\, ,$$ where $m_i (p)\,:=\, \dim (E_{p,i} /E_{p,i+1})$ is the multiplicity of the parabolic weight $ \alpha_i(p)$. The parabolic bundle is called *stable* (respectively, *semistable*) if for all subbundles $0\,\not=\,V\, \subsetneq\, E$, $$\label{stab}
\operatorname{par\mu \,}(V)\,<\, \operatorname{par\mu \,}(E) \qquad \text{(respectively, $\operatorname{par\mu \,}(V)\,\leq\, \operatorname{par\mu \,}(E)$)}$$ where $V$ has the induced parabolic structure. Given rank and degree the system of parabolic weights is called *generic* if every semistable parabolic bundle is stable. We note that the semistability condition describes hyperplanes (or *walls*) in the space of weights. Hence the genericity condition means that the parabolic weights lie in the interior of the chambers defined by the walls.
We denote by $M_{X}(d,r,\alpha)$ the moduli space of stable parabolic bundles over $X$ with degree $d$, rank $r$ and generic weights $\alpha$. This moduli space is a smooth projective variety with $$\dim M_{X}(d,r,\alpha) \,= \,
r^2 (g-1) +1 + \frac{1}{2}\sum_{p\in D}\sum_{i=1}^{r(p)} (r^2- m_{i}(p)^{2})\, .$$ For notational convenience we assume that the flag is *full* that is, $m_i(p)\,=\,1$ for all $p$ and $i$, so $r(p)\,=\,r$ for all $p$, but all the results generalize to non full flags case. Henceforth, we will only consider full flags. Therefore, $$\dim M_{X}(d,r,\alpha)\,=\,r^2(g-1)+1+ \frac{1}{2} nr (r-1) \; .$$
An endomorphism of a parabolic bundle $E$ is called *non-strongly parabolic* if, for all $p\in D$ and $i$, $$\varphi (E_{p,i})\,\subset\, E_{x,i},$$ and it is called *strongly parabolic* if $$\varphi (E_{p,i})\,\subset\, E_{p,i+1}\, .$$ The sheaves of non-strongly and strongly parabolic endomorphisms are denoted by $\operatorname{ParEnd \,}(E)$ and $\operatorname{SParEnd\,}(E)$ respectively.
A *parabolic Higgs bundle* is a pair $(E,\Phi)$ where $E$ is a parabolic bundle and $$\Phi\,:\,E\,{\longrightarrow}\, E\otimes K(D)\,=\, E\otimes K\otimes {\mathcal O}_X(D)$$ is a *strongly parabolic* homomorphism, i.e., $$\Phi(E_{x,i}) \,\subset\, E_{x,i+1}\otimes K(D)_x$$ for each point $x\,\in \,D$ and all $i$. A parabolic Higgs bundle is *stable* (respectively, *semistable*) if the inequality is satisfied for those $V$ with $\Phi(V)\, \subset\,
V\otimes K(D)$.
Let ${\mathcal{M}}_{X}(d,r,\alpha)$ denote the moduli space of stable parabolic Higgs bundles with degree $d$, rank $r$ and generic weights $\alpha$. It is a smooth quasiprojective variety that satisfy $$\dim {\mathcal{M}}_{X}(d,r,\alpha)\,=\, 2r^2 (g-1) +2 + nr(r-1)
\,=\, 2\cdot \dim M_{X}(d,r,\alpha)$$ (recall that the quasiparabolic flags are full).
For any $E\in M_{X}(r,d,\alpha)$ the tangent space at $E$, $T_{E}M_{X}(r,d,\alpha)$, is $H^{1}(\operatorname{ParEnd \,}(E))$. Also, the parabolic version of Serre duality gives an isomorphism $$H^1(\operatorname{ParEnd \,}(E))^\ast \,\cong\, H^0 (\operatorname{SParEnd\,}(E) \otimes K(D)) \; .$$ Therefore, the total space of the cotangent bundle $T^\ast M_{X}(r,d,\alpha)$ is a Zariski open subset of ${\mathcal{M}}_{X}(r,d,\alpha)$.
The moduli space of parabolic Higgs bundles is endowed with a $\mathbb{C}^*$ action, where $t\in \mathbb{C}^*$ acts as scalar multiplication on the Higgs field $$\label{action}
(E,\Phi) \longmapsto (E,t\cdot \Phi)$$ The total space of the cotangent bundle $T^\ast M_{X}(r,d,\alpha)$ also has a canonical $\mathbb{C}^*$ action given by scalar multiplication on the fibers. Both actions are compatible, in the sense that the inclusion of the cotangent in the moduli space of Higgs bundles is $\mathbb{C}^*$ equivariant.
The Hitchin system
==================
Let $\mathcal{K(D)}$ denote the total space of the line bundle $K(D)$ over $X$, and let $\gamma\,:\,\mathcal{K(D)}\,{\longrightarrow}\, X$ be the natural projection. Let $$\widetilde{x}\,\in\, H^0(\mathcal{K(D)},\, \gamma^* K(D))$$ be the tautological section whose evaluation at any point $z$ is $z$ itself. The characteristic polynomial of a Higgs field $\Phi$ is $$\label{eq:char_pol}
\det(\widetilde{x}\cdot\operatorname{Id\,}- \gamma^* \Phi) \,=\, \widetilde{x}^r + {\widetilde}{s}_1 \widetilde{x}^{r-1} +
{\widetilde}{s}_2 \widetilde{x}^{r-2} +\cdots +{\widetilde}{s}_r.$$ The sections $\widetilde{s}_{i}$, descent to $X$, meaning there are sections $s_i\,\in \,H^0(X,\,K^i(iD))$ such that ${\widetilde}{s}_i\,=\,\gamma^* s_i$. Since $\Phi$ is strongly parabolic its residue at each parabolic point is nilpotent, and hence $s_i\in H^0(X,\,K^i((i-1)D))$. Therefore, there is a morphism, called the *Hitchin map*, $$\label{eqn:hitchin}
H\,:\, {\mathcal{M}}_{X}(d,r,\alpha)\,\longrightarrow\,
{\mathcal{U}}\,:=\,\bigoplus_{i=1}^{r} H^{0}(X,K^{i}((i-1)D))\, .$$ This morphism is proper [@Hi], and it induces an isomorphism on globally defined algebraic functions, i.e., the lower arrow in the following commutative diagram is an isomorphism $$\label{eq:globalsections}
\xymatrix{
{{\mathcal{M}}_{X}(d,r,\alpha)} \ar[d]^{a} \ar[r]^-{H} & {{\mathcal{U}}} \ar@{=}[d]\\
\operatorname{Spec}\Gamma({\mathcal{M}}_{X}(d,r,\alpha)) \ar[r]^-{\cong}& \operatorname{Spec}\Gamma({\mathcal{U}})
}$$ The variety ${\mathcal{M}}_{X}(d,r,\alpha)$ has a natural holomorphic symplectic structure, and the Hitchin map defines an algebraically complete integrable system, in particular, the fibers of $H$ are Lagrangians (these is explained in [@GL]).
When the parabolic set is empty ($n\,=\, 0$), Hausel proved that the nilpotent cone $H^{-1}(0)$ coincides with the downwards Morse flow on ${\mathcal{M}}_{X}(d,r,\alpha)$ giving a deformation retraction of ${\mathcal{M}}_{X}(d,r,\alpha)$ to $H^{-1}(0)$ [@hausel Theorem 5.2]. The proof in [@hausel] can be translated into the parabolic situation word by word.
The fiber of $H$ over a point $u \,\in\, {\mathcal{U}}$ is canonically isomorphic to the Jacobian of a curve called the *spectral curve*; we now recall its construction.
Given a point $u \,=\, (s_1\, ,\cdots\, ,s_r)\,\in\, {\mathcal{U}}$ consider the curve $X_u\,\subset\,\mathcal{K(D)}$ defined by the equation $$\widetilde{x}^r + s_1 \widetilde{x}^{r-1} + s_2 \widetilde{x}^{r-2} +\cdots +s_r\,=\,0$$ (compare it with (\[eq:char\_pol\])). Note that when $X_u$ is reduced, the projection $$\rho\,:=\, \gamma\vert_{X_u}\,:\, X_u\,\longrightarrow\, X$$ is a ramified covering of $X$ of degree $r$ which is completely ramified over the parabolic points. Denote by $R_u$ the ramification divisor on $X_u$. Denote by ${\mathcal{S}}$ the family of spectral curves over ${\mathcal{U}}$.
For any $u\,\in\,{\mathcal{U}}$ such that the corresponding spectral curve $X_u$ is smooth, the fiber $H^{-1}(u)$ is identified with ${\rm Pic}^{d+r(r-1)(2g-2+n)/2}(X_u)$.
It follows from the proof of Proposition 3.6. in [@BNR].
Let ${\mathcal{E}}$ be a universal bundle on $X\times {\mathcal{M}}_{X}(d,r,\alpha)$ and let $q$ be the projection to ${\mathcal{M}}_{X}(d,r,\alpha)$. Fix a point $x\in X$ of the curve. Let $\chi=\chi(E)$ (since we have fixed the rank and degree, this does not depend on the particular $E$ chosen, and can be calculated by the Riemann-Roch formula). There is a line bundle ${\mathcal{L}}^x$ defined as follows [@KM] $$\label{linebundle}
{\mathcal{L}}^x= {\rm det}(Rq_* {\mathcal{E}})^{-r} \otimes
(\wedge^{r} {\mathcal{E}}|_{x\times {\mathcal{M}}})^{\chi}$$ where the presence of the second factor is a normalization that guarantees that this does not depend on the choice of universal bundle. Note that this determinant line bundle can also be defined for the moduli space $M_{X}(d,r,\alpha)$ without Higgs bundle.
We remark that this line bundle is invariant under the standard $\mathbb{C}^*$ action and we can choose a lift of this action.
The fiber of this line bundle over a point corresponding to a Higgs bundle $(E,\Phi)$ is canonically isomorphic to $$\Big[
(\wedge^{top} H^0(X,E))^* \otimes
(\wedge^{top} H^1(X,E))
\Big]^{\otimes r}
\otimes
({\wedge} E_x)^{\chi}$$ Since the curve $X$ is connected, the Néron-Severi class $NS({\mathcal{L}}^x)$ of the line bundle does not depend on the choice of the point $x\in X$.
Let $u\in \mathcal{U}$ is a point in the Hitchin space corresponding to a smooth curve, then the restriction of the line bundle ${\mathcal{L}}^x$ to the fiber $H^{-1}(u)={\rm Pic}^{d+r(r-1)(2g-2+n)/2}(X_u)$ is a multiple of the principal polarization of the Jacobian $J(X_u)$ of the spectral curve $X_u$
Let $(E,\Phi)$ be a point in the moduli space ${\mathcal{M}}$. If it is in the fiber $H^{-1}(u)$, then there is a line bundle $\eta$ on the spectral curve $\pi:X_u\longrightarrow X$ such that $E=\pi_* \eta$. Then the fiber of ${\mathcal{L}}^x$ over this point is canonically isomorphic to $$\Big[
(\wedge^{top} H^0(X,\pi_* \eta))^* \otimes
(\wedge^{top} H^1(X,\pi_* \eta))
\Big]^{\otimes r}
\otimes
({\wedge} (\pi_* \eta)_x)^{\chi}
=$$ $$\Big[
(\wedge^{top} H^0(X_s, \eta))^* \otimes
(\wedge^{top} H^1(X_s, \eta))
\Big]^{\otimes r}
\otimes
({\wedge} \eta_{\pi^{-1}(x)})^{\chi}$$ This is the fiber of a line bundle defining a multiple of a principal polarization of the Jacobian. The last factor is just a normalization, and the Néron-Severi class of the line bundle does not depend on the choice of the point.
In [@Hu], Hurtubise considers (local) integrable systems $$\mathbb{H}\,:\,\mathbb{J}\,\longrightarrow\, \mathbb{U}\, ,$$ where $\mathbb{U}$ is an open subset of ${\mathbb{C}}^m$ and $\mathbb{J}$ is a $2m$-dimensional symplectic variety with holomorphic symplectic form $\Omega$, such that the fibers of $\mathbb{H}$ are Lagrangian. Furthermore, suppose there is a family of curves $$\mathbb{H}'\,:\, \mathbb{S}\,\longrightarrow\, \mathbb{U}$$ such that for each $u\,\in\, \mathbb{U}$, the fiber $J_u\,=\,\mathbb{H}^{-1}(u)$ is isomorphic to the Jacobian of $S_u\,=\,\mathbb{H}'{}^{-1}(u)$. To define the Abel map $$I\,:\,\mathbb{S}\,\longrightarrow\, \mathbb{J}$$ we need a section of $\mathbb{H}'$. This can be done locally on $\mathbb{U}$. Under the assumption that $$\label{i20}
I^*\Omega \wedge I^*\Omega\,=\,0$$ Hurtubise proves that for the embedding $I$ the variety $\mathbb{S}$ is coisotropic, and the quotienting of $\mathbb{S}$ by the null foliation results a surface $Q$. The form $I^*\Omega$ descends to $Q$, and the descended form on $Q$, which we will denote by $\omega$, is a holomorphic symplectic form [@Hu Theorem 1.11]. He also proves that, choosing a different Abel map $I'$ with $I'{}^*\Omega \wedge I'{}^*\Omega\,=\,0$, we have $I^*\Omega\,=\,I'{}^*\Omega$ when $m\,\geq\, 3$, so that the surface $Q$ depends only on $\mathbb{S}$ and it is independent of the Abel map. We summarize:
\[thm:hurtubise\] For an integrable system $${\mathbb{H}}\,:\,\mathbb{J}\,\longrightarrow\, \mathbb{U}
\,\subset\, {\mathbb{C}}^{m}\, ,$$ with maps ${\mathbb{H}}'\,: \,\mathbb{S}\,\longrightarrow\, \mathbb{U}$, and $I\,:\,\mathbb{S}\,\longrightarrow\, \mathbb{J}$, as described above, there is an invariant surface $Q$ which only depends on $\mathbb{S}$ and not on the Abel map $I$, whenever $m\,\ge\, 3$.
In [@Hu Example 4.3] he shows that all these conditions are satisfied for the usual moduli space of Higgs bundles (i.e., no parabolic points), but restricted to the open subset $U$ of the Hitchin space $\mathcal{U}$ corresponding to smooth spectral curves $$\mathbb{H}: {\mathcal{M}}_X|_U \longrightarrow U \; .$$ Let $q$ be the projection $q:\mathbb{S}\longrightarrow \mathcal{K}$ sending each point on a spectral curve to the total space of the cotangent bundle and let $\omega$ the natural symplectic form on the cotangent. Hurtubise shows that $$I^* \Omega = q^* \omega$$ It follows that the surface $Q$ is $\mathcal{K}$.
The conditions of the theorem also hold for the moduli space of strongly parabolic Higgs bundles equipped with the Hitchin map, and in this case the surface $Q$ is the image of $\mathbb{S}\longrightarrow \mathcal{K(D)}$. Note that all spectral curves go through zero on the fibers over the parabolic points, because the eigenvalues of the residues are zero. Therefore, we obtain the following Corollary, which will be our main tool in the proof of the Main Theorem.
Note that the integer $m$ in the statement of Theorem \[thm:hurtubise\] is the genus of the spectral curve, which is equal to $\dim M_X(d,r,\alpha)$ and hence, under the assumptions on genus and rank of Theorems \[theoremPH\] and \[theoremP\] we always have $m\geq 3$ and hence can apply the Theorem of Hurtubise.
\[MainCor\] Let $$\mathbb{H}: {\mathcal{M}}_{X}(d,r,\alpha)|_U \longrightarrow U \; .$$ be the restriction of the Hitchin map on the moduli space of parabolic Higgs bundles with generic weights $\alpha$ to the open set $U$ corresponding to nonsingular curves (cf. Lemma ). Then this integrable system satisfies the conditions of the Theorem of Hurtubise and the surface $Q$ is the image of $\mathcal{K}$ in $\mathcal{K(D)}$ under the injective morphism of sheaves $K\longrightarrow K(D)$.
Proof of the Theorems
=====================
Let $$h\,:\,T^{\ast}M_{X}(d,r,\alpha)\,\longrightarrow \,{\mathcal{U}}\,=\, \bigoplus_{i=1}^{r}H^{0}(X,K^{i}((i-1)D))$$ be the restriction to the cotangent bundle of the moduli space of stable bundles of the Hitchin integrable system in (\[eqn:hitchin\]). To each point $u\,\in\, {\mathcal{U}}$ we associated its spectral curve $X_u\,\subset\, {\mathcal{S}}$.
\[lem:U\] If $g \geq 2$, then the Zariski open subset $U$ of ${\mathcal{U}}$ that parametrizes the smooth spectral curves is non-empty.
If $K^{r}((r-1)D)$ has a section without multiple zeros, then the above open subset $U$ is nonempty (cf. [@BNR Remark 3.5]). A holomorphic line bundle on $X$ of degree at least $2g+1$ is very ample (cf. [@Ha IV Corollary 3.2]), and hence $U$ is non-empty whenever $r(2g-2)+(r-1)n\,\ge\, 2g+1$, and this holds when $g\geq 2$.
Define ${\mathcal{J}}\,:=\,H^{-1}(U)$, where $H$ is the Hitchin map for the moduli of Higgs bundles and $U$ is the open subset in Lemma \[lem:U\]. Let $$H_{{\mathcal{J}}}\,:\,{\mathcal{J}}\,\longrightarrow\, U$$ be the restriction of $H$. Let $$H_{{\mathcal{S}}}\,:\,{\mathcal{S}}\, \longrightarrow\, U$$ be the total space for the family of spectral curve over $U$, so that the fiber of $H_{{\mathcal{S}}}$ over any $u\,\in\, U$ is the spectral curve $X_{u}$.
As we have seen in Corollary \[MainCor\], the surface $Q$ given by the Theorem of Hurtubise in this setting is the image of $\mathcal{K}$ in $\mathcal{K(D)}$ under the injective morphism of sheaves $K\longrightarrow K(D)$. In particular, $Q$ is singular.
The moduli space of parabolic Higgs bundles is known to be a Kähler manifold provided with a ${\mathbb{C}}^{\ast}$ action whose restriction to an ${\rm S}^1$ action preserves the Kähler structure $$\begin{aligned}
\label{eqn:c-action}
\tau:{\mathbb{C}}^{\ast} \times {\mathcal{M}}_{X}(r,d,\alpha) &\longrightarrow& {\mathcal{M}}_X(r,d,\alpha)\\
(t\, ,(E\, ,\Phi))&\longmapsto& (E\, ,t\Phi).\notag\end{aligned}$$ This $\mathbb{C}^*$ action is compatible with scalar multiplication in the fibers of the cotangent bundle $T^\ast M_X(r,d,\alpha)$ under the inclusion of this cotangent bundle in the moduli of parabolic Higgs bundles. It induces a ${\mathbb{C}}^{\ast}$ action on ${\mathcal{S}}$: $$\begin{aligned}
\label{eq:CQ}
{\mathbb{C}}^{\ast} \times {\mathcal{S}}&\longrightarrow& {\mathcal{S}}\\
(t\, ,x\in X_{u})&\longmapsto& (tx \in X_{t\cdot u}),\notag\end{aligned}$$ where $t\cdot (s_1\, , \cdots\, , s_r)\,=\,
(t s_1\, , t^2 s_2\, , \cdots\, , t^{r}s_r)$ (see ), and the multiplication $tx$ is defined using the embedding of the spectral curve $X_u$ in the total space of $K(D)$. This action of ${\mathbb{C}}^{\ast}$ on ${\mathcal{S}}$ evidently produces an action of ${\mathbb{C}}^{\ast}$ on the quotient surface $Q$. Let $$Q^{{\mathbb{C}}^{\ast}}\subset Q$$ be the fixed point locus for the above ${\mathbb{C}}^{\ast}$ action on $Q$.
The subset $Q^{{\mathbb{C}}^{\ast}}$ is the zero section of the fibration $\mathcal{K(D)}\longrightarrow X$.
Since the natural inclusion $K \, \hookrightarrow\, K(D)$ of ${\mathcal O}_X$–modules commutes with the multiplicative action of ${\mathbb{C}}^{\ast}$, the surface $Q$, which is the image of the total space of $K$ in $\mathcal{K}(\mathcal{D})$, is preserved by the action of ${\mathbb{C}}^{\ast}$ on $\mathcal{K}(\mathcal{D})$. Therefore, the action of ${\mathbb{C}}^{\ast}$ on $\mathcal{K}(\mathcal{D})$ produces an action of ${\mathbb{C}}^{\ast}$ on $Q$. This action of ${\mathbb{C}}^{\ast}$ on $Q$ coincides with the action on $Q$ induced by (\[eq:CQ\]). The lemma follows from this.
\[lem:curve\] The curve $X$ coincides with $Q^{{\mathbb{C}}^{\ast}}$.
\[prop:points\] The set of parabolic points coincides with the subset of $Q^{{\mathbb{C}}^{\ast}}$ through which every spectral cover pass.
Since the residue of $\Phi$ on the parabolic points is nilpotent, all spectral curves $X_{u}$ totally ramify over the parabolic points, and they intersect the fibre over the parabolic points at zero.
Conversely, let $x\in X$ be a point which is not parabolic. There exists a section $s_r\in H^0(K^r((r-1)D))$ which does not vanish at $x$ since this linear systems is base point free (recall that we are assuming $g\geq 2$). Furthermore, this section has still no zero on $x$ when considered as a section of $H^0(K^r(rD))$ because $x$ is not a parabolic point. Therefore, the spectral curve $\widetilde{x}^r +s_r=0$ on $\mathcal{K(D)}$ intersects the fibre over $x$ away from zero, and the spectral curve $\widetilde{x}^r=0$ intersects it only at zero, so there is no point over the fibre of $x$ through which every spectral cover passes.
Proof of Theorem \[theoremPH\]
------------------------------
We are given the moduli space ${\mathcal{M}}$ as an abstract algebraic variety with a holomorphic symplectic form, a line bundle ${\mathcal{L}}$ and a algebraic $\mathbb{C}^*$ action on ${\mathcal{M}}$ with a linearization on ${\mathcal{L}}$. Looking at global functions on ${\mathcal{M}}$ $$\alpha:{\mathcal{M}}\longrightarrow \operatorname{Spec}\Gamma({\mathcal{M}})$$ we obtain a morphism $\alpha$ which is isomorphic to the Hitchin fibration (cf. ) and the fibers are Lagrangians with respect to the given holomorphic symplectic form. The subset $U\subset
\operatorname{Spec}\Gamma({\mathcal{M}})$ of points corresponding to smooth spectral curves can be recovered as the points whose fibers are abelian varieties. Let $\beta$ be the restriction of $\alpha$ over $U$ $$\xymatrix{
{{\mathcal{J}}} \ar@{^{(}->}[r] \ar[d]_{\beta}& {{\mathcal{M}}} \ar[d]^{\alpha} \\
{U} \ar@{^{(}->}[r] & {\operatorname{Spec}\Gamma{{\mathcal{M}}}}\\
}$$ The line bundle ${\mathcal{L}}$ restricts to a (multiple of) a principal polarization on these abelian varieties, and then the classical Torelli theorem gives us a family of curves ${\mathcal{S}}\longrightarrow U$, such that the fiber ${\mathcal{J}}_u$ over $u\in U$ is the Jacobian of ${\mathcal{S}}_u$. Locally on $U$ there is an Abel-Jacobi map $I:{\mathcal{S}}\longrightarrow {\mathcal{J}}$.
The $\mathbb{C}^*$ action on ${\mathcal{M}}$ restricts to a $\mathbb{C}^*$ action on the family of Jacobians ${\mathcal{J}}$. This family of Jacobians has a family of principal polarizations given by the line bundle ${\mathcal{L}}$. The $\mathbb{C}^*$ action has a lift to ${\mathcal{L}}$ hence we have an action on the family of principal polarized Jacobians.
By the proof given by Weil of the Torelli theorem [@We Hauptsatz, p. 35], an isomorphism $\psi:({\mathcal{J}}_u,\theta_u)\longrightarrow ({\mathcal{J}}_{u'},\theta_{u'})$ of principal polarized Jacobians induces an isomorphism $f:{\mathcal{S}}_u\longrightarrow {\mathcal{S}}_u'$ of the corresponding curves, and this provides an action of $\mathbb{C}^*$ on the family of curves ${\mathcal{S}}$.
Now we apply Corollary \[MainCor\] to obtain a surface $Q$ as a quotient of ${\mathcal{S}}$. The action on ${\mathcal{S}}$ we have just defined clearly coincides with the action given in , therefore by Corollary \[lem:curve\] we recover $X$, and by Proposition \[prop:points\] we recover the parabolic points $D$, thus proving our main theorem.
Proof of Theorem \[theoremP\]
-----------------------------
We are given the moduli space as a smooth algebraic variety $M$ with a line bundle ${\mathcal{L}}$. We consider the total space of the cotangent bundle $T^*M$. This has a canonical holomorphic symplectic structure, and a $\mathbb{C}^*$ given by scalar multiplication on the fibers. The pullback of the line bundle to $T^* M$ is trivial along the fibers, so there is a canonical lift of the $\mathbb{C}^*$ action to the pullback of the line bundle ${\mathcal{L}}$ to $T^* M$.
We claim that the generic fiber of the morphism given by global sections $$h : T^* M \longrightarrow \operatorname{Spec}(\Gamma(T^* M))$$ is an open subset of an abelian variety. Indeed, we know that $M$ is the moduli space for some algebraic curve $X$ (which we want to find), so we know that $T^* M$ is an open subset of a moduli space of parabolic Higgs bundles ${\mathcal{M}}$, and by Corollary \[codimfibre\] we know that the codimension of the complement of this open set is at least two. Therefore, global section on $T^*M$ extend uniquely to global sections on ${\mathcal{M}}$ and the morphism $h$ is the restriction of the morphism of global sections of some moduli space of Higgs bundles ${\mathcal{M}}$ $$\xymatrix{
{T^* M} \ar[r]^-{h} \ar@{^{(}->}[d] & {\operatorname{Spec}\Gamma (T^* M)} \ar@{=}[d]\\
{{\mathcal{M}}} \ar[r]^-{H} & {\operatorname{Spec}\Gamma ({\mathcal{M}})} \\
}$$ The compactification of the fiber over $u$ to an abelian variety is unique, because birational abelian varieties are isomorphic. Therefore, the isomorphism class of ${\mathcal{J}}:=H^{-1}(U)$ is uniquely defined by the isomorphism class of $M$, and does not depend on the choice of ${\mathcal{M}}$.
Since the codimension of the complement of the inclusion $T^* M|_U\subset {\mathcal{J}}$ is at least two, all the structure that we have on $T^* M$ extends uniquely to ${\mathcal{J}}$, namely the determinant line bundle ${\mathcal{L}}$, the $\mathbb{C}^*$ action with the lift to ${\mathcal{L}}$ and the holomorphic symplectic form. Therefore we can now use the same arguments as in the proof of the main theorem to recover the curve $X$ and the parabolic points.
Codimension computation
=======================
In this section we compute the codimension of the complement of $T^{\ast}M(d,r,\alpha) $ inside ${\mathcal{M}}(d,r,\alpha)$ fiber-wise following the arguments in [@biswas-gothen-logares Section 5]. This complement is $${\mathcal{V}}=\{(E,\Phi)\in {\mathcal{M}}(d,r,\alpha)\,\mid\, E\,\mathrm{is\, not\,
stable}\,\} \; .$$
Recall from (\[eqn:c-action\]) that the moduli space of parabolic Higgs bundles is known to be a Kähler manifold provided with a ${\mathbb{C}}^{\ast}$ action, whose restriction to a ${\rm S}^1$ action preserves the Kähler structure.
This action provide us with two stratifications of the moduli space. The first one is the Bia[ł]{}ynicki-Birula stratification consisting of subsets of ${\mathcal{M}}(d,r,\alpha)$ such $$U^{+}_{\lambda}:=\{p\in {\mathcal{M}}_X(d,r,\alpha) ; \lim_{t\rightarrow 0} tp\in F_{\lambda}\}$$ and $$U^{-}_{\lambda}:=\{p\in {\mathcal{M}}_X(d,r,\alpha) ;\lim_{t\rightarrow \infty}tp\in F_{\lambda}\}$$ where $F_\lambda$ are the disjoint connected components of the fixed pointed set $F$ for the ${\mathbb{C}}^{\ast}$-action on ${\mathcal{M}}(d,r,\alpha)$.
The second one is known as the Morse stratification and comes from the restriction of the ${\mathbb{C}}^{\ast}$-action to an $S^{1}$-action. The last also preserves the Kähler form, hence it give us a circle Hamiltonian action on ${\mathcal{M}}(d,r,\alpha)$ with associated moment map, $$\begin{aligned}
\mu:{\mathcal{M}}_X(r,d,\alpha)&\longrightarrow& {\mathbb{R}}\\
(E,\Phi)&\longmapsto & ||\Phi||^{2}\end{aligned}$$ which is proper, bounded below and has a finite number of critical submanifolds. So this map is a Morse-Bott map.
For any component $F_{\lambda}$, we recall the definition of the upwards Morse strata, $\widetilde{U}^{+}_{\lambda}$, and the downwards Morse strata, $\widetilde{U}^{-}_{\lambda}$, that is $$\widetilde{U}^{+}_{\lambda}:=\{p\in {\mathcal{M}}_X(d,r,\alpha); \lim_{t\rightarrow -\infty} \psi_{t}(p)\in F_{\lambda} \}$$ and $$\widetilde{U}^{-}_{\lambda}:=\{ p\in {\mathcal{M}}_X(d,r,\alpha); \lim_{t\rightarrow +\infty} \psi_{t}(p)\in F_{\lambda} \}$$ Recall that this stratifications were proven to be equal $U^{+}=\widetilde{U}^{+}$ and $U^{-}=\widetilde{U}^{-}$ by Kirwan in [@kirwan Theorem 6.16].
The union $N=\bigcup_{\lambda}\widetilde{U}^{-}_{\lambda}$ is known as *downwards Morse flow*.
The inverse over the $0$ point of the Hitchin map $H^{-1}(0)$ is called *nilpotent cone*, and it coincides with the downwards Morse flow, i.e. $N=H^{-1}(0)$ [@garcia-prada-gothen-munoz Theorem 3.13].
The following proposition takes the same steps as Proposition 5.1 in [@biswas-gothen-logares] provided that in a family of parabolic bundles the Harder-Narasimhan type increases under specialization.
\[V=V’\] Let ${\mathcal{V}}$ be the complement of the cotangent bundle of $M_{X}(d,r,\alpha)$ in ${\mathcal{M}}_{X}(d,r,\alpha)$ and let ${\mathcal{V}}'$ be the Bia[ł]{}ynicki-Birula flow which does not converge to $M_X(d,r,\alpha)$ , that is $${\mathcal{V}}= \{(E,\Phi)\in{\mathcal{M}}_{X}(d,r,\alpha) : E \,\mathrm{is\, not\,
stable}\}\quad \mathrm{and}\quad {\mathcal{V}}'=
\{(E,\Phi): \lim_{t\rightarrow 0} (E,t\Phi)\notin M_X(d,r,\alpha)\}.$$ Then $${\mathcal{V}}'={\mathcal{V}}.$$
Let $(E,\Phi)\notin {\mathcal{V}}$ that is $E$ is stable, then $\lim_{t\rightarrow 0}(E,t\Phi)=(E,0)\in M_X(d,r,\alpha)$, so it proves that ${\mathcal{V}}'\subset {\mathcal{V}}$. To prove the converse, take $(E,\Phi)$ where $E$ is not stable. There exists a Harder-Narasimhan filtration for $E$, that is $$E=E_{m}\supset E_{m-1}\supset \cdots\supset E_{1}\supset 0.$$ Following Atiyah and Bott [@AB] we define the type of the Harder-Narasimhan filtration, as the following vector $(\mu_{1},\ldots,\mu_{r})$ where $\mu_{i}=\deg(F_{i})/\operatorname{rk \,}(F_{i})$ and $F_{i}=E_{i}/E_{i+1}$.
In a family of parabolic bundles Nitsure [@nitsure Proposition 1.10] proved that the Harder-Narasimhan type increases under specialization. Hence, as the Hitchin map is proper, and its composition with the map given by the ${\mathbb{C}}^{\ast}$-action, is such that $\lim_{t\rightarrow 0}h(E,t\Phi)=0$ then $\tau$ extends to a morphism $$\widetilde{\tau}:{\mathbb{C}}\times {\mathcal{M}}(d,r,\alpha) \longrightarrow {\mathcal{M}}(d,r,\alpha).$$ This maps gives, by pullback, that a family over ${\mathbb{C}}^{\ast}$ of non semistable parabolic bundles specializes to a non semistable parabolic bundle. That is, for ${\mathbb{E}}$, the universal bundle over ${\mathcal{M}}(d,r,\alpha)\times X$ when pullback $(\widetilde{\tau}\times
1_{X})^{\ast}({\mathbb{E}})$ gives a family of bundles over $X$ parametrized by ${\mathbb{C}}$. If $(\widetilde{\tau}\times 1_{X})^{\ast}({\mathbb{E}})|_{t}\times X=E$ for $t\neq 0$, is not semistable, then the specialization result says that $(\widetilde{\tau}\times 1_{X})^{\ast}({\mathbb{E}})|_{{0}\times X}$ is not semistable. This completes the proof.
The following facts are recovered from the literature on parabolic Higgs bundles.
Let $(E,\Phi)$ be a fixed point for the circle action, we have an isomorphism $(E,\Phi)\cong (E,e^{{i}\theta}\Phi)$ for $\theta\in
[0,2\pi)$ yielding the following commutative diagram. $$\xymatrix{
E \ar[r]^{\hspace*{-20pt}\Phi}\ar[d]^{\psi_{\theta}} & E\otimes K(D)\ar[d]^{\psi_{\theta}\otimes 1_{K(D)}}\\
E \ar[r]^{\hspace*{-20pt}e^{{i}\theta}\Phi} & E\otimes K(D).
}$$
\[fixed\_points:decomposition\] If $(E,\Phi)$ belongs to a critical subvariety $F_{\lambda}$ for the circle action on ${\mathcal{M}}_{X}(d,r,\alpha)$ then $E$ splits $$E=\bigoplus_{l=0}^{m} E_{l}$$ and $\Phi\in H^{0}(\operatorname{SParHom\,}(E_{l},E_{l+1})\otimes K(D))$.
The parabolic Higgs bundle in this case $(E,\Phi)$ is called Hodge bundle.
The deformation theory of the moduli space of parabolic Higgs bundles was worked out in [@y]. It is given by the following complex of bundles, $$C^{\bullet}(E): \operatorname{ParEnd \,}(E)\stackrel{\Phi:=[\cdot,\Phi]}{\longrightarrow} \operatorname{SParEnd\,}(E)\otimes K(D).$$
The tangent space of the moduli space ${\mathcal{M}}_X (d, r, \alpha)$ at a stable point $(E,\Phi)$ is then the first cohomology group ${\mathbb{H}}^{1} (C^{\bullet}(E))$ of this complex. Hence for a fixed point $(E,\Phi)$ of the ${\mathbb{C}}^{\ast}$-action, the decomposition in Proposition \[fixed\_points:decomposition\] induces a decomposition of the deformation complex and of the tangent space at the fixed point. That is, we define $$C_{k}:= \bigoplus_{j-i=k}\operatorname{ParHom \,}(E_{i},E_{j}) \quad \mathrm{and} \quad \widehat{C}_{k+1}:=\bigoplus_{j-i=k}\operatorname{SParHom\,}(E_{i},E_{j})$$ so then $$C^{\bullet}(E)_{k}: C_{k}\stackrel{\Phi_{k}}{\longrightarrow} \widehat{C}_{k+1}\otimes K(D),$$ and $$C^{\bullet}(E)=\bigoplus_{k=-m-1}^{k=m} C^{\bullet}(E)_{k}.$$
For this deformation complex, there is a long exact sequence $$\begin{aligned}
&0&\rightarrow {\mathbb{H}}^{0}(C^{\bullet}(E)_k)\rightarrow H^{0}(\bigoplus_{j-i=k}\operatorname{ParHom \,}(E_{i},E_{j})\rightarrow H^{0}(\bigoplus_{j-i=k}\operatorname{SParHom\,}(E_{i},E_{j})\otimes K(D))\\
&&\rightarrow {\mathbb{H}}^{1}(C^{\bullet}(E)_k)\rightarrow H^{1}(\bigoplus_{j-i=k}\operatorname{ParHom \,}(E_{i},E_{j})\rightarrow H^{1}(\bigoplus_{j-i=k}\operatorname{SParHom\,}(E_{i},E_{j})\otimes K(D))\\
&&\rightarrow {\mathbb{H}}^{2}(C^{\bullet}(E)_{k})\rightarrow 0.\end{aligned}$$
- There is a natural isomorphism $${\mathbb{H}}^{1}(C^{\bullet}(E)_{k})\simeq {\mathbb{H}}^{1}(C^{\bullet}(E)_{-k-1})^{\ast}$$ and hence a natural isomorphism $$T_{(E,\Phi)}{\mathcal{M}}_{k}\simeq (T_{(E,\Phi)}{\mathcal{M}}_{1-k})^{\ast}$$
- If $(E,\Phi)$ is stable, then we have $${\mathbb{H}}^{0}(C^{\bullet}(E)_{k})=\left\lbrace \begin{array}{ll} {\mathbb{C}}& if \; k=0\\ 0 & otherwise,\end{array}\right.$$ and $${\mathbb{H}}^{2}(C^{\bullet}(E)_{k})=\left\lbrace \begin{array}{ll} {\mathbb{C}}& if \; k=-1\\ 0 & otherwise.\end{array}\right.$$
$\Box$
Hence,
$$\dim {\mathbb{H}}^{1}(C^{\bullet}(E)_{k})=\left\lbrace \begin{array}{ll} 1-\chi(C^{\bullet}(E)_{k}) & if \; k=0\\ -\chi(C^{\bullet}(E)_{k}) & otherwise.\end{array}\right.$$
$\Box$
The function $\mu:{\mathcal{M}}_{X}(r,d,\alpha)\longrightarrow {\mathbb{R}}$ defined by $\mu(E,\Phi)=\| \Phi\|^{2}$ is a perfect Bott–Morse function. A parabolic Higgs bundle represents a critical point of $\mu$ if and only if it is a parabolic complex variation of Hodge structure, i.e. $E=\bigoplus _{k=0}^{m} E_k$ with $\Phi_{k}=\Phi|_{E_{k}}:\,E_{l}\longrightarrow E_{k+1}\otimes K(D)$ strongly parabolic (where $\Phi=0$ if and only if $m=0$). The tangent space to ${\mathcal{M}}_{X}(r,d,\alpha)$ at a critical point $(E,\Phi)$ decomposes as $$T_{(E,\Phi)}{\mathcal{M}}_{X}(r,d,\alpha)=\bigoplus_{k=-m}^{m+1} T_{(E,\Phi)}{\mathcal{M}}_{X}(r,d,\alpha)_{k}$$ where the eigenvalue $k$ subspace of the Hessian of $\mu$ is $$T_{(E,\Phi)}{\mathcal{M}}_{X}(r,d,\alpha)_{k} \cong {\mathbb{H}}^{1}(C^{\bullet}(E)_{k} ).$$
$\Box$
For a critical point $(E,\Phi)$ of $\mu$ we denote by $T_{(E,\Phi)}{\mathcal{M}}_{X}(r,d,\alpha)_{<0}$ the subspace of the tangent space on which the Hessian of $\mu$ has negative eigenvalues. The real dimension of this subspace is called the Morse index at the point $(E,\Phi)$.
\[prop:codim\] The codimension of the complement of $T^\ast M_{X}(d,r,\alpha)$ in ${\mathcal{M}}_X(d,r,\alpha)$ is equal to half of the minima of the Morse indexes at points $(E,\Phi)\in F_{\lambda}$ for $\lambda\neq 0$.
The complement of $T^\ast M_X(d,r,\alpha)$ is equal to ${\mathcal{V}}$ which is also equal to ${\mathcal{V}}'$ from Proposition \[V=V’\]. Bott–Morse theory give us that ${\mathcal{V}}'=\bigcup_{\lambda\neq 0}U^{+}_{\lambda}$, so we conclude that $$\operatorname{codim}({\mathcal{V}})=\min_{\lambda \neq 0} \operatorname{codim}U^{+}_{\lambda}.$$ From Bott–Morse theory we also know that the dimension of the upwards Morse flow is such that $$\dim U^{+}_{\lambda}+\dim T_{(E,\Phi)}\,{\mathcal{M}}_{X}(d,r,\alpha)_{<0}=\dim {\mathcal{M}}_{X}(d,r,\alpha),$$ where $T_{E}{\mathcal{M}}_{X}(d,r,\alpha)_{<0}$ is the negative eigenspace for the Hessian of the perfect Bott–Morse function $\mu$ for an $E\in U^{+}_{\lambda}$. As the Morse index $\mu_{\lambda}= 2 \dim T_{E}{\mathcal{M}}_{X}(d,r,\alpha)_{<0}$, $$\operatorname{codim}U^{+}_{\lambda}= \frac{1}{2}\mu_{\lambda}.$$ Our statement is then $$\operatorname{codim}({\mathcal{V}})=\min_{\lambda\neq 0} \frac{1}{2}\mu_{\lambda}$$ that is, $$\operatorname{codim}({\mathcal{V}})=\min_{\lambda\neq 0} \dim T_{E}{\mathcal{M}}_{X}(d,r,\alpha)_{<0}$$
$$T_{(E,\Phi)}{\mathcal{M}}_{X}(d,r,\alpha)_{<0}=\sum_{k>0} -\chi(C^{\bullet}(E)_{k}.$$
$ \Box$
So, we need to bound the Euler characteristic for any $k$.
\[prop:chi\] $$-\chi(C^{\bullet}(E)_{k})\ge (g-1)(\operatorname{rk \,}(C_{k}-\operatorname{rk \,}(\widehat{C}_{k+1}))$$
Recall that
$$\begin{aligned}
\label{eq:euler}
\chi(C^{\bullet}(E)_{k})&=&\dim H^{0}(C_{k})-\dim H^{1}(C_{k})-\dim H^{0}(\widehat{C}_{k+1}\otimes K(D))+\dim H^{1}(\widehat{C}_{k+1}\otimes K(D))\notag\\
&=&\deg(C_{k})-\deg(\widehat{C}_{k+1})-\operatorname{rk \,}(\widehat{C}_{k+1})\deg(K(D))+(\operatorname{rk \,}(C_{k})-\operatorname{rk \,}(\widehat{C}_{k+1}))(1-g).\end{aligned}$$
We first bound $\deg(C_k)-\deg(\widehat{C}_{k+1})$. Consider the following short exact sequences of bundles, $$\begin{aligned}
&&0\longrightarrow \ker(\Phi_{k})\longrightarrow C_k \longrightarrow \operatorname{im \,}(\Phi_{k})\longrightarrow 0\\
&&0\longrightarrow \operatorname{im \,}(\Phi_{k})\longrightarrow \widehat{C}_{k+1}\otimes K(D) \longrightarrow \operatorname{coker\,}(\Phi_{k})\longrightarrow 0.\end{aligned}$$ then $$\label{eq:bound1}
\deg(C_{k})-\deg(\widehat{C}_{k+1})=\deg(\ker(\Phi_{k}))+\deg(K(D))\operatorname{rk \,}(\widehat{C}_{k+1})-\deg(\operatorname{coker\,}(\Phi_{k})).$$
The $\ker(\Phi_{k})\subset\operatorname{ParEnd \,}(E)$ is a subbundle of the bundle of parabolic endomorphisms of $E$, which we claim is semistable whenever $E$ is stable (see Lemma \[lem:ss\]). Hence $\operatorname{pardeg \,}(\ker(\Phi_{k}))\le 0$ and this implies $\deg(\ker(\Phi_{k}))\le 0$.
Hence $$\deg(C_{k})-\deg(\widehat{C}_{k+1})\le \deg(K(D))\operatorname{rk \,}(\widehat{C}_{k+1})-\deg(\operatorname{coker\,}(\Phi_{k})).$$ We also get that $$\label{eq:coker}
-\deg(\operatorname{coker\,}(\Phi_{k}))\le (2-2g)(\operatorname{rk \,}(\widehat{C}_{k+1})-\operatorname{rk \,}(\Phi_{k})),$$ so that
$$\label{eq:bound-deg}
\deg(C_{k})-\deg(\widehat{C}_{k+1})\le n\operatorname{rk \,}(\widehat{C}_{k+1})+(2g-2)\operatorname{rk \,}(\Phi_{k})$$
in the following way.
Note that for any two parabolic bundles $E$, $F$, then $\operatorname{ParHom \,}(E,F)^{\ast}=\operatorname{SParHom\,}(F,E)\otimes {\mathcal{O}}(D)$. So then $C_{k}^{\ast}=(\bigoplus_{j-i=k} \operatorname{ParHom \,}(E_{i},E_{j}))^{\ast}=\bigoplus_{j-i=k} \operatorname{SParHom\,}(E_{j},E_{i})\otimes {\mathcal{O}}(D)=\widehat{C}_{k}\otimes{\mathcal{O}}(D)$.
Consider the adjoint map $$\Phi_{k}^{t}:(\widehat{C}_{k+1}\otimes K(D))^{\ast}\longrightarrow (C_{k})^{\ast}$$ then $$\ker(\Phi_{k}^{t})\hookrightarrow (\widehat{C}_{k+1}\otimes K(D))^{\ast}\cong C_{-1-k}\otimes K^{-1}.$$ Dualizing again we get a surjective homomorphism, $$\widehat{C}_{k+1}\otimes K(D)\longrightarrow (\ker(\Phi_{k}^{t}))^{\ast}$$
Define the homomorphism $$f: \operatorname{coker\,}(\Phi_{k})\longrightarrow \ker(\Phi_{k}^{t})^{\ast}$$ which makes the following diagram commutative $$\xymatrix{
0\ar[r]&\operatorname{im \,}(\Phi_{k})\ar[r] & \widehat{C}_{k+1}\otimes K(D)\ar[r]\ar@{=}[d]&\operatorname{coker\,}(\Phi_{k})\ar[r] \ar[d]^{f}&0\\
0\ar[r]&(\operatorname{im \,}(\Phi_{k}^{t}))^{\ast}\ar[r]& \widehat{C}_{k+1}\otimes K(D)\ar[r]&(\ker(\Phi_{k}^{t}))^{\ast}\ar[r]&0
}$$ Note that $f$ is surjective and $\ker(f)$ is a torsion subsheaf. Hence, $$0\longrightarrow \ker(f)\longrightarrow\operatorname{coker\,}(\Phi_{k})\longrightarrow (\ker(\Phi_{k}^{t}))^{\ast}\longrightarrow 0,$$ and $$\deg(\operatorname{coker\,}(\Phi_{k}))\ge \deg(\ker(\Phi_{k}^{t})^{\ast}).$$ As $\ker(\Phi_{k}^{t})$ is a sub bundle of $C_{k+1}\otimes K$, $$\label{eq:bound2}
-\deg(\operatorname{coker\,}(\Phi_{k}))\le \deg(\ker(\Phi_{k}^{t}))$$
Note that there are isomorphisms, making the following diagram commutative $$\xymatrix{
(\widehat{C}_{k+1}\otimes K(D))^{\ast} \ar[d]^{\cong} \ar[r]^{\Phi^{t}_{k}} &
(C_{k})^{\ast}\ar[d]^{\cong}\\
C_{-1-k}\otimes K^{-1}\ar[r]_{\Phi_{-1-k}\otimes 1_{K^{-1}}} &
\widehat{C}_{-k}\otimes {\mathcal{O}}(D),
}$$ therefore $$\Phi_{k}^{t}\cong \Phi_{-1-k}\otimes 1_{K^{-1}}$$ so $$\ker(\Phi^{t}_{k})=\ker(\Phi_{-1-k})\otimes K^{-1}$$ and $$\deg(\ker(\Phi^{t}_{k}))=\deg(\ker(\Phi_{-1-k}))+(2-2g)\operatorname{rk \,}(\ker(\Phi_{-1-k}))$$ Notice that $\operatorname{rk \,}(\Phi_{-1-k})=\operatorname{rk \,}(\Phi^{t}_{k})=\operatorname{rk \,}(\Phi_{k})$ and $\operatorname{rk \,}(\widehat{C}_{k+1})=\operatorname{rk \,}((\widehat{C}_{k+1})^{\ast})=\operatorname{rk \,}(C_{-1-k})$. Then $\operatorname{rk \,}(\ker(\Phi_{-1-k}))=\operatorname{rk \,}(\widehat{C}_{k+1})-\operatorname{rk \,}(\Phi_{k})$, so that, equation (\[eq:bound2\]) becomes $$-\deg(\operatorname{coker\,}\Phi_{k})\le \deg(\ker (\Phi_{-1-k}))+(2-2g)(\operatorname{rk \,}(\widehat{C}_{k+1})-\operatorname{rk \,}(\Phi_{k}))$$ and finally, by stability (see Lemma \[lem:ss\]), $$-\deg(\operatorname{coker\,}\Phi_{k})\le (2-2g)(\operatorname{rk \,}(\widehat{C}_{k+1})-\operatorname{rk \,}(\Phi_{k})).$$ This provides equation (\[eq:coker\]).
Putting together equations (\[eq:coker\]) and (\[eq:bound-deg\]) we get $$\chi(C^{\bullet}(E)_{k})\le (1-g)(\operatorname{rk \,}(C_{k})-\operatorname{rk \,}(\widehat{C}_{k+1})),$$ hence, $$-\chi(C^{\bullet}(E)_{k})\ge (g-1)(\operatorname{rk \,}(C_{k})-\operatorname{rk \,}(\widehat{C}_{k+1})),$$ as we wanted.
\[lem:ss\] Let $(E,\Phi)$ be a stable parabolic Higgs bundle, then $(\operatorname{ParEnd \,}(E),\operatorname{ad\,}(\Phi))$ is semistable.
The proof follows the arguments in [@garcia-prada-logares-munoz Proposition 6.7] adapted to the parabolic situation. That is, the vector bundle $\operatorname{ParEnd \,}(E)$ has a natural parabolic structure induced by the parabolic structure of $E$. In fact $\operatorname{ParEnd \,}(E)$ as a parabolic bundle is the parabolic tensor product of the parabolic bundle $E$ and the parabolic dual of $E$ (see [@y]), and hence its parabolic degree is $0$. With respect to this parabolic structure $(\operatorname{ParEnd \,}(E),\operatorname{ad\,}(\Phi))$, where $\operatorname{ad\,}(\Phi): \operatorname{ParEnd \,}(E)\to
\operatorname{SParEnd\,}(E)\otimes K(D)$, is, again, a parabolic Higgs bundle. Now, the stability of $(E,\Phi)$ implies the polystability of $(\operatorname{ParEnd \,}(E),\operatorname{ad\,}(\Phi))$.
$$\operatorname{codim}({\mathcal{V}})\ge \frac{1}{2}(r-1)(g-1).$$
Propositions \[prop:codim\] and \[prop:chi\] give $$\operatorname{codim}({\mathcal{V}})\,=\,\min_{F_{\lambda}}\left\{ \sum_{k>0} -\chi(C^{\bullet}(E)_{k})\right\}$$ $$\ge \,\min_{F_{\lambda}}\left\{\sum_{k>0}(g-1)(\operatorname{rk \,}(C_{k}-\operatorname{rk \,}(\widehat{C}_{k+1}))\right\}
\,=\, \operatorname{rk \,}(C_{1})(g-1)\, ,$$ where $\operatorname{rk \,}(C_{1})=\operatorname{rk \,}(\oplus_{j-i=1}\operatorname{ParHom \,}(E_{i},E_{j})$ so if we denote $r_{i}=\operatorname{rk \,}(E_{i}$ the rank of each piece is $\operatorname{rk \,}(\operatorname{ParHom \,}(E_{i},E_{j})=\frac{1}{2}r_{i}r_{j}$, so then $\operatorname{rk \,}(C_{1})=\frac{1}{2} (r_{1}r_{2}+\cdots +r_{m-1}r_{m})$ which is definitely $\operatorname{rk \,}(C_{1})\ge \frac{1}{2}(r-1)$.
\[codimfibre\] For $g= 2$ and $r\ge 5$ or $g= 3$ and $r\ge 3$ $g\ge 5$ and $r\ge 2$, and $u$ a generic point in ${\mathcal{U}}$, the codimension of the fiber $h^{-1}_0(u)$ in $h^{-1}(u)$ is greater or equal to 2.
The codimension does not depend on the number of marked points, as in [@biswas-gothen-logares] it did not depend on the degree of the line bundle $L$, which was twisting the Higgs bundle and in this case is $K(D)$.
We also obtain the following.
For $g= 2$ and $r\ge 5$ or $g= 3$ and $r\ge 3$ $g\ge 5$ and $r\ge 2$, the moduli space of parabolic bundles has the same number of irreducible components as the moduli space of parabolic Higgs bundles.
[ZZZZ]{}
M.F. Atiyah and R. Bott, The Yang-Mills equations over Riemann surfaces, *Phil. Trans. R. Soc. London* **308** (1982) 523–615.
V. Balaji, S. del Baño and I. Biswas, A Torelli type theorem for the moduli space of parabolic vector bundles over curves, *Math. Proc. Cambridge Philos. Soc.* **130** (2001), 269–280.
I. Biswas and T.L. Gómez, A Torelli theorem for the moduli space of Higgs bundles on a curve, *Quart. Jour. Math.* **54** (2003), 159–169.
I. Biswas, P.B. Gothen and M. Logares, On moduli spaces of Hitchin pairs, *Math. Proc. Cambridge Philos. Soc.* **151** (2011), 441–-457.
I. Biswas, Y. Holla and C. Kumar, On moduli spaces of parabolic vector bundles of rank 2 over ${\mathbb{C}}{\mathbb{P}}^{1}$, *Michigan Math. Jour.* **59** (2010), 467–479.
A. Beauville, M.S. Narasimhan and S. Ramanan, Spectral curves and the generalized theta divisor, *J. reine angew. Math.* **398**, (1989), 169–179.
C. Ciliberto, P. Ribenboim and E. Sernesi, Collected papers of Ruggiero Torelli, *Queen’s Papers in Pure and Applied Mathematics*, 101. Queen’s University, Kingston, 1995.
O. García-Prada, P.B. Gothen and V. Muñoz, Betti numbers for the moduli space of rank $3$ parabolic Higgs bundles, *Mem. Amer. Math. Soc.* **187** (2007), no. 879.
O. García-Prada, M. Logares and V. Muñoz, Moduli spaces of parabolic $\operatorname{U}(p,
q)$-Higgs bundles, *Quart. Jour. Math.* **60** (2009), 183–233.
T.L. Gómez and M. Logares, Torelli theorem for the moduli space of parabolic Higgs bundles, *Adv. Geom.* **11** (2011), 429–-444.
F. Knudsen and D. Mumford, The projectivity of the moduli space of stable curves I: preliminaries on “det” and “div”. *Math. Scand.*, **39** (1976), 19–55.
R. Hartshorne, *Algebraic Geometry*, Graduate Texts in Mathematics, No. 52. Springer-Verlag, New York-Heidelberg, 1977.
T. Hausel, Compactification of the moduli of Higgs bundles, *Jour. Reine Angew. Math.* **503** (1998), 169–192.
N.J. Hitchin, Stable bundles and integrable systems, *Duke Math. Jour.* **54** (1987), 91–114.
J.C. Hurtubise, Integrable systems and algebraic surfaces, *Duke Math. Jour.* **83** (1996), 19–49.
F.C. Kirwan, *Cohomology of quotients in symplectic and algebraic geometry*, Mathematical Notes 31, Princeton University Press, 1984.
M. Logares and J. Martens, Moduli of parabolic Higgs bundles and Atiyah algebroids, *Jour. reine angew. Math.* **649** (2010), 89–116.
D. Mumford and P. Newstead, Periods of a moduli space of bundles on curves, *Amer. Jour. Math.* **90** (1968), 1200–1208.
N. Nitsure, Cohomology of the moduli of parabolic vector bundles, *Proc. Indian Acad. Sci, (Math. Sci.)* **95** (1986), 61–77.
M.S. Narasimhan and S. Ramanan, Deformations of the moduli space of vector bundles over an algebraic curve, *Ann. Math.* **101** (1975), 391–417.
R. Sebastian, Torelli theorems for moduli of logarithmic connections and parabolic bundles, *Manuscr. Math.* **136** (2011), 249–271.
C.S. Seshadri, *Fibrés vectorieles sur les courbes algébriques*, *Asterisque* 96, 1982.
C. Simpson, Harmonic bundles on non compact curves, *Jour. Amer. Math. Soc.* **3** (1990), 713–770.
A.N. Tjurin, An analogue of the Torelli theorem for two-dimensional bundles over an algebraic curve of arbitrary genus, *Izv. Akad. Nauk SSSR Ser. Mat.* **33** (1969), 1149–1170.
A. Weil, Zum beweis des Torelli satzes, *Nachr. Akad. Wiss. Göttingen Math.-Phys. Kl. II* 2 (1957) 33–53.
K. Yokogawa, Infinitesimal deformation of parabolic Higgs sheaves, *Inter. Jour. Math.* **6** (1995), 125–148.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report on the electronic properties of few interacting electrons confined in a parabolic quantum dot based on a theoretical approach developed to investigate the influence of Bychkov-Rashba spin-orbit (SO) interaction on such a system. We note that the spin-orbit coupling profoundly influences the energy spectrum of interacting electrons in a quantum dot. Here we present accurate results for the energy levels and optical-absorption spectra for parabolic quantum dots containing upto four interacting electrons, in the presence of spin-orbit coupling and under the influence of an externally applied, perpendicular magnetic field. We have described in detail about a very accurate numerical scheme to evaluate these quantities. We have evaluated the effects of SO coupling on the Fock-Darwin spectra for quantum dots made out of three different semiconductor systems, InAs, InSb, and GaAs. The SO coupling on the single-electron spectra manifests itself by primarily lifting of the degeneracy at zero magnetic field, rearrangement of some of the energy levels at small magnetic fields, and level repulsions at high fields. These are explained as due to mixing of different spinor states for increasing strength of SO coupling. As a consequence, the corresponding absorption spectra reveal anticrossing structures in the two main lines of the spectra. For the interacting many-electron systems we observed the appearence of discontinuities, anticrossings, and new modes that appear in conjunction with the two main absorption lines. These additional features arise entirely due to the SO coupling and are a consequence of level crossings and level repulsions in the energy spectra. An intricate interplay between the SO coupling and the Zeeman energies are shown to be responsible for these new features seen in the energy spectra. Optical absorption spectra for all three types of quantum dots studied here show a common feature: new modes appear, mostly near the upper main branch of the spectra around 2 tesla that become stronger with increasing SO coupling strength. Among the three types of systems considered here, optical signatures of the SO interaction is found to be the strongest in the absorption spectra of the GaAs quantum dot, but only at very large values of the SO coupling strength, and appears to be the weakest for the InSb quantum dot. Experimental observation of these new modes that appear solely due to the presence of SO coupling would provide a rare glimpse on the role of SO coupling in nanostructured quantum systems.'
author:
- 'Pekka Pietiläinen$^{\ast\dag}$ and Tapash Chakraborty$^{\ast\ddag}$'
title: 'Energy levels and magneto-optical transitions in parabolic quantum dots with spin-orbit coupling'
---
=1.2truecm
Introduction
============
Impressive developments in nanofabrication technology have made it now possible to design the quantum dots and coupled quantum dots at the nanoscale. These systems comprise of a few electrons that are quantum confined, for our present purpose, at the semiconductor interface to form the zero-dimensional systems. What is more remarkable is that the electronic states in these systems can be precisely controlled via the external voltages [@nanobio]. Magneto-optical studies of parabolic quantum dots (described as [*artificial atoms*]{} [@makchak] by us in 1990) have been intensely explored for more than a decade [@makchak; @comment; @qdbook; @merkt; @farinfrared] in order to understand its unique electronic and optical properties and as promising candidates for optoelectronic devices, applications in optical quantum information technology, etc. [@david; @shields]. As a result of these studies, a very good theoretical and experimental understanding of the single-electron states in the dot has already been achieved. At the most basic level, the solution of the Schrödinger equation for an electron confined by a harmonic potential, $v_c=\frac12m^*\omega_0 r^2$, $\omega_0$ is the confinement potential strength, in the presence of an external perpendicular magnetic field is well known since the beginning of quantum mechanics [@qdbook; @fock]. The eigenvalues in this case are given by $$E_{nl}=\left(2n+\left|l\right|+1\right)\hbar\Omega-\tfrac12
l\hbar\omega_c$$ where $n=0,1,2,...$ and $l=0,\pm1...,$ are the principal and azimuthal quantum numbers respectively, $\Omega^2=\left[\omega_0^2+
\frac14\omega_c^2\right]$, and $\omega_c$ is the cyclotron frequency. Dipole-allowed transitions among these energy levels will have energies [@makchak; @qdbook; @merkt] $$\Delta E_{\pm}=\hbar\Omega\pm\tfrac12\hbar\omega_c.$$ This relation has been verified to great accuracy by a variety of experiments [@comment; @qdbook; @farinfrared]. Interestingly, however, the observed magnetic field dependent FIR absorption in quantum dots containing more than one electron was found to be essentially [*independent*]{} of the number of electrons confined and instead was dominated by the above relation for $\Delta E_{\pm}$ [@merkt]. It was a rather puzzling result because according to this, magneto-optics was clearly incapable of providing any relevant information about the effect of mutual interactions of the confined electrons. The puzzle was later resolved by Maksym and Chakraborty [@makchak; @comment; @qdbook; @bert], who pointed out that for a parabolic QD in an external magnetic field, the dipole interaction is a function of the center-of-mass (CM) coordinate alone and the inter-electron interaction does not play any role. Despite this somewhat disappointing performance of a parabolic dot, FIR spectroscopy of QDs (parabolic or otherwise) has generated enormous interest for over a decade that is yet to subside [@farinfrared]. In this paper we include another interesting element into the problem, the spin-orbit (SO) interaction. We demonstrate here that in the presence of spin-orbit coupling the magneto-optical transitions show many new and interesting features that can be tuned by the SO coupling. Based on these results, we propose that magneto-optical transitions are best suited to determine optically the unique effect of SO coupling in quantum dots described below [@condmat].
Interest on the role of the spin-orbit coupling in nanostructured systems is now at its peak, due largely to its relevance to spin transport in low-dimensional electron channels [@spintro; @ohno]. The intriguing possibility of tuning the SO field and thereby coherently manipulate electron spins in quantum dots has sparked major activities in the past years [@kuan; @vam_PRB; @vam_JAP; @governale; @cremers; @quasi_exact; @tsitsishvili; @manuel_041; @manuel_042; @manuel_043; @manuel_02; @lucignano; @destefani; @debald; @koenemann; @bellucci; @fransson; @spinorbit]. It is hoped that an improved understanding of spin dynamics in the QDs might pave the way for future electronic and information processing, especially in quantum computing and quantum communication [@loss]. Spin degree of freedom is perhaps more advantageous than charge because unlike charge, spin is not coupled to electromagnetic noise and therefore have much longer coherence time [@bandyo]. Improved knowledge of the influence of spin-orbit coupling in quantum dots is therefore quite essential in this pursuit. While the majority of experimental efforts has focused on magneto-transport measurements [@expt], here we present the results for the optical absorption spectra that are experimentally observable and could, in principle, provide an important probe of SO coupling in few-electron quantum dots.
The spin-orbit interaction in semiconductor heterostructures can be caused by an electric field perpendicular to the two-dimensional electron gas (2DEG). Riding on an electron, this electric field will be [*felt*]{} as an effective magnetic field lying in the plane of the 2DEG, perpendicular to the wave vector $k$ of the electron. The effective Zeeman interaction of the electron spin with the field lifts the spin degeneracy (internal Zeeman effect). This is usually referred to in the literature as the Bychkov-Rashba mechanism. This results in an isotropic spin splitting energy $\Delta_{\rm SO}$ at $B=0$ proportional to $k$ [@rashba].
Let us consider an electron in the 2DEG moving with a velocity $\vec v$ in the presence of an electric field $\vec E$. In the rest frame of the electron, this transforms (relativistically) into an effective magnetic field ${\vec B}_{\rm eff}$, $${\vec B}_{\rm eff}=-\frac1{2c^2}{\vec v}\times {\vec E}$$ where $c$ is the speed of light. The magnetic moment of the electron will then couple to ${\vec B}_{\rm eff}$. The resulting spin-orbit interaction is, $$\begin{aligned}
{\cal H}_{\rm SO} &=& {\vec\mu}\cdot{\vec B}_{\rm eff} \\
&=& -\frac1{2c^2}{\vec \mu}\cdot ({\vec v}\times{\vec E})=\dfrac{e
\hbar^2}{4m_0^2c^2}\vec\sigma\cdot(\vec k\times\vec E)\\\end{aligned}$$ since, ${\vec\mu}=-\dfrac{e\hbar}{2m_0}\vec\sigma$, and $\vec v=\hbar\vec k/m_0$. It can be rewritten as $${\cal H}_{\rm SO} = \dfrac{e \hbar^2}{4m_0^2c^2}\vec\sigma\cdot(
\vec k\times\vec E)=\alpha'\langle E_z\rangle(-{\rm i}\vec\nabla\times
\vec\sigma)_z,
\label{original}$$ where the electric field is aligned along the $z$ axis.
Alternatively, a general spin-orbit Hamiltonian that stems directly from the quadratic in $v/c$ expansion of the Dirac equation is [@bethe] $${\cal H}_{\rm SO}=\dfrac{e \hbar}{(2m_0c)^2}
{\vec\nabla}V(\vec r)\cdot(\vec\sigma\times{\vec p}).
\label{gradient}$$ The electric field associated with $V(\vec r)$ is ${\vec E}(\vec r)={\vec\nabla} V(\vec r)$, and is directed along the $z$ direction. The spin-orbit interaction Hamiltonian $$\begin{aligned}
{\cal H}_{\rm SO}&=& \dfrac{e \hbar^2}{(2m_0c)^2}\langle E_z\rangle
\frac1{\hbar}(\vec\sigma\times{\vec p})\cdot{\hat n}
= \alpha'\langle E_z\rangle\frac1{\hbar}(\vec\sigma
\times{\vec p})\cdot{\hat n} \\
&=& \alpha'\langle E_z\rangle(\vec\sigma\times\vec k)
\cdot{\hat n}, \\\end{aligned}$$ where $\alpha'= e (\hbar/2m_0c)^2$ is then identical to that in Eq. (\[original\]). An important point to note here is that a non-vanishing gradient in Eq. (\[gradient\]) requires that the system must have inversion asymmetry. In the present case that arises from the structural inversion asymmetry [@zawadzki]. The spin-orbit interaction that we are here concerned with is therefore described by the Hamiltonian $${\cal H}_{\rm SO}=\alpha (\vec k\times\vec\sigma)_z
={\rm i}\alpha \left(\sigma_y\frac{\partial}{\partial
x}-\sigma_x\frac{\partial}{\partial y}\right),$$ where the $z$ axis is chosen perpendicular to the 2DEG (in the $xy$-plane), $\alpha$ is the spin-orbit coupling constant, which is sample dependent and is proportional to the interface electric field that confines the electrons in the $x-y$ plane, $\vec{\sigma}=(\sigma_x,
\sigma_y, \sigma_z)$ denotes the Pauli matrices, and $\vec k$ is the planar wave vector. This is the Bychkov-Rashba Hamiltonian [@rashba] that has been receiving of late rather widespread attention [@expt].
The single-electron Hamiltonian for the 2DEG including the Bychkov-Rashba term has the form $$\begin{aligned}
{\cal H}&=&\frac{{\vec p}^2}{2m^*}+\frac{\alpha}{\hbar}
\left(\vec\sigma\times{\vec p}\right)_z \\
&=&-\frac{\hbar^2}{2m^*}{\vec\nabla}^2+{\rm i}\alpha\left(\sigma_y
\frac{\partial}{\partial x}-\sigma_x\frac{\partial}{\partial y}\right)\\
&=&{\left( \begin{array}{lr}
-\dfrac{\hbar^2}{2m^*}{\vec\nabla}^2 & \alpha\nabla^- \\
-\alpha\nabla^+ & -\dfrac{\hbar^2}{2m^*}{\vec\nabla}^2 \\
\end{array} \right)} \\\end{aligned}$$ where $\vec{\nabla}^2=\partial^2/\partial x^2+\partial^2/
\partial y^2$ and $\nabla^{\pm}=\partial/\partial
x\pm{\rm i}\partial/\partial y$.
Since the operators ${\hat p}_x$ and ${\hat p}_y$ commute with the Hamiltonian, we can search for $\alpha\ne0$ eigenstates of the form $$\Psi(k_x,k_y)={\rm e}^{{\rm i}k_xx+{\rm i}k_yy}\sum_{\sigma}
C^\sigma\vert\sigma\rangle
={\rm e}^{{\rm i}k_xx+{\rm i}k_yy}\left(\begin{array}{c}
C^+\\C^-\end{array}\right),$$ with $\vert\sigma\rangle=\left(\begin{array}{c}1\\0\end{array}\right)$ (spin up) or $\left(\begin{array}{c}0\\1\end{array}\right)$ (spin down). Solutions of $${\cal H}\Psi(k_x,k_y)={\cal E}\Psi(k_x,k_y)$$ are readily obtained as $$\Psi^{\pm}(k_x,k_y)=\frac1{\sqrt2}\left(\begin{array}{c}1\\
\dfrac{\pm k_y\mp{\rm i}k_x}k\end{array}
\right){\rm e}^{{\rm i}k_xx+{\rm i}k_yy}.$$
The energy dispersion then consists of two branches $${\mathcal E}^{\pm}(k)=\frac{\hbar^2}{2m^*}k^2 \pm \alpha k$$ with an energy separation $\Delta_{\rm SO}={\mathcal E}^+-{\mathcal
E}^-=2\alpha k$ for a given $k$. The spin parts of the wave functions $\chi^\pm(k_x,k_y)$ are mutually orthogonal and $\langle\chi^\pm|\sigma_z|
\chi^\pm\rangle=0.$ Therefore in the states $\Psi^\pm$ the spins of the electrons lie in the $xy$-plane and point in opposite directions. In addition, $$\langle\chi^\pm|\sigma_x|\chi^\pm\rangle=\frac{2k_y}k, \qquad
\langle\chi^\pm|\sigma_y|\chi^\pm\rangle=-\frac{2k_x}k,$$ i.e., the spins are [*perpendicular*]{} to the momentum $(k_x,k_y)$. Spatial alignment of spins therefore depends on the wave vector [@spintro; @rashba]. The Fermi surface is a pair of concentric circles with radii $k_{F,max}$ and $k_{F,min}$. In the present paper, we are dealing with systems having rotational symmetry. The formalism in that case is derived in detail in Sect.II.
Several experimental groups [@expt] investigating the Shubnikov-de Haas (SdH) oscillations in a 2DEG confined at the heterojunctions with a narrow-gap quantum well (e.g., InGaAs/InAlAs, InAs/GaSb, etc.) have already established that lifting of spin degeneracy results from inversion asymmetry of the structure which invokes an electric field perpendicular to the layer. Experimentally observed values of the SO coupling strength $\alpha$ lie in the range of 5 – 45 meV.nm [@expt]. Energy levels of two interacting electrons confined in a parabolic quantum dot in an external magnetic field were recently reported by us for this range of SO coupling strength [@spinorbit]. In the absence of SO coupling, electron-electron interaction causes the ground state energy to jump from one angular momentum value to another as the magnetic field is increased [@makchak; @qdbook]. The influence of the SO coupling is primarily to move the energy level crossings to weaker fields [@spinorbit].
Our paper is organized as follows: In Sect. II, we present the essential formalism for our study of non-interacting electrons in parabolic QDs in the presence of spin-orbit interaction. The single-electron basis and the dipole matrix elements are derived here. We also explain why the dipole-allowed transitions are so significantly influenced by the presence or absence of the SO interaction. The classic Fock-Darwin spectra for three different quantum dots systems (InAs, InSb, and GaAs) with or without the SO coupling are presented and discussed in detail. The formalism for the many-electron system in a parabolic QD with SO coupling is presented in Sect. III. The complexities of introducing the SO coupling, in particular for interacting electrons, are made clear in Sect. III. The task of finding a suitable numerical technique is even more challenging, and an approach that is appropriate for our purpose is described in the Appendix. Numerical results for the energy levels and the optical absorption spectra for the three types of QDs containing upto four interacting electrons are presented and discussed in Sect. IV. It should be pointed out that the low-lying energy levels calculated here for the single- and multi-electron quantum dots can, in principle, be observed in transport [@rolf_review], or capacitance spectroscopy [@ashoori]. Given the accute interest on the influence of SO coupling in nanostructured systems and the resulting intense activities on this topic, it is no surprise that many different theoretical techniques have been put forward in the literature. To view our work in proper perspective, we present a brief review of many of those theoretical papers in Sect. V. We conclude with a brief outlook for future work along this direction in Sect. VI. For a brief account of our earlier work on SO coupling effects in parabolic QDs, see Refs. [@condmat; @spinorbit].
Single-electron picture
=======================
The Hamiltonian for an electron in a parabolic confinement and under external magnetic field is given by $$\begin{aligned}
{\cal H}_0
&=&
\frac1{2m^*}\left(\vec p -\frac ec\vec A\right)^2+\frac12 m^*\omega_0^2r^2
\nonumber \\
&+&\frac {\alpha}{\hbar}\left[\vec\sigma\times\left(\vec p-\frac ec\vec A
\right)\right]_z+\frac12 g\mu_BB\sigma_z.
\label{spham}\end{aligned}$$ Here $\vec\sigma$ is the vector of Pauli matrices, i.e. $$\begin{aligned}
\vec\sigma&=&\sigma_x\vec i+\sigma_y\vec j+\sigma_z\vec k
\nonumber \\
&=&\left(\begin{array}{cc}0&1 \\ 1&0\end{array}\right)\vec i
+\left(\begin{array}{cc}0&-i \\ i&0\end{array}\right)\vec j
+\left(\begin{array}{cc}1&0 \\ 0&-1\end{array}\right)\vec k.
\label{pauli}\end{aligned}$$ We work in the symmetric gauge and the vector potential corresponding to the external perpendicular magnetic field is $$\vec A=\frac B2(-y,x,0).
\label{symmgauge}$$ The term $\frac {\alpha}{\hbar}\left[\vec\sigma\times
\left(\vec p - \frac ec\vec A\right)\right]_z$ in the Hamiltonian is the spin-orbit (SO) coupling due to the inhomogenous potential confining the electrons to the 2D plane and possible external gate voltages applied on the top of the dot. The parameter $\alpha$ determines the strength of this coupling and, in case of external gate voltages its magnitude can be varied. Finally, the last term $\frac12 g\mu_BB\sigma_z$ is the ordinary Zeeman coupling $g$ being the effective Lande’ $g$-factor.
The eigenstates of the single particle problem $${\cal H}_0\phi=\varepsilon\phi
\label{sparteq}$$ are clearly two-component spinors $$|\lambda\rangle=\left(\begin{array}{c}\phi^{\uparrow} \\
\phi^{\downarrow}\end{array}\right).
\label{spinordef}$$ Writing the equation (\[sparteq\]) in polar coordinates and substituting a trial wave function of the form $$\phi=\left(\begin{array}{c}
f^{\uparrow}(r)\,e^{i\ell^{\uparrow}\theta} \\
f^{\downarrow}(r)\,e^{i\ell^{\downarrow}\theta}
\end{array}\right)
\label{trialwf}$$ it is easy to see that the quantum numbers $\ell^{\uparrow}$ and $\ell^{\downarrow}$ must be integers and that they depend on each other in the way $$\ell^{\uparrow}=\ell^{\downarrow}-1.
\label{mqncond}$$ Hence we need only one quantum number for the angular motion, i.e. solutions of the single particle equation (\[sparteq\]) are of the form $$|\lambda\rangle=|k,\ell\rangle=\left(\begin{array}{c}
f^{\uparrow}_{k,\ell}(r)\,e^{i\ell\theta} \\
f^{\downarrow}_{k,\ell}(r)\,e^{i(\ell+1)\theta}
\end{array}\right).
\label{twocspin}$$ Here the quantum number $k$ is associated with radial motion (and not to be confused with the wave vector described in Sect.I). The form of the spinor (\[twocspin\]) simply restates the fact that under SO coupling the good quantum numbers are related to $\vec L+\vec S$. In our case the conserved quantity is $$j=\ell^{\uparrow,\downarrow} + s^{\uparrow,\downarrow}_z=\ell+\tfrac12
\label{spartj}$$ where $s_z=\pm\frac12$ depending on the component of the spinor, i.e., $+\frac12$ for the upper component and $-\frac12$ for the lower one.
In order to find the radial wavefunctions $f^{\uparrow,\downarrow}$ we transform to dimensionless units by setting $$\begin{aligned}
\omega_c&=&\frac{eB}{m^*c},\quad
a^2=\frac{\hbar}{m^*\omega_0
\left(1+{\omega_c^2}/{4\omega_0^2}\right)^{\frac12}}, \\
x&=&\frac{r^2}{a^2}, \quad
\beta=\frac{m^*\alpha a}{\hbar^2}, \\
b{^{\vphantom{\dagger}}}_R&=&\frac{ea^2B}{\hbar^2}, \quad
\eta^\pm=1\pm 2b{^{\vphantom{\dagger}}}_R, \\
\nu_\ell^{\uparrow,\downarrow}&=&
\frac{\ell\omega_c}{4\omega_0
\left(1+{\omega_c^2}/{4\omega_0^2}\right)^{\frac12}}
\pm\frac{g\mu_BB}{4\hbar\omega_0
\left(1+{\omega_c^2}/{4\omega_0^2}\right)^{\frac12}}, \\
\varepsilon&=&2\hbar\omega_0
\left(1+\frac{\omega_c^2}{4\omega_0^2}\right)^{\frac12}\nu,
g^{\uparrow,\downarrow}(x)=f^{\uparrow,\downarrow}(r).\end{aligned}$$ Substituting these into the Hamiltonian (\[spham\]) the radial part of the equation (\[sparteq\]) takes the form $$\begin{aligned}
\label{upeq}
x{g^{\uparrow}}'' &+& {g^{\uparrow}}' +\left(\nu-\frac{\ell^2}{4x}
-\frac x4+\nu^{\uparrow}_\ell\right)g^{\uparrow}
\nonumber \\
&-&\beta x^{1/2}\left({g^{\downarrow}}'+\frac{\ell+1}{2x}g^{\downarrow}
+b{^{\vphantom{\dagger}}}_Rg^{\downarrow}\right)=0 \\
x{g^{\downarrow}}'' &+& {g^{\downarrow}}'+\left(\nu-\frac{(\ell+1)^2}{4x}
-\frac x4+\nu^{\uparrow}_{\ell+1}\right) g^{\downarrow}
\nonumber \\
&+&\beta x^{1/2}
\left({g^{\uparrow}}'-\frac{\ell}{2x}g^{\uparrow}-
b{^{\vphantom{\dagger}}}_Rg^{\uparrow}\right)=0
\label{downeq}\end{aligned}$$ of two coupled differential equations. For $\beta=0$ (i.e., $\alpha=0$) these equations describe radial motions of two independent two-dimensional harmonic oscillators with eigenvalues $$\nu_{n\ell}^{\uparrow,\downarrow}
=n+\frac{|\ell|+1}2+\nu^{\uparrow,\downarrow}_\ell
\label{nunm}$$ and the eigenfunctions $$g_{n\ell}=\sqrt{\frac{n!}{(n+|\ell|)!}}\,e^{-x/2}x^{|\ell|/2}L_n^{|\ell|}(x).
\label{gnm}$$ Here $L_n^{|\ell|}$ is the associated Laguerre polynomial defined, for example, by the formula $$L_n^{|\ell|}(x)=e^xx^{-|\ell|}\frac{d^n}{dx^n}\left(e^{-x}x^{n+|\ell|}\right).$$ Therefore it is logical to seek the solution for Eqs. (\[upeq\], \[downeq\]) in the form of the expansion [@kuan] $$g^{\uparrow,\downarrow}=\sum_{n=0}c^{\uparrow,\downarrow}_{n,\ell}g_{n,\ell}.
\label{gnmexp}$$
In spinor language (\[trialwf\]) this corresponds to the expansion $$\begin{aligned}
|\lambda\rangle &=&
\left(\begin{array}{c}
\sum_{n=0}c^{\uparrow}_{n,\ell}g_{n,\ell}(x)\,e^{i\ell\theta} \\
\sum_{n=0}c^{\downarrow}_{n,\ell+1}g_{n,\ell+1}(x)\,e^{i(\ell+1)\theta}
\end{array}\right) \nonumber \\
&=&
e^{i\ell\theta}\sum_{n=0}c^{\uparrow}_{n,\ell}
\left(\begin{array}{c} g_{n,\ell}(x) \\ 0 \end{array}\right)
+ \nonumber \\
&&e^{i(\ell+1)\theta}\sum_{n=0}c^{\downarrow}_{n,\ell+1}
\left(\begin{array}{c} 0 \\ g_{n,\ell+1}(x) \end{array}\right).
\label{spdecomp}\end{aligned}$$ The coefficents $c^{\uparrow,\downarrow}_{n,\ell}$ can be obtained by minimizing the expectation value $\langle\lambda|{\cal H}_0|\lambda
\rangle$. At this point it is usefull to relabel our spinors. For example, for non-negative values of $\ell$ we set $$\begin{aligned}
u_{2n}&=&e^{i\ell\theta}
\left(\begin{array}{c} g_{n,\ell}(x) \\ 0 \end{array}\right) \\
u_{2n+1}&=&e^{i(\ell+1)\theta}
\left(\begin{array}{c} 0 \\ g_{n,\ell+1}(x) \end{array}\right) \\
z_{2n}&=&c^{\uparrow}_{n,\ell} \label{evenn} \\
z_{2n+1}&=&c^{\downarrow}_{n,\ell+1}. \label{oddn}\end{aligned}$$
The minimization of $$\langle\lambda|{\cal H}_0|\lambda\rangle =
\sum_{n,n'}z_{n'}z_n\langle u_{n'}|{\cal H}_0|u_n\rangle$$ leads now to the diagonalization of the symmetric tridiagonal matrix ${\cal H}$ with diagonal $$\mbox{diag}\, {\cal H}=(\nu_{0,\ell}^{\uparrow},\nu_{0,\ell+1}^{\downarrow},
\ldots,\nu_{n,\ell}^{\uparrow},\nu_{n,\ell+1}^{\downarrow},\ldots)
\label{posdiag}$$ and subdiagonal $$\begin{aligned}
\mbox{subdiag}\,{\cal H}=&&\frac{\beta}2\left(0,\sqrt{\ell+1}\eta^+,
\eta^-,\ldots,\sqrt n \eta^-, \right. \nonumber \\
&&\left. \sqrt{n+\ell+1}\eta^+,\sqrt{n+1}\eta^-,\ldots\right).
\label{possubd}\end{aligned}$$
The diagonalization yields a set of solutions: the set $\{\nu^{(k,\ell)}\}$ of eigenvalues and the set $\{z^{(k,\ell)}_n\}$ of eigenvectors indexed by the particular solution $k$ and the fixed angular momentum $\ell$. These are the energies of the spinors and the expansion coefficients (\[evenn\],\[oddn\]). For negative values of $\ell$ the tridiagonal matrix consists of the diagonal $$\mbox{diag}\, {\cal H}=(\nu_{0,\ell+1}^{\downarrow},\nu_{0,\ell}^{\uparrow},
\ldots,\nu_{n,\ell+1}^{\downarrow},\nu_{n,\ell}^{\uparrow},\ldots)
\label{negdiag}$$ and the subdiagonal $$\begin{aligned}
\mbox{subdiag}\,{\cal H}=&&-\frac{\beta}2(0,\sqrt{|\ell|}\eta^-,\eta^+,\ldots,
\sqrt n \eta^+, \nonumber \\
&&\sqrt{n+|\ell|}\eta^-,\sqrt{n+1}\eta^+,\ldots).
\label{negsubd}\end{aligned}$$ The spinors thus obtained comprise our single-particle basis $$\begin{aligned}
{\cal B}_S&=&\big\{|\lambda_i\rangle\big|\,i=0,1,2,\ldots\big\}
\nonumber \\
&=&\big\{|k,\ell\rangle\big|\,k=0,1,2,\ldots;\,
\ell=0,\pm1,\pm2,\ldots\big\}. \nonumber \\
\label{spbas}\end{aligned}$$
Dipole matrix elements
----------------------
According to the Fermi golden rule the intensity of absorption in dipole approximation is proportional to the square of the matrix element $$I=\langle f|\sum_{i=1}^N r_ie^{\pm i\theta_i}|i\rangle$$ when the transition goes from the initial $N$-particle state $|i\rangle$ to the final state $|f\rangle$. To evaluate this we need to know the dipole matrix elements $d_{\lambda',\lambda}$ between the spinor states $|\lambda'\rangle$ and $|\lambda\rangle$, i.e. the elements $$d_{\lambda',\lambda}
=\langle\lambda'|r\,e^{\pm i\theta}|\lambda\rangle
=\langle k',\ell'|r\,e^{\pm i\theta}|k,\ell\rangle.
\label{dipel}$$
For simplicity we consider only the circular polarization $e^{+i\theta}$ (the other circular polarization is obatined by reversing the roles of $\lambda$ and $\lambda'$). Substituting expansions (\[gnmexp\]) into the above expression (\[dipel\]) we get $$\begin{aligned}
d_{\lambda',\lambda}&=&a\delta_{\ell',\ell+1}\sum_n\left[
{c'}_n^\uparrow c_n^\uparrow\sqrt{n+\ell+1}-{c'}_{n-1}^\uparrow c_n^\uparrow
\sqrt n\right. \nonumber \\
&+&\left. {c'}_n^\downarrow c_n^\downarrow\sqrt{n+\ell+2}
-{c'}_{n-1}^\downarrow c_n^\downarrow\sqrt n
\right]
\label{posmdip}\end{aligned}$$ when $\ell\geq0$, $$\begin{aligned}
d_{\lambda',\lambda}
&=&a\delta_{\ell',0}\sum_n\left[{c'}_n^\uparrow c_n^\uparrow\sqrt{n+1}
-{c'}_{n+1}^\uparrow c_n^\uparrow\sqrt{n+1}\right. \nonumber \\
&+& \left. {c'}_n^\downarrow c_n^\downarrow\sqrt{n+1}
-{c'}_{n-1}^\downarrow c_n^\downarrow\sqrt n \right]
\label{n1mdip}\end{aligned}$$ when $\ell=-1$ and $$\begin{aligned}
d_{\lambda',\lambda}&=&a\delta_{\ell',\ell+1}\sum_n\left[
{c'}_n^\uparrow c_n^\uparrow\sqrt{n-\ell} -{c'}_{n+1}^\uparrow
c_n^\uparrow\sqrt{n+1}\right. \nonumber \\
&+&\left. {c'}_n^\downarrow c_n^\downarrow\sqrt{n-\ell-1}
-{c'}_{n+1}^\downarrow c_n^\downarrow\sqrt{n+1} \right]
\label{negmdip}\end{aligned}$$ when $\ell<-1$. The intensity is then obtained from ${\cal I}\propto \vert d_{\lambda_1\lambda_2}\vert^2$ [@halonen]. In all our figures for the absorption spectra, the size of the points is proportional to the calculated intensity.
The reason why dipole-allowed transitions in a parabolically confined quantum dot can be very different in the presence of SO interaction is explained as follows. When subjected to the radiation field with amplitude $a$ and polarization $\vec\epsilon$, the vector potential $\vec A$ in the single particle Hamiltonian $$\begin{aligned}
{\cal H}_0&=&\frac1{2m^*}\left(\vec p-\frac ec\vec A\right)^2
+\frac12 m^*\omega_0^2r^2\\
&&+\frac{\alpha}{\hbar}\left[\vec\sigma\times
\left(\vec p-\frac ec\vec A\right)\right]_z+\frac12 g\mu_BB\sigma_z \end{aligned}$$ must be replaced with the potential $$\vec A\rightarrow\vec A+ {\vec A}_\omega,
\vec A_\omega=\vec\epsilon a \,e^{i\vec k\cdot\vec r-i\omega t}.$$ In the dipole approximation we assume that $$A_\omega\approx\vec\epsilon a \,e^{-i\omega t}$$ and correspondingly the Hamiltonian will be [@lipparini] $${\cal H}\approx{\cal H}_0- {\cal H}'\, e^{-i\omega t},$$ where $${\cal H}'=\frac {ea}{m^*c}\vec\epsilon\cdot\left(
\vec p-\frac ec\vec A\right)
+\frac{\alpha ea}{\hbar c}\left[\vec\sigma
\times\vec\epsilon\right]_z.$$ In a many-body system when $\alpha=0$ the first term generates the CM density excitations where mutual interactions play no role. Consequently (in dipole approximation) only transitions between these modes are possible. When $\alpha$ is different from zero, the second term ($\propto \sigma_x\epsilon_y
-\sigma_y\epsilon_x$) in ${\cal H}'$ can create spin-density oscillations and interactions have effect on their properties. It is to be noted that, in SO coupled systems the dipole operator still retains its familiar form, $\hat Q=\dfrac{ea}c\vec\epsilon\cdot\vec r$, as is easily verified by evaluating its commutator with the Hamiltonian ${\cal H}_0$ $$[\hat Q,{\cal H}_0]=i\hbar{\cal H}'.$$ Dipole operator is independent of the electron spin. The dipole-allowed optical transitions are always between the same spin states, but the angular momenta must differ by unity. In the presence of SO coupling, neither the dipole operator nor the selection rule changes, but the SO interaction mixes the neighboring angular momentum values ($l$ and $l+1$) as well as the spin and hence the selection rule now applies to the total angular momentum $J$ as well. Therefore, transitions from other states that are not allowed without the SO coupling, are now allowed.
Fock-Darwin spectra
-------------------
In our numerical investigations, we choose InAs, InSb and GaAs quantum dots with parameters, $m^*/m_0=0.042, \epsilon=14.6, g=-14$, $m^*/m_0=0.014,
\epsilon=17.88, g=-40$, and $m^*/m_0=0.063, \epsilon=12.9, g=-0.44$, respectively. While the InAs quantum structures have been the system of choice for investigation of spin-related phenomena [@expt], InSb quantum dots are interesting for their very high $g$ values and a relatively large $\alpha\ (\sim 14$ meV nm) [@khodaparast]. For the GaAs quantum dots, the observed value of $\alpha$ is $\sim6$ meV.nm [@koenemann]. In all these systems we consider the confinement potential strength to be $\hbar\omega_0=7.5$ meV. Some of the low-lying states of the Fock-Darwin spectra of the InAs, InSb and GaAs QDs are shown in Figs. \[fig:fock1\] – \[fig:fock3\] respectively and the corresponding optical absorption spectra in these systems are presented in Figs. \[fig:optics1\] – \[fig:optics3\].
As compared to the Fock-Darwin spectra of quantum dots without the SO coupling (shown in panels Figs. \[fig:fock1\] – \[fig:fock3\] \[a\]) the most outstanding features in the energy spectra of quantum dots with SO coupling are the [*lifting of degeneracy*]{} at zero magnetic field, the rearrangement of some of the levels at small fields and level repulsions at higher magnetic fields. To get some insight into the mechanism causing this kind of behavior, let us have a closer look, as an example, at the energy levels involved in the lowest absorption lines of the InAs dot \[curves labelled “0” – “3” in Fig. \[fig:fock1\] (a)\].
----------------------------------------------------------------------------------------
$g<0$ $g>0$
--- ------------------------------------------- ----------------------------------------
0 $\displaystyle\left(\begin{array}{c} $\displaystyle\left(\begin{array}{c}
\mbox{\rule{1ex}{1.5ex}}\,e^{i0} \\ 0e^{-i\theta} \\
0e^{i\theta} \mbox{\rule{1ex}{1.5ex}}\,e^{i0}
\end{array}\right) $ \end{array}\right)$
1 $\displaystyle\left(\begin{array}{c} $\displaystyle\left(\begin{array}{c}
0e^{-i\theta} \\ \mbox{\rule{1ex}{1.5ex}}\,e^{i0} \\
\mbox{\rule{1ex}{1.5ex}}\,e^{i0} 0e^{i\theta}
\end{array}\right)$ \end{array}\right)$
2 $\displaystyle\left(\begin{array}{c} $\displaystyle\left(\begin{array}{c}
\mbox{\rule{1ex}{1.5ex}}\,e^{-i\theta} \\ 0e^{-2i\theta} \\
0e^{i0} \mbox{\rule{1ex}{1.5ex}}\,e^{-i\theta}
\end{array}\right) $ \end{array}\right)$
3 $\displaystyle\left(\begin{array}{c} $\displaystyle\left(\begin{array}{c}
\mbox{\rule{1ex}{1.5ex}}\,e^{i\theta} \\ 0e^{i0} \\
0e^{2i\theta} \mbox{\rule{1ex}{1.5ex}}\,e^{i\theta}
\end{array}\right) $ \end{array}\right)$
----------------------------------------------------------------------------------------
: Schematic spinors corresponding to four Fock-Darwin levels, marked 0 – 3 in Fig. \[fig:fock1\] (a) of an InAs dot without SO coupling. The black rectangles stand for non-zero radial wave functions.[]{data-label="spinortable"}
\
In the absence of the SO coupling energies of these levels are given by the formula (\[nunm\]). The corresponding spinors (schematic) are depicted in Table \[spinortable\], where the numbers in the first column refer to the labels in Fig. \[fig:fock1\] (a). The spinors for electrons with negative and positive Lande’ $g$-factors are shown in the middle and third columns respectively. In actual physical systems, conventionally only the spinors with $g<0$ are of any interest.
The spinor states of electrons on lines 0 and 1 of Fig. \[fig:fock1\] (a) differ only by the orientation of the spin: on line 0 the spin is parallel to the magnetic field ($\|\hat z$) while on line 1 the spin is antiparallel to the field. Thus the energy difference between these states is the Zeemen splitting. The total single-particle angular momenta (\[spartj\]) are correspondingly $j=\pm\frac12$. Since under the SO coupling $j$ is a good quantum number these two states will never mix even when the SO coupling is on. When the coupling strength $\alpha$ increases, the higher lying states with $j=\frac12$ couple to the state 0 as well as states with $j=-\frac12$ couple with the state 1. In Fig. \[fig:fock1\] this shows up as an increasing splitting of the lines 0 and 1 when going through the panels from \[a\] to \[d\]. It should be noted, however, that the mixing has very minor effect on the ground state since the other states with $j=\frac12$ are energetically very far from that.
![ Optical absorption spectra for non-interacting electrons confined in a InAs quantum dot for various values the SO coupling strength (in meV.nm), $\alpha=0$ \[a\], $\alpha=20$ \[b\], $\alpha=30$ \[c\], and $\alpha=40$ \[d\]. []{data-label="fig:optics1"}](FD_OA_InAs.eps){width=".45\textwidth"}
Turning now our attention to electrons on lines 1 and 2 we see from the Table \[spinortable\] (column $g<0$) that their spinor states both have the same angular momentum $j=\ell+\frac12=-\frac12$. Consequently the SO interaction can mix these states. The mixing is particularly pronounced when the states are nearly degenerate, i.e. in the vicinity of the crossing point of lines 1 and 2. At moderate coupling strengths this mixing leads to level repulsions as shown in panels \[b\] and \[c\]. When the coupling is very strong, energetically higher states with $j=-\frac12$ also become important in the mixing, leading to the imperceptibility of the level repulsion in panel \[d\].
![ Optical absorption spectra for non-interacting electrons confined in a InSb quantum dot for various values the SO coupling strength (in meV.nm), $\alpha=0$ \[a\], $\alpha=20$ \[b\], $\alpha=30$ \[c\], and $\alpha=40$ \[d\]. []{data-label="fig:optics2"}](FD_OA_InSb.eps){width=".45\textwidth"}
We also mentioned lifting of degeneracies and rearrangements of energy levels as features of the SO coupling under small magnetic fields. Since we are interested in the absorption the most important states for us are the ones which can be reached from the ground state ($j=\frac12$) respecting the dipole transition selection rule $\Delta j=\pm1$. These are the lowest states with $j=-\frac12$ and $j=\frac32$ corresponding to the lines 2 and 3 in Fig. \[fig:fock1\] (a) and to the spionrs 2 and 3 (with $g<0$) in Table \[spinortable\], respectively, for vanishing SO coupling. Since as the SO interaction is stronger the larger is the (angular) momentum of the electron, the spinor 3 possessing maximum orbital angular momentum 2 is affected more than the spinor 2. Hence, although the energies of both spinors are decreased by the SO coupling the effect on the spinor 3 is larger.
![ Optical absorption spectra for non-interacting electrons confined in a GaAs quantum dot for various values the SO coupling strength (in meV.nm), $\alpha=0$ \[a\], $\alpha=20$ \[b\], $\alpha=30$ \[c\], and $\alpha=40$ \[d\]. []{data-label="fig:optics3"}](FD_OA_GaAs.eps){width=".45\textwidth"}
Keeping the above discusssions in mind it is now easy to interpret the features introduced by the SO coupling into the absorption spectra \[Figs. (\[fig:optics1\] – \[fig:optics3\])\]. Firstly, although the lower absorption branch consists mainly of transitions from the state 0 to the state 2 it shows an anticrossing at moderate coupling strengths. This is a direct consequence of the mixing of the spinor states 1 and 2 which results in two spinors, both with nonzero upper component. Thus we can see transitions from the ground state 0 to both of these. Furthermore, as mentioned earlier the level repulsion between states 1 and 2 resumes the form of level crossing when the SO interaction becomes very strong. This causes the anticrossing in the absorption spectra to disappear. In the InAs dot this happens already at $\alpha=40$ while in the InSb dot the anticrossing still persists. Since the Lande’ $g$-factor of GaAs is very small the Zeeman split state 1 does not meet the state 2 within the range of the magnetic field under consideration. Consequently we see at the lower right corners of Fig. \[fig:optics3\] (b) – (d) only the beginnings of the lower branches of these anticrossings.
Secondly, the upper absorption branch corresponds mainly to transitions from the state 0 to the state 3 modified by the SO coupling. The small magnetic field however makes an exception. As we discussed above, at small fields the state 3 is energetically lower than the state 2 due to the SO interaction. Thus we get a crossing of spectra at small fields.
From Eq. (\[nunm\]), the separation between the states 2 and 3, and hence also the gap between the absorption line branches is roughly proportional to the cyclotron frequency $\omega_c$ which in turn is linearly proportional to the magnetic field and inversely proportional to the effective mass of the electron. Thus, due to the very small effective mass of the electron in an InSb dot the energy of the spinor 3 exceeds the energy of the spinor 2 already at very small magnetic fields. Consequently also the crossing of absorption lines of an InSb dot occurs at very small magnetic fields as can be seen in Figs. \[fig:optics2\] (b)-(d).
At this point it may be worth mentioning that the energetics of the single-electron quantum dot under the influence of the SO coupling and subjected to an external magnetic field depends strongly on the sign of the Lande’ $g$-factor. This is contrary to the case without SO coupling where the energy spectrum is independent on the sign of $g$ although, of course, the orientation of spin is determined by it. Let us consider, for example, the lowest absorption branch in the case $g>0$. From Table \[spinortable\] we can deduce that also in this case the transitions mainly take the spinor 0 to the spinor 2. Now, however, there is no spinor level which would cross or even come close to the energy of the state 2 and mix with it. Consequently the anticrossing described above would not be observable in this case.
This concludes our discussion of the energy levels and optical absorption spectra in a non-interacting QD in a magnetic field and in the presence of SO interaction. In what follows, we describe the theory for a interacting few-electron parabolic quantum dot.
Many-electron systems
=====================
The basis ${\cal B}_N$ for $N$ interacting electrons in a QD is constructed as a direct antisymmetrized product of single-particle basis ${\cal B}_S$ (\[spbas\]) of the form $$\begin{aligned}
{\cal B}_N
&=&{\cal A}\bigotimes_{j=1}^N{\cal B}_S
=\{|\Lambda_i\rangle|\,i=1,2,\ldots\}\nonumber \\
&=&\big\{|\lambda_{i_1};\lambda_{i_2};\ldots;\lambda_{i_N}\rangle
\big|\,i_j=0,1,2,\ldots\big\} \nonumber \\
&=&\big\{|k_1,\ell_1;\ldots;k_N,\ell_N\rangle
\big|\,k_j=0,1,\ldots;\, \nonumber \\
&& \ell_j=0,\pm1,\ldots\big\},
\label{mpbas}\end{aligned}$$ where ${\cal A}$ stands for the antisymmetrization operator. It is also understood that the notations such as $|\lambda_{i_1};\lambda_{i_2};\ldots;\lambda_{i_N}\rangle$ represent the antisymmetrized direct products, i.e., $$|\Lambda_q\rangle=
|\lambda_{i_1};\lambda_{i_2};\ldots;\lambda_{i_N}\rangle
={\cal A}\left[
|\lambda_{i_1}\rangle\otimes|\lambda_{i_1}\rangle\otimes
\cdots\otimes|\lambda_{i_1}\rangle\right].
\label{asdirprod}$$ Usually it is possible to restrict the size of ${\cal B}_N$ using the conservation laws. For example, in a rotationally invariant system the total angular momentum is a good quantum number. Therefore, for example, we fix it to $J$, and accept into the basis only those states that satisfy $$\sum_{i=1}^Nj_i=J.$$
The states of the interacting system are expressed as superposition of the non-interacting states taken from the basis set (\[mpbas\]) $$|\Psi\rangle=\sum_{i=1}c_i|\Lambda_i\rangle.
\label{iasuppos}$$ To extract the coefficients $c_i$, we again resort to the minimization, i.e. we minimize the Rayleigh quotient $$\rho=\frac{\langle\Psi|{\cal H}|\Psi\rangle}{\langle\Psi|\Psi\rangle},$$ where $\cal H$ is the total many-body Hamiltonian. Again this leads to the diagonalization of the Hamiltonian matrix with elements $\langle\Lambda_i\vert{\cal H}\vert\Lambda_j\rangle$. The eigenvectors are the desired expansion coefficients and the eigenvalues the corresponding energies of the interacting system. It will be clear from the Appendix that both these tasks, construction of the Hamiltonian matrix and its diagonalization are numerically quite challenging.
Coulomb matrix elements
-----------------------
We write the total Hamiltonian $\cal H$ as a sum of the single-particle operators (\[spham\]) and two-body operators $V(\vec r, \vec r')$ as $${\cal H}=\sum_{i=1}^N{\cal H}_0(\vec r_i)
+\frac12\sum_{i\not=j}^NV(\vec r_i,\vec r_j).
\label{totham}$$ Since our basis states $|\Lambda_i\rangle$ are diagonal by construction in ${\cal H}_0$ we only need to evaluate the matrix elements of the latter sum in (\[totham\]). Our many-body states $|\Lambda_i\rangle$ are expressed in occupation representation language (\[asdirprod\]), and therefore it is natural to proceed in the occupation number space. This means that for the interaction part we have to evaluate the two-body terms $$V_{\lambda_1\lambda_2\lambda_3\lambda_4}
=\langle\lambda_1\lambda_2|V|\lambda_3\lambda_4\rangle.
\label{twbterm}$$ In our system the mutual interaction between the electrons is taken to be purely Coulombic, i.e. $$V(\vec r, \vec r')=\frac{e^2}{\epsilon|\vec r-\vec r'|}
\label{coulpot}$$ where $\epsilon$ is the effective dielectric constant of the material. The interaction operator is thus diagonal in spin. Recalling that our single-particle states $|\lambda\rangle$ were two-component spinors (\[spinordef\]), the two-body term (\[twbterm\]) consists of sum of four terms and is of the form $$V_{\lambda_1\lambda_2\lambda_3\lambda_4}
=\sum_{\sigma,\sigma'}\int d\vec r\,d\vec r'
{\phi_1^{\sigma}}^\ast(\vec r){\phi_2^{\sigma'}}^\ast(\vec r')
V(\vec r, \vec r')
\phi_3^{\sigma'}(\vec r')\phi_3^{\sigma}(\vec r)
\label{twbtermdec}$$ where the summation indices take values $\uparrow$ and $\downarrow$. Furthermore, since we expressed the spatial components $\phi^\sigma$ as superpositions of functions $g_{n\ell}\,e^{i\ell\theta}$ \[Eq. (\[gnmexp\])\], $$g^{\uparrow,\downarrow}=\sum_{n=0}c^{\uparrow,\downarrow}_{n,\ell}g_{n,\ell},$$ we are ultimately led to evaluate the Coulomb matrix elements in the oscillator wavefunction $$w_{n\ell}(\vec r) = g_{n\ell}\,e^{i\ell\theta}=\sqrt{\frac{n!}{(n+|\ell|)!}}
\,e^{-x/2}x^{|\ell|/2}L_n^{|\ell|}(x)\,e^{i\ell\theta}$$ basis. These matrix elements can be expressed in terms of finite sums as [@qdbook] $$\begin{aligned}
&&{\cal A}{^{\vphantom{\dagger}}}_{{\scriptstyle n_1 n_2 n_3 n_4
\atop \scriptstyle \ell_1 \ell_2 \ell_3 \ell_4}}
=\langle w_{n_1\ell_1}w_{n_2\ell_2}|\frac{e^2}{\epsilon|\vec r -\vec r'|}
|w_{n_3\ell_3}w_{n_4\ell_4}\rangle
\nonumber \\
&&=\delta_{\ell_1+\ell_2, \ell_3+\ell_4}
\dfrac{\sqrt2 e^2}{\epsilon a}\left[\dfrac{n_1!}{(n_1+|\ell_1|)!}
\right]^{\frac12}
\nonumber \\
&&\times
\left[\dfrac{n_2!}{(n_2 +|\ell_2|)!}\right]^{\frac12}
\left[\dfrac{n_3!}{(n_3+|\ell_3|)!}\right]^{\frac12}
\left[\dfrac{n_4!}{(n_4+|\ell_4|)!}\right]^{\frac12}
\nonumber \\
&&\times\sum_{\kappa_1=0}^{n_1}\sum_{\kappa_2=0}^{n_2}
\sum_{\kappa_3=0}^{n_3}\sum_{\kappa_4=0}^{n_4}
\left[\kappa_1+\kappa_4+\tfrac12 (|\ell_1|+|\ell_4|-k)\right]!
\nonumber \\
&&\times\left[\kappa_2+\kappa_3+\tfrac12(|\ell_2|+|\ell_3|-k)\right]!
\nonumber \\
&&\times\dfrac{(-1)^{\kappa_1+\kappa_4}}{\kappa_1!\kappa_4!}
\dfrac{(n_1+|\ell_1|)! (n_4+|\ell_4|)!}{(n_1-\kappa_1)!(|\ell_1|+
\kappa_1)!(n_4-\kappa_4)!(|\ell_4|+\kappa_4)!}
\nonumber \\
&&\times\dfrac{(-1)^{\kappa_2+\kappa_3}}{\kappa_2!\kappa_3!}
\dfrac{(n_2 + |\ell_2|)! (n_3+|\ell_3|)!}{(n_2-\kappa_2)!(|\ell_2|+
\kappa_2)!(n_3-\kappa_3)! (|\ell_3|+\kappa_3)!}
\nonumber \\
&&\times\sum_{s=0}^{\kappa_{14}}
\dfrac{\left[\kappa_1+\kappa_4+\frac12(|\ell_1|+|\ell_4|+k)\right]!}{
\left[\kappa_1+\kappa_4+\tfrac12(|\ell_1|+|\ell_4|-k)-s\right]!(k+s)!}
\nonumber \\
&&\times\sum_{t=0}^{\kappa_{23}}
\dfrac{\left[\kappa_2+\kappa_3+\frac12(|\ell_2|+|\ell_3|+k)\right]!}{
\left[\kappa_2+\kappa_3+\frac12(|\ell_2|+|\ell_3|-k)-t\right]!(k+t)!}
\nonumber \\
&&\times\dfrac{(-1)^{s+t}}{s!t!}\cdot\dfrac{\Gamma(k+s+t+
\frac12)}{2^{k+s+t+1}},
\label{coulelsum}\end{aligned}$$ where $\kappa_{14}=\kappa_1+\kappa_4+\tfrac12(|\ell_1|+|\ell_4|-k)$, $\kappa_{23}=\kappa_2+\kappa_3+\frac12(|\ell_2|+|\ell_3|-k)$ and $k=|\ell_1-\ell_4|=|\ell_2-\ell_3|$. Numerical techniques to evaluate these two-body terms and to diagonalize the resulting Hamiltonian matrix is described in the Appendix.
Results and discussions
=======================
For numerical evaluation of the energy spectra and the optical absorption spectrum for QDs with a few interacting electrons, we have considered the InAs, InSb and GaAs quantum dots. Parameters of these systems are already given in Sect.II. The energy spectra and the optical absorption spectra for these systems are described in the subsections below.
![Same as in Fig. \[fig:OA2\_InAs\], but for a three-electron InSb quantum dot. []{data-label="fig:OA3_InSb"}](OA3_InSb.eps){width=".45\textwidth"}
InAs quantum dots
-----------------
Our numerical results for energy spectra and absorption spectra (dipole-allowed) for 2 – 4 electrons are presented in Figs. \[fig:E2\_InAs\]-\[fig:OA4\_InAs\], and for various values of the SO coupling strength (in meV.nm), $\alpha$. As in the case of the non-interacting electron system, we have considered the following parameters for the InAs quantum dot: $m^*/m_0=0.042, \epsilon=14.6, g=-14$ and $\hbar\omega_0=7.5$ meV.
A striking feature visible in the absorption spectra (Figs. \[fig:OA2\_InAs\], \[fig:OA3\_InAs\], and \[fig:OA4\_InAs\]) is the appearance of discontinuities, anticrossings and new modes in addition to the two main ($\alpha=0$) absorption lines. These optical signatures of the SO interaction are consequences of the multitude of level crossings and level repulsions that occur in the energy spectra (Figs. \[fig:E2\_InAs\], \[fig:E3\_InAs\], and \[fig:E4\_InAs\]). The latter ones can be attributed to an interplay between the SO and Zeeman couplings. In order to understand their origin, let us first examine the case of the two-electron system (Figs. \[fig:E2\_InAs\],\[fig:OA2\_InAs\]). In our spinor notation the main contribution to the ground state at zero magnetic field comes from the two-electron state $|\lambda_{\ell_1},\lambda_{\ell_2}\rangle=|\lambda_0,\lambda_{-1}\rangle$, where $|\lambda_{\ell_1}\rangle$ is a spinor with $j_1=\ell_1+1/2=1/2$, $d_n^{\lambda_1}=0$, and $|\lambda_{\ell_2}\rangle$ a spinor with $j_2=-1/2$ and $u_n^{\lambda_2}=0$, i.e., both electrons have zero orbital angular momenta with opposite spins (corresponding to the spinors 0 and 1 in Table \[spinortable\] with $J=j_1+j_2=0$). When we increase the magnetic field the spin triplet configuration will become, due to the interaction, energetically more favorable. If the Lande’ $g$-factor is negative then the electrons would like to occupy states with orbital angular momenta 0 and $-1$ with both spins up (i.e., states 0 and 2 of Table \[spinortable\]). In the spinor picture this means that $|\lambda_{\ell_2}\rangle$ still has $\ell_2=-1$ ($J=0$) but now $u_n^{\lambda_2}\not=0$ and $d_n^{\lambda_2}=0$. The SO interaction mixes these two configurations which results in a level repulsion. On the other hand, when the strength of the SO coupling is further increased, the relative significance of the Zeeman contribution to ${\cal H}_0$ decreases. The energy shifts to states with $J\not=0$ will then become energetically feasible and we again have crossings of levels. For increasing number of electrons in the dot, the energy spectra is more dense and exhibit additional level crossings (Figs. \[fig:E3\_InAs\] – \[fig:OA4\_InAs\]). As a consequence, the ground state angular momentum also changes more frequently as compared to that of the two-electron case. It should be pointed out that in many-electron dots these level crossings and repulsion are to be attributed, at least partly to the mutual Coulomb interactions. The level crossings/repulsions we saw earlier \[e.g. levels 1 and 2 in Fig. \[fig:fock1\] (a)\] in the single-particle particle picture are due to the Zeeman splitting whereas in interacting systems crossings occur even in the limit of vanishing Zeeman coupling. In the present InAs dot the Coulomb interaction brings, for example the singlet-triplet transition to much lower magnetic field ($B\approx 2$T) as compared to the field required in a noninteracting system ($B\approx 4$T).
![Same as in Fig. \[fig:OA2\_InAs\], but for a four-electron InSb quantum dot. []{data-label="fig:OA4_InSb"}](OA4_InSb.eps){width=".45\textwidth"}
At moderate SO coupling strengths the absorption spectra do not essentially differ from the single-particle spectrum. But when the coupling strength increases the deviation from the pure parabolic confinement also increases which in turn implies that the lowest final states of dipole allowed transitions are not any more achievable by adding $\hbar\Omega\pm\frac12\hbar\omega_c$ to the initial state energies. In particular, this results in discontinuities and anticrossing behaviors as well as appearence of new modes. As an illustration, let us consider the absorptions that at a magnetic field of $B=1$T take the two-electron system from the ground state to excited states. In the absence of the SO coupling the ground state is a spin-singlet state $S =0$ with total angular momentum $J=0$. According to the dipole selection rules absorptions cause transitions to states $J=\pm1$ and $S=0$ with energies $\Delta E_\pm$ above the ground state. In Fig. \[fig:OA2\_InAs\] (d), we note that in addition to the two main lines there are now two additional lines (at around $B=1$ T) of appreciable intensity at the SO coupling strength $\alpha=40$ (mev.nm). Further analysis reveals that the ground states still have $J=0$ and that the expectation value of the spin $z$-component is $\langle\sigma_z\rangle=0$. The excited states also have $J=\pm1$, as before. However, the final spin states can no longer be classified as singlets: the expectation values $\langle\sigma_z\rangle$ vary between $-0.03$ and 0.39. When the number of electrons increases the number of these additional modes also increases but at the same time the relative intensities decrease (at each $B$ we have normalized the total intensity to unity). On the other hand, the discontinuities as consequences of deviations from a parabolic confinement become more pronounced (Figs. \[fig:E3\_InAs\] – \[fig:OA4\_InAs\]). This is because there are higher angular momenta involved in the dipole transitions. As a consequence of this the upper absorption branch now exhibits a rich structure while in the single-particle picture it is practically featureless.
InSb quantum dots
-----------------
As mentioned above, in addition to the InAs quantum dots, investigation of InSb quantum dots are also thought to be interesting, particularly in the context of SO coupling effects due to the large values of $\vert g\vert$ and $\alpha$ [@khodaparast]. We have considered the following parameters for the InSb quantum dot: $m^*/m_0=0.014,
\epsilon=17.88, g=-40$ and $\hbar\omega_0=7.5$ meV. The energy levels for InSb quantum dots containing 2 – 4 interacting electrons are plotted in Figs. \[fig:E2\_InSb\], \[fig:E3\_InSb\], \[fig:E4\_InSb\] and for various values of the SO coupling strength $\alpha$. The corresponding optical absorption spectra are presented in Figs. \[fig:OA2\_InSb\], \[fig:OA3\_InSb\], and \[fig:OA4\_InSb\]. As compared to the spectra of InAs dots a clear difference is the almost total absence of anticrossings and discontinuities. This is partly due to the very large Zeeman coupling which practically nullifies the SO interaction at the coupling strengths $\alpha$ we are concerned with. Another reason is the large kinetic energies due to the very small electron effective mass. Because the strength of the Coulomb interaction is somewhat smaller than in InAs ($\epsilon_{\rm InSb}
> \epsilon_{\rm InAs}$) correlations caused by the mutual electronic interactions are effectively much smaller in InSb than in InAs. For the exploration of the SO coupling via absorption spectroscopy, InSb quantum dots do not seem to be a very promising system.
![Same as in Fig. \[fig:OA2\_InAs\], but for a two-electron GaAs quantum dot. []{data-label="fig:OA2_GaAs"}](OA2_GaAs.eps){width=".45\textwidth"}
GaAs quantum dots
-----------------
The results for GaAs quantum dots, ones that are most intensely explored in the absence of SO coupling, are presented here primarily as an academic interest. The parameters that we have used here are: $m^*/m_0=0.063, \epsilon=12.9,
g=-0.44$ and $\hbar\omega_0=7.5$ meV. Clearly, the very low value of the $\vert g \vert$-factor perhaps makes the GaAs QDs unsuitable for any observable effect due to the SO coupling. Interestingly, however, among all the three types of QDs studied here for optical absorptions, GaAs QDs show the most spectracular effects for large values of $\alpha$. The energy levels for GaAs quantum dots containing 2 – 4 interacting electrons are plotted in Figs. \[fig:E2\_GaAs\], \[fig:E3\_GaAs\], and \[fig:E4\_GaAs\] for various values of the SO coupling strength $\alpha$. The corresponding optical absorption spectra are presented in Figs. \[fig:OA2\_GaAs\], \[fig:OA3\_GaAs\], and \[fig:OA4\_GaAs\] respectively. As mentioned above, the only observed value of $\alpha$ for GaAs QD reported as yet is $\alpha\sim6$ meV.nm [@koenemann]. Reversing the arguments presented in the previous subsection, i.e. in GaAs a very small Zeeman coupling and a rather large effective electron mass but practically equal strength of Coulomb interaction help us understand why the absorption spectra of our GaAs dots exhibit a remarkably rich structure as opposed to those of the InSb and InAs dots. Finding an appropriate set up to generate a large $\alpha$ for GaAs quantum dot would be a major (but worthwhile) experimental endeavor.
A brief review of earlier theoretical works
===========================================
In this section, we present a critical review of earlier theoretical reports on how the Bychkov-Rashba spin-orbit coupling in parabolic quantum dots were treated [@kuan; @governale; @vam_PRB; @vam_JAP; @destefani; @debald; @koenemann; @quasi_exact; @tsitsishvili; @manuel_041; @manuel_042; @manuel_043; @manuel_02; @lucignano; @cremers; @bellucci; @fransson]. There are quasi-exact solutions available for electrons confined in a parabolic quantum dot in the presence of the SO interaction, but without the inter-electron interaction [@quasi_exact], and exact analytical results are also reported in the case of a circular quantum dot with hard walls [@tsitsishvili], again for a non-interacting system, but with the Bychkov-Rashba spin-orbit interaction included. However, for the realistic systems of parabolic quantum dots with interacting electrons these methods are prohibitively complicated and evaluation of the energy spectrum can only be done numerically. Among the theoretical papers dealing with the SO interaction in quantum dots discussed below, Kuan et al. [@kuan] presented the best treatment of the single-electron states. They looked at the energy levels of parabolically confined quantum dots with Bychkov-Rashba SO coupling and in the presence of zero and nonzero magnetic fields. They solved the single-particle equation correctly by expanding the solution spinors in terms of the eigenfunctions of QDs without the SO interaction, i.e. the Laguerre functions. We have used a similar approach to construct the basis states for our multi-electron QDs \[Sects. II, III\].
Voskoboynikov et al. [@vam_PRB] studied the effect of SO interaction on the energy spectrum of cylindrical semiconductor QDs in an externally applied magnetic field. They considered the Bychkov-Rashba SO coupling due to the parabolic confinement, i.e. an in-plane field conserving orbital and spin angular momenta. As a consequence, the SO coupling has no qualitative effects on, for example, the absorption spectra. The electron-electron interaction was not included in this scheme. In Ref. [@vam_JAP], they studied the magnetization and magnetic susceptibility in few-electron parabolic QDs with SO coupling. As in their earlier paper, they handled the SO term only due to the parabolic confinement and therefore the single-particle Hamiltonian is diagonal in spin space. They neglected the mutual electronic Coulomb interaction.
Governale [@governale] investigated the effects of SO coupling on the addition energy and on the spin properties of few-electron QDs in the absence of the external magnetic field. He introduced the SO coupling into single-particle states perturbatively but also compared the resulting energies to the ones obtained by numerical diagonalization technique. At the small SO coupling strength that he considered, the perturbation approach seems to be valid. Electron correlations were handled by using the spin-density functional approach. Cremers et al. [@cremers] studied conductance and its fluctuations in the presence of SO interaction, Zeeman coupling, externally applied magnetic field, in a (single-electron) QD. They solved the single-particle equation applying an approximate unitary transformation which in leading order takes the Hamiltonian to a diagonal form in spin space.
![Same as in Fig. \[fig:OA2\_InAs\], but for a three-electron GaAs quantum dot. []{data-label="fig:OA3_GaAs"}](OA3_GaAs.eps){width=".45\textwidth"}
Valin-Rodriguez [@manuel_041] considered a single electron in a parabolic QD with Bychkov-Rashba SO coupling. He performed a unitary transformation to transform the Hamiltonian in spin space to a diagonal form up to second order in SO and Zeeman coupling parameters. He showed that the effective SO interaction is influenced by the interplay between Zeeman and SO couplings. In Ref. [@manuel_042], Valin-Rodriguez et al. introduced a spatially modulated (in radial direction) Bychkov-Rashba coupling in single-electron (disk) QDs. They solved the two-component spinor equation and numerically evaluated the spin density. They concluded that it is possible to confine electrons spatially with appropriate structural modulation. These authors also investigated the SO couplings in deformed parabolic quantum dots [@manuel_043]. They solved the single-particle equations using the approximate unitary transformations mentioned above. They were interested in the effects of spatial deformations to the spin splitting oscillations. They estimated the Coulomb interaction contribution using the time-dependent local-spin-density approximation [@manuel_02].
Lucignano et al. [@lucignano] studied the few-electron QDs including the mutual electron-electron interaction and under the influence of an externally applied magnetic field. They applied an exact diagonalization method (but with a rather restricted basis: 28 single-particle states deducing from their earlier paper [@jouault]. They particularly looked at the possibility to use the SO coupling to control the excitations under the magnetic fields which polarize the ground state, i.e., close to the final [*single-triplet*]{} transition. The SO coupling is included in the many-electron Hamiltonian, but not in the basis states. They evaluated the dipole matrix elements for absorption from the ground state to the lowest dipole-allowed excited state. They claimed that there is an increase in intensity close to the transition to the fully polarized ground state.
![Same as in Fig. \[fig:OA2\_InAs\], but for a four-electron GaAs quantum dot. []{data-label="fig:OA4_GaAs"}](OA4_GaAs.eps){width=".45\textwidth"}
Destefani et al. [@destefani] reported numerical results for energy levels and spin polarizations for one and two-electron parabolic QDs under a magnetic field and the SO coupling. For the two-electron system, Coulomb interaction is also included. They do not construct the correct single-electron states in the two-electron QD, the SO coupling is, in fact, taken into account only in the many-electron Hamiltonian. Debald et al. [@debald] studied oscillations in few-electron parabolic QD in a magnetic field, between states where the degeneracy is lifted by the SO coupling, i.e. at the level repulsion points. The Coulomb interaction was taken into account only approximately, because the many-body effects were claimed to play only a minor role in the very small magnetic field considered in that work. Könemann et al. [@koenemann] considered the SO coupling in single electron QDs. They showed that there is an anisotropy between spin splittings due to magnetic fields parallel and perpendicular to the dot. The anisotropy was shown to be proportional to the strength of the SO coupling. Bellucci and Onorato [@bellucci] studied the influence of SO coupling on the charge and spin polarization in a vertical disk-shaped QD under a strong perpendicular magnetic field. They treated the SO coupling perturbatively (upto second order). They handled the Coulomb interaction within the Hartree-Fock approach. They studied the energy splittings due to the SO coupling. Finally, Fransson et al. [@fransson] studied transport through QDs with spin dependent couplings to the contacts. They evaluated the QD energy levels using a (first principles) density functional theory. They calculated the transport properties of (a) non-interacting electrons taking into account the few levels closest to the Fermi level, and (b) interacting electrons using an approximate Hamiltonian with the levels closest to the Fermi level. We would like to note here that, in the light of all these theoretical approaches, our method of including the SO coupling for interacting electrons in a parabolic QD seems to be the most accurate one. Although, given the fact that our approach involves extensive numerical computations, some of the approaches discussed above, such as the one by Lucignano et al. [@lucignano], seem to be very promising.
Conclusions
===========
In conclusion, we have studied the energy levels and optical-absorption spectra for parabolic quantum dots containing upto four interacting electrons, in the presence of spin-orbit coupling and under the influence of an externally applied, perpendicular magnetic field. We have presented a very accurate numerical scheme to evaluate these quantities. We have presented results for the Fock-Darwin spectra in the presence of SO coupling for quantum dots made out of three different semiconductor systems, InAs, InSb, and GaAs. The effects of SO coupling on the single-electron spectra are primarily to lift the degeneracy at $B=0$, rearrangement of some of the energy levels at small magnetic fields, and level repulsions at high fields. These are explained as due to mixing of different spinor states for increasing strength of SO coupling. As a consequence, the corresponding absorption spectra reveal anticrossing structures in the two main lines $(\alpha=0)$ of the spectra. For the interacting many-electron systems we observed the appearence of discontinuities, anticrossings, and new modes that appear in conjunction with the two main absorption lines. These additional features arise entirely due to the SO coupling and are a consequence of level crossings and level repulsions in the energy spectra. An intricate interplay between the SO coupling and the Zeeman energies are shown to be responsible for these new features seen in the energy spectra. Our accurate results for the low-lying energy levels for the SO coupled QDs can also be measured, in principle, by transport [@rolf_review] or capacitance [@ashoori] spectroscopy, which have been successfully employed earlier to map out the energy spectra of parabolic quantum dots. Optical absorption spectra for all three types of quantum dots containing a few interacting electrons that are studied here show a common feature: new modes appear, mostly near the upper main branch of the spectra around 2 tesla that become stronger with increasing $\alpha$. Among the three types of systems considered here, optical signatures of the SO interaction is found to be the strongest in the absorption spectra of a GaAs quantum dot, but only at very large values of the SO coupling strength, and appears to be the weakest for the InSb quantum dots. Experimental observation of these new optical modes that appear solely due to the presence of SO coupling would be very exciting because that would be a major step forward in our quest to manipulate the spin dynamics in nanostructured systems via the SO coupling.
Our future works along this line will be to explore coupled QDs, or QD molecules [@austing]. Our primary goal will be to generate accurate results for energy levels and optical absorption spectra for coupled (laterally [@laterally] or vertically [@vertically]) quantum dots with spin-orbit interaction. In addition to being important for fundamental studies, these results would be interesting from the point of view of quantum computations [@coupled_bits] as well. It is now well recognized that semiconductor quantum dots have the potential to become the building blocks for solid state quantum computation. Quantum states of single, double and even triple coupled [@triple_coupled] quantum dots have been explored for this purpose. In a quantum computer, information is stored in a two-level system. Hence a promising candidate system to realize the quantum bits, the fundamental unit of information in a quantum computer, is a quantum dot where the single-electron states can be used for that purpose. In a magnetic field, the Zeeman splitting of electron spin can provide a two-level system (for an odd number of electrons). Alternatively, the spatial wave function of a single-electron state in a double quantum dot (allowing for electron tunneling between the two dots) can also represent a two-level system. Spin-orbit coupling in the coupled quantum dot systems could perhaps be used to perform quantum computation (using the spin rotation, for example). Accurate results for the energy levels of the coupled-dot system might be beneficial in that direction of research. The effects of SO coupling on the energy levels and absorption spectra for a more complex system such as coupled QDs are, however, important and interesting in their own ways. These would be the subjects of our future publications.
ACKNOWLEDGMENTS
===============
The work of T.C. has been supported by the Canada Research Chair Program and the Canadian Foundation for Innovation (CFI) Grant. We wish to thank Dr. Alex Voskoboynikov for a critical reading of the manuscript.
Diagonalization of monster matrices
===================================
We have mentioned in Sect. III about the challenging task of construction of the Hamiltonian matrix and its numerical diagonalization. Here we present a brief discussion about the numerical method that we believe the most accurate (and appropriate) for that task. While in principle, evaluation of Eq. (\[coulelsum\]) is straightforward it turns out to be numerically highly unstable, primarily due to the expansion (\[gnmexp\]) extending to Laguerre polynomials of large degree and large angular momenta which in turn, leads to large terms of alternating sign [@stone]. A remedy for this is to employ multiple precision arithmetics such as, for example implemented in the Gnu arbitrary precision GMP library. However, if we apply multiple precision arithmetic directly into the sixfold summation (\[coulelsum\]) the time consumed to evaluate these becomes insurmountable. To circumvent this obstacle we note that many terms in the sums actually depend on very few parameters, the range of these parameters is restricted and the same functional forms repeat themselves. Thus a natural solution is to tabulate these forms and the subsums. In our Coulomb matrix element code we used the tabulated functions $$\begin{aligned}
D(i) &=& i! \\
F(n,\ell,\kappa)&=&\frac{(n+\ell)!}{(n-\kappa)!(\ell+\kappa)!}\\
&=&\frac{D(n+\ell)}{D(n-\kappa)D(\ell+\kappa)} \\
G(n,\ell,\kappa)&=&\frac{(n+\ell)!}{\kappa!(n-\kappa)!(\ell+\kappa)!}
=\frac{F(n+\ell)}{D(\kappa)} \\
H(s)&=&\frac{\Gamma(s+\frac12)}{2^{s+1}} \\
I(q_1,q_2,\ell)&=&
D(q_1)D(q_2)\sum_{s=0}^{q_1}(-1)^sG(q_1,\ell,s) \nonumber \\
&&\times\sum_{t=0}^{q_2}(-1)^tG(q_2,\ell,t)H(t+s+\ell).\end{aligned}$$ Now the summations $\Sigma$ in the expression (\[coulelsum\]) can be written as $$\begin{aligned}
\Sigma&=&\sum_{\kappa_1=0}^{n_1}(-1)^{\kappa_1}G(n_1,\ell_1,\kappa_1)
\sum_{\kappa_2=0}^{n_2}(-1)^{\kappa_2}G(n_2,\ell_2,\kappa_2)
\nonumber \\
&&\times
\sum_{\kappa_3=0}^{n_3}(-1)^{\kappa_3}G(n_3,\ell_3,\kappa_3)
\sum_{\kappa_4=0}^{n_4}(-1)^{\kappa_4}G(n_4,\ell_4,\kappa_4)
\nonumber \\
&&\times
I(\kappa_1+\kappa_4+\tfrac12(\ell_1+\ell_4-k),
\kappa_2+\kappa_3 \nonumber \\
&&+\tfrac12(\ell_2+\ell_3-k)).
\label{couleltabsum}\end{aligned}$$ Although the summation here is still fourfold it is nevertheless several orders of magnitude faster than the original one (\[coulelsum\]) and as such fast enough for our purposes.
Under the influence of the SO coupling the total spin $S$ of our many-electron system is not a conserved quantity. As a consequence of this we cannot fix the total $S_z$ of the many-body basis. This degree of freedom tends to make the number of non-interacting many-body states of the basis very large even for a small number of electrons, and even if the conservation of the total angular momentum $J_z=L+S_z$ is taken into account. For example, to achieve a convergence for a four electron system in the parabolic confinement with harmonic potential ($\hbar\omega_0$) of few meV the size of the basis must be [*of the order of a million*]{}. Furthermore, since we want to study properties of the eigenstates, such as polarization and dipole matrix elements between states, we also need the relevant eigenvectors. Clearly the sheer size of the matrix prohibits a full diagonalization and we have to resort to an iterative scheme aimed to search a given number of energetically lowest eigenvectors. Of course the algorithm should be fast and hopefully also robust.
The algorithm proposed by Davidson and Liu (DL) to evaluate the eigenvalues and eigenvectors of “monster matrices” [@davidson] seem to fit our criteria. Like any other iterative method it transforms the diagonalization of a matrix $A$ to minimization of the Rayleigh quotient $$\lambda=\frac{x^{\scriptsize T} Ax}{x^{\scriptsize T}x},
\label{minprinc}$$ where $x$ represents the column vector of the coefficients in the superposition of the basis states. Also, as in many other methods, the only operations involving the matrix $A$ are vector multiplications. This allows us to exploit fully the sparseness of $A$, i.e. we have to store only the non-zero elements.
The key idea behind practically any iterative method subjecting the matrix only to multiplication is to search the minimum of the quotient (\[minprinc\]) in a (very small) subspace and to update this subspace in each iteration step. How this updating is performed depends on the method. For example, in the common conjugate gradient method the subspace is two-dimensional and spanned by the gradient $g$ at current position $x$ and a vector $s$ conjugate to it with respect to $A$ (i.e. $s^{\scriptsize T} A g=0$).
In the DL method the dimension of the search space varies from step to step. Suppose that at a given step our search space $\cal S$ is spanned by the orthonormal vectors $s_1,\ldots,s_K$ (the dimension $K$ must, of course, be greater than the number of required eigenstates). Finding the minimum of (\[minprinc\]) in this subspace corresponds to the diagonalization of the $K\times K$ matrix ${\cal S}^{\scriptsize T}A{\cal S}$ (we take $\cal S$ to represent also the matrix with columns $s_k$). As a result we get $K$ eigenvalues $\mu_k$ and $K$ eigenvectors $z_k$, each of dimension $K$. The expanded vectors $$y_k=\sum_{i=1}^Kz_{k,i}s_i$$ will approximate the eigenvectors and the quantities $\mu_k$ the eigenvalues we are seeking for. The next task in the iteration step is to update the subspace $\cal S$. For that purpose we pick up a certain number (a parameter depending for example on the size of the computer memory) of the residuals $$r_k=(A-\mu_k)y_k$$ with largest norms. The selected residuals are orthonormalized with respect to the space $\cal S$ and then appended to it. So, in each step the dimension of the search space and the size required to store it increases and we may eventually exhaust all the memory. At this point we compress the space $\cal S$ to its bare minimum comprising only as many vectors as we are required to find. These vectors are selected from the set $\{y_k\}$ and are the ones with smallest $\mu_k$.
[99]{} Electronic mail: [email protected] T. Chakraborty, F Peeters, and U. Sivan (Eds.), [*Nano-Physics [&]{} Bio-Electronics: A New Odyssey*]{} (Elsevier, 2002). P.A. Maksym and T. Chakraborty, Phys. Rev. Lett. [**65**]{}, 108 (1990). T. Chakraborty, Comments Condens. Matter Phys. [**16**]{}, 35 (1992); V. Gudmundsson, A. Manolescu, R. Krahne, and D. Heitmann, in Ref. [@nanobio], Ch. 7. T. Chakraborty, [*Quantum Dots*]{} (North-Holland, Amsterdam, 1999). C. Sikorski and U. Merkt, Phys. Rev. Lett. [**62**]{}, 2164 (1989). See, for example, I. Magnusdottir and V. Gudmundsson, Phys. Rev. B [**60**]{}, 16591 (1999); R. Krahne, et al., [*ibid.*]{} [**63**]{}, 195303 (2001); E. Lipparini, et al. [*ibid.*]{} [**56**]{}, 12375 (1997); P.A. Maksym, Physica B [**184**]{}, 385 (1993); T. Seki, Y. Kuramoto, and T. Nishino, J. Phys. Soc. Jpn. [**65**]{}, 3945 (1996); M. Wagner, A.V. Chaplik, and U. Merkt, [*ibid.*]{} [**51**]{}, 13817 (1995); B. Meurer, D. Heitmann, and K. Ploog, Phys. Rev. Lett. [**68**]{}, 1371 (1992); T. Chakraborty, V. Halonen, and P. Pietiläinen, Phys. Rev. B [**43**]{}, 14289 (1991). D. Mowbray and J. Finley, in Ref. [@nanobio], Ch. 3; M.S. Skolnick and D.J. Mowbray, Physica E [**21**]{}, 155 (2004). A.J. Shields, et al., in Ref. [@nanobio], Ch. 4. V. Fock, Z. Phys. [**47**]{}, 446 (1928); C.G. Darwin, Proc. Cambridge Philos. Soc. [**27**]{}, 86 (1930). For a similar result in a parabolic quantum well, see, L. Brey, N.F. Johnson, and B.I. Halperin, Phys. Rev. B [**40**]{}, 10647 (1989). T. Chakraborty and P. Pietiläinen, Phys. Rev. Lett. (2005). D.D. Awschalom, D. Loss, and N. Samarth (Eds.), [*Semiconductor Spintronics and Quantum Computation*]{} (Springer, 2002); D. Grundler, Phys. World [**15**]{}, 39 (2002); S.A. Wolf, et al., Science [**294**]{}, 1488 (2001); G.A. Prinz, Phys. Today [**48**]{}, 58 (1995). , edited by H. Ohno \[Physica E [**10**]{} (2001)\]; G. Schmidt, C. Gould, and L.W. Molenkamp, Physica E [**25**]{}, 150 (2004). W.H. Kuan, C.S. Tang, W. Xu, J. Appl. Phys. [**95**]{}, 6368 (2004). O. Voskoboynikov, C.P. Lee, and O. Tretyak, Phys. Rev. B [**63**]{}, 165306 (2001). O. Voskoboynikov, O. Bauga, C.P. Lee, and O. Tretyak, J. Appl. Phys. [**94**]{}, 5891 (2003) M. Governale, Phys. Rev. Lett. [**89**]{}, 206802 (2002). J.H. Cremers, P.W. Brouwer, and V.I. Falko, Phys. Rev. B [**68**]{}, 125329 (2003). H. Tütüncüler, R. Koc, and E. Olgar, J. Phys. A [**37**]{}, 11431 (2004). E. Tsitsishvili, G.S. Lozano, and A.O. Gogolin, Phys. Rev. b [**70**]{}, 115316 (2004). M. Valin-Rodriguez, Phys. Rev. B [**70**]{}, 033306 (2004). M. Valin-Rodriguez, A. Puente, and L. Serra, Phys. Rev. B [**69**]{}, 153308 (2004). M. Valin-Rodriguez, A. Puente, and L. Serra, Phys. Rev. B [**69**]{}, 085306 (2004). M. Valin-Rodriguez, A. Puente, and L. Serra, Phys. Rev. B [**66**]{}, 165302 (2002). P. Lucignano, B. Jouault, A. Tagliacozzo, and B.L. Altshuler, Phys. Rev. B [**71**]{}, 121310 (2005); P. Lucignano, B. Jouault, and A. Tagliacozzo, [*ibid.*]{} [**69**]{}, 045314 (2004). C.F. Destefani, S.E. Ulloa, and G.E. Marques, Phys. Rev. B [**69**]{}, 125302 (2004); [**70**]{}, 205315 (2004). S. Debald and C. Emary, Phys. Rev. Lett. [**94**]{}, 226803 (2005). J. Könemann, R.J. Haug, D.K. Maude, V.I. Falko, and B.L. Altshuler, Phys. Rev. Lett. [**94**]{}, 226404 (2005). S. Bellucci and P. Onorato, Phys. Rev. B [**72**]{}, 045345 (2005). J. Fransson, et al., Phys. Rev. B [**67**]{}, 205310 (2003). T. Chakraborty and P. Pietiläinen, Phys. Rev. B [**71**]{}, 113305 (2005). M.A. Eriksson, et al., Quantum Information Processing [**3**]{}, 133 (2004); D. Loss, G. Burkard, and D.P. DiVincenzo, J. Nanoparticle Res. [**2**]{}, 401 (2000); R. Hanson et al., Phys. Rev. Lett. [**94**]{}, 196802 (2005); D. Stepanenko and N.E. Bonesteel, Phys. Rev. Lett. [**93**]{}, 140501 (2004). S. Bandyopadhyay, Phys. Rev. B [**61**]{}, 13813 (2000). D. Grundler, Phys. Rev. Lett. [**84**]{}, 6074 (2000); C.-M. Hu, J. Nitta, T. Akazaki, H. Takayanagi, J. Osaka, P. Pfeffer and W. Zawadzki, Physica E [**6**]{}, 767 (2000); C.-M. Hu, et al., Phys. Rev. B [**60**]{}, 7736 (1999); Y. Sato, T. Kita, S. Gozu, and S. Yamada, J. Appl. Phys. [**89**]{}, 8017 (2001); J. Nitta, T. Akazaki, and H. Takayanagi, Phys. Rev. Lett. [**78**]{}, 1335 (1997). Y.A. Bychkov and E.I. Rashba, J. Phys. C [**17**]{}, 6039 (1984). H.A. Bethe and E.E. Salpeter, [*Quantum Mechanics of One- and Two-Electrons Atoms*]{} (Springer, Berlin, 1957). W. Zawadzki and P. Pfeffer, Semicond. Sci. Technol. [**19**]{}, R1 (2004). R.J. Haug, J. Weis, R.H. Blick, K von Klitzing, K. Eberl, and K. Ploog, Nanotechnology [**7**]{}, 381 (1996); T Schmidt, R.J. Haug, K. von Klitzing, A. Förster, and H. Lüth, Phys. Rev. Lett. [**78**]{}, 1544 (1997). R.C. Ashoori, et al., Phys. Rev. Lett. [**71**]{}, 613 (1993). V. Halonen, P. Pietiläinen, and T. Chakraborty, Europhys. Lett. [**33**]{}, 377 (1996). P. Tonello and E. Lipparini, Phys. Rev. B [**70**]{}, 081201 (2004). G.A. Khodaparast, R.E. Doezema, S.J. Chung, K.J. Goldammer, and M.B. Santos, Phys. Rev. B [**70**]{}, 155322 (2004). B. Jouault, G. Santoro, and A. Tagliacozzo, Phys. Rev. B [**61**]{}, 10242 (2000). D.G. Austing, et al., in Ref. [@nanobio], Ch. 2. T. Chakraborty, V. Halonen, and P. Pietiläinen, Phys. Rev. B [**43**]{}, 14289 (1991); A. Wensauer, O. Steffens, M. Suhrke, and U. Rössler, [*ibid.*]{} [**62**]{}, 2605 (2000); J. Kolehmainen, S.M. Reimann, M. Koskinen, and M. Manninen, Eur. Phys. J. B [**13**]{}, 731 (2000); A. Harju, S. Siljamäki, and R.M. Nieminen, Phys. Rev. Lett. [**88**]{}, 226804 (2002); Z. Dai, et al., Eur. Phys. J. B [**29**]{}, 141 (2002); Appl. Phys. Lett. [**80**]{}, 2577 (2002); A.J. Markvoort, P.A.J. Hilbers, and R. Pino, J. Phys.: Condens. Matter [**15**]{}, 6977 (2003); B. Szafran, F.M. Peeters, and S. Bednarek, Phys. Rev. B [**70**]{}, 205318 (2004); T. Hatano, et al., Phys. Rev. Lett. [**93**]{}, 066806 (2004); L.-X. Zhang, et al., Phys. Rev. B [**69**]{}, 245301 (2004); B. Szafran and F.M. Peeters, [*ibid.*]{} [**71**]{}, 245314 (2005). J.J. Palacios and P. Hawrylak, Phys. Rev. B [**51**]{}, 1769 (1995); J.H. Oh, K.J. Chang, G. Ihm, and S.J. Lee, [*ibid.*]{} [**53**]{}, 13264 (1996); W.-Y. Ruan and H.-F. Cheung, Eur. Phys. J. B [**3**]{}, 407 (1998); H. Imamura, P.A. Maksym, and H. Aoki, Phys. Rev. B [**53**]{}, 12613 (1996); [**59**]{}, 5817 (1999); Y. Tokura, D.G. Austing, and S. Tarucha, J. Phys.: Condens. Matter [**11**]{}, 6023 (1999); B. Partoens, A. Matulis, and F.M. Peeters, Phys. Rev. B [**59**]{}, 1617 (1999); G. Burkard, G. Seelig, and D. Loss, [*ibid.*]{} [**62**]{}, 2581 (2000); B. Partoens and F.M. Peeters, Europhys. Lett. [**56**]{}, 86 (2001); W. Xie and P. Sun, J. Phys.: Condens. Matter [**14**]{}, 7245 (2002); S. Bednarek, T. Chwiej, J. Adamowski, and B. Szafran, Phys. Rev. B [**67**]{}, 205316 (2003); D. Bellucci, et al., [*ibid.*]{} [**69**]{}, 201308 (2004); M. Rontani, et al., [*ibid.*]{} [**69**]{}, 085327 (2004); G. Ortner, et al., Phys. Rev. Lett. [**94**]{}, 157401 (2005); D. Bellucci, F. Troiani, G. Goldoni, and E. Molinari, J. Lumin. [**112**]{}, 109 (2005). X.-Q. Li and Y. Arakawa, Phys. Rev. A [**63**]{}, 012302 (2000); J.M. Elzerman, et al., Phys. Rev. B [**67**]{}, 161308 (2003); X.-Q. Li and Y. Yan, [*ibid.*]{} [**65**]{}, 205301 (2002); H. Qin, et al., [*ibid.*]{} [**64**]{}, 241302 (2001); T. Hayashi, et al., Phys. Rev. Lett. [**91**]{}, 226804 (2003); S. Vorojitsov, E.R. Mucciolo, and H.U. Baranger, Phys. Rev. B [**69**]{}, 115329 (2004); J. Gorman, D.G. Hasko, and D.A. Williams, Phys. Rev. Lett. [**95**]{}, 090502 (2005). H. Sasakura, et al., Semicond. Sci. Technol. [**19**]{}, S409 (2004). M. Stone, H.W. Wyld, and R.L. Schult, Phys. Rev. B [**45**]{}, 14156 (1992). E.R. Davidson, Computers in Physics [**7**]{}, 519 (1993).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We examine the Bose-Hubbard model in the Penrose lattice based on inhomogeneous mean-field theory. Since averaged coordination number in the Penrose lattice is four, mean-field phase diagram consisting of the Mott insulator (MI) and superfluid (SF) phase is similar to that of the square lattice. However, the spatial distribution of Bose condensate in the SF phase is significantly different from uniform distribution in the square lattice. We find a fractal structure in its distribution near the MI-SF phase boundary. The emergence of the fractal structure is a consequence of cooperative effect between quasiperiodicity in the Penrose lattice and criticality at the phase transition.'
author:
- 'Rasoul Ghadimi, Takanori Sugimoto, Takami Tohyama'
bibliography:
- 'myreference.bib'
title: 'Mean-field study of the Bose-Hubbard model in Penrose lattice'
---
*Introduction*– Quasicrystals have aperiodic structure different from fully disordered one. Although translational symmetry is absent, the presence of sharp spots in Brag reflection indicates long-range order [@PhysRevLett.53.1951; @PhysRevLett.53.2477]. Quasicrystals can be realized even in bilayer graphene [@Yao6928] and photonic lattices [@doi:10.1063/1.4754136]. In addition to various characteristics due to aperiodicity [@Steurer:ib5056; @doi:10.1146annurev.matsci.38.060407.130318], recent new findings expand the field of quasicrystal to include superconductivity [@Kamiya2018], quantum criticality [@Deguchi2012], and topology [@PhysRevLett.108.220401; @PhysRevB.100.014510; @doi:10.7566/JPSJ.85.073712; @takemori2018intersite; @doi:10.7566/JPSJ.83.083707; @PhysRevLett.119.215304; @rasoul2020-3; @PhysRevB.100.081405; @PhysRevResearch.1.022002; @PhysRevLett.121.126401; @Jagannathan2012; @PhysRevB.79.172406; @PhysRevLett.111.185304; @PhysRevLett.120.060407]. In general, self-similarity in quasicrystals dictates fractal structure in wavefunction, phase diagram, and so on [@Rasoul2017; @PhysRevX.6.011016]. This characteristic is justified by the presence of the inflation and deflation rules to construct quasicrystals [@PhysRevB.39.9904].
One of the well-known two-dimensional (2D) quasicrystals is the so-called Penrose lattice [@DEBRUIJN198139]. One can construct the lattice using inflation, projection, or multi-grade rules. The Penrose lattice has been studied intensively [@PhysRevB.77.104427; @PhysRevX.6.011016; @PhysRevB.93.075141; @Takemori_2015] and its structure dictates thermodynamically degenerate states in energy spectrum [@PhysRevB.38.1621; @PhysRevB.37.2797]. The presence of degeneracy is similar to the Lieb-lattice and causes a singularity in the density of state [@PhysRevB.58.13482], being crucial for understanding antiferromagnetism at half-filling [@PhysRevB.96.214402; @PhysRevB.39.9904].
Ultracold gases in optical lattices provide us an ideal playground of strong correlation [@Bloch2005] and also quasicrystals [@PhysRevA.92.063426; @PhysRevLett.79.3363; @PhysRevA.72.053607; @Corcovilos:19], which allows us to investigate the interplay of strong correlation and aperiodicity. A typical strongly correlated system in optical lattice is the Bose-Hubbard model, where phase transition between Mott insulator (MI) to superfluid (SF) phase appears [@Dutta_2015] as experimentally observed [@PhysRevLett.81.3108; @Greiner2002]. Recent achievements in establishing an eight-fold rotationally symmetric optical lattice attract new attention [@PhysRevLett.122.110404], in connection with theoretical investigation of an extended Bose-Hubbard with quasicrystalline confined potential [@PhysRevA.100.053609], where spontaneous breaking of underlying eight-fold symmetry is observed. However, the effect of aperiodicity in the Bose-Hubbard model is not yet fully understood both theoretically and experimentally.
In this Letter, we investigate the phase diagram of the Bose-Hubbard model in the Penrose lattice. We use the inhomogeneous mean-field theory (IMFT) and find that the distribution of Bose condensate in the Penrose lattice exhibits a fractal structure near the MI-SF boundary. We attribute the appearance of the fractal structure to a consequence of the divergence of correlation length seen in any phase transition. Therefore, the fractal structure is a common signature of phase transition in aperiodic systems.
$\alpha$ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----- -----
$M_1$ 3 3 3 3 3 3 3 3 3 3 3 4 4 4 5 5 5 5 5 5 5 5 5 5 5 6 7
$m_1^{(3)}$ 0 0 0 0 0 0 0 0 0 0 0 2 2 2 5 5 5 4 4 4 4 4 3 2 0 3 6
$m_1^{(4)}$ 0 0 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 2 0 0 0
$m_1^{(5)}$ 3 3 3 1 1 1 2 2 2 2 2 2 2 2 0 0 0 1 1 1 1 0 0 0 5 3 1
$m_1^{(6)}$ 0 0 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0
$m_1^{(7)}$ 0 0 0 0 1 1 0 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
$M_2$ 15 15 15 15 16 16 16 17 17 17 17 16 16 16 15 15 15 17 17 17 17 18 19 21 25 24 23
$m_2^{(3)}$ 11 10 9 9 12 12 11 15 14 14 13 6 5 4 0 0 0 0 0 0 0 3 5 10 20 10 2
$m_2^{(4)}$ 2 3 4 0 0 0 0 0 0 1 1 4 5 6 0 0 0 2 2 2 0 2 0 0 0 4 4
$m_2^{(5)}$ 0 0 0 6 4 3 5 2 3 1 1 2 2 2 15 13 11 11 11 11 13 9 12 11 5 4 10
$m_2^{(6)}$ 2 1 0 0 0 1 0 0 0 1 2 4 2 0 0 0 0 4 2 0 0 0 0 0 0 6 0
$m_2^{(7)}$ 0 1 2 0 0 0 0 0 0 0 0 0 2 4 0 2 4 0 2 4 4 4 2 0 0 0 7
$M_3$ 53 55 57 57 56 57 58 55 57 57 60 68 71 74 75 79 83 87 89 91 93 90 89 85 85 102 121
\[tab\]
*Model and method*– The Hamiltonian of the single-band Bose-Hubbard model is defined by: $$\label{Bose-Hubbard}
H_{BH}=-J\sum_{<i,j>}({\hat{b}}^{\dagger}_i{\hat{b}}_j+{\hat{b}}^{\dagger}_j{\hat{b}}_i)-\mu\sum_i {\hat{n}}_i+\frac{U}{2}\sum_i {\hat{n}}_i({\hat{n}}_i-1),$$ where ${\hat{b}}_i$ and ${\hat{b}}^{\dagger}_i$ are annihilation and creation of bosons at site (vertex) $i$ and the number operator ${\hat{n}}_i={\hat{b}}^{\dagger}_i{\hat{b}}_i$. The summation $\langle i,j\rangle$ represents nearest-neighbor (NN) links in the Penrose lattice shown in Fig. \[fig:crystall\](a). $J$, $\mu$, and $U$ in Eq. (\[Bose-Hubbard\]) are the hopping energy of boson, the chemical potential, and on-site Coulomb interaction, respectively.
![(a) Part of Penrose lattice. The number and color in each vertex indicate the index $\alpha$ of vertexes given in the Table \[tab\]. (b) Perpendicular space of Penrose lattice for $\mathcal{Z}=1$, and (c) that for $\mathcal{Z}=2$. The number is the same as (a). Different colors in (b) and (c) distinguish different sections in perpendicular space. []{data-label="fig:crystall"}](fig1.png){width="\linewidth"}
Because of the presence of the hopping term in Eq. (\[Bose-Hubbard\]), the exact solution is inaccessible. Therefore, we use a mean-field technique and decouple the hopping term using local condensation amplitude $\langle {\hat{b}}_i\rangle$. The resulting mean-field Hamiltonian is given by $H_\mathrm{MF}=\sum_i H_i+E_0$ with $$\label{IMFT}
H_i=-J\left( \psi_i^* {\hat{b}}_i+H.c.\right)-\mu {\hat{n}}_i+\frac{U}{2} {\hat{n}}_i({\hat{n}}_i-1),$$ where $\psi_i=\sum_{j\in \text{NN.i}}\langle{\hat{b}}_j\rangle$ with summation over NN links connected to the vertex $i$ and $E_0=J\sum\psi^*_i\langle {\hat{b}}_i\rangle$.
In order to obtain a self-consistent solution of Eq. (\[IMFT\]) in the local Hilbert space containing maximally $n_\mathrm{b}$ bosons, we start with an initial $\psi_i$ and then calculate $\langle {\hat{n}}_i\rangle$ and $\langle {\hat{b}}_i\rangle$ using the ground-state wavefunction for each vertex. We continue updating $\psi_i$ until convergence of $\langle {\hat{n}}_i\rangle$ and $\langle {\hat{b}}_i\rangle$ is obtained within a certain tolerance ($10^{-9}$ in our case). This procedure gives rise to inhomogeneous distribution of $\langle {\hat{n}}_i\rangle$ and $\langle {\hat{b}}_i\rangle$ on the Penrose lattice. Therefore, we call this mean-field technique the IMFT. We take $n_\mathrm{b}=7$. Within IMFT, we generally find the MI and SF phases in the Bose-Hubbard model. In the MI phase, all sites have equal integer number of bosons and thus $\langle {\hat{b}}_i\rangle=0$. On the other hand, $\langle {\hat{b}}_i\rangle$ is nonzero for the SF phase. We note that this self-consistent procedure gives moderately accurate results as compared with quantum Monte Carlo simulations and gives equivalent results with variational Gutzwiller method [@PhysRevA.83.053608; @PhysRevA.86.013623; @Barman_2014; @PhysRevLett.98.260405; @PhysRevLett.75.4075; @PhysRevA.91.043632; @PhysRevB.86.054520; @PhysRevB.44.10328; @PhysRevB.45.3137].
In the Penrose lattice, open boundary condition breaks the symmetries of Penrose tiling. On the other hand, we can construct a supercell with periodic boundary conditions (PBC), where the local symmetry of Penrose lattice remains intact throughout. Although a few defects appear in the supercell, their effect is very small. This supercell called approximant can be found by approximating golden ratio with $\mathrm{F}(g+1)/\mathrm{F}(g)$, where $\mathrm{F}(g)$ is the $g$-th sequence of Fibonacci number. Here, we set $g=11$ and obtain a supercell with the total number of vertexes $N=4F(2g+1)+3F(2g)=167761$ [@entin1988penrose; @BABALIEVSKI199027; @PhysRevB.43.1378; @PhysRevB.43.8879; @oitmaa1990antiferromagnetic].
In the Penrose lattice, we can classify any vertexes in terms of their local environment. For this classification, we first find the number of NN links, $M_1$, i.e., the total number of paths using one link (the second row in Table \[tab\]), which is equivalent to coordination number for each vertex in Fig. \[fig:crystall\](a). $M_1$ changes from $3$ to $7$. This means that all of sites are indexed by five kinds of vertexes. Next, we count the number of NN vertexes having $l$ links, $m_1^{(l)}$, and make a list of them (the third-seventh rows in Table \[tab\]). From the list of $m_1^{(l)}$ together with $M_1$, we find fourteen types of configurations, meaning that all of sites are indexed by fourteen kinds of vertexes. The total number of paths using two links from a given vertex is then expressed as $M_2=\sum_{l=3}^7 m_1^{(l)} l$, which is listed in the eighth row. We repeat this listing for the vertexes accessed by using the two links form a given vertex, which is shown in the ninth-thirteenth rows as $m_2^{(l)}$. In the last row of Table \[tab\], the total number of paths using three links from a given vertex ($M_3=\sum_{l=3}^7 m_2^{(l)} l$) is listed. Performing this procedure for all vertexes in our supercell, we find that there are twenty-seven kinds of vertexes, by which almost the whole system is covered [^1]. They are indexed as $\alpha$ in the first row of Table \[tab\]. In this manner, we can classify all vertexes for a given $k$. We thus define the number of classes (NoC) determined by the given $k$. For example, NoC is equal to 5, 14, and 27 for $k=1$, 2, and 3, respectively. We can increase $k$ as many as possible, but we stop $k$ up to $k=3$ in Table \[tab\] in order to make underlying physics in our model more transparent. We find NoC $\propto k^{1.93}$ in the large $k$ region (see Fig. \[fig:MeanFieldComparison\](c)). We will come back to this point later.
![Phase diagram of the Bose-Hubbard model on the Penrose lattice obtained by IMFT. The white area with the shape of lobes corresponds to the MI phase with $n_0$ bosons in all vertexes, denoted by MI$n_0$ ($n_0=1,2,\cdots$). In the SF phase, the number of gap (NoG) for a threshold value of $10^{-7}$ defined in the text is plotted with color scale. The analytical MI-SF mean-field phase boundary for the square lattice in Eq. (\[Eq:SquareLattice\]) is plotted by the orange dashed curve.[]{data-label="fig:phaseDiagram"}](fig2.png){width="0.9\linewidth"}
![ (a) Averaged order parameters $\overline{b}_\alpha/\overline{b}$ and mean deviation $\delta b_\alpha/\overline{b}$ as a function of $\bar{z}J/U$ along the horizontal dotted line in Fig. \[fig:phaseDiagram\]. The colored curves represent $\overline{b}_\alpha/\overline{b}$. The bars centered at each curve represent $\delta b_\alpha/\overline{b}$. The color scheme is the same as Fig. \[fig:crystall\](a) and the number denoted at the right-hand side indicates each class $\alpha$ in Table \[tab\]. Inset shows averaged order parameter $\overline{b}$. (b) Log-log plot of the number of gaps (NoG) defined in the text as a function of $\bar{z}(J-J_c)/U$ along the horizontal dotted line in Fig. \[fig:phaseDiagram\]. The brown, purple, blue and red circles represent NoG for threshold values of $10^{-5}$, $5\times10^{-6}$, $10^{-6}$, $5\times10^{-7}$, respectively. The lines represent fitting function denoted by the corresponding color, where $x=\bar{z}(J-J_c)/U$. (c) Log-log plot of the number of classes (NoC) as a function of the number of links $k$. The blue line represents a fitting function shown in the figure.[]{data-label="fig:MeanFieldComparison"}](fig3.png){width="\linewidth"}
Vertexes in the Penrose lattice can be labeled with five integers, originated from cut and projection of five dimensional cubic lattice [@DEBRUIJN198153; @PhysRevB.96.214402]. One can construct original Penrose lattice by mapping those labels. However, using another mapping, one finds four different 2D structures, called perpendicular space, where we assign them $\mathcal{Z}=1$, 2, 3, and 4 \[see Figs. \[fig:crystall\](b) and \[fig:crystall\](c) for $\mathcal{Z}=1$ and 2, respectively\]. We can divide perpendicular space into symmetric sections, where each section represents vertex with similar local circumstances. Therefore, one notices the index $\alpha$ in Table \[tab\] mapped to different sections in the perpendicular space \[see Figs. \[fig:crystall\](b) and \[fig:crystall\](c)\]. Considering the bipartite properties of Penrose lattice, we realize that $\mathcal{Z}=1,3$ and $\mathcal{Z}=2,4$ are related to different subsystems, though the same $\alpha$ are shared each other.
{width="\linewidth"}
*Results*–We first examine the phase diagram of the Bose-Hubbard model on the Penrose lattice. Since there are two order parameters (per vertex) $\langle {\hat{n}}_i\rangle$ and $\langle {\hat{b}}_i\rangle$ in IMFT, we expect two phases: one is MI with $\langle {\hat{b}}_i\rangle=0$ and $\langle {\hat{n}}_i\rangle=n_0$ ($n_0=1,2,\cdots$, corresponding to bosonic occupation number at each vertex), and the other is SF with $\langle {\hat{b}}_i\rangle\neq 0$ and $\langle {\hat{n}}_i\rangle=0$ as is the case of the square lattice. Figure \[fig:phaseDiagram\] shows the phase diagram, where we find MI phases denoted by MI$n_0$ and SF. Since averaged coordination number in the Penrose lattice is $\bar{z}=4$, which is the same as the coordination number $z=4$ in the square lattice, the phase boundary between MI and SF is expected to be similar to that of the square lattice. This is the case as shown by the dashed orange curve along MI lobes in Fig. \[fig:phaseDiagram\], which is mean-field phase boundary for the square lattice given analytically [@PhysRevA.63.053601; @Freericks_1994] by $$\label{Eq:SquareLattice}
zJ_{c}/U=\frac{-\frac{\mu}{U}-(\frac{\mu}{U})^2+s+2\mu s-s^2}{1+\frac{\mu}{U}},$$ where $s=\text{round}(\mu/U+1/2)$. The similarity indicates small effect of aperiodicity on the phase boundary.
In the SF phase of square lattice, the condensation amplitude $\langle {\hat{b}}_i\rangle$ is uniform, i.e., independent of $i$, for any region in the phase diagram. On the other hand, nonuniform distribution of $\langle {\hat{b}}_i\rangle$ in the Penrose lattice is easily expected from the presence of different types of vertexes as discussed in Table \[tab\]. Then, an arising question is how its nonuniform distribution changes in the phase diagram. To see this, we define an $\alpha$ dependent average of $\langle {\hat{b}}_i\rangle$ as $\overline{b}_\alpha=N_\alpha^{-1}\sum_{i\in\alpha}\langle {\hat{b}}_i\rangle$, where $N_\alpha$ is the number of $\alpha$-type vertex in the whole lattice. This quantity can distinguish the twenty-seven classes of vertexes. However, each class should have further internal structure coming from possible extension of $M_k$ for $k\ge 4$. To recognize this structure, we also define a mean deviation of condensation amplitude distribution as $\delta b_\alpha=\sqrt{N_\alpha^{-1}\sum_{i\in\alpha}(\langle {\hat{b}}_i\rangle-\overline{b}_\alpha)^2}$. The larger $\delta b_{\alpha}/\overline{b}$ is, the deeper the internal structure is.
In Fig. \[fig:MeanFieldComparison\](a), we plot $\overline{b}_\alpha/\overline{b}$ and $\delta b_\alpha/\overline{b}$ as a function of $\bar{z}J/U$ along the horizontal dotted line in Fig. \[fig:phaseDiagram\], where $\overline{b}=\sum_i\langle {\hat{b}}_i\rangle/N$ with $N=\sum_\alpha N_\alpha$. We note that $\delta b_\alpha/\overline{b}$ is denoted by the length of bars for each $\overline{b}_\alpha/\overline{b}$. At large $\bar{z}J/U$ far from the phase boundary, $\overline{b}_\alpha/\overline{b}$ is tend to be grouped accompanied by negligibly small $\delta b_\alpha/\overline{b}$. In the limit of $\bar{z}J/U\rightarrow\infty$, $\overline{b}_\alpha/\overline{b}$ is grouped into five classes depending on $M_1$. This means that, if correlation effect is small, the coordination number controls physical properties as expected. On the other hand, with approaching $\bar{z}J/U$ to the phase boundary, the mean deviation $\delta b_\alpha/\overline{b}$ becomes large. This means that the number of distinct vertexes increases with approaching to the boundary. In other words, long-distant correlation becomes important in order to obtain critical behaviors near the phase transition.
In order to make critical behaviors visible, we introduce a new quantity that can characterize distinct number of vertexes more than 27 listed in Table \[tab\]. We use $\langle {\hat{b}}_i\rangle$ for this purpose. We i) make shifting and scaling for $\langle {\hat{b}}_i\rangle$ to be located within $[0,1]$, ii) sort the scaled $\langle {\hat{b}}_i\rangle$ from 0 to 1, iii) calculate the difference of $\langle {\hat{b}}_i\rangle$ between $i$ and $i+1$ from $i=1$ to $i=N-1$, and iv) count the number of the difference (gap) whose magnitude is more than a given small threshold value. We call this number the number of gap (NoG). For example, NoG is zero for the square lattice because $\langle {\hat{b}}_i\rangle$ is independent of $i$. In the Penrose lattice, we have four NoG in the large limit of $\bar{z}J/U$ since there are five distinct values of $\langle {\hat{b}}_i\rangle$. We show log-log plot of NoG in Fig. \[fig:MeanFieldComparison\](b) along the horizontal dotted line in Fig. \[fig:phaseDiagram\], where four different threshold values, $10^{-5}$ , $5\times10^{-6}$, $10^{-6}$, and $5\times10^{-7}$ are used. With approaching to the phase boundary at $\bar{z}J/U=0.084$, NoG increases, indicating the increase of distinct vertexes. Interesting is that, with decreasing the threshold value, NoG rapidly increases near the boundary and shows a diverging behavior with an approximate exponent around $-0.9$, i.e., NoG $\propto (J-J_c)^{-0.9}$. This resembles to a critical behavior toward continuous phase transition as suggested from the vanishing of averaged order parameter $\bar{b}$ \[see inset of Fig \[fig:MeanFieldComparison\](a)\].
In order to understand this diverging behavior more, we focus on the fact that the increase of NoG corresponds to the increase of distinct vertexes. The latter is measured by NoC, whose large region is proportional to $k^{-1.93}$ as shown in Fig. \[fig:MeanFieldComparison\](c). Therefore, diverging behavior in NoG is directly connected to diverging behavior in NoC at large $k$. Since $k$ represents the number of links from a given vertex, we may regard $k$ as a measure of correlation length $\xi$ from a given vertex. Based on this reasoning, we have NoG $\propto$ NoC $\propto k^{-1.93} \propto \xi^{-1.93}$. Since $\xi \propto (J-J_c)^{-0.5}$ for the mean-field phase transition, we finally expect that NoG $\propto (J-J_c)^{-0.96}$, whose exponent is not far from the calculated one in NoG, $-0.9$. This indicates that diverging behavior in NoG is a consequence of criticality in the mean-field phase transition. We note that this critical behavior does not appear if $\mu/U=n_0$ and $J/U\rightarrow 0$.
Usefulness of perpendicular space presentation has already been found in considering magnetism on the Penrose lattice [@PhysRevB.96.214402; @PhysRevB.77.104427]. Therefore, we show the perpendicular space representation of $\langle {\hat{b}}_i\rangle$ in Fig. \[fig:PrependicularSpace\] for two sets of parameters at the end of the red dashed line in Fig. \[fig:phaseDiagram\]. We recognize notable differences in the two cases. For the parameter far from the phase boundary, we find fourteen distinct sections in Figs. \[fig:PrependicularSpace\](c) and \[fig:PrependicularSpace\](d). The number corresponds to the number of distinct vertexes obtained by taking into account $M_1$ and $m_1^{(l)}$ as discussed above. On the other hand, for the parameter close to the phase boundary, we can see a fractal structure in Figs. \[fig:PrependicularSpace\](a) and \[fig:PrependicularSpace\](b). For example, we find a various size of star structure inside stars. We can understand the emergence of the fractal structure near the phase transition as follows. Because of diverging behavior in NoG near the MI-SF phase boundary, all distances become relevant. We have found from the previous discussion that tracing far distant links by increasing $k$ enhances NoC dramatically. Therefore we can expect further distinguishable sections in the perpendicular space, resulting in fractal nature. In other words, a combination of criticality leading to phase transition and aperiodicity is a key for emergence of fractal structure.
*Conclusion*–\[conclusion\] We have obtained mean-field phase diagram in the Penrose-Bose-Hubbard model. We have found that the Penrose lattice does not change the MI-SF boundary drastically in comparison with square lattice, because of the same averaged coordination number. However, the spatial distribution of Bose condensate is unequal, and indeed fractal structure appears in the perpendicular representation of condensation amplitude near the MI-SF phase transition. This is a consequence of the cooperative effect of criticality leading to phase transition and quasiperiodicity, which is expected to be a common feature in aperiodic strongly correlated systems.
This work was supported by Challenging Research Exploratory (Grant No. JP17K18764), Grant-in-Aid for Scientific Research on Innovative Areas (Grant No. JP19H05821).
[^1]: Vertexes around defects caused by PBC introduce other types. However, we can ignore them since the density of the defects is around 0.0001 for the total vertexes.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study categories of $d$-dimensional cobordisms from the perspective of [@Tillmann] and [@GMTW]. There is a category $\mathcal{C}_\theta$ of closed smooth $(d-1)$-manifolds and smooth $d$-dimensional cobordisms, equipped with generalised orientations specified by a fibration $\theta : {\mathbf{X}}\to BO(d)$. The main result of [@GMTW] is a determination of the homotopy type of the classifying space $B\mathcal{C}_\theta$. The goal of the present paper is a systematic investigation of subcategories $\mathcal{D} \subseteq \mathcal{C}_\theta$ with the property that $B\mathcal{D} \simeq B\mathcal{C}_\theta$, the smaller such $\mathcal{D}$ the better.
We prove that in most cases of interest, $\mathcal{D}$ can be chosen to be a homotopy commutative monoid. As a consequence we prove that the stable cohomology of many moduli spaces of surfaces with $\theta$-structure is the cohomology of the infinite loop space of a certain Thom spectrum ${\bold{MT \theta}}$. This was known for certain special $\theta$, using homological stability results; our work is independent of such results and covers many more cases.
address:
- |
Department of Mathematics\
Stanford University\
Stanford CA, 94305
- |
Mathematical Institute\
24-29 St Giles’\
Oxford\
OX1 3LB\
United Kingdom
author:
- 'S[ø]{}ren Galatius'
- 'Oscar Randal-Williams'
bibliography:
- 'MMM.bib'
title: Monoids of moduli spaces of manifolds
---
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Ian Moult,'
- 'Lina Necib,'
- and Jesse Thaler
bibliography:
- 'ECFVariant.bib'
title: New Angles on Energy Correlation Functions
---
Introduction {#sec:intro}
============
With the Large Hadron Collider (LHC) rapidly acquiring data at a center-of-mass energy of $13$ TeV, jet substructure observables are playing a central role in a large number of analyses, from Standard Model measurements [@Chatrchyan:2012sn; @CMS:2013cda; @Aad:2015cua; @Aad:2015lxa; @ATLAS-CONF-2015-035; @Aad:2015rpa; @Aad:2015hna; @ATLAS-CONF-2016-002; @ATLAS-CONF-2016-039; @ATLAS-CONF-2016-034; @CMS-PAS-TOP-16-013; @CMS-PAS-HIG-16-004] to searches for new physics [@CMS:2011bqa; @Fleischmann:2013woa; @Pilot:2013bla; @TheATLAScollaboration:2013qia; @Chatrchyan:2012ku; @CMS-PAS-B2G-14-001; @CMS-PAS-B2G-14-002; @Khachatryan:2015axa; @Khachatryan:2015bma; @Aad:2015owa; @Aaboud:2016okv; @Aaboud:2016trl; @Aaboud:2016qgg; @ATLAS-CONF-2016-055; @ATLAS-CONF-2015-071; @ATLAS-CONF-2015-068; @CMS-PAS-EXO-16-037; @CMS-PAS-EXO-16-040; @Khachatryan:2016mdm; @CMS-PAS-HIG-16-016; @CMS-PAS-B2G-15-003; @CMS-PAS-EXO-16-017].[^1] As the field of jet substructure matures [@Abdesselam:2010pt; @Altheimer:2012mn; @Altheimer:2013yza; @Adams:2015hiv], observables are being designed for increasingly specific purposes, using a broader set of criteria to evaluate their performance beyond simply raw discrimination power. Continued progress relies on achieving a deeper understanding of the QCD dynamics of jets, allowing for more subtle features within a jet to be exploited. This understanding has progressed rapidly in recent years, due both to advances in explicit calculations of jet substructure observables [@Feige:2012vc; @Field:2012rw; @Dasgupta:2013ihk; @Dasgupta:2013via; @Larkoski:2014pca; @Dasgupta:2015yua; @Seymour:1997kj; @Li:2011hy; @Larkoski:2012eh; @Jankowiak:2012na; @Chien:2014nsa; @Chien:2014zna; @Isaacson:2015fra; @Krohn:2012fg; @Waalewijn:2012sv; @Larkoski:2014tva; @Procura:2014cba; @Bertolini:2015pka; @Bhattacherjee:2015psa; @Larkoski:2015kga; @Dasgupta:2015lxh; @Frye:2016okc; @Frye:2016aiz; @Kang:2016ehg; @Hornig:2016ahz] as well as to the development of techniques for understanding the dominant properties of substructure observables using analytic [@Walsh:2011fz; @Larkoski:2014gra; @Larkoski:2014zma] and machine learning [@Cogan:2014oua; @deOliveira:2015xxd; @Almeida:2015jua; @Baldi:2016fql; @Guest:2016iqz; @Conway:2016caq; @Barnard:2016qma] approaches.
A particularly powerful method for constructing jet substructure observables is power counting, introduced in . Given a basis of infrared and collinear (IRC) safe observables, power counting can identify which combinations are optimally sensitive to specific parametric features within a jet.[^2] Furthermore, power counting elucidates the underlying physics probed by the observable. This approach was successfully applied to the energy correlation functions [@Larkoski:2013eya], leading to a powerful 2-prong discriminant called $D_2$ [@Larkoski:2014gra]. Vital to the power counting approach, though, is a sufficiently flexible basis of IRC safe observables to allow the construction of discriminants with specific properties.
In this paper, we exploit the known properties of IRC safe observables to systematically identify a useful basis for jet substructure, which we call the generalized energy correlation functions. These observables—denoted by ${{_{v}e_{n}^{(\beta)}}}$ and defined in —are an extension of the original energy correlation functions with a more flexible angular weighting.[^3] Specially, these new observables correlate $v$ pairwise angles among $n$ particles, whereas the original correlators were restricted to $v$ equaling $n$ choose 2. Using these generalized correlators, we apply power counting to identify new jet substructure observables for each of the major jet substructure applications at the LHC: 3-prong boosted top tagging, 2-prong boosted $W/Z/H$ tagging, and 1-prong quark/gluon discrimination. In each case, our new observables exhibit improved performance over traditional observables when tested with parton shower generators.
The flexibility of our basis, combined with insights from power counting, allows us to tailor our observables for specific purposes, beyond those that have been previously considered. As an interesting example, we are able to specifically design observables for use on groomed jets [@Butterworth:2008iy; @Ellis:2009su; @Ellis:2009me; @Krohn:2009th; @Dasgupta:2013via; @Dasgupta:2013ihk]. While grooming procedures are heavily used at the LHC to remove jet contamination from initial state radiation, underlying event, and pileup, most LHC analyses apply observables that were designed for use on ungroomed jets. Here, by understanding the impact of grooming on soft radiation, we introduce a 2-prong discriminant, ${M_{2}}$, which exhibits almost no discrimination power on ungroomed jets, but outperforms traditional observables when measured on groomed jets. This observable therefore acts both as a probe of the grooming procedure and as a powerful discriminant. We also show how the use of groomed observables leads to remarkably stable distributions as a function of the jet mass and $p_T$, even for distributions that are unstable before grooming, such as $D_2$. This has recently been emphasized as a desirable feature for substructure observables, particularly to facilitate sideband calibration and produce smooth mass distributions for backgrounds [@Dolen:2016kst]; observables modified to achieve stability have been used by both ATLAS and CMS [@ATL-PHYS-PUB-2015-033; @CMS-PAS-EXO-16-030].
The generalized energy correlation functions allow us to introduce a wide variety of new substructure observables, though we focus on three series with particularly nice properties. The first is the ${M_{i}}$ series, defined via the ratio $$\begin{aligned}
{M_{i}^{(\beta)}}=\frac{{{_{1}e_{i+1}^{(\beta)}}}}{{{_{1}e_{i}^{(\beta)}}}}\,.\end{aligned}$$ These observables identify jets with $i$ hard prongs, but, as mentioned above, are only effective for discrimination on suitably groomed jets. The second is the ${N_{i}}$ series, defined via the ratio $$N_i^{(\beta)} = \frac{{{_{2}e_{i+1}^{(\beta)}}}}{({{_{1}e_{i}^{(\beta)}}})^2}\,,$$ which are designed to mimic the behavior of the $N$-subjettiness ratio ${\tau_{i,i-1}}$ [@Thaler:2010tr; @Thaler:2011gf]. The ${N_{i}}$ observables are defined without respect to subjet axes, and therefore exhibit improved behavior compared to $N$-subjettiness, particularly in the transition to the unresolved region, where the definition of subjet axes becomes ambiguous. The third is the ${U_{i}}$ series, defined as $$\begin{aligned}
{U_{i}^{(\beta)}}={{_{1}e_{i+1}^{(\beta)}}}\,,\end{aligned}$$ which probe multiple emissions within 1-prong jets and can be used to improve quark/gluon discrimination. In all cases, the parameter $\beta$ controls the overall angular scaling of these observables, and the ${}^{(\beta)}$ superscript will often be dropped when clear from context.
To guide the reader, we summarize the particular applications studied in this paper, so that the (un)interested reader can skip to the relevant section. These observables will be made available in the `EnergyCorrelator` [<span style="font-variant:small-caps;">FastJetcontrib</span>]{} [@Cacciari:2011ma; @fjcontrib] starting in version 1.2.0.
- **Boosted Top Tagging** ():
- ${N_{3}}$: An axes-free observable which reduces to the $N$-subjettiness ratio ${\tau_{3,2}}$ in the resolved limit, but exhibits improved performance in the unresolved limit on groomed jets.
- **Boosted $W/Z/H$ tagging** ():
- ${M_{2}}$: A 2-prong discriminant specifically designed for use on groomed jets.
- ${N_{2}}$: An axes-free observable which reduces to the $N$-subjetttiness ratio ${\tau_{2,1}}$ in the resolved limit, but exhibits improved performance on both groomed and ungroomed jets.
- ${D_{2}^{(\alpha,\beta)}}$: A generalization of the standard $D_2$ observable [@Larkoski:2014gra] specifically designed for groomed jets, which exhibits improved performance when $\alpha=1$, $\beta=2$.
- **Quark/Gluon Discrimination** ():
- ${U_{i}}$: A new series of observables for quark/gluon discrimination which probes the structure of multiple soft gluon emissions from the hard jet core, leading to improved performance over the standard ${C_{1}}$ observable [@Larkoski:2013eya].
The specific form of these observables, and the origin of their discrimination power, will be analyzed using power counting. We verify all power-counting predictions using parton shower generators and compare the performance of our newly introduced observables to traditional observables for each of the above applications.
The remainder of this paper is organized as follows. In , we review standard substructure and grooming techniques as well as the power counting approach for understanding soft and collinear scaling. In , we discuss the general structure of IRC safe observables and introduce the generalized energy correlation functions, ${{_{v}e_{n}}}$, as well as the $M_i$, $N_i$, and $U_i$ series. The three key case studies bulleted above appear in . We conclude in and discuss possible future directions for improving our understanding of jet substructure at the LHC.
Review of Substructure Approaches {#sec:review}
=================================
In this section, we review a number of standard jet substructure techniques that will be used throughout this paper. We begin in by defining the energy correlation functions [@Larkoski:2013eya] and $N$-subjettiness ratios [@Thaler:2010tr; @Thaler:2011gf], both of which are widely used in jet substructure. In , we review the soft drop/modified mass drop [@Larkoski:2014wba; @Dasgupta:2013ihk; @Dasgupta:2013via] algorithm, which we use as our default grooming procedure. Finally in , using the 2-point energy correlation function as an example, we review the power-counting approach for analyzing jet substructure observables, which features heavily in later discussions. Readers familiar with these topics can safely skip to , though we recommend reviewing the logic of .
Energy Correlation Functions and $N$-subjettiness {#sec:ecfs}
-------------------------------------------------
The energy correlation functions [@Larkoski:2013eya] are a convenient basis of observables for probing multi-prong substructure within a jet. In this paper, we use the 2-, 3-, and 4-point energy correlation functions, defined as[^4] $$\begin{aligned}
\label{eq:ECFs}
{{e_{2}^{(\beta)}}} &= \sum_{1\leq i<j\leq n_J} z_i z_j \, \theta_{ij}^\beta \, ,\nonumber \\
{{e_{3}^{(\beta)}}} &= \sum_{1\leq i<j<k\leq n_J} z_i z_{j}z_{k} \, \theta_{ij}^\beta \theta_{ik}^\beta \theta_{jk}^\beta \, , \nonumber \\
{{e_{4}^{(\beta)}}} &=\sum_{1\leq i<j<k<\ell\leq n_J} z_{i}z_{j}z_{k}z_{\ell} \, \theta_{ij}^\beta \theta_{ik}^\beta \theta_{jk}^\beta \theta_{i\ell}^\beta \theta_{j\ell}^\beta \theta_{k\ell}^\beta \,,\end{aligned}$$ where $n_J$ is the number of particles in the jet. The generalization to higher-point correlators is straightforward, though we will not use them here. For simplicity, we often drop the explicit angular exponent $\beta$, writing the observable as ${e_{n}}$. This simplified notation will also be used for other observables introduced in the text.
It is convenient to work with dimensionless observables, written in terms of a generic energy fraction variable, $z$, and a generic angular variable, $\theta$. The precise definitions of the energy fraction and angle can be chosen depending on context and do not affect our power-counting arguments. For the case of $pp$ collisions at the LHC, which is the focus of our later studies, we work with longitudinally boost-invariant variables, $$\begin{aligned}
\label{eq:ptratio}
z_i\equiv\frac{p_{Ti}}{\sum_{j \in \text{jet}} p_{Tj}}\,, \qquad \theta_{ij}^2\equiv R_{ij}^2 = (\phi_i-\phi_j)^2+(y_i-y_j)^2\,,\end{aligned}$$ where $p_{Ti}$, $\phi_i$, and $y_i$ are the transverse momentum, azimuthal angle, and rapidity of particle $i$, respectively. Two other measures intended for $e^+e^-$ collisions are available in the `EnergyCorrelator` [<span style="font-variant:small-caps;">FastJetcontrib</span>]{} [@Cacciari:2011ma; @fjcontrib]. The first is a definition based strictly on energies and opening angles, $$\begin{aligned}
\label{eq:Eratio}
z_i\equiv\frac{E_{i}}{E_{J}}\,, \qquad \theta_{ij}^2\equiv \Theta_{ij}^2 \,,\end{aligned}$$ where $E_J$ is the total jet energy, and $\Theta_{ij}$ is the Euclidean angle between the 3-momenta $\vec p_i$ and $\vec p_j$. There is an alternative definition in terms of energies and Mandelstam invariants, $$\begin{aligned}
\label{eq:mandelstamratio}
z_i\equiv\frac{E_{i}}{E_{J}}\,, \qquad \theta_{ij}^2\equiv\frac{2p_i \cdot p_j}{E_i E_j} \,,\end{aligned}$$ which reduces to in the collinear limit but is easier for analytic calculations.
From , we see that the $n$-point energy correlation functions vanish in the soft and collinear limits, and therefore are natural resolution variables for $(n-1)$-prong substructure. A number of powerful 2-prong discriminants have been formed from the energy correlation functions [@Larkoski:2013eya; @Larkoski:2014gra], namely $$\begin{aligned}
{C_{2}^{(\beta)}}=\frac{{{e_{3}^{(\beta)}}}}{({{e_{2}^{(\beta)}}})^{2}}\,,\qquad {D_{2}^{(\beta)}}=\frac{{{e_{3}^{(\beta)}}}}{({{e_{2}^{(\beta)}}})^{3}}\,, \qquad {D_{2}^{(\alpha,\beta)}}=\frac{{{e_{3}^{(\alpha)}}}}{({{e_{2}^{(\beta)}}})^{3\alpha/\beta}}\,.\end{aligned}$$ Beyond their discrimination power, these observables have nice analytic properties. First, since they can be written as a sum over particles in the jet without reference to external axes, they are automatically “recoil-free” [@Catani:1992jc; @Dokshitzer:1998kz; @Banfi:2004yd; @Larkoski:2013eya; @Larkoski:2014uqa]. Second, since they have well-defined behavior in various soft and collinear limits, they are amenable to resummed calculations; in , ${D_{2}}$ was calculated to next-to-leading-logarithmic (NLL) accuracy in $e^+e^-$ for both signal (boosted $Z$) and background (QCD) jets.
The basic structure of the ${e_{2}}$, ${e_{3}}$ phase space is shown in and discussed in more detail in . Signal jets which have resolved 2-prong structure live in the region of phase space satisfying ${e_{3}}\ll ({e_{2}})^3$, whereas QCD background jets with 1-prong structure live in the phase space region defined by $({e_{2}})^3 \ll{e_{3}}\ll ({e_{2}})^2$. The observable ${D_{2}}$ is designed to define contours which cleanly separate the 1-prong and 2-prong regions of phase space, and therefore identifies the extent to which a jet is 1- or 2-prong-like.
Observables for boosted top tagging have also been proposed using the energy correlation functions, namely the $C_3$ observable [@Larkoski:2013eya], $$\begin{aligned}
\label{eq:C3_def}
{C_{3}^{(\beta)}}=\frac{ {{e_{4}^{(\beta)}}} {{e_{2}^{(\beta)}}} }{ ( {{e_{3}^{(\beta)}}} )^2 }\,,\end{aligned}$$ and the $D_3$ observable [@Larkoski:2014zma], $$\begin{aligned}
\label{eq:D3_full_def}
D_3^{(\alpha,\beta,\gamma)}= \frac{ {{e_{4}^{(\gamma)}}} \left ({{{e_{2}^{(\alpha)}}}}\right)^{\frac{3\gamma}{\alpha}} }{ \left( {{e_{3}^{(\beta)}}}\right )^{\frac{3\gamma}{\beta}} } +x \frac{ {{e_{4}^{(\gamma)}}} \left ({{e_{2}^{(\alpha)}}}\right)^{\frac{2\gamma}{\beta}-1} }{ \left ({{e_{3}^{(\beta)}}}\right )^{\frac{2\gamma}{\beta}} } +y \frac{ {{e_{4}^{(\gamma)}}} \left ({{e_{2}^{(\alpha)}}}\right)^{\frac{2\beta}{\alpha}-\frac{\gamma}{\alpha}} }{ \left ({{e_{3}^{(\beta)}}}\right )^{2} }\,.\end{aligned}$$ Here, $x$ and $y$ are constants given in that depend on the jet mass and $p_T$. The ${C_{3}}$ observable does not exhibit particularly good discrimination power, and while ${D_{3}}$, which was constructed using the power counting approach, performs well, it has a complicated functional form. For the boosted top study in , we compare to a simplified version of the ${D_{3}}$ observable $$\begin{aligned}
\label{eq:D3_simp_def}
D_3^{(\beta)}= \frac{ {{e_{4}^{(\beta)}}} \left ({{{e_{2}^{(\beta)}}}}\right)^{3} }{ \left( {{e_{3}^{(\beta)}}}\right )^{3} } \,,\end{aligned}$$ obtained by setting $x=y=0$, which behaves well on groomed jets. Unlike its more complicated cousin, this simplified $D_3$ has only a single angular exponent.
We also find it interesting to compare our new observables to $N$-subjettiness. The (normalized) $N$-subjettiness observable ${\tau_{N}}$ [@Thaler:2010tr; @Thaler:2011gf] is defined as[^5] $$\label{eq:nsubdef}
{\tau_{N}^{(\beta)}} = \sum_{1\leq i \leq n_J} z_{i}\min\left\{
\theta_{i1}^\beta,\dotsc,\theta_{iN}^\beta
\right\} \ .$$ Here, the angle $\theta_{iK}$ is measured between particle $i$ and subjet axis $K$ in the jet. As for the case of the energy correlation functions, a number of different possible measures can be used to define $\theta_{iK}$. For our LHC studies, we take $\theta_{iK}=R_{iK}$, analogously to .
Unlike the energy correlation functions of , which correlate groups of $n$ particles within the jet, $N$-subjettiness divides a jet into $N$ sectors and correlates the particles in each sector with their corresponding axis. Thus, implicit in the definition of $N$-subjettiness in is the definition of appropriate $N$-subjettiness axes. Different definitions of the axes can lead to different behaviors of the observable, particularly away from the resolved limit [@Larkoski:2015uaa]. A natural definition is to choose the axes that minimize the value of ${\tau_{N}}$ itself [@Thaler:2011gf], as is done for the classic $e^+e^-$ event shape thrust [@Farhi:1977sg]. Exact minimization is computationally challenging, though, so a number of definitions which approximate the minimum are used instead, which are provided in the `Nsubjettiness` [<span style="font-variant:small-caps;">FastJetcontrib</span>]{} [@Cacciari:2011ma; @fjcontrib].
The relevant $N$-subjettiness ratio observables are $$\begin{aligned}
{\tau_{2,1}^{(\beta)}}= \frac{{\tau_{2}^{(\beta)}}}{{\tau_{1}^{(\beta)}}}, \qquad {\tau_{3,2}^{(\beta)}}= \frac{{\tau_{3}^{(\beta)}}}{{\tau_{2}^{(\beta)}}} \ .\end{aligned}$$ Here, ${\tau_{2,1}}$ is designed to be small when a jet has well-resolved 2-prong substructure, making it useful for boosted $W$/$Z$/$H$ tagging. Similarly, ${\tau_{3,2}}$ is designed to be small in the 3-prong limit, useful for boosted tops. The observable ${\tau_{2,1}}$ was calculated in $e^+e^-$ collisions for signal (boosted $Z$) jets at N$^3$LL accuracy [@Feige:2012vc].
The phase space for ${\tau_{1}}$, ${\tau_{2}}$ is shown schematically in , along with contours of constant ${\tau_{2,1}}$. Background QCD jets are defined by the linear scaling ${\tau_{2}}\sim {\tau_{1}}$, whereas signal jets are defined by ${\tau_{2}}\ll {\tau_{1}}$. This phase space structure is different from that of the $e_2$ and $e_3$ observables shown in , where the phase space for background QCD jets is defined by two boundaries with distinct scalings. It is this fact which ultimately leads to many of the differences seen between ${D_{2}}$ and ${\tau_{2,1}}$, including the fact that the ${\tau_{2,1}}$ distribution is more stable as a function of jet mass and $p_T$. The phase space for ${\tau_{3,2}}$ is similar to ${\tau_{2,1}}$, to be contrasted with the complicated phase space for ${D_{3}}$ [@Larkoski:2014zma]. Using the generalized energy correlation functions, we can define new axes-free observables that mirror the phase space structures of ${\tau_{2,1}}$ and ${\tau_{3,2}}$, thereby exhibiting similar scaling and stability behaviors, particularly for groomed jets. This will be discussed for ${\tau_{3,2}}$ in and for ${\tau_{2,1}}$ in .
Soft Drop Grooming {#sec:soft_drop}
------------------
Two powerful tools which have emerged from the study of jet substructure are groomers [@Butterworth:2008iy; @Ellis:2009su; @Ellis:2009me; @Krohn:2009th; @Dasgupta:2013via; @Dasgupta:2013ihk] and pileup mitigation techniques [@Cacciari:2007fd; @Alon:2011xb; @Soyez:2012hv; @Tseng:2013dva; @Krohn:2013lba; @Cacciari:2014gra; @Bertolini:2014bba], both of which remove soft radiation from a jet. Groomers have proven to be useful both for removing jet contamination as well as for identifying hard multi-prong substructure within a jet. In this paper, we use the soft drop [@Larkoski:2014wba] groomer with $\beta=0$, which coincides with the modified mass drop procedure [@Dasgupta:2013ihk; @Dasgupta:2013via] with $\mu = 1$. The soft drop groomer exhibits several theoretical advantages over other groomers; in particular, it removes non-global logarithms [@Dasgupta:2001sh] to all orders, and it mitigates the process dependence of jet spectra. The soft-dropped groomed jet mass has recently been calculated to NNLL accuracy [@Frye:2016okc; @Frye:2016aiz].
Starting from a jet identified with an IRC safe jet algorithm (such as anti-$k_t$ [@Cacciari:2008gp]), the soft drop algorithm is defined using Cambridge/Aachen (C/A) reclustering [@Dokshitzer:1997in; @Wobisch:1998wt; @Wobisch:2000dk]. Specializing to the case of $\beta=0$, the algorithm proceeds as follows:
1. Recluster the jet using the C/A clustering algorithm, producing an angular-ordered branching history for the jet.
2. Step through the branching history of the reclustered jet. At each step, check the soft drop condition $$\begin{aligned}
\label{eq:sd_cut}
\frac{\min\left[ p_{Ti}, p_{Tj} \right]}{p_{Ti}+p_{Tj}}> \zcut \,.\end{aligned}$$ Here, $\zcut$ is a parameter defining the scale below which soft radiation is removed. If the soft drop condition is not satisfied, then the softer of the two branches is removed from the jet. This process is then iterated on the harder branch.
3. The soft drop procedure terminates once the soft drop condition is satisfied.
Given a jet that has been groomed with the soft drop procedure, we can then measure any IRC safe observable on this jet and it will remain IRC safe. As we will see, because soft drop removes soft radiation from a jet, power-counting arguments for groomed jets can be dramatically different than those for ungroomed jets. This is previewed in , where the phase space for ${D_{2}}$ is substantially modified by the removal of soft radiation.
More general groomers are expected to give rise to similar power-counting modifications. For example, the soft drop condition in can be generalized to include an angular weighting exponent $\beta$, which controls the aggressiveness of the groomer, and we expect deviations away from our default of $\beta = 0$ to yield similar behavior, so long as the groomer continues to remove parametrically soft particles. We also expect that other groomers such as trimming [@Krohn:2009th], which is used heavily by the ATLAS experiment, will behave similarly for the same value of $\zcut$. We leave a detailed study of other groomers to future work.
Power Counting the Soft/Collinear Behavior {#sec:power_counting}
------------------------------------------
An efficient approach for studying jet substructure is power counting [@Larkoski:2014gra], which allows one to determine the parametric scaling of observables. This parametric behavior is determined by the soft and collinear limits of QCD and is robust to hadronization or modeling in parton shower generators. Here, we briefly review the salient features of power counting, using the 2-point energy correlator as an example. We refer readers interested in a more detailed discussion to the original paper.
High-energy QCD jets are dominated by soft and collinear radiation, a language which will be used frequently throughout this paper. Since QCD is approximately conformal, there is no intrinsic energy or angular scale associated with this radiation.[^6] By applying a measurement to a jet, though, one introduces a scale, which then determines the scaling of soft and collinear radiation. The simple observation that all scales are set by the measurement itself allows for a powerful understanding of the jet’s energy and angular structure. Arguments along these lines are ubiquitous in the effective field theory (EFT) community. For example, in Soft Collinear Effective Theory (SCET) [@Bauer:2000ew; @Bauer:2000yr; @Bauer:2001ct; @Bauer:2001yt], they are used to identify the appropriate EFT modes required to describe a particular set of measurements.
In the context of power counting, soft and collinear emissions are defined by their parametric scalings. A soft emission, denoted by $s$, is defined by $$z_s \ll1 \ , \qquad \theta_{sx}\sim 1 \,.$$ Here, $z_s$ is the momentum fraction, as defined in , and $\theta_{sx}$ is the angle to any other particle $x$ in the jet, including other soft particles. The scaling $\theta_{sx}\sim 1$ means that $\theta_{sx}$ is not assigned any parametric scaling associated with the measurement. A collinear emission, denoted by $c$, is defined by $$z_c\sim1 \ , \qquad \theta_{cc}\ll 1\,, \qquad \theta_{cs}\sim 1\, .$$ Here, $\theta_{cc}$ is the angle between two collinear particles, while $\theta_{cs}$ is the angle between a collinear particle and a soft particle. In an EFT context, overlaps between soft and collinear regions are systematically removed using the zero-bin procedure [@Manohar:2006nz], but this is not relevant for the arguments here. The soft and collinear modes are illustrated in and their scalings are summaried in .
$\qquad$
We now use the simple example of ${e_{2}}$ to demonstrate how an applied measurement sets the scaling of soft and collinear radiation.[^7] The analysis of more general observables proceeds analogously. Repeating for convenience, the 2-point energy correlation function is $$\label{eq:2pt_ex}
{{e_{2}^{(\beta)}}} = \sum_{1\leq i<j\leq n_J} z_i z_j \, \theta_{ij}^\beta\,.$$ If we only consider regions of phase space where ${e_{2}}\ll1$, such that we have a well-defined collimated jet, all particles in the jet either have small $z_i$ or small $\theta_{ij}$. In this phase space region, the observable is indeed dominated by soft and collinear emissions.
To determine the scaling of $z_s$ and $\theta_{cc}$ in terms of the observable, we can consider the different possible contributions to ${e_{2}}$: soft-soft correlations, soft-collinear correlations, and collinear-collinear correlations. Parametrically, ${e_{2}}$ can therefore be written as $${{e_{2}^{(\beta)}}} \sim \sum_{s} z_{s}z_{s} \, \theta_{ss}^\beta + \sum_{s,c} z_{s}z_{c} \, \theta_{cs}^\beta + \sum_{c} z_{c}z_{c} \, \theta_{cc}^\beta \,.$$ Expanding this result to leading order in $z_s$ and $\theta_{cc}$, we find $${{e_{2}^{(\beta)}}} \sim \sum_{s} z_s +\sum_{c} \theta_{cc}^\beta \,.$$ For simplicity, we drop the summation symbol, writing $$\label{eq:fact_e2}
{{e_{2}^{(\beta)}}} \sim z_s + \theta_{cc}^\beta \,.$$
Since we have only measured a single observable, ${e_{2}}$, it sets the only scale in the jet, and there is no measurement to further distinguish the scalings of soft and collinear particles. We therefore find the scaling of $z_s$ and $\theta_{cc}$ in terms of the observable, $$z_s \sim {{e_{2}^{(\beta)}}}\,, \qquad \theta_{cc} \sim \left({{e_{2}^{(\beta)}}}\right)^{1/\beta} \,.$$ More generally, after identifying all parametrically different modes that can contribute to a set of measurements, the scaling of those modes is determined by the measured observables.
In this paper, we are interested not only in jets with soft and collinear radiation, but also in jets which have well-resolved substructure. In addition to the strictly soft and collinear modes which are found in , a jet with well-resolved substructure also includes radiation emitted from the dipoles within the jet, shown in orange for the particular case of a 2-prong jet in . This radiation is referred to as “collinear-soft" (or just “c-soft”) as it has a characteristic angle $\theta_{12}$ defined by the opening angle of the subjets, as well as a momentum fraction $z_{cs}\ll1$, both of which are set by the measurement. The appropriate EFT description for multi-prong substructure is referred to as SCET$_+$ [@Bauer:2011uc; @Procura:2014cba; @Larkoski:2015zka; @Pietrulewicz:2016nwo], and the scaling of the collinear-soft mode is summarized in . Using the mode structure of multi-prong jets, it is straightforward to apply power-counting arguments to a wide variety of $n$-prong jet substructure observables, as demonstrated in .
We also apply power-counting arguments to groomed jets after soft drop has been applied. The effect of the grooming algorithm is not just to remove jet contamination, but also to modify the power counting in interesting, and potentially useful, ways. As discussed in , soft drop with $\beta=0$ is defined with a single parameter $\zcut$, which determines the scale below which soft radiation is removed. To perform a proper power-counting analysis, one should also incorporate the scale $\zcut$ and consider different cases depending on the relative scaling of $\zcut$ and $z_s$. For simplicity, we ignore this complication through most of this paper and assume that the soft drop procedure simply removes the soft modes. That said, the residual soft scaling will matter for the quark/gluon study in . For a more detailed discussion, and a proper treatment of the scale $\zcut$ involving collinear-soft modes, see .
Enlarging the Basis of Jet Substructure Observables {#sec:defn}
===================================================
An important goal of jet substructure is to design observables that efficiently identify particular features within a jet. A popular, and theoretically well-motivated, approach is to construct observables from combinations, often ratios, of IRC safe jet shapes.[^8] Such observables are widely employed at the LHC, and have proven to be both experimentally useful and theoretically tractable. Indeed, the observables reviewed in —${\tau_{2,1}}$, ${\tau_{3,2}}$, ${C_{2}}$, and ${D_{2}}$—are all of this form.
Essential to this approach is a flexible basis of IRC safe observables from which to build discriminants. While the original energy correlators are indeed a useful basis, they are still somewhat restrictive. For example, the phase space structure of $e_2$ and $e_3$ in is completely fixed, as are all of the parametric properties inherited from this structure, such that $D_2$ is the only combination that parametrically distinguishes 1- and 2-prong substructure.
In this section, we enlarge the basis of jet substructure observables by defining generalizations of the energy correlation functions, allowing for a more general angular dependence than considered in . These new observables are flexible building blocks, which we use in the rest of this paper to identify promising tagging observables using power-counting techniques.[^9]
General Structure of Infrared/Collinear Safe Observables {#sec:gen_IRC}
--------------------------------------------------------
![Schematic depiction of a hard scattering event. A general IRC safe observable can be constructed by summing over all energy deposits, $E_i$, in an event, with a symmetric angular weighting function depending on the dimensionless unit vectors $\hat p_i$. []{data-label="fig:sphere_observable"}](figures/sphere_observable.pdf){width="6.5cm"}
In order to engineer the phase space structure of observables to have specific properties, we first need to systematically understand the structure of IRC safe observables that probe $n$-particle correlations. The general structure of an IRC safe observable is shown schematically in , where any IRC safe observable can be constructed from the energy deposits and angular information on the sphere. In the $pp$ case, of course, one typically uses the longitudinally boost-invariant quantities $p_T$ and $R_{ij}$, but the following argument is insensitive to that coordinate change.
As shown in , any IRC safe observable can be constructed from the following (complete) basis of observables[^10] $$\begin{aligned}
\label{eq:gen_obs}
F_N(\{p_i\})=\sum_{i_1} \sum_{i_2} \ldots \sum_{i_N} E_{i_1} E_{i_2} \ldots E_{i_N} \, f_N( \hat p_{i_1}, \hat p_{i_2}, \ldots, \hat p_{i_N} )\,,\end{aligned}$$ where $E_i$ is the energy of particle $i$, $\hat p_i$ is a dimensionless unit vector describing its direction, and $f_N$ is a symmetric function of its arguments. For IRC safety, we must further demand that the function $f_N$ vanishes when any two particles become collinear. Note that is a linear function of the momenta of the particles and a symmetric function of the angles. This basis of observables are referred to in the literature as $C$-correlators [@Tkachov:1995kk; @Sveshnikov:1995vi; @Cherzor:1997ak; @Tkachov:1999py].
Since the above discussion is completely general, it is not immediately obvious that it is useful for jet substructure studies. Still, has the interesting feature that, while the dependence on the energies is fixed by IRC safety, the angular function $f_N$ is much less restricted and can be chosen for specific purposes. The original energy correlators in are a specific case of , where, up to an overall normalization, the angular weighting function is $$\begin{aligned}
\label{eq:eNangles}
{{e_{N}^{(\beta)}}}: \qquad f_N( \hat p_{i_1}, \hat p_{i_2}, \ldots, \hat p_{i_N} )=\prod_{s < t \in \{i_1, i_2 , \dots, i_n \}} \theta_{st}^\beta\,.\end{aligned}$$ The key observation is that by considering alternative angular weighting functions for $n$-point correlators beyond , we can define a more flexible basis of observables for jet substructure studies.
New Angles on Energy Correlation Functions: ${{_{v}e_{n}^{(\beta)}}}$ {#sec:gen_ecf}
---------------------------------------------------------------------
There are many known decompositions of the angular function $f_N$—including Fox-Wolfram moments [@Fox:1978vu; @Fox:1978vw] and orthogonal polynomials on the sphere [@GurAri:2011vx]—but these are not necessarily optimal for jet substructure. The reason is that jets with well-resolved subjets exhibit a hierarchy of distinct angular scales, so we need to design $f_N$ to identify hierarchical—instead of averaged—features within a jet.
As seen in , the original energy correlation functions do capture multiple angular scales, but they do so all at once; it would be preferable if $f_N$ could identify one angular scale at a time in order to isolate different physics effects. Furthermore, to make power-counting arguments more transparent, we want $f_N$ to exhibit homogeneous angular scaling, such that each term in has a well-defined scaling behavior without having to perform a non-trivial expansion in the soft and collinear limits.
With these criteria in mind, we can now translate the general language of IRC safe observables into a useful basis for jet substructure studies. The angular function $f_N$ has to be symmetric in its arguments, and the simplest symmetric function that preserves homogeneous scaling is the $\min$ function.[^11] This leads us to the generalized energy correlation functions, which depend on $n$ factors of the particle energies and $v$ factors of their pairwise angles, $$\label{eq:ecf_gen}
{{_{v}e_{n}^{(\beta)}}} = \sum_{1 \leq i_1 < i_2 < \dots < i_n \leq n_J} z_{i_1} z_{i_2} \dots z_{i_n} \prod_{m = 1}^{v} \min^{(m)}_{s < t \in \{i_1, i_2 , \dots, i_n \}} \left\{ \theta_{st}^{\beta} \right\},$$ where $\min^{(m)}$ denotes the $m$-th smallest element in the list. For a jet consisting of fewer than $n$ particles, ${{_{v}e_{n}}}$ is defined to be zero. More explicitly, the three arguments of the generalized energy correlation functions are as follows.
- The subscript $n$, appearing to the right of the observable, denotes the number of particles to be correlated. This plays the same role as the $n$ subscript for the standard $e_n$ energy correlators in .
- The subscript $v$, appearing to the left of the observable, denotes the number of pairwise angles entering the product. By definition, we take $v \leq \binom{n}{2}$, and the minimum then isolates the product of the $v$ smallest pairwise angles.
- The angular exponent $\beta>0$ can be used to adjust the weighting of the pairwise angles, as in .
For the special case of $v = \binom{n}{2}$, the generalized energy correlators reduce to the standard ones in , with ${{_{1}e_{2}}}\equiv{e_{2}}$, ${{_{3}e_{3}}}\equiv{e_{3}}$, ${{_{6}e_{4}}}\equiv{e_{4}}$, and so on for the higher-point correlators.
Compared to the original energy correlators, the generalization in allows more flexibility in the angular scaling; this simplifies the construction of useful ratios and extends the possible applications of energy correlators. In the case of boosted top tagging, for example, the standard ${e_{4}}={{_{6}e_{4}}}$ observable involves six different pairwise angles. A decaying boosted top quark, however, does not have six characteristic angular scales, so most of these angles are redundant and only serve to complicate the structure of the observable. This is reflected in the definition of ${D_{3}}$ in , which involves three distinct terms [@Larkoski:2014zma].
To make more explicit the definition in , we summarize the particular correlators used in our case studies below. For boosted 2-prong tagging in , we use the 2-point energy correlation function $$\begin{aligned}
\label{eq:explicit_twopointvar}
{{_{1}e_{2}^{(\beta)}}}&\equiv{{e_{2}^{(\beta)}}}=\sum_{1\leq i<j\leq n_J} z_{i}z_{j} \, \theta_{ij}^\beta\ ,\end{aligned}$$ whose definition is unique, since it only involves only a single pairwise angle. We also need the 3-point correlators, which have three variants probing different angular structures:[^12] $$\begin{aligned}
\label{eq:explicit_ecfvar}
{{_{1}e_{3}^{(\beta)}}}&=\sum_{1\leq i<j<k\leq n_J} z_{i}z_{j}z_{k} \min \left\{ \theta_{ij}^\beta\,, \theta_{ik}^\beta\,, \theta_{jk}^\beta \right\} \ , \nonumber \\
{{_{2}e_{3}^{(\beta)}}}&=\sum_{1\leq i<j<k\leq n_J} z_{i}z_{j}z_{k} \min \left\{\theta_{ij}^\beta \theta_{ik}^\beta\,, \theta_{ij}^\beta \theta_{jk}^\beta\,, \theta_{ik}^\beta \theta_{jk}^\beta \right\} \ , \nonumber \\
{{e_{3}^{(\beta)}}}\equiv{{_{3}e_{3}^{(\beta)}}}&=\sum_{1\leq i<j<k\leq n_J} z_{i}z_{j}z_{k} \, \theta_{ij}^\beta \theta_{ik}^\beta \theta_{jk}^\beta \,.\end{aligned}$$ Interestingly, we are able to construct powerful observables from each of these three 3-point correlators, resulting in different tagging properties.
For boosted top tagging in , we also need the 4-point correlators. There are six possible variants, but we only study three of them in the body of the text: $$\begin{aligned}
{{_{1}e_{4}^{(\beta)}}} &= \sum_{1\leq i<j<k<\ell\leq n_J} z_{i}z_{j}z_{k}z_{\ell} \min \left\{ \theta_{ij}^\beta, \theta_{ik}^\beta, \theta_{jk}^\beta, \theta_{i\ell}^\beta, \theta_{j\ell}^\beta, \theta_{k\ell}^\beta \right\} \ , \nonumber\\
{{_{2}e_{4}^{(\beta)}}} &= \sum_{1\leq i<j<k<\ell\leq n_J} z_{i}z_{j}z_{k}z_{\ell} \min \left\{ \theta_{ij}^\beta, \theta_{ik}^\beta, \theta_{jk}^\beta, \theta_{i\ell}^\beta, \theta_{j\ell}^\beta, \theta_{k\ell}^\beta \right\} \nonumber\\
& \hspace{4.5cm} \times \min^{(2)} \left\{ \theta_{ij}^\beta, \theta_{ik}^\beta, \theta_{jk}^\beta, \theta_{i\ell}^\beta, \theta_{j\ell}^\beta, \theta_{k\ell}^\beta \right\} \ , \nonumber\\
&\hspace{0.2cm}\vdots \nonumber \\
{{e_{4}^{(\beta)}}}\equiv{{_{6}e_{4}^{(\beta)}}} &= \sum_{1\leq i<j<k<\ell\leq n_J} z_{i}z_{j}z_{k}z_{\ell} \theta_{ij}^\beta \theta_{ik}^\beta \theta_{jk}^\beta \theta_{i\ell}^\beta \theta_{j\ell}^\beta \theta_{k\ell}^\beta \,,
\label{eq:4pointcases}\end{aligned}$$ where $\min^{(2)}$ is again the second smallest element in the list. Here, we see the simplicity in the angular structure of ${{_{1}e_{4}}}$ and ${{_{2}e_{4}}}$, as compared to ${{_{6}e_{4}}}$ which involves all six angles. The vertical dots denote other 4-point correlation functions; we have not found them to be particularly useful, but they might have applications in (and beyond) jet substructure.
When constructing jet substructure observables, it is often desirable to work with ratios that are approximately boost invariant. Since the different generalized correlators probe a different number of energy fractions and pairwise angles, each scales differently under Lorentz boosts. Under a boost $\gamma$ along the jet axis and assuming a narrow jet, the energies and angles scale as $$\begin{aligned}
z_i \to z_i \,, \qquad \theta_{ij} \to \gamma^{-1} \theta_{ij}\,.\end{aligned}$$ This implies that the transformation of ${{_{v}e_{n}}}$ under boosts along the jet axis is determined solely by the $v$ index, $$\begin{aligned}
\label{eq:boost}
{{_{v}e_{n}^{(\beta)}}}\to \gamma^{-v\beta} {{_{v}e_{n}^{(\beta)}}}\,.\end{aligned}$$ Therefore, another way of interpreting the different ${{_{v}e_{n}}}$ is as ways of probing $n$ particle correlations with different properties under Lorentz boosts. The $v$ index therefore broadens the set of boost-invariant combinations that can be formed.
Finally, we remark that the definition in is certainly not unique, and we explore a few alternative definitions in that reduce to the $\min$ function in collinear limits. To further generalize while maintaining homogeneous scaling, one could use different angular exponents depending on the ordering of the angles. For the cases that we consider, though, we find that ${{_{v}e_{n}}}$ is sufficiently general to provide excellent performance while keeping the form of the observable (relatively) simple. That said, we expect alternative $f_N$ functions to also be useful, and their performance could be studied using the same power-counting techniques pursued here.
New Substructure Discriminants {#sec:mandn_series}
------------------------------
Our case studies are based primarily on three series of observables formed from the generalized correlators. We summarize their definitions here, and study their discrimination power in the forthcoming sections using both power-counting arguments and parton shower generators.
### The $M_i$ Series {#sec:m_series}
The $M_i$ series of observables is defined as $$\begin{aligned}
{M_{i}^{(\beta)}}=\frac{{{_{1}e_{i+1}^{(\beta)}}}}{{{_{1}e_{i}^{(\beta)}}}}\,.\end{aligned}$$ This observable is dimensionless, being formed as a ratio of dimensionless observables. As can be seen from , it is also invariant to boosts along the jet axis, since one angular factor appears in both the numerator and denominator.
These observables are constructed to identify $i$ hard prongs, but due to their limited angular structure, they are only effective when acting on suitably groomed jets. The main example of the $M_i$ series that we will consider explicitly in this paper is $$\begin{aligned}
{M_{2}^{(\beta)}}=\frac{{{_{1}e_{3}^{(\beta)}}}}{{{_{1}e_{2}^{(\beta)}}}}\,,\end{aligned}$$ which provides an example of a 2-prong substructure observable that only performs well after grooming. In , we briefly discuss the behavior of ${M_{3}}$ for boosted top tagging, where we argue that a more aggressive grooming strategy would be needed to make ${M_{3}}$ perform well.
### The $N_i$ Series {#sec:n_series}
We also define the $N_i$ series of observables as $$N_i^{(\beta)} = \frac{{{_{2}e_{i+1}^{(\beta)}}}}{({{_{1}e_{i}^{(\beta)}}})^2}.$$ As with the $M_i$ series, the $N_i$ series is dimensionless, and from , it is boost invariant, as two angular factors appear in both the numerator and denominator. Indeed, the fact that the 2-point correlation function appears squared in the denominator is fixed by boost invariance.
Two particular examples we find useful for this paper are $$\begin{aligned}
{N_{2}}=\frac{ {{_{2}e_{3}^{(\beta)}}} }{ ( {{_{1}e_{2}^{(\beta)}}} )^2}\,,\end{aligned}$$ which is a powerful boosted $W/Z/H$ tagger, and $$\begin{aligned}
{N_{3}}=\frac{ {{_{2}e_{4}^{(\beta)}}} }{ ( {{_{1}e_{3}^{(\beta)}}} )^2}\,,\end{aligned}$$ which is a powerful boosted top tagger on groomed jets. More generally, $N_i$ should be effective as an $i$-prong tagger, as discussed in , at least for groomed jets.
The $N_i$ observables take their name from the fact that in the limit of a resolved jet, they behave parametrically like the $N$-subjettiness ratio observables, as discussed in . Despite their similarity to $N$-subjettiness, the $N_i$ observables achieve their discrimination power in a substantially different manner, which has both theoretical and experimental advantages.
### The ${U_{i}}$ Series {#sec:qvsg_series}
Finally, we consider the ${U_{i}}$ series of observables defined as $$\begin{aligned}
{U_{i}^{(\beta)}}={{_{1}e_{i+1}^{(\beta)}}}\,,\end{aligned}$$ which are designed for quark/gluon discrimination. Note that unlike ${M_{i}}$ and ${N_{i}}$, the ${U_{i}}$ observables are not boost invariant. For the case $i=1$, $U_1$ coincides with the usual quark/gluon discriminants formed from the energy correlation functions [@Larkoski:2013eya], namely $$\begin{aligned}
{U_{1}^{(\beta)}}={C_{1}^{(\beta)}}={{_{1}e_{2}^{(\beta)}}}={{e_{2}^{(\beta)}}}\,,\end{aligned}$$ which probe single soft particle correlations within the jet. For $i>1$, the ${U_{i}}$ observables probe multi-particle correlations within the jet in a specific way that is useful for quark/gluon discrimination.
Simplifying Observables for Boosted Top Tagging {#sec:tops}
===============================================
Boosted top tagging has achieved significant attention at the LHC, with a large number of proposed observables to distinguish 3-prong hadronic top jets from the QCD background [@Kaplan:2008ie; @Thaler:2008ju; @Almeida:2008yp; @Almeida:2008tp; @Plehn:2009rk; @Plehn:2010st; @Almeida:2010pa; @Thaler:2010tr; @Thaler:2011gf; @Jankowiak:2011qa; @Soper:2012pb; @Larkoski:2013eya; @Anders:2013oga; @Freytsis:2014hpa; @Larkoski:2015yqa; @Kasieczka:2015jma; @Lapsien:2016zor]. In addition to $b$-tagging one of the subjets [@ATLAS-CONF-2012-100; @CMS:2013vea; @ATL-PHYS-PUB-2014-014; @ATL-PHYS-PUB-2015-035; @CMS-PAS-BTV-15-001; @CMS-PAS-BTV-15-002; @ATLAS-CONF-2016-001; @ATLAS-CONF-2016-002] and requiring the (groomed) jet mass to be close to $m_t \simeq 172~{\text{GeV}}$, two of the most effective tagging observables are shower deconstruction [@Soper:2011cr; @Soper:2012pb; @Soper:2014rya] and the $N$-subjettiness ratio ${\tau_{3,2}}$ [@Thaler:2010tr; @Thaler:2011gf]. Shower deconstruction works by testing the compatibility of a QCD shower model with the observed shower pattern; it is an extremely powerful discriminant, particularly at lower efficiencies. Jet shapes like $N$-subjettiness are also powerful discriminants, particularly at higher efficiencies. For a detailed discussion and experimental study, see, for example, .
In this section, we use the generalized energy correlation functions to construct $N_3$, a simple but powerful boosted top tagger designed for use on groomed jets. Unlike ${\tau_{3,2}}$, $N_3$ is defined without reference to external axes, allowing it to achieve better background rejection at high signal efficiencies. Interestingly, in the limit of well-resolved subjets and acting on groomed jets, $N_3$ has identical power counting to $N$-subjettiness. The behavior on ungroomed jets is discussed in .
Constructing the $N_3$ Observable {#sec:tops_makeobs}
---------------------------------
Mode Energy Angle
----------- ----------- ---------------
soft $z_s$ $1$
collinear 1 $\theta_{cc}$
c-soft $z_{cs}$ $\theta_{12}$
cc-soft $z_{ccs}$ $\theta_{23}$
: A summary of the modes in which enter the power-counting analysis for boosted top quarks. []{data-label="tab:pc_top"}
To detect boosted top jets with hard 3-prong substructure, we can use combinations of 2-point, 3-point, and 4-point correlators. Due to the large number of possible combinations, the power counting approach becomes essential to systematically study the behavior of these observables.
In order to use power counting to probe the boundary between the 3-prong (signal) and 2-prong (background) regions of phase space, one must analyze signal configurations which approach this boundary. For this reason, we consider not only the case of three subjets with equal energies and opening angles, as shown in , but also the strongly-ordered limit, shown in , where two of the three prongs become collinear. When the opening angles are hierarchical, the emission modes for each of the dipoles are distinct and must be treated separately, as discussed in . For lack of a better name, we call these additional modes collinear-collinear-soft modes (shown in magenta in ) to distinguish them from collinear-soft modes (shown in orange). A summary of these different modes, and the scaling of their angles and energies, are given in . These modes satisfy the relations $$\begin{aligned}
\label{eq:3prongscaling}
z_{cs}\ll z_{ccs} \ll 1\,, \qquad \theta_{cc} \ll \theta_{23} \ll \theta_{12} \ll 1\,.\end{aligned}$$ Note the reversal of the energy and angle hierarchies: collinear-collinear-soft modes have smaller angles but higher energies than collinear-soft modes. With this slight modification, the power-counting analysis proceeds identically to the simpler case shown in .
Many experimental analyses use jet shapes as measured on groomed jets, even if the original jet shapes were proposed without grooming. Grooming has the advantage of making jet properties resistant to pileup contamination and it also leads to observables that are more stable as the jet mass and $p_T$ are varied. More generally, grooming techniques minimize sensitivity to low momentum particles and the corresponding experimental uncertainties associated with their reconstruction. It is also possible to use a combination of groomed and ungroomed (or lightly groomed) substructure discriminants [@gregory_talk; @gregory_paper]. Here, we design our observable specifically for use on groomed jets, since it will help us identify discriminants that are both stable and high-performing. From the perspective of power counting, grooming simplifies the scaling properties of observables, since we can ignore regions of phase space with soft wide-angle subjets. In the past, such regions caused complications in designing top tagging observables based on energy correlators [@Larkoski:2014zma], as seen in the definition of $D_3$ in . After jet grooming, we can drop soft radiation (shown in green in ) for the purposes of power counting.
From , we have six 4-point correlators we could use to form ratio observables with the 2- and 3-point correlators. To reduce the number of possibilities, we restrict our attention to boost-invariant combinations, but this still leaves many ratios to test. In , we outline a systematic strategy to isolate the most promising 3-prong discriminants using power counting. Here, we focus on the best performing observable, $$\begin{aligned}
{N_{3}^{(\beta)}}=\frac{ {{_{2}e_{4}^{(\beta)}}} }{ ({{_{1}e_{3}^{(\beta)}}})^2 }\,,\end{aligned}$$ which was presented in as a member of the $N_i$ series.
To understand why ${N_{3}}$ is a powerful discriminant on groomed jets, we need to contrast the phase space for 3-prong signal jets versus 2-prong background jets. For the 3-prong top signal, it is sufficient to study the strongly-ordered limit in , since the balanced case of can be obtained by setting $z_{ccs} = z_{cs}$ and $\theta_{23} = \theta_{12}$. Using the methods of on the modes from , we find the following parametric scaling: $$\begin{aligned}
\text{3-prong signal (groomed):}\qquad {{_{1}e_{3}^{(\beta)}}}&\sim \theta_{23}^\beta\,, \nn \\
{{_{2}e_{4}^{(\beta)}}}& \sim z_{cs} \theta_{12}^\beta \theta_{23}^\beta + z_{ccs} \theta_{23}^{2\beta} + \theta_{23}^\beta \theta_{cc}^\beta \,.\end{aligned}$$ The dominant background to boosted top quarks are gluon and quark jets, particularly bottom quarks when subjet $b$-tagging is used [@ATLAS-CONF-2012-100; @CMS:2013vea; @ATL-PHYS-PUB-2014-014; @ATL-PHYS-PUB-2015-035; @CMS-PAS-BTV-15-001; @CMS-PAS-BTV-15-002; @ATLAS-CONF-2016-001; @ATLAS-CONF-2016-002]. While we ordinarily think of these as being 1-prong backgrounds (see ), they are mainly relevant when they feature 2-prong substructure from a hard parton splitting. Therefore, the phase space configuration we have to consider for the background is that of . Using the modes from , we find $$\begin{aligned}
\text{2-prong background (groomed):}\qquad{{_{1}e_{3}^{(\beta)}}}& \sim z_{cs} \theta_{12}^\beta + \theta_{cc}^\beta\,, \nn \\
{{_{2}e_{4}^{(\beta)}}}& \sim z_{cs}^2 \theta_{12}^{2\beta} + z_{cs} \theta_{12}^\beta \theta_{cc}^\beta + \theta_{cc}^{2\beta}\,.\end{aligned}$$ From these power-counting relations, we now want to derive the scaling of ${{_{2}e_{4}}}$ versus ${{_{1}e_{3}}}$.
For signal jets, ${{_{2}e_{4}}}$ is always smaller than ${{_{1}e_{3}}}$, since they share a factor of $\theta_{23}^\beta$, but each term in ${{_{2}e_{4}}}$ is also multiplied by parametrically small quantity. In particular, $\theta_{cc} \ll \theta_{23}$ by the assumption of , so we have the parametric relation $$\begin{aligned}
\text{3-prong signal (groomed):}\qquad {{_{2}e_{4}^{(\beta)}}} \ll ({{_{1}e_{3}^{(\beta)}}})^2\,.\end{aligned}$$ A much more detailed derivation of this scaling, and an illustration of how it can be identified systematically, is presented in . For background jets, each term in ${{_{2}e_{4}}}$ is the product of two terms in ${{_{1}e_{3}}}$, so we have the relation $$\begin{aligned}
\text{2-prong background (groomed):}\qquad {{_{2}e_{4}^{(\beta)}}} \sim ({{_{1}e_{3}^{(\beta)}}})^2\,.\end{aligned}$$ This shows that the particular combination chosen to define ${N_{3}}$ is indeed appropriate, since we can isolate the top signal region by making a cut of ${N_{3}} \ll 1$. These phase space relations are shown in .
To further improve our understanding, it is instructive to compare this with the $N$-subjettiness ratio ${\tau_{3,2}}$, whose phase space is shown in . For strongly-ordered 3-prong substructure, we find $$\begin{aligned}
\text{3-prong signal (groomed):}\qquad {\tau_{2}^{(\beta)}}&\sim \theta_{23}^\beta \sim {{_{1}e_{3}^{(\beta)}}}\,, \nn \\
{\tau_{3}^{(\beta)}}&\sim z_{cs} \theta_{12}^\beta + z_{ccs} \theta_{23}^\beta + \theta_{cc}^\beta \sim \frac{{{_{2}e_{4}^{(\beta)}}}}{ \theta_{23}^\beta}\,.\end{aligned}$$ For 2-prong background jets, we find $$\text{2-prong background (groomed):}\qquad {\tau_{2}^{(\beta)}} \sim {\tau_{3}^{(\beta)}} \sim z_{cs} \theta_{cs}^\beta + \theta_{cc}^\beta \sim {{_{1}e_{3}^{(\beta)}}}\,.$$ Remarkably, in both cases, this leads to the relations $$\begin{aligned}
{{_{1}e_{3}^{(\beta)}}}&\sim {\tau_{2}^{(\beta)}}\,, \nn \\
{{_{2}e_{4}^{(\beta)}}}&\sim {\tau_{2}^{(\beta)}} {\tau_{3}^{(\beta)}} \label{eq:2e4factorize}\,.\end{aligned}$$ Therefore, on groomed jets, the ${N_{3}}$ and ${\tau_{3,2}}$ observables are parametrically identical: $$\begin{aligned}
\label{eq:N3_related_Nsub}
{N_{3}^{(\beta)}}=\frac{{{_{2}e_{4}^{(\beta)}}}}{({{_{1}e_{3}^{(\beta)}}})^2}\sim \frac{ {\tau_{2}^{(\beta)}} {\tau_{3}^{(\beta)}} }{ ( {\tau_{2}^{(\beta)}})^2 } = \frac{ {\tau_{3}^{(\beta)}} }{ {\tau_{2}^{(\beta)}} }\,.\end{aligned}$$ This result is quite surprising. By summing over groups of four particles and taking double products of their pairwise angles, we have achieved an observable that behaves parametrically like an $N$-subjettiness ratio.
The observables $N_3$ and ${\tau_{3,2}}$ achieve their discrimination power in substantially different ways, as shown schematically in . Each term in ${{_{2}e_{4}}}$ is sensitive to multiple energies and angles and contains cross terms like $ \theta_{12}^\beta \theta_{cc}^\beta$. By contrast, $N$-subjettiness does not contain such cross terms; after determining the axes, each term in the $N$-subjettiness sum is independent of the presence of other subjets. Despite these differences, shows that the 4-point correlation function factorizes into a product of lower-point $N$-subjettiness observables, yielding the same parametric behavior in the resolved limit.
While there are no parametric difference between $N_3$ and ${\tau_{3,2}}$, our parton shower study will show that ${N_{3}}$ exhibits improved discrimination power on groomed jets, particularly at high efficiencies. Part of the reason this occurs is because $N_3$ is defined without respect to subjet axes. This not only offers the practical advantage of not needing to specify an axes-finding algorithm, but it also has an effect on the behavior of $N_3$ away from the power-counting regime. Recall that $N$-jettiness was originally designed to isolate regions of phase space where there are $N$ well-resolved jets [@Stewart:2010tn]. In this limit, the axes are well defined and independent of the particular axes definition up to power corrections. When used in jet substructure, however, $N$-subjettiness is used both in the limit of well-resolved subjets as well as in the limit of unresolved subjets. Indeed, in many substructure analyses, relatively loose requirement are placed on $N$-subjettiness, such that the ${\tau_{3,2}}$ cut is placed precisely in the unresolved region. Here, $N$-subjettiness can exhibit pathological behavior related to the axes choice [@Larkoski:2015uaa]. By contrast, the $N_3$ observable, being composed simply as sums over the jet constituents, is well behaved throughout the entire jet spectrum, and this will be reflected in its improved performance.
Performance in Parton Showers {#sec:N3_MC}
-----------------------------
\
\
Having understood the power counting of ${N_{3}}$ on groomed jets, we now study its behavior in parton shower generators, comparing ${N_{3}}$ with both ${\tau_{3,2}}$ and the simplified version of ${D_{3}}$ defined in . The comparison to ${\tau_{3,2}}$ is particularly interesting, since the parametrics in suggest it should perform similarly to ${N_{3}}$ in the resolved limit.
For our parton shower study, we generate background QCD jets from $pp\to jj$ events, where we consider separately the cases of $j=g$ (gluon) and $j=u$ (representative of light quarks). We also consider the case of $b$-quark backgrounds, which are interesting to treat separately due to recent advances in $b$-tagged substructure [@ATLAS-CONF-2012-100; @CMS:2013vea; @ATL-PHYS-PUB-2014-014; @ATL-PHYS-PUB-2015-035; @CMS-PAS-BTV-15-001; @CMS-PAS-BTV-15-002; @ATLAS-CONF-2016-001; @ATLAS-CONF-2016-002]; heavy quarks were generated from the process $pp\to b \bar b$. The boosted top signal is generated from $pp\to t \bar t$ events, with both tops decaying hadronically.
Events were generated with [<span style="font-variant:small-caps;">MadGraph52.3.3</span>]{} [@Alwall:2014hca] at the $13$ TeV LHC and showered with [<span style="font-variant:small-caps;">Pythia8.219</span>]{} [@Sjostrand:2006za; @Sjostrand:2007gs] with underlying event and hadronization implemented with the default settings. Anti-$k_T$ [@Cacciari:2008gp] jets with radius $R=1.0$ were clustered in [<span style="font-variant:small-caps;">FastJet3.2.0</span>]{} [@Cacciari:2011ma] using the Winner Take All (WTA) recombination scheme [@Larkoski:2014uqa; @Larkoski:2014bia].[^13] The energy correlation functions and $N$-subjettiness ratio observables were calculated using the `EnergyCorrelator` and `Nsubjettiness` [<span style="font-variant:small-caps;">FastJetcontrib</span>]{}s [@Cacciari:2011ma; @fjcontrib]. For $N$-subjettiness, we use one-pass WTA minimization with $\beta=1$. As a concrete example of a groomer, we use $\beta=0$ soft drop [@Larkoski:2014wba] (a.k.a. modified mass drop with $\mu =1$ [@Dasgupta:2013ihk; @Dasgupta:2013via]) with $\zcut=0.1$, though our general observations should be independent of the particular choice of groomer.
As discussed in , we focus on the behavior of the observables on groomed jets, where ${N_{3}}$ was designed to perform well and where ${N_{3}}$ behaves parametrically like ${\tau_{3,2}}$. In , we study boosted top tagging without grooming, where ${N_{3}}$ is still a reasonably powerful discriminant on ungroomed jets, but not as strong as ${\tau_{3,2}}$. We also discuss the behavior of ${M_{3}}$ in , using power-counting arguments to show why it is a poor discriminant with standard groomers, but might perform better with a more aggressive grooming strategy.
In , we show distributions for groomed ${N_{3}}$, comparing the top jet signal to the backgrounds of $b$-quark, light quark, and gluon jets. A groomed mass cut of $m_{\rm SD}>80$ GeV is applied, following a recent ATLAS study [@ATL-PHYS-PUB-2015-053]. Here, we use $\beta=2$ as the angular exponent for ${N_{3}}$; power counting does not, in this case, predict a preferred value of $\beta$, so it could be optimized for experimental performance. The behavior of these distributions is quite interesting, particularly for $p_{TJ}>500$ GeV in , where the top quarks are truly boosted. The signal distribution drops off sharply above ${N_{3}} \simeq 1.5$, while the background distribution extends to larger values for all three samples. This behavior leads to excellent performance at high signal efficiencies, and is quite different than for ${\tau_{3,2}}$ (see in ). Note that these distributions are calculated after the soft drop mass cut, so the region where ${N_{3}}$ exhibits improved performance is the one directly relevant for LHC searches.
In , we show signal efficiency versus background rejection (ROC) curves for boosted top discrimination against $b$-quark, light quark, and gluon jets. In these and all subsequent ROC curves, the efficiency and mistag rates are given *after* applying a baseline mass requirement, in order to show just the gain in performance from adding a substructure cut. The baseline efficiencies for the $m_{\rm SD}>80$ GeV mass selection are $$\begin{aligned}
\text{$p_{TJ}=200$ GeV}: & \quad \mathcal{E}_t =61\%, \quad \mathcal{E}_b = 2.4\%, \quad \mathcal{E}_q = 2.8\%, \quad \mathcal{E}_g = 6.6\%, \\
\text{$p_{TJ}=500$ GeV}: & \quad \mathcal{E}_t =87\%, \quad \mathcal{E}_b = 10\%,\phantom{.} \quad \mathcal{E}_q = 10\%, \quad \mathcal{E}_g = 19\%,\end{aligned}$$ and the final efficiencies and mistag rates are obtained by multiplying these baseline values by those shown in . Comparing $p_{TJ}=200$ GeV and $p_{TJ}=500$ GeV, we conclude that the behavior of $N_3$ is reasonably robust as a function of $p_{TJ}$ (see for higher $p_{TJ}$ values). The simplified version of $D_3$ with this choice of angular exponent gives rather poor discrimination power, especially for gluon jets; the apparent negative discrimination power for certain ROC curves in is due to the use of a (non-optimal) one-sided cut. It is also satisfying to see the behavior predicted from the power-counting analysis. At lower top efficiencies, where there are well-resolved jets, $N_3$ and ${\tau_{3,2}}$ exhibit similar discrimination power, but at higher efficiencies, where there are not well-resolved jets, the structure of $N_3$ leads to considerably improved performance. It would be interesting to see whether these parton shower predictions remain true in LHC data.
Finally, another important feature of the soft-dropped $N_3$ observable is its stability as the mass and $p_{T}$ of the jet are varied. This has recently been emphasized in as a highly desirable feature of jet substructure observables, as it removes mass sculpting. In , we show the signal and background distributions for three different values of the jet $p_{TJ}$, namely $p_{TJ}= \{200, 500, 1000\}$ GeV following . Remarkable stability of the $N_3$ distribution is seen, with the main distortion appearing for the top sample in the lowest $p_{TJ}$ bin, where the $R = 1.0$ jet radius is not always large enough to capture all of the top decay products. Between $p_{TJ}=500$ GeV and $1000$ GeV, there are almost no changes to either the signal or background distributions. Though not shown here, ${\tau_{3,2}}$ exhibits comparable stability to $N_3$, as expected since they share the same power counting. We conclude that soft-dropped $N_3$ is a powerful boosted top tagger that exhibits many experimentally desirable features.
New Observables for 2-prong Substructure {#sec:2prong}
========================================
Jet substructure techniques have played an increasingly important role in recent LHC searches, especially for new resonances with decays involving boosted $W/Z/H$ bosons [@Aad:2015owa; @Aaboud:2016okv; @Aaboud:2016trl; @Aaboud:2016qgg; @ATLAS-CONF-2016-055; @ATLAS-CONF-2015-071; @ATLAS-CONF-2015-068; @ATLAS-CONF-2016-082; @ATLAS-CONF-2016-083; @Khachatryan:2015bma; @CMS:2016pfl; @CMS:2016mvc; @CMS:2016wev]. In order to understand any possible hint of new physics in diboson analyses, it is essential to have exceptional control over the behavior of jet substructure discriminants, to allay concerns about possible analysis artifacts [@Goncalves:2015yua; @Martin:2016jdw]. In our view, echoing the perspective of , properties like stability with jet $p_T$ and resilience to mass sculpting are just as important as (and perhaps more so than) absolute tagging performance.
In this section, we use the generalized correlators to construct 2-prong taggers that are robust and perform well. This is an application where the original energy correlators have already proven useful through the ${C_{2}}$ and ${D_{2}}$ ratios [@Larkoski:2013eya; @Larkoski:2014gra]. Here, we propose three new ratios: $$\begin{aligned}
{M_{2}^{(\beta)}}=\frac{{{_{1}e_{3}^{(\beta)}}}}{{{_{1}e_{2}^{(\beta)}}}}\,, \qquad {N_{2}^{(\beta)}}=\frac{ {{_{2}e_{3}^{(\beta)}}} }{ ( {{_{1}e_{2}^{(\beta)}}} )^2}\,,
\qquad {D_{2}^{(\alpha,\beta)}}=\frac{{{_{3}e_{3}^{(\alpha)}}}}{({{_{1}e_{2}^{(\beta)}}})^{3\alpha/\beta}}\,,\end{aligned}$$ corresponding to the three variants of the 3-point correlator in . Each of these observables is sensitive to different angular structures within the jet and therefore achieves its discrimination power in a different manner. This fact is highlighted in their different behavior under grooming, where ${M_{2}}$ and ${D_{2}^{(1,2)}}$ were constructed to only perform well on groomed jets. Therefore, these observables are probes not only of 2-prong jet substructure but also of any grooming procedure applied to the jet.[^14]
Power-Counting Analysis and Observable Phase Space {#sec:twoprong_PC}
--------------------------------------------------
The power counting for 2-prong discriminants follows straightforwardly from , using the modes summarized in and . Since the phase space is much simpler than in the 3-prong case, we can study the behavior of ${M_{2}}$, ${N_{2}}$, and ${D_{2}^{(\alpha,\beta)}}$ both before and after jet grooming.
To begin, we consider the 1-prong background in and power count the contributions to ${e_{2}}$ and ${{_{v}e_{3}}}$ from every possible triplet of soft and collinear modes. We do the same for the 2-prong signal in , where we also have to consider collinear-soft modes, though we do not show the power-suppressed triplets for brevity. These tables show that while the 3-point correlators have similar behavior for soft particles, they have different behavior for correlations among collinear particles (cf. the first row of and the second and third row of ). This is expected given the different number of pairwise angles in the definition of each ${{_{v}e_{3}}}$. We discuss the consequences of this power counting for each of the proposed ratios in the following subsections.
{width="5cm"}
Modes ${{e_{2}^{(\beta)}}}$ ${{_{1}e_{3}^{(\beta)}}}$ ${{_{2}e_{3}^{(\beta)}}}$ ${{_{3}e_{3}^{(\beta)}}}$
------- ----------------------------- --------------------------- --------------------------- ---------------------------
$CCC$ $\theta_{cc}^{\beta}$ $\theta_{cc}^{\beta}$ $\theta_{cc}^{2\beta}$ $\theta_{cc}^{3\beta}$
$CCS$ $z_s + \theta_{cc}^{\beta}$ $z_s \theta_{cc}^{\beta}$ $z_s \theta_{cc}^{\beta}$ $z_s \theta_{cc}^{\beta}$
$CSS$ $z_s$ $z_s^2$ $z_s^2$ $z_s^2$
$SSS$ $z_s^2$ $z_s^3$ $z_s^3$ $z_s^3$
: Parametric contributions to ${e_{2}}$ and the 3-point correlators, ${{_{v}e_{3}}}$, in the case of a jet with 1-prong substructure. The different contributions arise from correlations among soft ($S$) and collinear ($C$) radiation. []{data-label="tab:pc"}
{width="5cm"}
Modes ${{e_{2}^{(\beta)}}}$ ${{_{1}e_{3}^{(\beta)}}}$ ${{_{2}e_{3}^{(\beta)}}}$ ${{_{3}e_{3}^{(\beta)}}}$
--------------- ----------------------- ----------------------------- ------------------------------------------- --------------------------------------------
$C_1C_2\,S$ $\theta_{12}^{\beta}$ $z_s \theta_{12}^{\beta}$ $z_s \theta_{12}^{\beta}$ $z_s \theta_{12}^{\beta}$
$C_1C_2\, C$ $\theta_{12}^{\beta}$ $ \theta_{cc}^{\beta}$ $\theta_{12}^{\beta} \theta_{cc}^{\beta}$ $\theta_{12}^{2\beta} \theta_{cc}^{\beta}$
$C_1C_2\,C_s$ $\theta_{12}^{\beta}$ $z_{cs}\theta_{12}^{\beta}$ $z_{cs}\theta_{12}^{2\beta}$ $z_{cs} \theta_{12}^{3\beta}$
: Same as , but for a jet with a resolved 2-prong substructure. The different contributions arise from correlations among soft ($S$), collinear ($C_i$), and collinear-soft ($C_s$) radiation. Power-suppressed contributions are not shown. []{data-label="tab:pc_ninja"}
### ${M_{2}}$
The observable ${M_{2}}$ is based on ${{_{1}e_{3}}}$: $$\begin{aligned}
{M_{2}^{(\beta)}}=\frac{{{_{1}e_{3}^{(\beta)}}}}{{{_{1}e_{2}^{(\beta)}}}}\,.\end{aligned}$$ We first consider its behavior on 1- and 2-prong jets without grooming. For 1-prong background jets from , we have $$\begin{aligned}
\text{1-prong background (ungroomed):} \quad {{e_{2}^{(\beta)}}} & \sim z_s + \theta_{cc}^\beta \,, \nn \\
{{_{1}e_{3}^{(\beta)}}} & \sim z_s^2 + \theta_{cc}^{\beta} \,.
\label{eq:1prongM2scaling}\end{aligned}$$ This exhibits a non-trivial phase space with boundaries ${{_{1}e_{3}}} \sim ({e_{2}})^2$ when the jet is dominated by soft radiation, and ${{_{1}e_{3}}} \sim {e_{2}}$ when the jet is dominated by collinear radiation. For 2-prong signal jets from , we have $$\begin{aligned}
\text{2-prong signal (ungroomed):} \quad {{e_{2}^{(\beta)}}} & \sim \theta_{12}^\beta \,, \nn \\
{{_{1}e_{3}^{(\beta)}}} &\sim z_{cs} \theta_{12}^{\beta} + \theta_{cc}^{\beta} \,.\end{aligned}$$ From the fact that $z_{cs}\ll 1$, and $\theta_{cc} \ll \theta_{12}$, one therefore finds the inequality ${{_{1}e_{3}}} \ll {e_{2}}$.
The phase space for ${M_{2}}$ is shown in . This power-counting analysis demonstrates that before any grooming has been applied, there is considerable overlap between the parametric phase space regions occupied by 1- and 2-prong jets. Therefore, ${M_{2}}$ has limited discrimination power on ungroomed jets. The power-counting analysis also makes clear why ${M_{2}}$ performs so poorly: 1-prong jets are dominated by soft radiation with scaling ${{_{1}e_{3}}} \sim ({e_{2}})^2$, which overlaps with the 2-prong signal region with ${{_{1}e_{3}}} \ll {e_{2}}$. The fact that this overlap is caused only by soft radiation also suggests that it can be eliminated by applying a jet grooming procedure to remove soft radiation.
In , we show the phase space for ${M_{2}}$ after grooming. Soft drop removes the $z_s$ contributions from , which pushes 1-prong background jets to the upper boundary of the phase space with ${{_{1}e_{3}}} \sim {e_{2}}$. By contrast, the parametric scaling of the signal jets is unaffected by the soft drop procedure.[^15] This yields a triangular phase space that resembles the case of ${\tau_{2,1}}$ in , where 1-prong background jets live on the upper boundary and 2-prong signal jets live in the bulk. Perhaps counterintuitively, the soft drop procedure pushes the background to larger values of ${M_{2}}$, achieving better discrimination power.
The ${M_{2}}$ observable therefore provides an interesting example of a discriminant that only performs well after grooming. It emphasizes the parametric effect that grooming procedures can have on radiation within a jet, beyond simply removing jet contamination. For this reason, we expect precision calculations of the ${M_{2}}$ distribution to provide useful insights into the behavior of such grooming procedures.
### ${N_{2}}$
The observable ${N_{2}}$ is based on ${{_{2}e_{3}}}$, $$\begin{aligned}
{N_{2}^{(\beta)}}=\frac{{{_{2}e_{3}^{(\beta)}}}}{ ({{_{1}e_{2}^{(\beta)}}})^2 }\,.\end{aligned}$$ The power-counting argument for ${N_{2}}$ closely parallels ${M_{2}}$. We will see that the phase space for ${N_{2}}$ is parametrically unmodified by the grooming procedure, making it perform well on both groomed and ungroomed jets.
We again begin by analyzing the parametric behavior of the observable on ungroomed jets. Using for 1-prong background jets, we find $$\begin{aligned}
\label{eq:N2_singleprong}
\text{1-prong background (ungroomed):} \quad {{e_{2}^{(\beta)}}} & \sim z_s + \theta_{cc}^\beta \,, \nn \\
{{_{2}e_{3}^{(\beta)}}} & \sim z_s^2 + z_s \theta_{cc}^\beta + \theta_{cc}^{2\beta} \,.\end{aligned}$$ In contrast to ${M_{2}}$, the 1-prong background jets exhibit only a single scaling, ${{_{2}e_{3}}} \sim ({e_{2}})^2$, for jets dominated by either soft or collinear radiation. Using for 2-prong signal jets, we find $$\begin{aligned}
\label{eq:N2_twoprong}
\text{2-prong signal (ungroomed):} \quad {{e_{2}^{(\beta)}}} & \sim \theta_{12}^\beta \,, \nn \\
{{_{2}e_{3}^{(\beta)}}} & \sim z_s \theta_{12}^{\beta} + z_{cs} \theta_{12}^{2\beta} + \theta_{cc}^{\beta} \theta_{12}^{\beta} \,.\end{aligned}$$ Signal jets satisfy the inequality ${{_{2}e_{3}}} \ll ({e_{2}})^2$, explaining the definition of the $N_2$ observable. The phase space before grooming is summarized in , where there is clear separation between 1-prong background jets, which live on the upper boundary of the phase space, and 2-prong signal jets, which live in the bulk of the phase space, again resembling the case of ${\tau_{2,1}}$ in .
Because the 1-prong background jets have a single scaling, removing $z_s$ from has no effect on the parametric phase space. Similarly, removing $z_s$ from does not change the parametrics of the 2-prong signal. Therefore, ${N_{2}}$ behaves more similarly to other 2-prong discriminants in the literature, since its discrimination power does not come entirely from the grooming procedure. The power-counting analysis also suggests that ${N_{2}}$ should be a powerful 2-prong discriminant both before and after grooming is applied; this will be verified in the parton shower studies below.
It is also interesting to contrast the ${N_{2}}$ phase space in with that of ${D_{2}}$ in . For ${D_{2}}$, the background jets are bounded by two different scaling behaviors: $$(e_2^{(\beta)})^3 < e_3^{(\beta)} < (e_2^{(\beta)})^2.$$ For ${N_{2}}$, by contrast, the background jets exhibit a single scaling and therefore live entirely on the boundary of phase space: $$_2e_3^{(\beta)} \sim (e_2^{(\beta)})^2.$$ Since this boundary is purely geometric, the ${N_{2}}$ distributions are remarkably insensitive to the mass or $p_T$ of the jet, even before grooming is applied.
Just as $N_3$ is related to ${\tau_{3,2}}$ (see ), $N_2$ behaves parametrically like ${\tau_{2,1}}$ in the resolved limit. The power-counting analysis proceeds identically as for $N_3$ and will not be repeated here; see for the general argument relating $N_i$ to ${\tau_{i,i-1}}$. We want to emphasize again that, in analogy to , $N_2$ exhibits ${\tau_{2,1}}$-like behavior without reference to any axes within the jet. It therefore does not exhibit the axes pathologies that arise for $N$-subjettiness in the limit of unresolved substructure, and $N_2$ can therefore be expected to have improved performance compared to ${\tau_{2,1}}$, particularly at high efficiencies.
### ${D_{2}^{(1,2)}}$
Our final example of a 2-prong discriminant is based on ${{_{3}e_{3}}} = {e_{3}}$, where we reconsider the ${D_{2}}$ observable with two distinct angular exponents, $$\label{eq:D2_def_genangle}
{D_{2}^{(\alpha, \beta)}} \equiv \frac{{{e_{3}^{(\alpha)}}}}{\left({{e_{2}^{(\beta)}}}\right)^{3\alpha/\beta}} \,.$$ The case of $\alpha=\beta$ was first defined in and analytically calculated in . While the phase space for ${D_{2}^{(\alpha, \beta)}}$ was discussed in detail in , we focus on the impact that $\alpha\not=\beta$ has on groomed jet discrimination.
With distinct angular exponents $\alpha$ and $\beta$, 1-prong background jets exhibit the scaling $$\begin{aligned}
\text{1-prong background (ungroomed):} \quad {{e_{2}^{(\beta)}}} & \sim z_s + \theta_{cc}^\beta \,, \nn \\
{{e_{3}^{(\alpha)}}} & \sim z_s^2 + z_s \theta_{cc}^\alpha + \theta_{cc}^{3\alpha} \,.\end{aligned}$$ The background therefore occupies a non-trivial phase space with boundaries ${{e_{3}^{(\alpha)}}} \sim ({{e_{2}^{(\beta)}}})^2$, when the jet is dominated by soft radiation, and ${{e_{3}^{(\alpha)}}} \sim ({{e_{2}^{(\beta)}}})^{3\alpha/\beta}$, when the jet is dominated by collinear radiation. The 2-prong signal has the parametric scaling $$\begin{aligned}
\text{2-prong signal (ungroomed):} \quad {{e_{2}^{(\beta)}}} & \sim \theta_{12}^\beta \,, \nn \\
{{e_{3}^{(\alpha)}}} & \sim z_s \theta_{12}^{\alpha} +z_{cs} \theta_{12}^{3\alpha} + \theta_{12}^{2\alpha} \theta_{cc}^{\alpha}\,,\end{aligned}$$ from which one can derive the relation ${{e_{3}^{(\alpha)}}} \ll ({{e_{2}^{(\beta)}}} )^{3\alpha/\beta}$. This demonstrates that the definition of ${D_{2}^{(\alpha, \beta)}}$ in is indeed appropriate for 2-prong substructure, confirming the expectation from boost invariance (see ).
As discussed in for ungroomed jets, the observable ${D_{2}^{(\alpha, \beta)}} $ only provides good discrimination between 1-prong and 2-prong jets for $$\begin{aligned}
\label{eq:constraint_alpha}
3\alpha >2\beta\,.\end{aligned}$$ When this relation is violated, the phase space regions for signal and background jets overlap. This is shown in , where contours of ${D_{2}^{(\alpha, \beta)}}$ cannot separate the 1- and 2-prong regions when is violated. After a grooming procedure is applied, though, the overlapping phase space region is removed, as shown schematically in . Now the constraint in no longer applies, and the angular exponents can be chosen with a particular focus on discrimination power on groomed jets.
The choice of $(\alpha,\beta)$ exponents could be tuned to optimize performance, but we advocate that $\alpha=1$, $\beta=2$ is a natural choice for groomed 2-prong discrimination. This choice explicitly violates , so ${D_{2}^{(1, 2)}}$ can only have good performance after grooming. The choice of $\beta=2$ is motivated by the relation $$\begin{aligned}
\label{eq:e2_mass_relation}
{{e_{2}^{(2)}}} \simeq \frac{p_{TJ}^2}{m_J^2}\,,\end{aligned}$$ such that a cut on the jet mass, or $p_T$, is effectively a cut on ${{e_{2}^{(2)}}}$. Without grooming, one would typically take $\alpha = 2$, but with grooming, one can lower the angular exponent $\alpha$ to $1$ to more directly probe collinear emissions. Importantly, by considering the observable with separate $\alpha$ and $\beta$ exponents, we are able to satisfy both the requirement that it behaves sensibly under a mass cut, as well as improve its sensitivity to collinear emissions. This is not possible with the $\alpha=\beta$ version of the ${D_{2}}$ observable, and indeed, ${D_{2}^{(1, 2)}}$ leads to improved discrimination power on groomed jets. From an analytic perspective, the choice of $\alpha=1$, $\beta=2$ simplifies calculations, hopefully facilitating precision calculations of ${D_{2}^{(1, 2)}}$ on groomed jets at the LHC.
Performance in Parton Showers {#sec:twoprong_MC}
-----------------------------
We now perform a parton shower study to verify the predictions of the above power-counting analysis. It is useful to briefly summarize our robust predictions regarding the behavior of ${M_{2}}$, ${N_{2}}$, and ${D_{2}^{(1,2)}}$ as boosted 2-prong taggers:
- The ${M_{2}}$ observable should provide little discrimination power before grooming, but will act as a powerful discriminant after the removal of wide-angle soft radiation.
- The ${N_{2}}$ observable will act as a powerful discriminant both before and after grooming, matching the behavior of ${\tau_{2,1}}$ in the resolved limit.
- The ${D_{2}^{(1,2)}}$ observable will behave similarly to ${M_{2}}$, providing good discrimination power only after grooming has been applied.
These predictions rely only on parametric scalings and are therefore independent of the implementation details of the perturbative parton shower or the hadronization model. For conciseness, we only show results generated with [<span style="font-variant:small-caps;">Pythia8.219</span>]{}, though we used [<span style="font-variant:small-caps;">Vincia2.0.01</span>]{} [@Giele:2007di; @Giele:2011cb; @GehrmannDeRidder:2011dm; @Ritzmann:2012ca; @Hartgring:2013jma; @Larkoski:2013yi; @Fischer:2016vfv] to check that the same results could be obtained with an alternative perturbative shower. We have not yet studied hadronization uncertainties, but we expect them to be small, particularly for groomed jets.
\
\
To verify these power-counting predictions, we use the same analysis and generation strategy as , again using a jet radius of $R = 1.0$. We generate background QCD jets from $pp\to Zj$ events, where we consider separately the cases of $j=g$ (gluon) and $j=u,d,s$ (light quark), letting the $Z$ decay leptonically to avoid additional hadronic activity. The 2-prong signal of boosted $Z$ bosons are generated from $pp\to ZZ$ events, with one $Z$ decaying leptonically, and the other to light quarks, $q=u,d,s$. We do not address in this paper the issue of sample dependence and the impact of color connections to the rest of the event. While it would be interesting to compare the discrimination power of ${N_{2}}$ against the more-prevalent $pp\to jj$ background, we expect the conclusions from $pp\to Zj$ to be robust, especially after grooming has been applied.
For concreteness, we always set the angular exponent in the energy correlator to $\beta=2$, such that a mass cut directly corresponds to a cut on the denominator of the observable, see . While this is a nice theoretical feature, it is by no means necessary, and the value of $\beta$ could be optimized for experimental performance. To focus on the phase space where tagging performance actually matters, we place a cut of $m \in [80,100]~$GeV for all of the ungroomed distributions and a cut of $m_{\rm SD} \in [80,100]~$GeV for all of the groomed distributions. We only present distributions with a cut of $p_T > 500$ GeV, though other $p_T$ ranges exhibit similar behaviors.
In , we show normalized distributions of ${M_{2}}$, ${N_{2}}$, and ${D_{2}^{(1,2)}}$ before and after soft drop grooming. Despite all being derived from 3-point correlators, they exhibit rather different behaviors. As expected, ${M_{2}}$ is a poor discriminant before grooming is applied; amusingly, the distributions of the $Z$ boson signal and quark jet background are essentially identical. As predicted by the power-counting analysis, the soft drop grooming procedure pushes the background ${M_{2}}$ distributions to larger values while leaving the signal distribution largely unmodified. We are not aware of another substructure discriminant with such a dramatic shift in behavior after jet grooming.
Turning to ${N_{2}}$, it exhibits good discrimination power both before and after grooming is applied, even though the shapes of the distributions are substantially modified by grooming. Before grooming, the ${N_{2}}$ distribution exhibits a sharp edge at its upper boundary. This arises because 1-prong background jets have a single parametric scaling and are therefore compressed along the upper boundary of the phase space (see ). After grooming, ${N_{2}}$ remains a powerful discriminant, as the phase space is parametrically unchanged by the grooming procedure. As expected, the peak values of the distributions decrease as soft radiation is groomed away, but the range spanned by the distribution remains approximately constant. This highlights the fact that parametric arguments give robust predictions about the boundaries of phase space but not the specific shapes of the distributions. In , we also verify that ${N_{2}}$ and ${\tau_{2,1}}$ exhibit the same parametric behaviors in the resolved limit.
Finally, the ${D_{2}^{(1,2)}}$ observable, while only a fair discriminant before grooming, exhibits good discrimination power after soft drop is applied. Therefore, we have seen that all of the power-counting predictions are observed in the parton shower generators, suggesting that parametric scalings dominate the behavior of these observables, at least for the purposes of 2-prong substructure tagging. From , we see that some of the observables behave quite differently for the quark and gluon samples. We revisit the possibilities of using ${{_{v}e_{3}}}$ for quark/gluon discrimination in , where we introduce the $U_2$ observable, which is based on ${{_{1}e_{3}}}$, similar to ${M_{2}}$.
To study the discrimination power more quantitatively, we show ROC curves before and after grooming in , considering the quark and gluon backgrounds separately. The baseline efficiencies for the ungroomed and groomed mass selections are $$\begin{aligned}
\text{$m \in [80,100]~$GeV}: & \quad \mathcal{E}_{Z} = 27\%, \quad \mathcal{E}_q = 17\%,\phantom{.} \quad \mathcal{E}_g = 15\%, \nonumber \\
\text{$m_{\rm SD} \in [80,100]~$GeV}: & \quad \mathcal{E}_{Z} = 37\%, \quad \mathcal{E}_q = 2.6\%, \quad \mathcal{E}_g = 4.3\%, \label{eq:2prong_eff}\end{aligned}$$ where we again normalize the ROC curves to show only the gains from the new 2-prong discriminants.[^16] We use ${D_{2}}$ (with $\beta = 2$) as a standard reference, since it is currently used by the ATLAS experiment for its excellent tagging performance [@ATLAS-CONF-2015-035; @Aad:2015rpa; @ATLAS-CONF-2015-068; @ATLAS-CONF-2015-071; @ATLAS-CONF-2015-073; @Aaboud:2016trl; @ATLAS-CONF-2016-016; @ATLAS-CONF-2016-039; @ATLAS-CONF-2016-055; @ATLAS-CONF-2016-082; @ATLAS-CONF-2016-083].[^17]
Of the three new observables, only ${N_{2}}$ is designed to act as a discriminant on ungroomed jets. In both , we see that ${N_{2}}$ outperforms the standard ${D_{2}}$ observable in discriminating against both quark and gluon jets. From power-counting arguments, we cannot predict the relative performance between the quark and gluon samples, but the fact that ${N_{2}}$ sees significant performance gains on the gluon sample is very encouraging. As discussed in , the discrimination power of ${N_{2}}$ is closely related to ${\tau_{2,1}}$ in the resolved limit, but with an improved behavior in the transition to the unresolved region. We discuss this relation in more detail in , showing that ${N_{2}}$ has slightly improved performance compared to ${\tau_{2,1}}$ on ungroomed jets, but considerably improved performance after grooming.
After jet grooming, shown in , all three new observables offer improved discrimination power over ${D_{2}}$. Comparing the results before and after grooming, we see dramatic gains in performance for ${M_{2}}$ and ${D_{2}^{(1,2)}}$, as expected from power counting. It is rather curious that after grooming, all three observable offer comparable discrimination power, even though they are based on ${{_{v}e_{3}}}$ correlators with different characteristic behaviors. It would be interesting to study the correlations between these observables to see if they are probing complementary physics effects. Such correlations go beyond the power-counting analysis of this paper, so we leave a study to future work.
Thus far, we have only considered observables measured entirely on either groomed or ungroomed jets. Experimentally, though, it may be desirable to measure ungroomed observables after the application of a groomed mass cut (see e.g. [@CMS-PAS-JME-14-002]); we refer to this as a “hybrid" strategy. In , we present ROC curves for ${M_{2}}$, ${N_{2}}$, ${D_{2}}$, and ${\tau_{2,1}}$ using this hybrid strategy and analyze their behavior using power counting. We leave a more detailed study of the optimal use of mixed groomed/ungroomed observables to future work.
Stability in Parton Showers {#sec:twoprong_MC_stable}
---------------------------
In addition to their absolute performance, our new 2-prong discriminants exhibit stable behavior, especially after grooming. As recently emphasized in , stability of background distributions as a function of mass and $p_T$ cuts is an important consideration when designing jet substructure observables. Excessive dependence on jet mass and $p_T$ can lead to mass sculpting, which can increase systematic uncertainties in sideband fits, counteracting gains from improved tagging performance.
To illustrate how the phase space structure controls the stability of the observable, it is interesting to study the stability of $D_2$, $M_2$, and $N_2$ before and after grooming. These three observables represent the three scaling behaviors we have encountered in this paper. Prior to grooming, we have:
- $D_2$ in : The background occupies a non-trivial phase space region that does not overlap with the signal.
- $M_2$ in : The background occupies a non-trivial phase space region overlapping with the signal.
- $N_2$ in : The background is confined to a single scaling on the boundary of phase space.
The ${D_{2}^{(1,2)}}$ observable has a similar phase space structure to $M_2$, and will therefore behave similarly, so we do not show it explicitly in this section. Note that ${\tau_{2,1}}$ has the same phase space structure as $N_2$, so it exhibits related stability properties.
In , we use parton showers to test the stability of $D_2$, $M_2$, and $N_2$ on the light quark background as the jet mass cut is varied.[^18] Prior to grooming, only the $N_2$ observable exhibits any degree of stability on the background. After grooming, all three observables have a nicely stable peak position and shape, and the residual variation could be compensated using the decorrelation technique of . We can now use a power-counting analysis to demonstrate how these behaviors are dictated by the form of the phase space. Although we focus on light quark jets in , similar stability properties are observed for gluon jets. This is also emphasized by the power-counting argument, which is insensitive to the quark or gluon nature of the jet.
\
\
We begin by considering the observables before grooming. For $D_2$ in , the background region is defined by two different scalings, one of which defines the upper boundary of the phase space and one of which defines the scaling of the boundary between the signal and background, and therefore the scaling of the desired cut value for discrimination. The upper boundary of the phase space is defined by the scaling ${e_{3}} \sim ({e_{2}})^2$, leading to the maximum value $$\begin{aligned}
D^{ \text{max}}_2\sim \frac{{e_{3}}}{({e_{2}})^3} \sim \frac{ ({e_{2}})^2}{({e_{2}})^3} \sim \frac{1}{{e_{2}}}\,.\end{aligned}$$ Simplifying to the case of $\beta=2$, and using , we have $$\begin{aligned}
\label{eq:D2max}
D^{(2), \text{max}}_2\sim \frac{p_{TJ}^2}{m_J^2}\,,\end{aligned}$$ which depends sensitively on $m_J$ and $p_{TJ}$. This behavior can be clearly seen in , where the $D_2$ distribution shifts dramatically with the jet mass cut, an undesirable feature for the purposes of sideband calibration.
For $M_2$ with a phase space given in , we see quite different behavior. In this case, the upper boundary of the phase space is defined by ${{_{1}e_{3}}} \sim {e_{2}}$, and therefore $M_2$ has a maximum value $$\begin{aligned}
\label{eq:M2max}
M^{ \text{max}}_2 \sim \frac{{{_{1}e_{3}}}}{{e_{2}}} \sim \frac{{e_{2}}}{{e_{2}}} \sim \text{const}\,,\end{aligned}$$ which is largely independent of the jet mass, $p_T$, and the angular exponent $\beta$. Stability of the maximal value (endpoint), though, is not sufficient to guarantee stability of the distribution. Indeed, the scaling of the lower boundary of the phase space for the background is ${{_{1}e_{3}}} \sim ({e_{2}})^2$, so we expect a sharp drop in the background, and therefore a peak in the distribution, around $$\begin{aligned}
\label{eq:M2peak}
M^{ \text{peak}}_2 \sim \frac{{{_{1}e_{3}}}}{{e_{2}}} \sim \frac{({e_{2}})^2}{{e_{2}}} \sim {e_{2}}\,.\end{aligned}$$ Simplifying again to the case of $\beta=2$, and using , we have $$\begin{aligned}
M^{(2), \text{peak}}_2\sim \frac{m_{J}^2}{p_{TJ}^2}\,,\end{aligned}$$ which depends sensitively on $m_J$ and $p_{TJ}$, but in exactly the opposite way as $D_2$. This behavior is observed in .
Finally, for $N_2$ shown in , the background region is defined by a single scaling, namely ${{_{2}e_{3}}} \sim ({e_{2}})^2$, which defines the upper boundary. Since there is a single scaling, we expect the peak for the background distribution to be defined by the same scaling. This means that $N_2$ has a maximum value and a peak location that both scale like $$\begin{aligned}
\label{eq:N2max}
N^{ \text{max,peak}}_2 \sim \frac{{{_{2}e_{3}}}}{({e_{2}})^2} \sim \frac{({e_{2}})^2}{({e_{2}})^2} \sim \text{const}\,,\end{aligned}$$ which is largely independent of the jet mass, $p_T$, and the angular exponent $\beta$. This is well verified in the parton shower analysis, as shown in . Thus, we see that by carefully engineering the phase space of an observable, one can achieve properties, such as stability, that are important experimentally. In this specific case, the stability of the full $N_2$ distribution gives further evidence that $N_2$ is a promising 2-prong tagger, even without grooming.
After grooming away soft radiation, we see from that all the distributions are stable, and from our power counting analysis, it is easy to understand why this is true. For $D_2$, grooming has a dramatic impact (note the change in the $x$-axis range), since it removes the region of phase space that leads to the undesired scaling behavior in (see also ). In this way, the endpoint for groomed $D_2$ (as well as the whole distribution) becomes remarkably robust to the jet mass cut. For the $M_2$ observable, the grooming removes the background in the bulk of phase space and pushes it to the upper boundary, as shown in , stabilizing the peak of the $M_2$ distribution but leaving the endpoint largely unchanged. After jet grooming, the parametric phase space for $N_2$ is unmodified, so the endpoint and peak scaling in should not change. Comparing , we see that the specific value of the $N_2$ endpoint and peak is modified, but the stability with varying mass cut is robust.
Therefore, in all cases after grooming, we have $$\begin{aligned}
\text{groomed}: \quad D^{ \text{max,peak} }_2 \sim \text{const}\,, \quad M^{ \text{max, peak}}_2\sim \text{const} \,, \quad N^{ \text{max,peak}}_2 \sim \text{const}\,.\end{aligned}$$ This demonstrates three distinct ways of generating a stable distribution: engineering the background phase space to directly have the desired boundary (e.g. $N_2$), or grooming soft radiation to the stabilize the boundary (e.g. $D_2$) or the peak (e.g. $M_2$) of the background distribution. It is important to emphasize that the power-counting analysis can only identify the power-law scaling of the distribution in $m_J$ or $p_{TJ}$. Removing this power-law scaling does not, however, guarantee complete numerical stability of the distribution. For this, techniques such as designing decorrelated taggers (DDT) [@Dolen:2016kst] can be used. We expect that methods like DDT will be most powerful when applied to variables that are already naturally stable, but we leave a study to future work.
Improving Quark/Gluon Discrimination {#sec:qvsg}
====================================
A major challenge in the field of jet substructure is reliable quark/gluon discrimination. Despite its many potential applications, there has been significant difficulty both in understanding the behavior of quark/gluon discriminants in parton showers, as well as in developing analytically-tractable observables which surpass the Casimir scaling limit (see below). For detailed discussions of these issues, we refer the reader to , as well as to studies in data [@CMS:2013wea; @CMS:2013kfa; @Aad:2014gea; @ATLAS-CONF-2016-034].
Quark/gluon discrimination has mostly been studied using IRC safe observables, such as the angularities [@Berger:2003iw; @Almeida:2008yp] or 2-point energy correlation functions $C_1 = {e_{2}}$ [@Larkoski:2013eya], which are set by a single emission at LL accuracy.[^19] At LL order, and ignoring nonperturbative effects, one can show that the discrimination power of such observables is set by the Casimir scaling relation $$\begin{aligned}
\label{eq:casimir_scaling}
\text{disc}(x)=x^{C_A/C_F}=x^{9/4}\,,\end{aligned}$$ where $x$ is the fraction of quarks retained by the cut and $\text{disc}(x)$ is the fraction of gluons retained. In this way, discrimination power is capped by the ratio of the gluon and quark color charges, $C_A/C_F = 9/4$. Casimir scaling arises because after a single emission, the discrimination power is set only by the color factor associated with the hard jet core, independent of the particular details of the observable.
Beyond LL accuracy, where one is sensitive to physics beyond the leading emission, improved discrimination power is observed. In , an analytic calculation of $C_1$ was performed at NLL accuracy, and a noticeable increase in discrimination power beyond the Casimir limit was found for $\beta<1$ (though not confirmed in an ATLAS study [@Aad:2014gea]). For small values of $\beta$, however, one is highly sensitive to nonperturbative effects, which must be modeled or extracted from data. Particularly for gluon jets, which are not well constrained by LEP event shape data [@Abdallah:2003xz; @Heister:2003aj; @Achard:2004sv; @Abbiendi:2004qz], this leads to significant discrepancies between distributions obtained from different parton shower generators.[^20] This in turn leads to rather large uncertainties in the predicted quark/gluon efficiencies; see for detailed studies.
Given the Casimir scaling limit of single-emission observables, a promising approach for improving quark/gluon discrimination is to design observables that are directly sensitive to multiple emissions within the jet, even at lowest order. In this section, we define a series of observables ${U_{i}}$ specifically intended for this purpose. Since these observables exhibit different behavior from standard single-emission observables, they may also prove useful in improving the parton shower description of quark and gluon jets. We will particularly emphasize the stability of their discrimination power as a function of the angular exponent $\beta$, which could be helpful for disentangling perturbative and nonperturbative effects.
Probing Multiple Emissions with ${U_{i}}$ {#sec:qvsg_setup}
-----------------------------------------
A standard observable for quark/gluon discrimination is the 2-point energy correlation function ${e_{2}}$, whose scaling was derived already in for 1-prong jets: $$\begin{aligned}
\label{eq:e2pc_qvsg}
{{e_{2}^{(\beta)}}}\sim z_s + \theta_{cc}^\beta\,.\end{aligned}$$ As discussed, ${e_{2}}$ is set at LL accuracy by a single emission from the hard core. Note that the scaling is the same for quarks and gluons, since $C_F = 4/3$ versus $C_A = 3$ is not a parametric difference between the samples.
To go beyond this single-emission behavior, we consider the 3-point correlators, ${{_{v}e_{3}}}$, which explicitly probe two emissions from the hard jet core. Using the modes in , we derive the following scalings (which were already given in ): $$\begin{aligned}
\label{eq:pc_qvsg}
{{_{1}e_{3}^{(\beta)}}} &\sim z_s^2+\theta_{cc}^\beta \,, \nonumber\\
{{_{2}e_{3}^{(\beta)}}} &\sim z_s^2+ z_s \theta_{cc}^\beta +\theta_{cc}^{2\beta}\,, \nonumber\\
{{_{3}e_{3}^{(\beta)}}} &\sim z_s^2+ z_s \theta_{cc}^\beta +\theta_{cc}^{3\beta}\,.\end{aligned}$$ We can draw a number of interesting conclusions from . First, in the majority of phase space there is a direct relationship between the last two 3-point correlators and the 2-point correlator: ${{_{2}e_{3}}}\sim ({e_{2}})^2$ and ${{_{3}e_{3}}}\sim ({e_{2}})^2$.[^21] We therefore do not expect ${{_{2}e_{3}}}$ or ${{_{3}e_{3}}}$ to yield improved quark/gluon discrimination power compared to ${e_{2}}$; this illustrates the importance of understanding parametric correlations between different observables. By contrast, ${{_{1}e_{3}}}$ does not obey such a relation to ${e_{2}}$, since only for ${{_{1}e_{3}}}$ is the cross term $\theta_{cc}^\beta z_s$ power suppressed. Since ${{_{1}e_{3}}}$ directly probes the double-soft limit of a jet, without soft/collinear cross talk at leading power, we can expect it to carry more information about the flavor of the jet’s initiating parton. This intuition will be verified in our parton shower study.
Another interesting feature of ${{_{1}e_{3}}}$ is the relative scaling between the collinear and soft modes, as can be seen from comparing to . To improve quark/gluon efficiency with ${e_{2}}$, one typically needs to use small values of the angular exponent $\beta$. Since ${{_{1}e_{3}}}$ already has a suppressed soft scaling, it can achieve good quark/gluon discrimination at comparatively higher values of the angular exponent. In the parton shower study below, we will find that the performance of ${{_{1}e_{3}}}$ with $\beta=2$ is comparable to ${e_{2}}$ with $\beta=0.2$. This relative scaling also modifies the structure of nonperturbative corrections, although we will not discuss this aspect further in this paper.[^22] Note that the discrimination power as a function of $\beta$ is not a prediction of power counting and can only be obtained by explicit calculations (or measurements) of the distributions.
Seeing the potential of ${{_{1}e_{3}}}$, it is natural to consider higher-point correlators. For an $n+1$ point correlator, we have $$\begin{aligned}
\label{eq:U_pc}
{{_{1}e_{n+1}^{(\beta)}}} \sim z_s^n+ \theta_{cc}^\beta \,,\end{aligned}$$ which probes the $n$-soft limit, again without soft/collinear cross talk at leading power. We are therefore led to define the ${U_{i}}$ series of observables, $$\begin{aligned}
{U_{i}^{(\beta)}}={{_{1}e_{i+1}^{(\beta)}}}\,,\end{aligned}$$ for quark/gluon discrimination. The reason one might expect ${U_{i}}$ to perform better with increasing $i$ is that higher-point correlators can effectively “count” more emissions than lower-point correlators. Since gluon jets generate more emissions than quark jets, on average by a factor of $C_A/C_F$, one expects improved quark/gluon contrast with each additional emission probed; this intuition will be borne out in the parton shower study below. More generally, we hope that these observables will prove useful for probing the structure of the QCD shower.
From the power counting in , we see that the scaling of the soft modes for ${U_{i}}$ depends on the index $i$ as $z_s^i$. One might therefore naively think that after grooming is applied, all the ${U_{i}}$ observables would be identical. This is not the case for a fixed value of $\zcut$, however, since the soft scale increases as a function of $i$. To emphasize this point, the average values of ${U_{i}}$ are typically $\langle {U_{2}} \rangle =0.05$ and $\langle {U_{3}} \rangle =0.01$ (see from our parton shower study below). By , these correspond to $z_s$ values of $z_s\simeq 0.25$ and $z_s\simeq 0.4$, respectively, both of which are well above the $\zcut=0.1$ scale that we use as our grooming benchmark. Therefore, the emissions that dominate the ${U_{2}}$ and ${U_{3}}$ distributions are not actually removed by our grooming procedure. Thus, the behavior of $U_i$ is expected to be more resilient to grooming for larger values of $i$.
Performance in Parton Showers {#sec:qvsg_mc}
-----------------------------
We now use a parton shower study to verify the above power-counting predictions and to assess quantitatively the potential improvements in quark/gluon discrimination power achievable using higher-point correlators. For reasons of computational time we restrict our study of the ${U_{i}}$ series to $i=1,2,3$.[^23] The quark and gluon jets are generated from the same <span style="font-variant:small-caps;">Pythia</span> $pp \to Z+j$ samples described in , and the same overall analysis strategy applies, though no cut is placed on jet masses. Furthermore, we use a smaller jet radius of $R=0.6$. Given known parton shower uncertainties, it would be interesting to study different shower and hadronization algorithms to understand the degree to which LHC measurements of ${U_{i}}$ could provide insight into quark/gluon tagging; we leave such studies to future work.
We begin by verifying the power-counting argument of , which suggested that ${{_{2}e_{3}}}$ and ${e_{3}}$ should be highly correlated with $U_1 = C_1 = {e_{2}}$. Even though ${{_{2}e_{3}}}$ and ${e_{3}}$ probe three particle correlations, they have a fixed scaling relation with respect to ${e_{2}}$, and are therefore not expected to provide new information for quark/gluon tagging. Taking ${{_{2}e_{3}}}$ as a representative example in , we compare the distributions of ${{_{2}e_{3}}}$ and $\frac{1}{2}({e_{2}})^2$; they are remarkably similar so we conclude that power counting is indeed capturing the dominant scaling relation. From the ROC curves in , we see that the discrimination power of ${e_{2}}$, ${{_{2}e_{3}}}$, and ${{_{3}e_{3}}}$ are very similar for the same value of $\beta$, with limited improvement observed by including 3-particle correlations. This emphasizes that probing multi-particle correlations does not, in and of itself, improve quark/gluon discrimination, since higher-point correlation functions can be correlated with lower-point correlation functions.
We now consider the behavior of ${U_{2}}$ and ${U_{3}}$, which were designed to exploit multi-particle correlations to improve quark/gluon discrimination. In , we show distributions of ${U_{2}}$ and ${U_{3}}$ with $\beta=0.2$, indicating good separation of the quark and gluon samples. This is quantified in , which shows ROC curves for ${U_{i}}$ comparing $i=1,2,3$. Recall that ${U_{1}}= C_1 = {e_{2}}$ is a standard quark/gluon discriminant and a useful baseline to assess performance gains (even if <span style="font-variant:small-caps;">Pythia</span> itself skews optimistic about quark/gluon separation power [@Larkoski:2013eya; @Larkoski:2014pca]). Going from ${U_{1}}$ to ${U_{2}}$ to ${U_{3}}$, the discrimination power at high efficiencies does increase with more emissions being probed, though the change is relatively small going from $i = 2$ to $i = 3$.
Beyond absolute performance gains, it is also interesting to study the relative performance of ${U_{i}}$ as a function of the angular exponent $\beta$. In , we show the gluon rejection at $70\%$ quark efficiency as a function of $\beta$.[^24] Unlike for $U_1 = {e_{2}}$, where the discrimination power falls off rapidly with increasing $\beta$, for ${U_{2}}$, and even more so for ${U_{3}}$, the discrimination power remains well above the Casimir scaling limit, even into the large $\beta$ regime where ${U_{i}}$ should be amenable to fixed-order or resummed perturbative calculations. We find this much flatter behavior of the discrimination power with respect to $\beta$ to be one of the most interesting features of these observables, suggestive that multiple soft emissions are just as important as hard collinear emissions for discriminating quarks from gluons. Full ROC curves for different values of the angular exponents are provided in .
It would be interesting to see if there is asymptotic behavior as $i \to \infty$, though this is likely only meaningful in the context of a comparative study of parton shower generators, since it depends sensitively on the assumptions made for correlated soft emissions. As a first step in this direction, in we compare $U_i$ to hadron multiplicity, which is known to be a powerful quark/gluon discriminant. Remarkably, the performance of the ${U_{i}}$ observables appears to asymptote to multiplicity as $i$ is increased, both in the shape of the ROC curves as well as in the behavior as a function of $\beta$. It would be interesting to understand whether this connection can be made formal, and whether the ${U_{i}}$ observables can be used to give an IRC safe definition of a multiplicity-like observable.
Finally, we want to test whether this improvement in quark/gluon discrimination power is robust to grooming. In , we compare the ${U_{2}}$ and ${U_{3}}$ distributions before and after grooming has been applied. At large values of the observables, relatively little difference is observed for our baseline grooming parameters, as expected from the power-counting analysis of . At smaller values of the observables, there is a distortion in the distributions due to the fact that grooming substantially decreases the overall particle multiplicity. In particular, there are expected features at ${U_{2}} = 0$ (${U_{3}} = 0$), from when the grooming gives less than three (four) particles in the jet. In this regime, power-counting arguments are no longer applicable since the distribution is dominated by nonperturbative effects. That said, as shown in , the ROC curves after grooming exhibit the same features as in the ungroomed case, with ${U_{2}}$ and ${U_{3}}$ outperforming ${U_{1}}$, indicating that this parametric prediction is still robust.
It would be of great interest to perform explicit calculations of ${U_{2}}$ to understand its exact dependence on the color Casimirs, as well as on the angular exponent $\beta$. A resummed calculation, in particular, would shed light onto the all-orders structure of multiple-emission observables, which have not been widely explored in the literature.[^25] It would also be useful to understand whether the measurement of multiple ${U_{i}}$ observables with different $\beta$ values could be used to improve quark/gluon discrimination. The multi-differential cross section for ${U_{1}}={e_{2}}$ with two different angular exponents was calculated in and the gains in performance for quark/gluon discrimination were studied in from the perspective of mutual information. In preliminary investigations, we find that correlations among the $U_i$ are indeed helpful, but we leave a detailed study to future work.
Conclusions {#sec:conc}
===========
Continued progress in jet substructure relies on the ability to devise observables that can probe increasingly detailed aspects of jets. In this paper, we used the known structures imposed by IRC safety to motivate the generalized energy correlation functions, ${{_{v}e_{n}}}$, a flexible basis for constructing new substructure discriminants. These generalized correlators incorporate an angular weighting function, allowing them to probe different angular structures within a jet. We presented a number of case studies of relevance to the jet substructure community—boosted top tagging, boosted $W/Z/H$ tagging, and quark/gluon discrimination—demonstrating the power of power-counting techniques to design discriminants for specific purposes. In each case, our newly-developed observables outperform standard jet shapes in parton shower studies.
The three series of observables introduced in this paper—${M_{i}}$, ${N_{i}}$, and ${U_{i}}$—exhibit new ways to probe the soft and collinear limits of QCD. The ${M_{i}}$ series is designed for tagging groomed jets, showing that the removal of soft radiation can dramatically change the phase space of $i$-prong discriminants. The ${N_{i}}$ series is designed to mimic $N$-subjettiness in the limit of resolved substructure, showing how to probe radiation patterns around collinear prongs without requiring external axes. Finally, the ${U_{i}}$ series is designed to evade the usual quark/gluon limitations imposed by Casimir scaling, showing the importance of multiple soft emissions for quark/gluon radiation patterns. Taken together, these observables widen the scope for jet substructure investigations, allowing more handles to optimally use jets at the LHC.
Given their tagging performance, it would be interesting to calculate these observables from first principles. This would provide insights into the impact of jet grooming on multi-prong observables, the difference between axes-based and axes-free observables, and the structure of multiple emissions within quark/gluon jets. We are particularly interested in the differences between groomed and ungroomed distributions, since jet grooming not only changes the power counting of observables, but it also changes the logarithmic structure and power corrections in analytic calculations [@Dasgupta:2013ihk; @Dasgupta:2013via; @Larkoski:2014wba; @Dasgupta:2015yua; @Frye:2016okc; @Frye:2016aiz]. Beyond jet substructure, we suspect that the generalized correlators could eventually be useful as a tool for performing NNLO calculations; powerful slicing schemes have been devised using $N$-jettiness [@Boughezal:2015dva; @Gaunt:2015pea] and ${{_{v}e_{n}}}$-based slicing could potentially be valuable in regimes where axes are inappropriate or cumbersome.
One aspect of jet substructure that has not been studied here is the correlations between discriminants. We did apply power-counting techniques to identify correlations among basis elements to define optimal discriminants, but we did not consider whether power-counting could reveal parametric relationships between different proposed discriminants. Along similar lines, we did not consider in detail the hybrid strategy of using both groomed and ungroomed observables. In preliminary investigations, we find that, not surprisingly, discriminants with the same power counting are highly correlated. When discriminants have different power counting, though, there appears to be additional information gained through multi-variate combinations. At the moment, our application of power counting does not tell us what these multi-variate correlations are or whether we can robustly predict high-performing combinations. We look forward to developing more sophisticated power-counting strategies to exploit these correlations in the future.
Finally, we want to emphasize the importance of first-principles calculations and unfolded experimental measurements of $U_1$, $U_2$, and $U_3$. While the expected tagging performance of 2- and 3-prong discriminants—like ${M_{2}}$, ${N_{2}}$, ${D_{2}^{(1,2)}}$, and ${N_{3}}$—can be seen directly from power-counting arguments, this is not the case for quark/gluon discriminants, since $C_F$ and $C_A$ are not parametrically different quantities. For 1-prong jets, power counting can tell us which soft/collinear features are probed by the $U_i$ series, but it cannot reliably predict their expected parametric behavior or relative performance. In parton shower studies, we do find that ${U_{2}}$ and ${U_{3}}$ exhibit improved performance over naive Casimir scaling, even in the larger $\beta$ regime where they are under better perturbative control, suggesting that the $U_i$ series is a sensitive probe of the QCD shower. Therefore, measurements of the $U_i$ series, along with comparisons to parton shower (and eventually analytic) predictions, are likely to lead to deeper understanding of jets in QCD.
We thank Philip Harris, Andrew Larkoski, Simone Marzani, Ben Nachmann, Sid Narayanan, Duff Neill, Sal Rappoccio, and Nhan Tran for helpful discussions, and we thank Matteo Cacciari, Gavin Salam, and Gregory Soyez for help developing the `EnergyCorrelator` [<span style="font-variant:small-caps;">FastJetcontrib</span>]{}. IM is supported by the U.S. Department of Energy (DOE) under cooperative research agreement DE-SC0011090. The work of LN and JT is supported by the DOE under grant contract numbers DE-SC-00012567 and DE-SC0015476. JT is also supported by a Sloan Research Fellowship from the Alfred P. Sloan Foundation. This work was performed in part at the Aspen Center for Physics, which is supported by National Science Foundation grant PHY-1066293.
Alternative Angular Weighting Functions {#app:alt}
=======================================
As discussed in , any symmetric function of the angles, $f_N( \hat p_{i_1}, \hat p_{i_2}, \ldots, \hat p_{i_N} )$, that vanishes in the collinear limits can in principle be used in . While we argued in that the $\text{min}$ function is particularly effective due to its ability to isolate hierarchical angular structures, other functional forms can certainly be used. In this appendix, we study two alternate definitions of the angular weighting function, which, from a power counting perspective, are identical to those considered in the text.
For concreteness, we study variants of the $N_2$ observable from , which was based on a 3-point correlator: $${{_{2}e_{3}^{(\beta)}}}=\sum_{1\leq i<j<k\leq n_J} z_{i}z_{j}z_{k} \min \left\{\theta_{ij}^\beta \theta_{ik}^\beta\,, \theta_{ij}^\beta \theta_{jk}^\beta\,, \theta_{ik}^\beta \theta_{jk}^\beta \right\} \qquad \Rightarrow \qquad N_2^{(\beta)}=\frac{_2e_3^{(\beta)}}{( {{_{1}e_{2}^{(\beta)}}})^2}.$$ One variant is to consider an angular weighting function that smoothly approximates the $\text{min}$ function.[^26] $$\begin{aligned}
\label{eq:e3j}
_2r_3^{(\beta)} = \sum_{1\leq i<j<k\leq n_J} z_{i} z_{j} z_{k} \left( \frac{1}{\theta_{ij} \theta_{ik}} + \frac{1}{\theta_{ij} \theta_{jk}} + \frac{1}{\theta_{ik} \theta_{jk} } \right)^{-\beta} \qquad \Rightarrow \qquad R_2^{(\beta)}=\frac{_2r_3^{(\beta)}}{( {{_{1}e_{2}^{(\beta)}}})^2}\,.\end{aligned}$$ Another variant is to use the geometric fact that in the collinear limit, the minimum product of pairwise distances is parametrically the same as the area of the triangle spanned by the three points[^27] $$\begin{aligned}
_2a_3^{(\beta)} = \sum_{1\leq i<j<k\leq n_J} z_i z_j z_k \left(\sqrt{s (s-\theta_{ij}) (s-\theta_{ik})(s-\theta_{jk})} \right)^\beta \qquad \Rightarrow \qquad A_2^{(\beta)}=\frac{_2a_3^{(\beta)}}{( {{_{1}e_{2}^{(\beta)}}})^2}\,,\end{aligned}$$ where $s = (\theta_{ij}+\theta_{jk}+\theta_{ik})/2$ comes from Heron’s formula. While $A_2$ is parametrically identical to $N_2$, it has the interesting property that it vanishes when the vectors defining the three particles are coplanar, similar to dipolarity introduced in .
Even though the ${N_{2}}$, $R_2$, and $A_2$ observables have identical power counting, their distributions could in principle differ by $\mathcal{O}(1)$ numbers, possibly allowing for improved discrimination power. In we compare the distributions of these three observables in <span style="font-variant:small-caps;">Pythia</span>, showing that they are rather similar. To aid the eye, we have rescaled the $R_2$ and $A_2$ distributions to match the $N_2$ distribution. Turning to the $Z$ versus quark ROC curve in , the performance is nearly identical. This further emphasizes that the behavior of the observables is dominated by parametric scalings. Since we did not find any gains from using these more complicated variants, we restricted the study in the text to the definition given in .
It is still an interesting question whether other choices of angular weighting functions might lead to improved performance in more complicated jet substructure applications. It seems unlikely, however, since for small radius jets, one can Taylor expand the angular function in the small $\theta$ limit, and observables with the same power counting must have the same lowest-order expansion. In practice, the use of smoother definitions which approximate the $\text{min}$ function might be useful for performing perturbative calculations.
Aspects of 3-prong Tagging
==========================
Challenges for $M_3$ {#app:add_plots_M3}
--------------------
In , we defined the general series of ${M_{i}}$ observables. We saw in that the ${M_{2}}$ observable was an effective boosted $W/Z/H$ tagger on groomed jets. One might therefore consider the ${M_{3}}$ observable, $$\begin{aligned}
{M_{3}^{(\beta)}}=\frac{{{_{1}e_{4}^{(\beta)}}}}{{{_{1}e_{3}^{(\beta)}}}}\,,\end{aligned}$$ as a possible boosted top tagger.
We can see from a power-counting analysis, however, that even with grooming, ${M_{3}}$ will not perform well. Following the notation of , a strongly-ordered 3-prong jet has $$\begin{aligned}
\text{3-prong signal (groomed):} \qquad {{_{1}e_{3}^{(\beta)}}}&\sim \theta_{23}^\beta\,, \nn \\
{{_{1}e_{4}^{(\beta)}}}&\sim z_{ccs} \theta_{23}^\beta + \theta_{cc}^\beta \,,\end{aligned}$$ while a 2-prong background jet has $$\begin{aligned}
\label{eq:M3_pc_bkg}
\text{2-prong background (groomed):} \qquad {{_{1}e_{3}^{(\beta)}}}&\sim z_{cs} \theta_{cs}^\beta +\theta_{cc}^\beta \,, \nn \\
{{_{1}e_{4}^{(\beta)}}}&\sim z_{cs}^2 \theta_{cs}^\beta + \theta_{cc}^\beta \,.\end{aligned}$$ For signal jets, we have the relation ${{_{1}e_{4}}}\ll {{_{1}e_{3}}}$, so we would like the background to satisfy ${{_{1}e_{4}}}\sim {{_{1}e_{3}}}$. That desired relation is violated, though, by contributions of the collinear-soft modes to ${{_{1}e_{4}}}$, due to the different $z_{cs}$ scalings in . We therefore predict from power counting that ${M_{3}}$ should be a poor discriminant.
In , we show the distribution of ${M_{3}}$ for boosted top jets compared to those from QCD jet backgrounds, where little discrimination power is observed. Similar to how ordinary grooming was required for ${M_{2}}$ to become an effective discriminant in the 2-prong case, it is likely that another layer of grooming is be needed to remove the undesired collinear-soft contributions to ${M_{3}}$ and make it an effective 3-prong tagger. While we do not pursue ${M_{3}}$ further in this paper, it would be interesting to consider alternative grooming methods designed to isolate 3-prong structure and mitigate both soft and collinear-soft radiation. As a starting point, one could consider doubly-soft-dropped boosted top jets, where after an initial application of soft drop, one reapplies soft drop to the two remaining prongs.
![Distributions of ${M_{3}}$, comparing the signal of boosted top jets to the background of $b$-quarks, light quarks, and gluons. As expected from power counting, limited discrimination is observed. []{data-label="fig:app_M3"}](figures/dist_M3_M3_2.pdf){width="6.5cm"}
$N_3$ Without Grooming {#app:add_plots_N3}
----------------------
In , we argued that on groomed jets with well-resolved substructure, ${N_{3}}$ behaves parametrically like ${\tau_{3,2}}$, but exhibits improved discrimination power in the transition to the unresolved region. On ungroomed jets, however, ${N_{3}}$ behaves differently from ${\tau_{3,2}}$, and in particular, it does not provide good discrimination in regions of phase space where there is a soft wide-angle subjet. This same issue was discussed in detail for the case of $D_3$ in ; the treatment of the soft subjet region of phase space required the addition of two extra terms to $D_3$, leading to the complicated form shown in . To avoid the soft subjet issue, and to advocate for the stability of groomed observables, we explicitly focused on the case of groomed top jets in .
Here, we compare ${N_{3}}$ and ${\tau_{3,2}}$ on ungroomed jets. Though ${N_{3}}$ was not designed for use on ungroomed jets, it still provides reasonably good discrimination power, though not as good as ${\tau_{3,2}}$. Distributions of ungroomed ${N_{3}}$ are shown in , where we use an alternative mass window cut of $m \in [160, 240]$ GeV. The discrimination performance for the top signal against the $b$-quark, light quark, and gluon jet backgrounds are shown in , respectively. The best performance is seen in rejecting quark jets, although ungroomed $N_3$ has worse performance on gluon jets. Interestingly, similar quark/gluon differences were seen for $D_3$ in , although the nature of this behavior is not understood and is not necessarily connected in any way to the use of energy correlators.
Though ${N_{3}}$ was designed for use on groomed jets, we believe that ${N_{3}}$ is a sufficiently good discriminant on ungroomed jets to merit further investigations. At minimum, ungroomed ${N_{3}}$ distributions could be measured as a baseline to test the impact of jet grooming. We offer a bounty to the first group that identifies an axes-free observable with the same power counting as ungroomed ${\tau_{3,2}}$.
For completeness, in , we show the $N$-subjettiness observable ${\tau_{3,2}}$ as measured on the same samples, both before and after grooming. As expected, excellent discrimination power is observed is observed before grooming. After grooming, the discrimination power is worsened primarily due to the behavior in the unresolved region, namely as ${\tau_{3,2}}\to 1$. It is in this region that ${N_{3}}$ exhibits improved performance, as seen already in the behavior of the distributions in and the performance in the ROC curve in .
![Distributions of the $N$-subjettiness ratio ${\tau_{3,2}}$ (a) before grooming and (b) after grooming for both the boosted top signal and the different QCD backgrounds. []{data-label="fig:app_tau32"}](figures/dist_tau3N3_q_b_g.pdf "fig:"){width="6.5cm"} ![Distributions of the $N$-subjettiness ratio ${\tau_{3,2}}$ (a) before grooming and (b) after grooming for both the boosted top signal and the different QCD backgrounds. []{data-label="fig:app_tau32"}](figures/dist_tau3_SD_q_b_g.pdf "fig:"){width="6.5cm"}
Identifying ${N_{3}}$ {#app:N3_identify}
---------------------
In , we considered the observable ${N_{3}}$ defined as $$\begin{aligned}
{N_{3}^{(\beta)}}=\frac{{{_{2}e_{4}^{(\beta)}}}}{({{_{1}e_{3}^{(\beta)}}})^2}\,.\end{aligned}$$ There are, however, a large number of other possible observables that could be formed from combinations of the different 2- , 3-, and 4-point correlators. In this appendix, we describe in more detail the justification for our focus on ${N_{3}}$. It is interesting that this process happens to identify an observable with the same parametric behavior as the $N$-subjettiness ratio ${\tau_{3,2}}$. As discussed in the text, we focus on the case of groomed jets. This means that we can ignore soft radiation for our power counting analysis.
For groomed boosted top jets, it is sufficient to consider a 3-prong configuration with hierarchical angles, as illustrated in . In particular, we do not have to consider the soft subjet phase space region from , which has hierarchical energies, since those configurations are removed by the grooming procedure. For the 3-prong signal, the scaling of the 2-point correlator is $$\begin{aligned}
{{e_{2}^{(\beta)}}}\sim \theta_{12}^\beta\,,\end{aligned}$$ the scalings of different 3-point correlators are $$\begin{aligned}
{{_{3}e_{3}^{(\beta)}}}&\sim \theta_{23}^\beta \theta_{12}^{2\beta}\,, \nn \\
{{_{2}e_{3}^{(\beta)}}}&\sim \theta_{23}^\beta \theta_{12}^{\beta}\,, \nn \\
{{_{1}e_{3}^{(\beta)}}}&\sim \theta_{23}^\beta \,.\end{aligned}$$ and the scalings of the different 4-point correlators are $$\begin{aligned}
\label{eq:app_pc_4point}
{{_{6}e_{4}^{(\beta)}}}&\sim z_{cs} \theta_{12}^{5\beta} \theta_{23}^{\beta} +z_{ccs} \theta_{12}^{3\beta} \theta_{23}^{3\beta}
+\theta_{23}^\beta \theta_{cc}^\beta \theta_{12}^{4\beta} \,, \nn \\
{{_{5}e_{4}^{(\beta)}}}&\sim z_{cs} \theta_{12}^{4\beta} \theta_{23}^{\beta} + z_{ccs} \theta_{12}^{2\beta} \theta_{23}^{3\beta}
+ \theta_{23}^\beta \theta_{cc}^\beta \theta_{12}^{3\beta}\,, \nn \\
{{_{4}e_{4}^{(\beta)}}}&\sim z_{cs} \theta_{12}^{3\beta} \theta_{23}^{\beta} + z_{ccs} \theta_{12}^{1\beta} \theta_{23}^{3\beta}
+\theta_{23}^\beta \theta_{cc}^\beta \theta_{12}^{2\beta}\,, \nn \\
{{_{3}e_{4}^{(\beta)}}}&\sim z_{cs} \theta_{12}^{2\beta} \theta_{23}^{\beta} +z_{ccs} \hphantom{\theta_{12}^{3\beta}} \theta_{23}^{3\beta}
+\theta_{23}^\beta \theta_{cc}^\beta \theta_{12}^{\beta}\,, \nn \\
{{_{2}e_{4}^{(\beta)}}}&\sim z_{cs} \theta_{12}^{\beta} \theta_{23}^{\beta} + z_{ccs} \hphantom{\theta_{12}^{3\beta}} \theta_{23}^{2\beta}
+\theta_{23}^\beta \theta_{cc}^\beta \hphantom{\theta_{12}^{4\beta}}\,, \nn \\
{{_{1}e_{4}^{(\beta)}}}&\sim 0\hphantom{z_c \theta_{12}^{\beta} \theta_{23}^{\beta} }
+z_{ccs} \hphantom{\theta_{12}^{5\beta}} \theta_{23}^{\beta}
+\hphantom{\theta_{23}^\beta} \theta_{cc}^\beta \hphantom{\theta_{12}^{4\beta}}\,,\end{aligned}$$ where the alignment and zero in the last line are there just to help guide the eye.
For 2-prong background jets, all we need is the scaling of the 2-point correlator, $$\begin{aligned}
{{e_{2}^{(\beta)}}}\sim \theta_{cs}^\beta\,,\end{aligned}$$ and the scalings of the different 3-point correlators, $$\begin{aligned}
{{_{3}e_{3}^{(\beta)}}}&\sim z_{cs} \theta_{cs}^{3\beta} +\theta_{cs}^{2\beta} \theta_{cc}^\alpha\,, \nn \\
{{_{2}e_{3}^{(\beta)}}}&\sim z_{cs} \theta_{cs}^{2\beta} +\theta_{cs}^{\beta} \theta_{cc}^\alpha\,, \nn \\
{{_{1}e_{3}^{(\beta)}}}&\sim z_{cs}\theta_{cs}^{\beta} + \theta_{cc}^\alpha \,.\end{aligned}$$ While the background scalings of the 4-point correlators would be needed to verify signal/background separation, as was done in , they are not needed to restrict the combinations under consideration. Since their form is not particularly illuminating, we do not show them here.
While there are a large number of observables listed above, the analysis can be simplified by noting that for both the signal and background, the information contained in the 3-point correlators ${{_{2}e_{3}}}$ and ${{_{3}e_{3}}}$ is redundant, since it can be expressed in terms of ${e_{2}}$ and ${{_{1}e_{3}}}$. Furthermore, any observable derived from power counting will be linear in the 4-point correlator and will have the 3-point correlator appearing in the denominator raised to some power. Finally, from , we see that $\theta_{23}$ appears at most raised to the third power; it therefore suffices to consider ${{_{1}e_{3}}}$ raised at most to the third power. The power of ${e_{2}}$ is then fixed by Lorentz invariance.
The above logic allows us to write down a parametrically complete set of potential 3-prong observables, $$\begin{aligned}
\mathcal{O}_{v,y} = \frac{{{_{v}e_{4}^{(\beta)}}} \left( {{e_{2}^{(\beta)}}} \right)^{y-v} }{ \left( {{_{1}e_{3}^{(\beta)}}} \right)^y}\,, \qquad v\in \{1,2,3,4,5,6\}\,, \qquad y\in \{1,2,3\}\,, \qquad y\leq v\,.\end{aligned}$$ At this point, one can then either power count each of these options explicitly to test for background isolation, or simply evaluate their performance in a parton shower generator. To limit the number of options to consider, one can apply the further constraint that ${e_{2}}$ should not appear explicitly in the observable, to mitigate correlations with the jet mass. This is equivalent to setting $y=v$, and gives $$\begin{aligned}
T_{v}=\frac{{{_{v}e_{4}^{(\beta)}}} }{ \left( {{_{1}e_{3}^{(\beta)}}} \right)^v}\,.\end{aligned}$$ Note that $v=1$ gives ${M_{3}}$ and $v=2$ gives ${N_{3}}$. Among all of the $\mathcal{O}_{v,y}$ observables, we found that the best performing one in <span style="font-variant:small-caps;">Pythia</span> was ${N_{3}}$, which then became the focus of our boosted top study.
Power Counting ${N_{3}}$ {#app:N3_pc}
------------------------
While the identification of the parametrically optimal discriminant is usually fairly straightforward given the parametric expressions for the observables, confusions can arise when the scalings have multiple terms. Here, we present more details for the signal analysis of the ${N_{3}}$ observable from , to illustrate how power counting can be performed systematically. This allows one to avoid potential confusions when there are competing parametric relations. This same approach can be used in the other examples studied in the paper, though for the 1- and 2-prong case studies, we find that the more heuristic treatment in the text is just as illuminating as the systematic strategy.
We begin by recalling the power counting for ${{_{1}e_{3}}}$ and ${{_{2}e_{4}}}$, considering the signal with (hierarchical) 3-prong substructure: $$\begin{aligned}
{{_{1}e_{3}^{(\beta)}}}&\sim \theta_{23}^\beta\,, \nn \\
{{_{2}e_{4}^{(\beta)}}}&\sim z_{cs} \theta_{12}^\beta\theta_{23}^\beta +z_{ccs} \theta_{23}^{2\beta} + \theta_{23}^\beta \theta_{cc}^\beta \,.
\label{eq:appB4scaling}\end{aligned}$$ We next need to identify which of the parameters—$\theta_{12}$, $\theta_{23}$, $\theta_{cc}$, $z_{cs}$, and $z_{ccs}$—are set by which measurements. Since most boosted top analyses apply a mass cut, we assume that $\theta_{12}$ is set by a mass measurement. This is not crucial, however, and the argument below can be generalized without the fixed-mass assumption. This leaves us with the task of determining the parametric relationship between $\{\theta_{23},\theta_{cc},z_{cs},z_{ccs}\}$ and $\{{{_{1}e_{3}}},{{_{2}e_{4}}}\}$. Clearly, the measurement of ${{_{1}e_{3}}}$ sets $\theta_{23}$. By assumption, there is no hierarchy between the three terms in ${{_{2}e_{4}}}$, yielding the following scaling of the kinematic variables: $$\begin{aligned}
\label{eq:appB4pararelation}
\theta_{23}^\beta \sim {{_{1}e_{3}^{(\beta)}}}\,, \qquad \theta_{cc}^\beta \sim \frac{{{_{2}e_{4}^{(\beta)}}}}{{{_{1}e_{3}^{(\beta)}}}}\,, \qquad z_{cs}\sim \frac{{{_{2}e_{4}^{(\beta)}}}}{\theta_{12}^\beta {{_{1}e_{3}^{(\beta)}}}} \,, \qquad z_{ccs}\sim \frac{{{_{2}e_{4}^{(\beta)}}}}{({{_{1}e_{3}^{(\beta)}}})^2}\,.\end{aligned}$$
Now, we want to derive an observable which distinguishes 3-prong jets from jets with fewer than 3 prongs. This can be accomplished by identifying the regions in phase space where the 3-prong EFT description breaks down, and translating that into constraints on the relationship between ${{_{1}e_{3}}}$ and ${{_{2}e_{4}}}$. As given in and illustrated in , 3-prong phase space is defined by the following four conditions: $$\begin{aligned}
{2}
&(a)~~ \theta_{23} \ll \theta_{12}\ll 1\,, \qquad &
(b)~~ \theta_{cc} \ll \theta_{23}\,, \nn \\
&(c)~~ z_{ccs}\ll 1\,, &
(d)~~ z_{cs} \ll z_{ccs}\,, \end{aligned}$$ Plugging into condition $(a)$, we find $$\begin{aligned}
(a)\implies {{_{1}e_{3}^{(\beta)}}} \ll {{e_{2}^{(\beta)}}}\ll 1\,.\end{aligned}$$ This just defines the region of validity of our analysis, but is not helpful in determining a relationship between ${{_{1}e_{3}}}$ and ${{_{2}e_{4}}}$. Turning to condition $(b)$, we find $$\begin{aligned}
(b) \implies \frac{{{_{2}e_{4}^{(\beta)}}}}{{{_{1}e_{3}^{(\beta)}}}} \ll {{_{1}e_{3}^{(\beta)}}}\implies \frac{{{_{2}e_{4}^{(\beta)}}}}{({{_{1}e_{3}^{(\beta)}}})^2} \ll 1\,.\end{aligned}$$ Note that the constraint $\theta_{cc}\ll1$ does not give as strong a bound. Condition $(c)$ gives the same constraint, $$\begin{aligned}
(c) \implies \frac{{{_{2}e_{4}^{(\beta)}}}}{({{_{1}e_{3}^{(\beta)}}})^2} \ll 1\,.\end{aligned}$$ Finally, we see that condition $(d)$ is already satisfied by condition $(a)$, $$\begin{aligned}
(d) \implies \frac{{{_{2}e_{4}^{(\beta)}}}}{\theta_{12}^\beta {{_{1}e_{3}^{(\beta)}}}} \ll \frac{{{_{2}e_{4}^{(\beta)}}}}{({{_{1}e_{3}^{(\beta)}}})^2} \implies {{_{1}e_{3}^{(\beta)}}} \ll \theta_{12}^\beta.\end{aligned}$$ and provides no extra information.
From this analysis, one finds that the strongest constraint on the breakdown of the 3-prong EFT is $$\begin{aligned}
\frac{{{_{2}e_{4}^{(\beta)}}}}{({{_{1}e_{3}^{(\beta)}}})^2} \ll 1\,,\end{aligned}$$ leading to the definition of the ${N_{3}}$ observable, $$\begin{aligned}
{N_{3}^{(\beta)}}=\frac{{{_{2}e_{4}^{(\beta)}}}}{({{_{1}e_{3}^{(\beta)}}})^2}\,.\end{aligned}$$ With practice, one can immediate infer this result from the scaling of the observables in , without having to explicitly consider each EFT constraint, but this example illustrates how the procedure can be performed systematically when confusions arises.
Relationship Between $N_i$ and $N$-subjettiness {#app:Nsub_Ni}
===============================================
In , we claimed that the ${N_{i}}$ observables and the $N$-subjettiness ratio observables are related for groomed jets. This was shown explicitly for the case of $i=3$ in . In this appendix, we show that this is generically true, suggesting that ${N_{i}}$ is indeed an appropriate observable for identifying $i$-prong substructure on groomed jets.
Since we work with groomed jets, we do not have to consider soft subjet configurations (i.e. $i$-prong jets with hierarchical energies). Instead, the power counting is determined by the generalization of with hierarchical angles, where a jet has $i$ subjets, two of which become collinear and approach an $(i-1)$-subjet configuration. We label the two subjets that approach each other by $1$ and $2$, such that $\theta_{12}$ denotes the angle between them. By assumption, $\theta_{12}$ is smaller than the angles between any other subjets (which we power count as $\theta_{st} \sim 1$), but larger than the typical collinear scale $\theta_{cc}$.
By considering the contributions from collinear modes aligned along subjets $1$ and $2$, we find the parametric relation $$\begin{aligned}
\label{eq:appC_1angle}
{{_{1}e_{i}^{(\beta)}}} \sim \theta_{12}^\beta \sim {\tau_{i-1}^{(\beta)}}\,,\end{aligned}$$ where all other pairwise combinations of modes are power suppressed. Here, we are assuming that the $N$-subjettiness axes are defined such that one axis is aligned with subjet 1 or 2, with the remaining $i-2$ axes aligned along the other subjets; this is indeed the configuration that minimizes ${\tau_{i-1}}$ in the small $\theta_{12}$ limit, assuming balanced energies. Adding an extra axis yields $${\tau_{i}^{(\beta)}} \sim \theta_{cc}^\beta,$$ where now the $i$ axes align with the $i$ subjets.
For the correlator involving two angles, the power-counting analysis yields $$\begin{aligned}
\label{eq:appC_2angle}
{{_{2}e_{i+1}^{(\beta)}}} &\sim \theta_{12}^\beta \left( \theta_{cc}^\beta +\ldots \right) \sim {\tau_{i-1}^{(\beta)}} \cdot {\tau_{i}^{(\beta)}}\,,\end{aligned}$$ where the ellipses denote contributions from collinear-soft modes, which depend on the other angles between the subjets. To understand the appearance of $\theta_{12}^\beta \theta_{cc}^\beta $, note that the largest contribution to ${{_{2}e_{i+1}}}$ comes from selecting two collinear modes from one subjet and one collinear mode from each of the remaining $i-1$ subjets; for that configuration, the two smallest pairwise angles are indeed $\theta_{cc}$ and $\theta_{12}$.
Generalizing the argument in , imply ${{_{2}e_{i+1}}} \ll ({{_{1}e_{i}}})^2$ on $i$-prong signal jets, such that the appropriate $i$-prong discriminant is $$\begin{aligned}
{N_{i}^{(\beta)}}=\frac{{{_{2}e_{i+1}^{(\beta)}}}}{({{_{1}e_{i}^{(\beta)}}})^2}\sim\frac{{\tau_{i}^{(\beta)}}}{{\tau_{i-1}^{(\beta)}}}\,,\end{aligned}$$ where the last relation should be understood in the power-counting sense. Therefore, as advertised, the $N_i$ observable is indeed related to the $N$-subjettiness ratio ${\tau_{i,i-1}}$, and both are expected to be good $i$-prong discriminants.
\
As an example to demonstrate this parametric relation, we consider the case $i = 2$, which was alluded to in . The relevant observables are shown schematically in . In , we show distributions of ${\tau_{2,1}}$ and ${N_{2}}$ before and after grooming for $\beta=2$, taking quarks as representative of the background. To aid in a visual comparison, we have rescaled the ${\tau_{2,1}}$ distributions by a common factor to match the ${N_{2}}$ endpoint. Before grooming, the shapes of the two distributions are quite different, with ${N_{2}}$ being much more peaked towards the endpoint for the background. After soft drop has been applied, the distributions for the two observables are quite similar, as predicted by the power-counting discussion above.
Still, there is a non-parametric difference between the ${\tau_{2,1}}$ and ${N_{2}}$ distributions, which leads to improved tagging performance for ${N_{2}}$. This can be seen by eye in the groomed plot in , where the background distribution for ${N_{2}}$ is pushed to higher values while the signal distribution is more rapidly falling toward the endpoint. More quantitatively, we can consider the ROC curves in . For the ungroomed case, the discrimination power is similar, with ${N_{2}}$ showing slightly improved behavior at higher efficiencies. For the groomed case, there are significant gains to be had in using ${N_{2}}$ instead of ${\tau_{2,1}}$.[^28]
Hybrid Strategies for 2-prong Observables {#app:hybrid}
=========================================
Throughout the text, we focused on discriminants formed from combinations (often ratios) of either groomed or ungroomed observables. It is also interesting to consider discriminants formed from mixtures of groomed and ungroomed observables [@gregory_talk; @gregory_paper], which we will refer to as a hybrid strategy. While we will not explore this topic in detail, we take as a simple example ungroomed 2-prong observables after the application of a groomed mass cut.
In , we show the ROC curves for boosted $Z$ discrimination, showing light quark and gluon backgrounds separately; this should be contrasted with . The behavior of these hybrid observables can be understood using the power-counting analysis of , where we analyzed the stability of the observables as a function of $m_J$ and $p_{TJ}$. For signal jets, a cut on the groomed mass has little effect due to the color singlet nature of the $Z$ boson, and therefore the hybrid observables should have a similar behavior to the ungroomed observables. For background QCD jets, however, applying a groomed mass cut in the same mass window enforces a higher effective cut on the ungroomed mass. This, in turn, enters the scaling relations for the background distributions given in : $$\begin{aligned}
\label{eq:app_2prong_scaling}
M^{(2), \text{peak}}_2\sim \frac{m_{J}^2}{p_{TJ}^2}\,, \qquad N^{(2), \text{max,peak}}_2 \sim \text{const}\,, \qquad D^{(2), \text{max}}_2\sim \frac{p_{TJ}^2}{m_J^2}\,.\end{aligned}$$ For ${M_{2}}$, and similarly for ${D_{2}^{(1,2)}}$, using a groomed mass cut has the interesting effect of pushing the ungroomed background distribution to higher values, thereby improving discrimination power. For ${N_{2}}$, the distribution is parametrically unmodified, and therefore similar discrimination power is expected for the ungroomed and hybrid observables. For ${D_{2}^{(2)}}$, larger effective mass values push the distribution to lower values, thereby worsening discrimination power. These power-counting predictions are seen clearly in .
The above behavior is perhaps counterintuitive, especially the poor performance of ${D_{2}^{(2)}}$ and the good performance of ${M_{2}}$, but it follows straightforwardly from the power-counting analysis. That said, the quantitative discrimination power depends crucially on the choice of mass window, and one must keep in mind that this study is based on a relatively narrow soft-dropped mass cut around $m_Z$. Further studies are therefore warranted to test whether discrimination performance can indeed be improved by simultaneously using information before and after grooming.
Supplemental Quark/Gluon Plots {#app:add_qg}
==============================
In , we emphasized the stability of $U_i$ for $i=2,3$ as a function of the angular exponent $\beta$. In , we show the full ROC curves for both ${U_{2}}$ and ${U_{3}}$ as a function of the angular exponent $\beta$. Neither observable asymptotes to the Casimir scaling prediction, even at high efficiencies or high $\beta$ values. Furthermore, the ${U_{3}}$ distributions exhibit stability as a function of $\beta$ throughout the whole ROC curve. This would be interesting to verify in an analytic calculation.
In , we show the ROC curves for ${U_{2}}$ and ${U_{3}}$ after grooming for $\beta = 0.2$, showing that the $U_i$ series continues to perform better for larger values of $i$. In , we show the performance as a function of $\beta$, demonstrating the stability of ${U_{3}}$, even after grooming.
[^1]: This is by no means a complete list. Other studies from the LHC using jet substructure can be found at <https://twiki.cern.ch/twiki/bin/view/AtlasPublic> and <http://cms-results.web.cern.ch/cms-results/public-results/publications/>.
[^2]: In this paper, we use “basis” to refer to any set of observables, even if they do not span the full space of IRC safe observables.
[^3]: The ${{_{v}e_{n}}}$ notation is inspired by the hypergeometric functions, which are similarly flexible.
[^4]: We use the normalized dimensionless definition denoted with a lower case $e$ [@Larkoski:2014gra]. This is related to the original dimensionful definition in by ${{e_{n}^{(\beta)}}} = \text{ECF}(n,\beta)/\left( \text{ECF}(1,\beta) \right)^n.$
[^5]: This observable is based on the global event shape $N$-jettiness [@Stewart:2010tn], which has recently been used to define the <span style="font-variant:small-caps;">XCone</span> jet algorithm [@Stewart:2015waa; @Thaler:2015xaa].
[^6]: We ignore the scale $\Lambda_{\text{QCD}}$ for this discussion, focusing on regions of phase space dominated by perturbative dynamics. While $\Lambda_{\text{QCD}}$ plays an important role in certain phase space regions, for IRC safe observables it contributes only a power-suppressed contribution away from singular limits.
[^7]: In this analysis, we do not consider the scale set by the jet radius, $R$. For $R\ll1$, the jet radius must also be considered in the power counting and the scale $R$ appears in perturbative calculations. For recent work on the resummation of logarithms associated with this scale, see .
[^8]: These ratios are not themselves IRC safe, but are instead Sudakov safe [@Larkoski:2013paa; @Larkoski:2015lea]. For a discussion of Sudakov safety for the case of ${D_{2}}$, see . For this reason, the ratio observables we construct in this paper cannot be written in the form of , even though their ${{_{v}e_{n}^{(\beta)}}}$ ingredients can.
[^9]: An alternative approach to identifying specific features within jets is machine learning, which has seen significant recent interest [@Cogan:2014oua; @deOliveira:2015xxd; @Almeida:2015jua; @Baldi:2016fql; @Guest:2016iqz; @Conway:2016caq; @Barnard:2016qma]. The contrast between these strategies has been dubbed “deep thinking” versus “deep learning". In the deep thinking approach pursued here, the goal is to identify the physics principles that lead to discrimination power, focusing on observables with desirable properties for first-principles calculations. In the deep learning approach, the goal is to use reliable training samples to optimize the discrimination power and, in many cases, visualize the underlying physics. Ultimately, one would want to merge these two approaches, which could help avoid theoretical blindspots in the cataloging of observables and mitigate modeling uncertainties inherent in training samples. Detailed studies in data, ideally with high purity samples, will also be needed for a complete understanding.
[^10]: With a completely generic angular weighting function, $f_N$, this basis is of course overcomplete.
[^11]: The appearance of $\min$ can also be viewed as the lowest-order Taylor expansion of a more generic observable, which should be a good approximation in the case of small radius jets. This can be seen explicitly in , where different functional forms are compared that give the same quantitative behavior as the $\min$ version here. Another motivation for the $\min$ definition is that it naively behaves more similarly to thrust [@Farhi:1977sg] or $N$-jettiness [@Stewart:2010tn], though we emphasize that ${{_{v}e_{n}}}$ does not rely on external axes.
[^12]: For ${{_{2}e_{3}}}$, note that $\min\{{a,b,c}\} \times \min^{(2)} \{a,b,c\} = \min\{ab,ac,bc\}$.
[^13]: WTA axes align with a hard prong within the jet. They are nice theoretically, as they avoid recoil due to soft emissions [@Catani:1992jc; @Dokshitzer:1998kz; @Banfi:2004yd; @Larkoski:2013eya; @Larkoski:2014uqa]. For low $p_T$ tops, however, the use of WTA axes can potentially lead to lopsided axes. We explicitly checked that are our results are unmodified if standard $E$-scheme recombination is used instead.
[^14]: See also for an example of an observable designed specifically to probe the grooming procedure by measuring non-global correlations, and for an example of improving discrimination power by understanding the behavior of the grooming procedure.
[^15]: As stated at the end of , for simplicity we do not power count the grooming parameter $z_{\rm cut}$. It is well understood how to properly incorporate $z_{\rm cut}$ into the power-counting analysis (see e.g. [@Frye:2016aiz; @Frye:2016okc]), but this has a negligible impact for understanding the qualitative behavior of 2-prong discriminants.
[^16]: Note the improved signal significance in the groomed case, which offsets the apparent decrease in discrimination performance when comparing the ungroomed and groomed ROC curves.
[^17]: Note that ATLAS uses ${D_{2}}$ after jet trimming [@Krohn:2009th], which has a similar parametric behavior to ${D_{2}}$ after soft drop in the region we are considering.
[^18]: We could alternatively vary the cut on the jet $p_{T}$. From the power-counting analysis, all stability properties are determined by functions of the ratio $m_J/p_{TJ}$, and therefore it is straightforward to understand the $p_{TJ}$ dependence from the $m_J$ dependence.
[^19]: Important exceptions are (IRC unsafe) multiplicity-based observables, which have a long history in QCD [@Brodsky:1976mg; @Konishi:1978yx; @Mueller:1983cq; @Malaza:1984vv; @Gaffney:1984yd; @Malaza:1985jd; @Catani:1991pm; @Catani:1992tm; @Dremin:1993vq; @Dremin:1994bj; @Capella:1999ms; @Bolzoni:2012ii] (see [@Aad:2016oit] for a recent experimental study), and more recently, shower deconstruction [@FerreiradeLima:2016gcz].
[^20]: This has been coined the “[<span style="font-variant:small-caps;">Pythia</span>]{}-[<span style="font-variant:small-caps;">Herwig</span>]{} sandwich”, with LHC data as the filling.
[^21]: Because of the $\theta_{cc}^{3\beta}$ term, this parametric relation is strictly speaking not true for ${{_{3}e_{3}}}$, but the difference is power suppressed in much of the phase space.
[^22]: Our reluctance to weigh in on nonperturbative corrections is because a standard shape function analysis [@Korchemsky:1999kt; @Korchemsky:2000kp; @Bosch:2004th; @Hoang:2007vb; @Ligeti:2008ac], which is applicable for ${e_{2}}$, does not hold for ${{_{1}e_{3}}}$. In future work, we might hope to extend the shape function logic to non-additive observables like ${{_{1}e_{3}}}$.
[^23]: For a jet with $n_J$ particles, the computational cost of ${U_{i}}$ scales like $n_J^{i+1}$. On a typical laptop, the analysis of a single jet takes around $\{0.14~\text{ms}, 0.86~\text{ms}, 11~\text{ms}\}$ for $\{U_1,U_2,U_3\}$.
[^24]: We chose 70% quark efficiency as a benchmark, since it was used in the recent study of , though the features emphasized in the text are largely independent of this particular choice. At very low quark efficiencies, deep in the nonperturbative regime, the different $U_i$ behaviors merge.
[^25]: See for discussions of factorization and resummation of such observables, and for fixed-order studies.
[^26]: The $r$ notation is motivated by the resistance formula for a set of parallel resistors.
[^27]: To mimic the behavior of ${{_{1}e_{3}}}$, one could consider the triangle area divided by its perimeter, which is parametrically related to the smallest distance in the collinear limit.
[^28]: While it is possible that different axes choices for $N$-subjettiness could provide improved performance, it seems to us that any axes definition will be ambiguous in the unresolved region. This also highlights the nice property that ${N_{2}}$ is defined without respect to subjet axes.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Even though end-to-end supervised learning has shown promising results for sensorimotor control of self-driving cars, its performance is greatly affected by the weather conditions under which it was trained, showing poor generalization to unseen conditions. In this paper, we show how knowledge can be transferred using semantic maps to new weather conditions without the need to obtain new ground truth data. To this end, we propose to divide the task of vehicle control into two independent modules: a control module which is only trained on one weather condition for which labeled steering data is available, and a perception module which is used as an interface between new weather conditions and the fixed control module. To generate the semantic data needed to train the perception module, we propose to use a generative adversarial network (GAN)-based model to retrieve the semantic information for the new conditions in an unsupervised manner. We introduce a master-servant architecture, where the master model (semantic labels available) trains the servant model (semantic labels not available). We show that our proposed method trained with ground truth data for a single weather condition is capable of achieving similar results on the task of steering angle prediction as an end-to-end model trained with ground truth data of 15 different weather conditions.'
author:
- Patrick Wenzel
- Qadeer Khan
- Daniel Cremers
- 'Laura Leal-Taixé'
bibliography:
- 'main.bib'
title: Modular Vehicle Control for Transferring Semantic Information Between Weather Conditions Using GANs
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In gauge-mediated supersymmetry (SUSY) breaking (GMSB) models the lightest supersymmetric particle (LSP) is the gravitino and the phenomenology is driven by the nature of the next-to-lightest SUSY particle (NLSP) which is either the lightest neutralino, the stau or mass degenerate sleptons. Since the NLSP decay length is effectively unconstrained, searches for all possible lifetime and NLSP topologies predicted by GMSB models in $e^+e^-$ collisions are performed on the data sample collected by OPAL at centre-of-mass energies up to 209 GeV at LEP.\
Results independent of the NLSP lifetime are presented for all relevant final states including direct NLSP pair-production and, for the first time, also NLSP production via cascade decays of heavier SUSY particles.\
None of the searches shows evidence for SUSY particle production. Cross-section limits are presented at the 95% confidence level both for direct NLSP production and for cascade decays, providing the most general, almost model independent results.\
These results are then interpreted in the framework of the minimal GMSB (mGMSB) model, where large areas of the accessible parameter space are excluded. In the mGMSB model, the NLSP masses are constrained to be $m_{\tilde\chi_1^0} > 53.5$ GeV$c^2$, $m_{\tilde\tau_1} > 87.4$ GeV$c^2$ and $m_{\tilde\ell} > 91.9$ GeV$c^2$ in the neutralino, stau and slepton co-NLSP scenarios, respectively.\
A complete scan on the parameters of the mGMSB model is performed, constraining the universal SUSY mass scale $\Lambda$ from the direct SUSY particle searches: $\Lambda>40,\,27,\,21,\,17,\,15$TeV/$c^2$ for messenger indices $N=1,\,2,\,3,\,4,\,5$, respectively, for all NLSP lifetimes.
author:
- Gabriele Benelli
title: 'Searches for Gauge-Mediated Supersymmetry Breaking Topologies in collisions at LEP2 '
---
[ address=[Department of Physics, The Ohio State University, 191 West Woodruff Avenue, Columbus, OH-43210-1117, U.S.A.]{} ]{}
Introduction
============
Supersymmetry[@susy], one of the proposed solution to the hierarchy problems of the Standard Model (SM), postulates the existence of a bosonic partner for each SM fermionic particle and viceversa. The discovery of these superpartners would be the most direct evidence for SUSY. Since these particles are not observed to have the same mass as their SM partners, SUSY must be a broken symmetry. In the most widely investigated scenarios, it is assumed that SUSY is broken in some [*hidden*]{} sector of new particles and is [*communicated*]{} to the [*visible*]{} sector of SM and SUSY particles by gravity or gauge interactions.
We present a study of gauge-mediated SUSY breaking topologies [@opal_gmsb] using the data collected by the OPAL detector at LEP up to the highest center-of-mass energies of 209 GeV.
Searches for GMSB topologies
============================
An attractive feature of GMSB models is that the hidden sector can lie at masses as low as $10^4$ GeV$/c^2$. In most current GMSB theoretical work [@theo], it is assumed that this sector is coupled to a messenger sector, which in turn couples to the visible sector through normal SM gauge interactions.
The minimal GMSB model introduces five new parameters and a sign: the SUSY breaking scale ($\sqrt{F}$), the SUSY particle mass scale ($\Lambda$), the messenger mass ($M$), the number of messenger sets ($N$), the ratio of the vacuum expectation values of the two Higgs doublets ($\tan\beta$) and the sign of the Higgs sector mixing parameter (sign($\mu$)).
In GMSB models the LSP is a light gravitino ($m_{\rm \tilde G}<$ 1 MeV$/c^2$), and the nature of the NLSP, which is either the lightest neutralino ($\tilde\chi^0_1$), stau ($\tilde\tau_1^{\pm}$) or mass-degenerate sleptons (${\tilde{\rm e}}_R^{\pm}$, ${\tilde{\mu}}_R^{\pm}$ and ${\tilde{\tau}}_1^{\pm}$), determines the phenomenology. As the gravitino couples very weakly to heavier SUSY particles, these will decay typically in a cascade to the NLSP which then decays via either $\tilde\chi^0_1 \rightarrow
\gamma \tilde\mathrm{G}$ or $\tilde\ell^{\pm} \rightarrow \ell^{\pm}
\tilde\mathrm{G}$. We study [@opal_gmsb] all relevant final states: both direct NLSP production and its appearance in the decay chain of heavier SUSY particles, like charginos, neutralinos and sleptons.
Since the decay length of the NLSP depends on $\sqrt{F}$ and is effectively unconstrained, the NLSP can decay inside or outside of the detector, so all possible lifetime topologies are searched for. With increasing decay length, the event signatures range from energetic leptons or photons and missing energy due to the undetected gravitino, to tracks with large impact parameters, kinked tracks, or heavy stable charged particles.
In total 14 different selections, each incorporating several signature variations, some based on the analyses described in [@opal:old], are implemented to cover all the GMSB topologies: slepton NLSP, neutralino NLSP, direct and cascade production, all lifetimes. In order to obtain lifetime independent results, the results from the various lifetime topologies are combined, with special attention to study the overlaps among the many channels. To achieve a good description of the selection efficiencies over the whole mass and lifetime range at all center-of-mass energies, without generating an excessive number of Monte Carlo samples, an interpolating function is determined. On Fig \[fig:gmsb\] it is demonstrated how the different selections contribute to the signal detection efficiency as a function of the NLSP lifetime.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![[**(A)**]{} Efficiencies for stau pair-production at $\sqrt{s}=208$ GeV. The symbols represent the efficiencies for ten simulated lifetimes while the curves show the interpolating efficiency functions of the searches for promptly decaying staus (dashed), large impact parameters (long dash-dotted), kinks (dotted) and stable staus (dash-dotted) together with the overlap efficiencies (filled areas). The total efficiency is shown by full line. [**(B)**]{} Observed and expected lower mass limits for pair-produced staus in the stau NLSP (a) and smuons (b), selectrons (c) in the slepton co-NLSP scenario as a function of the particle lifetime using the direct ${\tilde{\ell}}^+{\tilde{\ell}}^-$ search. For staus the observed and expected lower limit are identical in the stau NLSP scenario and in the slepton co-NLSP scenario. The mass limits are valid for a messenger index N$\leq 5$. For the stau NLSP and slepton co-NLSP scenarios, the NLSP mass limits are set by the stau mass limit and by the smuon mass limit, respectively.[]{data-label="fig:gmsb"}](pr409_06.epsi "fig:") ![[**(A)**]{} Efficiencies for stau pair-production at $\sqrt{s}=208$ GeV. The symbols represent the efficiencies for ten simulated lifetimes while the curves show the interpolating efficiency functions of the searches for promptly decaying staus (dashed), large impact parameters (long dash-dotted), kinks (dotted) and stable staus (dash-dotted) together with the overlap efficiencies (filled areas). The total efficiency is shown by full line. [**(B)**]{} Observed and expected lower mass limits for pair-produced staus in the stau NLSP (a) and smuons (b), selectrons (c) in the slepton co-NLSP scenario as a function of the particle lifetime using the direct ${\tilde{\ell}}^+{\tilde{\ell}}^-$ search. For staus the observed and expected lower limit are identical in the stau NLSP scenario and in the slepton co-NLSP scenario. The mass limits are valid for a messenger index N$\leq 5$. For the stau NLSP and slepton co-NLSP scenarios, the NLSP mass limits are set by the stau mass limit and by the smuon mass limit, respectively.[]{data-label="fig:gmsb"}](pr409_19.epsi "fig:")
**[(A)]{} & **[(B)]{}****
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
None of the searches shows evidence for SUSY particle production. To interpret the results, a detailed scan of the minimal GMSB parameter space is performed with the gravitino mass fixed to 2 eV, corresponding to $\sqrt{F} \approx
100$ TeV, motivated by the requirement that the branching ratio of the next-to-NLSP to the gravitino is small. If that is the case, the cross-sections and branching ratios do not depend on the gravitino mass. One should note that $\sqrt{F}$ can be eliminated from the scan as all limits are computed independent of the NLSP lifetime, and $\sqrt{F}$ has no significant effect on other particle masses.
“Model independent” cross-section limits are derived for each topology as a function of the NLSP lifetime. For direct NLSP decays, this is done by taking the worst limit for a given NLSP mass from the generated GMSB parameter scan points. For cascade channels, the cross-section evolution is assumed to be $\beta/s$ for spin-1/2 and $\beta^3/s$ for scalar SUSY particles, respectively, and the highest bound for all intermediate particle masses is retained. The maximum limit valid for all lifetimes is then quoted as the “lifetime independent” cross-section limit. In the neutralino NLSP scenario this is typically better than 0.04 pb for direct NLSP production, 0.1 pb for selectron and smuon production, 0.2 pb for stau production and 0.3 pb for chargino production. In the stau and slepton co-NLSP scenarios, the limit on direct NLSP production is 0.05 pb for smuons, 0.1 pb for selectrons and staus. For the cascade decays the bounds are typically better than 0.1 pb for neutralino, 0.2 for chargino and in the stau NLSP scenario 0.4 for selectron and smuon production.
The cross-section limits can be turned into constraints on the NLSP mass. For sleptons, the lowest mass limits are found for very short lifetimes, except for selectrons, as shown in Figure \[fig:gmsb\], where searches using d$E$/d$x$ measurements lose efficiency for particles with momenta around 65 GeV. The lifetime independent limits are $m_{\tilde\mathrm{e}_\mathrm{R}} > 60.1$ GeV, $m_{\tilde\mu_\mathrm{R}} >
93.7$ GeV and $m_{\tilde\tau_1} > 87.4$ GeV. The limit on the stau mass is the same in the stau and the slepton co-NLSP scenarios. In the slepton co-NLSP scenario, the best limit can be used to derive a universal limit on the slepton masses $m_{\tilde\ell} = m_{\tilde\mu_\mathrm{R}} - m_\tau
> 91.9$ GeV, where by definition the mass differences between the different slepton flavors are smaller than the lepton masses. For neutralino NLSP, no lifetime independent NLSP mass limit can be set directly. For short lifetimes ($\tau < 10^{-9}$ s) a mass limit of 96.8 GeV is derived. For the first time limits on the production cross-section for all GMSB search topologies, including cascade, are presented.
The GMSB parameter space is constrained by our results as shown in Figure \[fig:mGMSB\] for $N=1$, $M=1.01\cdot\Lambda$ and sign($\mu$)>0. The universal SUSY mass scale is $\Lambda> 40, 27, 21, 17, 15$ TeV for messenger indices $N=1, 2, 3, 4, 5$, respectively, independent of $M, \tan\beta$, sign($\mu$) and the NLSP lifetime ($\sqrt{F}$). The constraints on $\Lambda$ imply lower limits on the neutralino mass in the neutralino NLSP scenario: $m_{\tilde\chi_1^0} > 53.5$ GeV for N=1 and $m_{\tilde\chi_1^0} > 94.0$ GeV for N=5, independent of the lifetime.
![Examples of regions in the $\Lambda-\tan{\beta}$ plane excluded by pair-production searches for different particles, with sign$(\mu)>0$ and valid for any NLSP lifetime for four different sets of parameters, $N=1$ or $3$ and $M=1.01\cdot\Lambda$ or $250\,$TeV/$c^2$.[]{data-label="fig:mGMSB"}](pr409_20.epsi){height=".45\textheight"}
\#1\#2\#3\#4[[\#1]{} [**\#2**]{} (\#3) \#4]{}
[9]{} H.P. Nilles, . G.F. Giudice and R. Rattazzi, , S. Ambrosanio, G.D. Kribs and S.P. Martin, , S. Dimopoulos, S. Thomas and J.D. Wells, . OPAL Collaboration, G. Abbiendi [*et al.*]{}, Phys. Lett. B [**572**]{} (2003) 8, OPAL Collaboration, G. Abbiendi [*et al.*]{}, Eur. Phys. J. C [**32**]{} (2004) 453, OPAL Collaboration, G. Abbiendi [*et al.*]{}, Phys. Lett. B [**602**]{} (2004) 167, OPAL Collaboration, G. Abbiendi [*et al.*]{}, Eur. Phys. J. C [**35**]{} (2004) 1, OPAL Collaboration, G. Abbiendi [*et al.*]{} Eur. Phys. J. C [**33**]{} (2004) 173. OPAL Collaboration, G. Abbiendi [*et al.*]{}, Eur. Phys. J. C [**46**]{}, 307 (2006).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Mahmoud Abo Khamis
- 'Hung Q. Ngo'
- Dan Olteanu
- Dan Suciu
bibliography:
- 'sample.bib'
title: Boolean Tensor Decomposition for Conjunctive Queries with Negation
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'With advances in Generative Adversarial Networks (GANs) leading to dramatically-improved synthetic images and video, there is an increased need for algorithms which extend traditional forensics to this new category of imagery. While GANs have been shown to be helpful in a number of computer vision applications, there are other problematic uses such as ‘deep fakes’ which necessitate such forensics. Source [*camera*]{} attribution algorithms using various cues have addressed this need for imagery captured by a camera, but there are fewer options for synthetic imagery. We address the problem of attributing a synthetic image to a specific generator in a white box setting, by inverting the process of generation. This enables us to simultaneously determine whether the generator produced the image [*and*]{} recover an input which produces a close match to the synthetic image.'
author:
- |
Michael Albright and Scott McCloskey\
Honeywell ACST\
1985 Douglas Drive North, Golden Valley, MN, USA\
[[email protected]]{}
bibliography:
- 'GAN-Inversion\_zoteroexport.bib'
title: Source Generator Attribution via Inversion
---
Introduction
============
Because of its use in ‘fake news’ and ‘revenge porn’, the implications of fully- or partially-synthetic imagery has recently become a matter of broad social concern. Underlying both of these is the technology of deep networks used to generate imagery, often of faces, that are increasingly realistic. Whereas traditional image forensics already include powerful techniques applicable to images captured with a wide range of cameras [@Lukas; @Kurusowa; @ChenCamcorder; @Choi; @Kharrazi], there are relatively fewer options available to forensic analysts operating on synthetic imagery.
We aim to close the capability gap between source [*camera*]{} attribution and source [*generator*]{} attribution, and to provide additional functionality for generator attribution. Source camera attribution methods vary considerably, but typically use low-level cues (such sensor non-uniformity) and statistical indicators; they are unable to re-create the image capture process, because the physical camera may be unattainable and the scene being photographed may be ephemeral. Neither of these restrictions apply to synthetically-generated imagery, so we expand the attribution problem to encompass both the determination of whether the generator produced the image [*and*]{} the inputs necessary to re-create the generation process. Mathematically, we describe the generator as a function $G$ which transforms a vector $z$ into an image $I_g$ as $$I_g = G(z).$$ When there are multiple (known) generators $G_1$, $G_2$, etc. the [**limited attribution problem**]{} is to determine for a ‘probe’ image $I$ the value of $i$ that - for some value $z$ - satisfies $G_i(z) = \tilde{I} \approx I$. The related [**inversion problem**]{}, given a specific generator $G$ and probe image $I$, is to estimate a latent vector $\tilde{z}$ such that $G(\tilde{z}) = \tilde{I} \approx I.$ In both, we use approximation to acknowledge that small differences should be expected due to quantization, dynamic range clipping, and perhaps compression applied to $I$.
![Image attribution by generator inversion: Given a synthetic probe image $I$ and pre-trained generators $G_i$, the generators are inverted to find latent vectors $\tilde{z}_i $ such that generated outputs $G_i(\tilde{z})$ approximate the probe. Attribution is assigned to the generator with the best reconstruction (red dashed line). []{data-label="fig:image_attribution_diagram"}](figures/diagram_inversion_attribution_v2.png "fig:"){width="0.9\linewidth"}\
In our [**extended attribution problem**]{} (Fig. \[fig:image\_attribution\_diagram\]), we consider the case where there are multiple (known) generators and, given a probe image, we determine [*both*]{} $i$ and $\tilde{z}$ such that $G_i(\tilde{z}) = \tilde{I} \approx I$. The utility of attribution, in the context of synthetically generated imagery, is that training GAN-based generators for high-resolution outputs is compute-intensive, data-intensive, and error-prone. As such, the majority of users are more likely to download and use publicly-available generators than they are to train their own from the ground up. We also believe that, beyond limited attribution, the ability to re-create the generation process (enabled by our estimation of $z$) is critical to improving the explainability of the attribution decision.
In our experiments, we demonstrate extended attribution on generators from two different domains, trained via two different methods. We first show that the generative layers of an auto-encoder, trained on MNIST [@Lecun98gradient-basedlearning] characters, can successfully be attributed in increasingly difficult scenarios, up to and including discrimination between generators trained on the same data, in the same order, but with different random seeds. We then show that generators trained to produce realistic facial images in a GAN can be attributed, and discuss some of the interesting similarities and differences between the probe and reconstructed image.
Related Work
============
Both the limited attribution and inversion problems described in the introduction have attracted attention in recent years. For attribution, Yu [@yu_attributing_2018] and Marra [@marra_attributing_2018] have recently presented methods which use noise-type signatures to attribute a given image to a specific generator, which is conceptually similar to how Photo Response Non-Uniformity (PRNU) [@Lukas] provides source [*camera*]{} identification. These methods are very successful, providing nearly perfect attribution performance, even being able to discriminate between two generators with the same architecture, having been trained on the same data. A related problem is the detection of whether a given image was generated by [*any*]{} GAN-type generator, which has successfully been demonstrated by several groups using various methods [@marra_detecting_2018; @nataraj_detecting_2019; @hsu; @guera; @albanyEyes; @self_detecting_2019].
The generator inversion problem has likewise attracted attention. Creswell and Bharath [@creswell_inverting_2018] develop a generator inversion method as a means to explore GAN performance as a function of attributes in the image space. Luo [@luo_learning_2017] solve inversion by training an encoder coupled to a pre-trained generator in an auto-encoder framework. Lipton and Tripathi [@lipton_precise_2017] introduce a method called stochastic clipping to recover arbitrarily precise approximations of $z$, even in the presence of simple types of noise. None of the prior inversion work has addressed attribution.
Another related work, by Kilcher[@kilcher_generator_2017], demonstrates that generators instantiated with random weights can produce blurry versions of real images from various benchmark datasets. This demonstrates that the convergence of inversion, by itself, does not signal that the image can be attributed to a generator.
Method
======
Our method for attributing a probe image $I$ to a generator $G_i$ is predicated on the assumption that other generators $G_{j }$ $(j \ne i)$ cannot generate $I$ as well as the true generator $G_i$, so the minimum reconstruction error will correspond to the true generator $i$. We offer a brief discussion on why one may reasonably expect that situation to often (though not always) hold: First, we note that generators map a low dimensional latent vector $z$ into very high dimensional output space $G_j(z)$. For instance, in the case of ProGAN [@karras2017progressive], dim($z$) = 512 and dim($I$) = 3,145,728 (3x1024x1024). Hence, the mapping from $z$ to $G_j(z)$ parameterizes a union of low dimensional manifolds (of dimension $\le$ dim($z)$) in the high dimensional space [@arjovsky_towards_2017]. Furthermore, there are a number of sources of inherent randomness in generator training, such as random initialization of network weights, ordering of training images, and absence or inclusion of particular images in training batches. Arjovksy and Bottou [@arjovsky_towards_2017] have shown that it is highly improbable for low-dimensional manifolds in high-dimensional spaces to perfectly overlap everywhere if subjected to random perturbations; this is in fact part of the motivation for using a Wasserstein metric in Wasserstein GANs [@arjovsky_towards_2017; @arjovsky_wasserstein_2017]. Furthermore, in the case of GANs, training failures such as *mode dropping* may make the generator incapable of generating some regions of the manifold of natural images. In the extended attribution problem, illustrated in Figure \[fig:image\_attribution\_diagram\], we start from a probe image $I$ which was generated by one of several generators $G_1$, $G_2$, …, $G_n$ , where we assume we know both the architecture and weights of all generators. We do not know a-priori which generator $G_i$ made the probe image, nor do we know the latent vector z, but we seek to determine them. Note that if we could perfectly identify $z$ and $i$, we could (in the ideal case) perfectly recreate the probe image $I = G_i(z)$. In practice, there may be some residual discrepancies due to post-processing of the generated image, e.g. quantization, clipping, image compression, etc., so we allow for for some small differences between the probe image and the generator output. Hence, we seek to estimate a latent vector $\tilde{z}$ such that $G_i(\tilde{z}) = \tilde{I}$ is as close as possible to the true probe $I$. For our experiments, we formalize generator inversion as an optimization problem, where we minimize the loss function $$L_j(z) = \frac{1}{MN} || I - G_j(z) ||^2 ,
\label{eq:loss_fcn}$$ with $N$ and $M$ being the number of pixels and color channels in the image, respectively.
We separately minimize Eq. \[eq:loss\_fcn\] for each generator $j$, using an optimization algorithm to find the best latent vector $z$ which minimizes the loss function. Attribution is assigned to the generator $i$ with the smallest residual error, and the estimated latent vector $\tilde{z}$ is the point of minimum loss $$\tilde{z} = \operatorname*{argmin}_z \frac{1}{MN} || I - G_i(z) ||^2 \,.$$
Because generators used in GANs and autoencoders are neural networks, they are readily differentiable via backpropagation, so the loss function in Eq. \[eq:loss\_fcn\] may be optimized efficiently using gradient-based methods; in this work, we found the Adam [@kingma2014adam] optimization algorithm to perform well for all tested generators. Because Eq. \[eq:loss\_fcn\] is not convex, there is no guarantee that the optimization will converge to a global minimum. Hence, to ensure low reconstruction errors are obtained, we perform a multi-start optimization, where we perform each optimization multiple times, each starting from different random initial starting guesses for the latent vector $z$, and we choose the result with the lowest residual error.
We note that since the magnitude of the minimum residual error $L_i^{\rm min} = L_i(\tilde{z} $) quantifies how well the generated output $G_i(\tilde{z}) = \tilde{I}$ matches the probe image $I$, one may use the residual errors to assess confidence in the attribution assignment: for a correct assignment, one would expect the minimum loss $L^{\rm min}_i$ to be very small, and for high confidence it should be significantly smaller than reconstruction errors $L^{\rm min}_j$ from other generators $j \ne i$.
In the special case where there are only two generators $i$ and $j$, we can summarize our attribution decision and confidence by a single numerical score: $$S = \frac{ L^{\rm min}_j - L^{\rm min}_i }{L^{\rm min}_j + L^{\rm min}_i} \, .
\label{eq:Sequation}$$ Note that $S \rightarrow 1$ when the $i^{\rm th}$ generator perfectly reconstructs the probe ($L_i^{\rm min} \rightarrow 0$), $S \rightarrow -1$ when the $j^{\rm th}$ generator perfectly reconstructs the probe ($L_j^{\rm min} \rightarrow 0$), and $S \rightarrow 0 $ when both generators reconstruct the probe equally well ($L_i^{\rm min} = L_j^{\rm min}$; no attribution possible). Hence, the score $S$ provides a natural way to evaluate attribution performance using Receiver Operating Characteristic (ROC) curves, effectively treating binary attribution as binary classification; we do so in Section \[subsec:MNIST\_experiments\].
In the next section, we show that differences in generator training indeed contribute to detectable differences in generated images, even in the case of identical generator architectures.
Experiments
===========
In this section, we describe experiments by which we demonstrate the utility of our generator attribution and inversion. They present a series of increasingly difficult attribution problems, distinguishing between pairs of generators with increasing similarity in how they are trained.
MNIST Experiments {#subsec:MNIST_experiments}
-----------------
The first set of experiments were carried out with the MNIST dataset, selected because the smaller image size allows for faster training and testing. In each MNIST experiment, we trained a pair of fully-connected auto-encoders with sigmoid activations and an L2 loss. The encoder and decoder parts are symmetric, each having 784 nodes at the input/output layers and hidden layers with 64 and 32 nodes each. Each auto-encoder was trained on approximately 30,000 MNIST digits for 20,000 steps in batches of 256 images using the Adam optimization algorithm with a learning rate of 0.01. Once the training was completed, the decoders were separated from their encoders and the weights were frozen, rendering them pre-trained generators which map 32 dimensional latent vectors into 784 dimensional outputs. We then performed optimization-based generator inversion for image attribution on a test set consisting of 500 digits generated from each generator and saved as a PNG. We again used the Adam optimization algorithm with a learning rate of 0.01, and ran it for 1000 steps per inversion. Inversion of the MNIST generators proved straightforward, but to reduce the chance of optimizers getting trapped in local minima and biasing the results, we implemented a multi-start optimization strategy, where optimization was repeated 10 times per image with 10 different random initial starting guesses for the latent vector $z$. We assess performance by treating attribution as a binary classification problem and plotting the Receiver Operating Characteristic (ROC) curve; we use the Area Under the Curve (AUC) as a performance summary statistic. We arbitrarily designated one of the generators as the target generator and measure the True Positive Rate (TPR) at which outputs from the target generator are classified as such. The False Positive Rate (FPR) measures the frequency with which non-target generator outputs are classified as target outputs. These values are computed over a range of thresholds on $S$ (as defined in Eq. \[eq:Sequation\]).
Finally, we recall that each trained generator is influenced by a number of sources of randomness in the training process, such as the initial values of network weights, the order of images in training, etc., so attribution performance may vary if the experiment is repeated with different random number generator seeds. To assess the uncertainty in attribution performance, we repeat each MNIST experiment five times and plot five ROC curves per figure.
### Non-overlapping Training Data
In this experiment, we trained the auto-encoders on two non-overlapping subsets of the MNIST digits. For simplicity, we train one using odd digits and the other with even digits, though we note that our attribution does [*not*]{} recognize the digit or its parity. Having thus trained even and odd digit generators, we attribute each of the test images by the method descried above. Figure \[fig:ROC\_oddEven\] shows that the performance of our attribution on this experiment is nearly perfect, despite the fact that the generators do a surprisingly good job synthesizing digits that they’ve never seen before. The bottom row of Figure \[fig:ROC\_oddEven\], for example, shows a ‘9’ synthesized by an odd generator (left column) and reconstructions from inverting both even and odd generators (center and right columns, respectively). Despite never having been trained on ‘9’s, the even number generator can produce a good approximation of this input, though less so for the odd generator and the ‘2’ in the row above. We also observe that variance in the area under the curve is quite low over the five different repetitions of the experiment, indicating the attribution performance is robust.
![(Top) ROC curve on attribution between two generators with non-overlapping training data. One generator is trained on only even MNIST digits, and the other on odd digits. For reference, the diagonal line shows the performance of random attribution. The experiment is repeated 5 times. (Bottom rows) Examples of two successful attributions. The probe image (left) was approximated with the even (middle) and odd (right) generators via inversion.[]{data-label="fig:ROC_oddEven"}](figures/ROC_oddEven.eps "fig:"){width="0.9\linewidth"}\
![(Top) ROC curve on attribution between two generators with non-overlapping training data. One generator is trained on only even MNIST digits, and the other on odd digits. For reference, the diagonal line shows the performance of random attribution. The experiment is repeated 5 times. (Bottom rows) Examples of two successful attributions. The probe image (left) was approximated with the even (middle) and odd (right) generators via inversion.[]{data-label="fig:ROC_oddEven"}](figures/examples_oddEven.png "fig:"){width="0.9\linewidth"}
### Same Training Data, Different Order {#sec:samedatadifforder}
In this experiment, we trained the auto-encoders on the same set of MNIST digits (comprised of both even and odd digits), but shuffled the order in which they’re used. Both generators have the same architecture, but their weights receive different random initializations at the start of training. Figure \[fig:ROC\_shuffle\] shows that the performance is reduced when our algorithm is asked to differentiate between generators trained with the same training data, and the variation in performance between trials has also increased, but the performance is still quite good in every trial. The figure also shows two examples of [*mis*]{}-attribution, illustrating that the outputs of the two generators are quite similar, visually, and have reconstruction errors which are quite close to one another. This portends difficulty in successfully attributing images subject to the normal forensic challenges such as compression and re-encoding.
![(Top) ROC curve on attribution between two generators with different permutations of the same training data. The experiment is repeated 5 times. (Bottom rows) Examples of two [*unsuccessful*]{} attributions. []{data-label="fig:ROC_shuffle"}](figures/ROC_shuffle.eps "fig:"){width="0.9\linewidth"}\
![(Top) ROC curve on attribution between two generators with different permutations of the same training data. The experiment is repeated 5 times. (Bottom rows) Examples of two [*unsuccessful*]{} attributions. []{data-label="fig:ROC_shuffle"}](figures/examples_shuffle.png "fig:"){width="0.9\linewidth"}
### Same Data, Same Ordering
In this experiment, the two auto-encoders are trained on the same subset of MNIST digits, and those training digits are presented in the same order. The only difference between the two generators is the initial (random) weights. Despite this high level of similarity, Fig. \[fig:ROC\_initialization\] shows that our approach can unambiguously attribute $> 70\%$ of the inputs with extremely high confidence for all five trials.
![ROC curve on attribution between two generators with the same training data, in the same order. Only the random initialization of the network differs between the two generators, despite which we can attribute their outputs well. The experiment is repeated 5 times.[]{data-label="fig:ROC_initialization"}](figures/ROC_initialization.eps){width="0.99\linewidth"}
### Same Training Data, Different Order, with Compression {#sec:compression}
In the last of our MNIST experiments, we quantify the impact of JPEG compression on the performance of the attribution algorithm. Compression is a well-known nuisance factor for many forensic algorithms. In the case of our attribution algorithm, compression introduces perturbations in the probe image which move it off of the manifold representing the span of the generator’s output. Note that even lossless compression does this, as the double precision output of the generator is quantized to an integer. Given that (a) compression is expected in most real-world uses of attribution, (b) Kilcher [@kilcher_generator_2017] showed that random GANs can approximate a probe image, and (c) our empirical observation that outputs from the wrong generator can be quite close to the probe image, it is important to understand how likely compression is to lead to mis-attribution.
In order to study this impact, we re-visit the experiment described in Sec. \[sec:samedatadifforder\], where we discriminate between two generators trained on the same MNIST data, but with different orderings of the training data. After training a pair of networks, the two generators are used to create digits which we compress using OpenCV’s imwrite function, using different values for the quality factor $q$ (which ranges from 0 to 100). Fig. \[fig:ROC\_compression\] shows ROC curves from this experiment, along with a reference ROC for images saved as PNGs without compression. We see that while increasing levels of JPEG compression do reduce performance, the degradation is graceful and attribution performance is still decidedly better than chance.
![ROCs demonstrating performance of our attribution as a function of JPEG compression quality factor. The blue curve represents performance on probe images saved as PNGs, and the other curves represent performance on JPEGs with a range of quality factors $q$. While increasing compression reduces performance, the reduction is graceful and performance is still significantly better than chance at a quality factor of 50.[]{data-label="fig:ROC_compression"}](figures/Compression.eps){width="0.99\linewidth"}
We emphasize that the different ROC curves shown in Fig. \[fig:ROC\_compression\] all originate from a *single* pair of generators, and the differences between the curves originate in different compression levels applied when saving a common set of test images. This is very different from Fig. \[fig:ROC\_shuffle\], which tested attribution performance variation across *different* pairs of generators trained with different random initializations and different orderings of training data (but without lossy compression).
In Fig. \[fig:CompressionErrorsAB\], we plot histograms of reconstruction errors (Eq. \[eq:loss\_fcn\]) at different compression levels. In each histogram, we show the errors obtained when inverting each of the generators on the test images produced by that generator. We see reconstruction errors increase as the quality factor $q$ decreases. However, the errors are still low enough to permit attribution, as evidenced by the large AUCs in Fig. \[fig:ROC\_compression\].
![Histograms of reconstruction errors (Eq. \[eq:loss\_fcn\]) at different compression levels. The plots show reconstruction errors obtained when inverting each of the generators from Sec. \[sec:compression\] on the test images created by that generator. []{data-label="fig:CompressionErrorsAB"}](figures/Compression_Errors_AB.png){width="1.01\linewidth"}
CelebA Experiments {#subsec:CelebA_experiments}
------------------
In the second set of experiments, we performed attribution on generators that were trained as Generative Adversarial Networks (GANs). The three GANs tested were ProGAN [@karras2017progressive], SAGAN [@zhang2018SAGAN], and SNGAN [@miyato2018spectral], all of which were modified from their original description to output 128x128 resolution images and were trained on face images from the CelebA dataset. We re-used pre-trained generator weights shared by Ning Yu [@yu_attributing_2018]. The SAGAN and SNGAN generators proved easy to invert via optimization algorithms, so multi-start was not strictly necessary, but for good measure we employed 10 random initial optimization starts per image. For each random start, we used the Adam optimization algorithm with a learning rate of 0.1 and 1000 optimization steps per image, and chose the best reconstruction with the lowest loss encountered. When inverting SAGAN, we found it helpful to employ an explicit learning rate reduction on plateau, to reduce oscillation in the loss as the optimizer approached the minimum; we used a learning rate shrink factor of 0.5 and a patience of 30. The ProGAN generator proved slightly more challenging to invert reliably, so we employed 20 random initial optimization starts per image. We also used the Adam optimizer, but with a learning rate of 0.9, 300 optimization steps per image, and no explicit learning rate reduction (besides those built into Adam).
![Histograms of $S_{i}$ scores for each generator. The top figure evaluates inversion of the ProGAN generator $(P)$, the middle figure evaluates SAGAN inversion $(SA)$, and the bottom figure evaluates SNGAN inversion $(SN)$. Each inversion is evaluated on 500 probes created by ProgGAN, SAGAN, and SNGAN. Note that $S_i \rightarrow 1$ when the algorithm is confident that the $i^{\rm th}$ generator was the source of the probe, and $S_i \rightarrow -1$ when the algorithm is confident generator $i$ was not the source.[]{data-label="fig:CelebA_histogram"}](figures/histogram_v4.png){width="0.9\linewidth"}
We tested inversion-based attribution on a dataset of 1500 synthetic images (500 of which were generated by ProGAN, 500 by SAGAN, and 500 by SNGAN) which were saved as PNGs. As summarized in Table \[table:GANaccuracy\], attribution accuracy was 100% for images generated by SAGAN and SNGAN and 95% for images generated by ProGAN. Examples of successful attributions of images generated by ProGAN and SAGAN are shown in Figures \[fig:ProGAN\_success\] and \[fig:SAGAN\_success\]. As can be seen in Figures \[fig:ProGAN\_success\] and \[fig:SAGAN\_success\], successful reconstruction of the probe image by a generator offers compelling, interpretable evidence that a probe image was created by the indicated generator. This ability to offer strong, interpretable evidence of a synthetic image’s source is unique to our method of generator attribution, compared to other attribution methods which tend to operate as black boxes without strong interpretability or confidence measures. In Figure \[fig:ProGAN\_fail\], we show three examples where attribution of images generated by ProGAN failed. From the figure, it is clear that the cause of the failure is that the optimization algorithm failed to converge to the best reconstruction, causing mis-attribution. In each of the three cases, it is obvious that the quality of the best image reconstruction is poor, with relatively large residual error.
We can analyze the invertibility of each generator by generalizing Eq. \[eq:Sequation\] to the situation of more than two generators. We define multiple $S_i$ scores, where each score $S_i$ characterizes the reconstruction error achieved by generator $i$ on a particular image relative to the errors of other generators: $$S_i = \frac{ min_{j \ne i}( L^{\rm min}_j ) - L^{\rm min}_i }{ min_{j \ne i}(L^{\rm min}_j) + L^{\rm min}_i} \, .
\label{eq:S_i_equation}$$ (Intuitively, this is describing one-vs-rest classification.) Note that with this extended formula, $S_i \rightarrow 1$ when generator $i$ perfectly reconstructs the probe, $S_i \rightarrow -1$ when a different generator $j \ne i$ perfectly reconstructs the probe, and $S_i \rightarrow 0$ when generator $i$’s reconstruction error matches the next-best reconstruction error (i.e., uncertain attribution). The scores are visualized in Fig. \[fig:CelebA\_histogram\]. From the bottom two histograms, we can see that inversion-based attribution worked very well for SAGAN and SNGAN—attribution was correct 100% of the time and with high confidence ($S_i \approx \pm 1 $) on targets from those generators. From the top histogram, we can see ProGAN attribution was more difficult—there were more errors ($S_P < 0$ on ProGAN targets) and lower attribution confidence ($|S_P| < 1$ on ProGAN targets), indicating that optimization-based inversion of the ProGAN generator on ProGAN targets sometimes struggled to produce extremely low reconstruction errors.
Source ProGAN SAGAN SNGAN
---------- -------- ------- -------
Accuracy 95% 100% 100%
: Inversion-based attribution accuracy, by data source, measured on a dataset of 1500 synthetic images—500 generated by ProGAN, 500 by SAGAN, and 500 by SNGAN.[]{data-label="table:GANaccuracy"}
![Examples of correct attribution of probe images generated by ProGAN. The top row shows probe images while the $2^{\rm nd}$ through $4^{\rm th}$ rows show the closest approximations found by inverting ProGAN, SAGAN, and SNGAN generators. Green boxes indicate the decision of our attribution algorithm. Note the subtle differences in the 2nd column between the probe and ProGAN reconstruction, indicating an imperfect inversion.[]{data-label="fig:ProGAN_success"}](figures/success_ProGAN_4x3_green.png){width="0.9\linewidth"}
![Examples of correct attribution of probe images generated by SAGAN. The top row shows probe images while the $2^{\rm nd}$ through $4^{\rm th}$ rows show the closest approximations found by inverting ProGAN, SAGAN, and SNGAN generators. Green boxes indicate the decision of our attribution algorithm.[]{data-label="fig:SAGAN_success"}](figures/success_SAGAN_4x3_green.png){width="0.9\linewidth"}
![Examples of incorrect attribution of probe images generated by ProGAN. Green boxes indicate the decision of our attribution algorithm. In each case, mis-attribution is caused by failures of the inversion, which is evident in the ProGAN row in the background in the first column, the face in the second column, and everything in the third column.[]{data-label="fig:ProGAN_fail"}](figures/fail_ProGAN_4x3_green.png){width="0.9\linewidth"}
In addition to the above work, we experimented with minimizing $L1$ loss functions, on the hope that it would encourage a sparser residual error vector $r_i = I - G_i(\tilde{z})$, but we saw no significant attribution benefit. We also tested the LBFGS optimization algorithm, which uses 2nd-order derivatives to accelerate convergence; similar to [@webster2019detecting], we found that significantly fewer optimization steps were required to invert an image, but we saw no significant benefit in the minimum loss obtained by the optimizer.
Discussion
==========
We have shown that generator inversion is a useful means by which to pursue synthetic image attribution, with the added advantage of being able to re-create the generation process by estimating the latent vector via inversion. Despite this increased functionality, we achieve similar results to black box systems such as Yu [@yu_attributing_2018]. Both our method and Yu’s are able to distinguish between generators that are quite similar, up to and including generators trained from the same data in the same order.
![Example inversion with semantically meaningful differences. In this case, the reconstruction (right) has lips that are slightly more closed and a smaller reflection in the hair (upper left) compared to the probe image (left). Note that this inversion was performed against the full-resolution (1024x1024) version of the CelebA-HQ ProGAN, whereas inversions in Sec. \[subsec:CelebA\_experiments\] were done against a lower-resolution version.[]{data-label="fig:semanticDifferences"}](figures/celeba1024_example.png){width="0.9\linewidth"}
While our attribution results correspond to a completely automated system, it is interesting to consider how our output image would be used and interpreted by a forensic analyst. As illustrated by Fig. \[fig:ProGAN\_fail\], inversion failures are visually salient and so the assigned attributions could be rejected by the analyst for failing to meet quality thresholds. We note, however, that designing such a threshold is non-trivial in light of semantically meaningful differences between the probe and reconstruction in the case of successful inversion, an example of which is illustrated in Fig. \[fig:semanticDifferences\].
Future Work
===========
Though our experiments have so far been limited to attribution between two or three generators, it can easily be extended to $N$-way attribution with the use of Eq. \[eq:S\_i\_equation\]; note that the dynamic range would need to be normalized for networks having different outputs, i.e. different than the $[-1,1]$ range used by ProGAN, SAGAN, and SNGAN. However, the computational complexity of inversion over $N$ generators can be quite high. Given the computational complexity of inverting the generator, a more efficient system could use an attribution-only method, i.e. one of [@yu_attributing_2018; @marra_attributing_2018], to determine the source generator and then apply inversion on only that generator. Another extension would be to train neural network encoders to explicitly perform the inverse mapping from an image to a latent vector, which could also be further fine-tuned by optimization, as in [@zhu_generative_2016].
One area of future work would be to extend our attribution capabilities into the black box domain, in order to handle generators which may be offered online as a service, but without published network weights. It has been shown that, for deep networks performing classification tasks, representative proxy networks can be trained based on a relatively small number of input/output pairs from the target network. It may be the case that we can develop proxies for target [*generators*]{} in order to support our white-box attribution method.
Thanks {#thanks .unnumbered}
======
We thank Ning Yu for sharing data and pre-trained face generators used in this work, and Asongu Tambo for helpful suggestions.
Acknowledgements {#acknowledgements .unnumbered}
================
This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA). The views, opinions and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We discuss and compare several geometric structures which imply an upper bound to the acceleration of a particle measured in its rest system. While all of them have the same implications on the motion of a point particle, they differ in other important respects. In particular, they have different symmetry groups, which influence in a different way the search for an underlying dynamical theory.
PACS numbers:
45.50.Dd (Dynamics and kinematics of a particle);
11.30.Cp (Lorentz and Poincaré invariance);
02.40.Ft (Convex sets and geometric inequalities).
author:
- |
M. Toller [^1]\
via Malfatti n. 8\
I-38100 Trento, Italy
title: 'Geometries of Maximal Acceleration.'
---
Introduction.
=============
Starting from a letter published by Caianiello in 1981 [@Caianiello1], several ideas have been proposed concerning a possible upper bound to the acceleration of a particle [@Toller1; @CDFMV; @Brandt1; @Brandt2; @Scarpetta; @Caianiello2; @CL; @SS; @GS; @Brandt3; @Toller2; @Brandt4; @WPI1; @WPI2; @Brandt5; @NFS; @Rama; @Castro; @Papini]. More recently, further important developments have appeared [@Schuller; @Schuller1; @SWG; @SP; @Schuller2]. For lack of space, we don’t cite here many other papers devoted to applications of the maximal acceleration idea.
There are two points of view, which are physically rather different, though many formal considerations can be applied to both. Some authors [@Brandt1; @Brandt2; @Scarpetta; @CL; @Brandt3; @Brandt4] assume a maximal acceleration valid for all the kinds of particles. If it has to be explained in terms of the known general physical principles, from a dimensional argument we find $$a_M \approx (G \hbar)^{-1/2} c^{7/2},$$ where $a_M$ is the maximal acceleration in the rest frame and $G$ is the gravitational coupling constant. The constants appearing in this formula suggest that an explanation should be based on a relativistic quantum theory of gravitation.
Other authors [@Caianiello2; @SS; @WPI1; @Papini] propose an upper bound depending on the mass $m$ of the particle. If we assume that gravitational phenomena are irrelevant, from a dimensional argument we find $$a_M \approx m \hbar^{-1} c^3.$$ The same assumption can equally well be described as a mass-dependent upper bound to the force (always measured in the rest system) given by $$f_M \approx m^2 \hbar^{-1} c^3.$$
These formulas suggest that one should explain these upper bounds in terms of a special-relativistic quantum theory and many arguments in this direction have been proposed. However, in our opinion, no formal rigorous proof has been given up to now. Presumably, a rigorous treatment should involve the locality principle, namely it should proceed in the framework of quantum field theory. Then, during an interaction the production of new particles is inavoidable and one may think that, when the acceleration is too large, the system cannot be considered any more as a single particle state. In this way, however, it is difficult to find a sharp limitation.
The aim of the present article is not to give a proof, or at least a new justification, of the maximal acceleration hypothesis, but to discuss its formulations in terms of geometric concepts. We think that this discussion is relevant, since various authors use rather different geometric constructions, and the connection between them is not trivial.
We use as a guide the familiar geometric description of the maximal velocity which appears in relativity theory. It is given in terms of the closed future cone $V^+$ in the Minkowski space-time. In the presence of gravitation, one has to consider a cone in each tangent space of a pseudo-Riemannian space-time $\mathcal{M}$. The vectors tangent to the world-line of a particle must belong to these cones.
In the following a [*wedge*]{} means a convex dilatation-invariant subset of a vector space and a [*cone*]{} is a wedge that does not contain straight lines. According to this definition, a cone is necessarily convex. The set $V^+$ has the following properties:
1. It is a closed cone with nonempty interior;
2. It is invariant under the rotation group.
One can easily show (see the appendix A) that these conditions determine $V^+$ completely up to the choice of the numerical value $c$ of the light velocity and of the direction of time.
We see that the convexity requirement is very powerful and it is physically well justified. It means that if a particle moves during two consecutive time intervals with two allowed constant velocities, the overall average velocity is also allowed.
We see that a cone satisfying the conditions a), b) listed above is necessarily symmetric with respect to the orthochronous Lorentz group $O^{\uparrow}(1, 3)$ and one may consider this argument as a derivation of the Lorentz symmetry from the rotation symmetry and the existence of a maximal velocity. The next step towards a relativistic theory is to assume that the Lorentz group is not only the symmetry group of $V^+$, but also the symmetry group of all the physical laws. This approach to special relativity is certainly less satisfactory than the usual one, based on the relativity principle, but it can be extended by analogy to a treatment of theories with a maximal acceleration.
In order to introduce a maximal acceleration, one has to consider a manifold with at least seven dimensions, which describes, besides time and position, the velocity (or the energy-momentum) of a particle. The maximal acceleration (or the maximal force) is described by cones in the tangent spaces. We shall show that in some cases, assuming the topological properties a) given above and Lorentz symmetry, one can determine these cones up to the choice of the numerical values of the light velocity $c$ and of the maximal acceleration $a_M$.
The cones obtained in this way have a symmetry group which acts linearly on the tangent spaces of the manifold considered (but not necessarily on the manifold itself). This group may be larger than $SO^{\uparrow}(1, 3)$ and, following the example of relativity, one may try to assume that it (or one of its subgroups larger than $SO^{\uparrow}(1, 3)$) also represents a symmetry of the dynamical equations. Of course, this higher symmetry, which is not physically observed, has to be a broken symmetry, namely the vacuum is only symmetric with respect to the Lorentz subgroup.
In the following three sections we consider some different geometric schemes of the kind described above, based on different manifolds, and we find that they give rise to different enlarged symmetry groups. This means that there is some deep physical difference between these approaches. In Section 5 we introduce a different physical interpretation which permits a better understanding of the problem. In the last section we draw some conclusions and we indicate some possible future developments.
The space-time-velocity manifold.
=================================
To the best of our knowledge, the first geometrical scheme which implies a maximal acceleration is given by Born’s duality theory [@Born1; @Born2]. It is based on the eight-dimensional relativistic phase space $\mathcal{W} = \mathbf{R}^8$ the coordinates of which are the space-time coordinates $x^k$ and the components $p^k$ of the energy-momentum four-vector. In the following, the indices $i, j, k$ take the values $0, 1, 2, 3$, we assume $c = 1$ and for the relativistic scalar product of two four-vectors we use the notation $x \cdot p = x^k p_k = x^0 p^0 - \mathbf{x} \cdot \mathbf{p}$. One introduces a metric of the form $$d\sigma^2 = dx \cdot dx + f_M^{-2} \, dp \cdot dp,$$ where $f_M$ is the maximal force. Here $p$ represents the canonical four-momentum, which satisfies the canonical Poisson brakets.
If we consider a free particle with mass $m$, we have $p = m u$, where $u$ is the four-velocity, and the same expression can be written in the form $$\label{Metric}
d\sigma^2 = dx \cdot dx + a_M^{-2} \, du \cdot du.$$ Since we are concerned with the acceleration, we have to consider interacting particles and the two expressions given above are not equivalent any more.
If a particle interacts with an electromagnetic field, the first expression is not gauge invariant and we prefer to concentrate our attention on the second expression and on a space $\mathcal{W} = \mathbf{R}^8$ with coordinates $x^k$ and $u^k$. It can be interpreted as the tangent bundle of the Minkowski space-time and many of the following considerations can be generalized to the tangent bundle of a pseudo-Riemannian space-time.
One can also introduce the kinetic four-momentum, which by definition is given by $mu$. In general, its Poisson brackets are more complicated than the canonical ones. In the present article we deal with the kinematical aspects of the problem and we do not enter into the details of the Hamiltonian dynamics.
We describe the world line of a massive particle by means of the functions $\tau \to x(\tau)$, where the parameter $\tau$ is the proper time, defined by $$d\tau^2 = dx \cdot dx.$$ The curves in $\mathcal{W}$ which describe the motion of a particle must satisfy the equation $$\frac{dx}{d\tau} = u,$$ which implies the constraints $$\label{Constr1}
u \cdot u = 1, \qquad u^0 \geq 1,$$ $$\label{Constr2}
d \mathbf{x} = (u^0)^{-1} \mathbf{u} dx^0.$$ An interesting consequence of them, considered in the next section, is $$\label{Constr3}
dx \cdot du = 0.$$
The constraint (\[Constr1\]) is [*holonomous*]{} and it defines a 7-dimensional submanifold $\mathcal{V}$ of $\mathcal{W}$, which we call the [*space-time-velocity manifold*]{}. Only this submanifold is involved in the description of the motion of a particle. From the dynamical point of view, we are dealing with a constrained Hamiltonian system [@Dirac]. The manifold $\mathcal{V}$ has a pre-symplectic structure [@Souriau], that determines the dynamics without any reference to the larger space $\mathcal{W}$. Borrowing a term from quantum field theory, one may say that $\mathcal{W}$ also describes “off shell” particles.
The constraint (\[Constr2\]) is [*anholonomous*]{}, since it involves the tangent vectors. More correctly, rather than a constraint, it should be considered ad a dynamical equation, which might take a different form in a modified dynamical scheme. In section 5 we discuss a physical interpretation which goes beyond the particle kinematics and does not require the constraint (\[Constr2\]), but only the holonomous constraint (\[Constr1\]).
It is natural to impose the condition $$\label{Cond1}
d\sigma^2 \geq 0.$$ Note that the expression (\[Metric\]) and the condition (\[Cond1\]) are invariant under the pseudo-orthogonal group $O(2, 6)$, which may be called the “maximal acceleration group” [@Brandt2]. We shall find other groups that deserve this denomination equally well.
If one restricts his attention to the canonical transformations, one is led to consider the intersection $O(2, 6) \cap Sp(4)$, which is isomorphic to the pseudo-unitary group $SU(1, 3)$ [@Brandt5; @Low1; @Low2; @Low3; @Rama; @Castro]. One has to be careful when dealing with interacting particles, since the symplectic group $Sp(4)$ acts on the canonical four-momentum, while the pseudo-orthogonal group $O(2, 6)$ acts on the four-velocity or on the kinetic four-momentum.
It is important to stress that the constraints (\[Constr1\]) and (\[Constr2\]) are not invariant under the group $O(2, 6)$, which does not transform a particle world line into another one. In other words, the dynamics, as we know it, has not this symmetry. The subset of a tangent space $W$ of $\mathcal{W}$ defined by eq. (\[Cond1\]) is invariant under the dilatation group $\mathbf{R}^*$, but it is not convex and it is not a cone.
If we take into account the holonomous constraint (\[Constr1\]), we have $$du^0 = (u^0)^{-1} \mathbf{u} \cdot d \mathbf{u}$$ and the condition (\[Cond1\]) takes the form $$\label{Cond2}
\|d \mathbf{x}\|^2 + a_M^{-2} \left(\|d\mathbf{u}\|^2 - (u^0)^{-2} (\mathbf{u} \cdot d \mathbf{u})^2\right) \leq (dx^0)^2.$$
If we add the condition $dx^0 \geq 0$, these inequalities define a closed cone with nonempty interior in a space tangent to $\mathcal{V}$. The symmetry group of this cone is the product of the dilatation group $\mathbf{R}^*$ and a subgroup of $O(2, 6)$ which leaves the time-like four-vector $u$ invariant and is isomorphic to $O^{\uparrow}(1, 6)$. It is easier to work in a reference frame in which $\mathbf{u} = 0$, which is a rest frame of the particle if the constraint (\[Constr2\]) is satisfied. In this frame the condition (\[Cond2\]) takes the simple form $$\label{Cond3}
\|d \mathbf{x}\|^2 + a_M^{-2} \|d\mathbf{u}\|^2 \leq (dx^0)^2.$$ We have obtained a formalism that agrees with the ideas proposed in the preceding section.
If we also impose the constraint (\[Constr2\]), the condition (\[Cond2\]) takes the form $$\label{Cond4}
a_M^{-2} \left(\|d\mathbf{u}\|^2 - (u^0)^{-2} (\mathbf{u} \cdot d \mathbf{u})^2\right) \leq \left(1 - (u^0)^{-2} \|\mathbf{u}\|^2\right)(dx^0)^2,$$ or in the rest frame $$\label{Cond5}
a_M^{-2} \|d\mathbf{u}\|^2 \leq (dx^0)^2,$$ which implies the upper bound $$\label{Bound}
\|\mathbf{a}\| = \left\|\frac{d\mathbf{u}}{d\tau}\right\| \leq a_M$$ for the acceleration measured in the rest frame. Eq. (\[Cond4\]) and the constraint (\[Constr2\]) define in the spaces tangent to $\mathcal{V}$ a closed cone with empty interior symmetric with respect to a group isomorphic to $O^{\uparrow}(1, 3) \times \mathbf{R}^*$ (not to be interpreted as the usual Lorentz group).
The Born-Infeld kinematics.
===========================
In this section we consider again the space $\mathcal{W} = \mathbf{R}^8$ with coordinates $x^k$ and $u^k$. We indicate by $W$ a tangent space of $\mathcal{W}$ and, guided by the considerations given in section 1, we try to describe the maximal acceleration by means of a cone $W^+ \subset W$ with the following properties:
1. It is a closed cone with nonempty interior;
2. It is invariant under the proper orthochronous Lorentz group;
3. It is invariant under the transformation $dx \to dx$, $du \to -du$.
The last property requires a symmetry (indicated by $C$ in the following) with respect to a change of the acceleration sign. We show in the appendix B that these conditions determine $W^+$ up to the choice of the parameter $a_M$ and up to a time inversion. It is described by the conditions $$\label{Cond6}
dx_+ = dx + a_M^{-1} \, du \in V^+, \qquad dx_- = dx - a_M^{-1} \, du \in V^+.$$
From the equality $$dx_+ \cdot dx_+ + dx_- \cdot dx_- = 2 d\sigma^2,$$ we see that the condition (\[Cond6\]) is stronger than eq. (\[Cond1\]). The symmetry group of $W^+$ is $C \times_s (O^{\uparrow}(1, 3) \times \mathbf{R}^* \times O^{\uparrow}(1, 3) \times \mathbf{R}^*$), where the Lorentz groups and the dilatation groups $\mathbf{R}^*$ act separately on the four-vectors $x_+$ and $x_-$. The reflection $C$, introduced above, exchanges $x_+$ and $x_-$.
As in the preceding section, we have to take into account the constraints (\[Constr1\]) and (\[Constr2\]). They are not invariant under the symmetry group, but their consequence (\[Constr3\]) is invariant, as we see from the equality $$dx_+ \cdot dx_+ - dx_- \cdot dx_- = 4a_M^{-1} dx \cdot du.$$
If we take into account the holonomous constraint (\[Constr1\]), we obtain a closed cone in a space tangent to $\mathcal{V}$ given by $$\|d\mathbf{x} \pm a_M^{-1} d \mathbf{u}\| \leq dx^0 \pm a_M^{-1} (u^0)^{-1} \mathbf{u} \cdot d \mathbf{u}.$$ In a frame in which $\mathbf{u} = 0$ it takes the form $$\|d\mathbf{x} \pm a_M^{-1} d \mathbf{u}\| \leq dx^0,$$ namely $$\label{Cond7}
\|d\mathbf{x}\|^2 + 2 a_M^{-1} |d \mathbf{x} \cdot d \mathbf{u}| + a_M^{-2} \|d \mathbf{u}\|^2 \leq (dx^0)^2,
\qquad dx^0 \geq 0.$$ Note that this cone is smaller than the one defined by eq. (\[Cond3\]). Its symmetry group is $C \times_s (O(3) \times O(3)) \times \mathbf{R}^*$.
If we also take into account the constraint (\[Constr2\]), we obtain exactly the condition (\[Cond4\]) or, in the rest frame, the condition (\[Cond5\]). It follows that, as far as the motion of point particles is concerned, the formalism of the present section is equivalent to the formalism of the preceding section. However, it presents some advantages: the geometry of $\mathcal{W}$ is described by a cone $W^+$, in agreement with the ideas of section 1, and the constraint (\[Constr3\]) has the same symmetry as the cone.
The formalism described above, and in particular the symmetry group $O^{\uparrow}(1, 3) \times O^{\uparrow}(1, 3)$, coincides with some principles of the “Born-Infeld kinematics” proposed in refs. [@Schuller; @Schuller1; @SWG; @SP; @Schuller2] and motivated by the Born-Infeld theory of electrodynamics [@BI]. Our considerations may provide an alternative partial justification of those ideas.
In refs. [@Schuller; @Schuller1; @SWG; @SP; @Schuller2] one finds a very elegant treatment obtained by considering the space $W$ as a four-dimensional free module over the ring of the pseudo-complex numbers, which is generated by the real field and by a pseudo-imaginary unit $I$ with the property $I^2 = 1$. We refer to the original papers for more details and for dynamical considerations. Some further comments are given at the end of the next section and in the last section.
The space of reference frames.
==============================
Another approach to the geometry of maximal acceleration is based on the principal fibre bundle of the Lorentz frames, which we indicate by $\mathcal{S}$. A point of $\mathcal{S}$ corresponds to a tetrad of four-vectors $e_0, \ldots, e_3$ in a tangent space of space-time with the properties $$e_i \cdot e_k = \eta_{ik},$$ where in the right hand side the usual diagonal metric tensor of special relativity appears. We assume that $e_0 \in V^+$ and that the other three four-vectors define a left-handed spatial frame
As in the preceding sections, we disregard gravitation; the extension to the general case does not present difficulties, it is sufficient to replace the infinitesimal translations by parallel displacements. Then we may choose a frame $s_0 \in \mathcal{S}$ and all the other frames have the form $s = gs_0$, where $g$ is an element of the proper orthochronous Poincaré group. In this way we can identify $\mathcal{S}$ with the Poincaré group.
An infinitesimal displacement of an element of $\mathcal{S}$ can be obtained by means of an infinitesimal Poincaré transformation. This means that the tangent spaces of $\mathcal{S}$ can be identified with the Poincaré Lie algebra $\mathcal{T}$. We introduce in $\mathcal{T}$ a basis formed by the vectors $A_i$ and $A_{ik} = -A_{ki}$ which represent, respectively, the generators of the space-time translations and of the Lorentz transformations. Under the Lorentz group, they transform as the components of a four-vector and of an antisymmetric tensor. They form a set of ten vector fields in the manifold $\mathcal{S}$.
The Hamiltonian dynamics of a particle in the space $\mathcal{S}$ has been discussed in ref. [@Kunzle]. As in the preceding sections, we only consider the kinematical aspects. To the particle motion we associate a trajectory in $\mathcal{S}$ given by $\tau \to s(\tau) \in \mathcal{S}$ and we write $$\label{Deriv}
\frac{ds(\tau)}{d \tau} = b^{i} A_{i} + 2^{-1} b^{ik} A_{ik}
\in \mathcal{T}, \qquad b^{ik} = - b^{ki}.$$ We also introduce the vectors $$\mathbf{b} =(b^1, b^2, b^3), \qquad
\mathbf{b}' =(b^{12}, b^{23}, b^{31}), \qquad
\mathbf{b}'' =(b^{10}, b^{20}, b^{30}),$$ which describe space translations, rotations and Lorentz boosts.
The analogy argument presented in Section 1 suggests that the maximal acceleration hypothesis can be formulated by requiring that the vector (\[Deriv\]) belongs to a cone $\mathcal{T}^+ \subset \mathcal{T}$ with the properties
1. $\mathcal{T}^+$ is a closed cone with nonempty interior;
2. It is invariant under the proper orthochronous Lorentz group.
It has been proved in refs. [@Toller1; @Toller2] that these properties determine $\mathcal{T}^+$ up to the value of the parameter $a_M$ and up to an inversion of all the coordinates. In the following we choose the units in such a way that $a_M = c = 1$.
A simple definition of $\mathcal{T}^+$ is based on a representation of the elements of $\mathcal{T}$ by means of $4 \times 4$ matrices given by $$\hat b = b^i \gamma_0 \gamma_i
+ i 2^{-1} b^{ik} \gamma_0 \gamma_{i} \gamma_{k},$$ where $\gamma_{i}$ are the Dirac matrices in the Majorana representation. One can show that the matrix $\hat b$ is real and symmetric and one defines $\mathcal{T}^+$ by requiring that it is positive semi-definite, namely that $$\label{Cond8}
\psi^T \hat b \psi \geq 0$$ for any choice of the real spinor $\psi$. This condition implies the inequalities $$\|\mathbf{b}\| \leq b^0, \qquad \|\mathbf{b}'\| \leq b^0, \qquad \|\mathbf{b}''\| \leq b^0,$$ but it is stronger than them.
It is clear that the cone $\mathcal{T}^+$ has the symmetry group $GL(4, \mathbf{R})$ acting as $$\hat b \to a \hat b a^T, \qquad a \in GL(4, \mathbf{R}).$$ This group represents a broken symmetry, since the structure constants of the Poincaré algebra, considered as external fields that describe the vacuum, are not invariant [@Toller1]. It may be interesting to recall that the special linear group $SL(4, \mathbf{R})$ is locally isomorphic to the pseudo-orthogonal group $SO(3, 3)$.
As in the approaches of the preceding sections, a line which represents the motion of a particle must satisfy some constraints. However, there are no holonomous constraints of the kind (\[Constr1\]) and all the points of $\mathcal{S}$ are physically relevant. We require that $s(\tau)$ is a rest frame of the particle. It follows that the four-vector $e_0$ is just the four-velocity $u$ introduced in the preceding sections and that $$\label{Constr4}
b^0 = 1, \qquad \mathbf{b} = 0.$$ Moreover, we require that the frame $s(\tau)$ is accelerated without rotation, namely it is Fermi-Walker transported. In this way we obtain the constraint $$\label{Constr5}
\mathbf{b}' = 0.$$
We remain with the components of the acceleration $$\mathbf{a} = (b^0)^{-1} \mathbf{b}'',$$ measured in the rest frame $s(\tau)$. Our condition means that the matrix $$1 + i \mathbf{a} \cdot \boldsymbol{\gamma}, \qquad
\boldsymbol{\gamma} = (\gamma^1, \gamma^2, \gamma^3),$$ is positive semi-definite and it follows that the acceleration has the upper bound (\[Bound\]).
If we consider a particle with spin or an extended particle, the cone $\mathcal{T}^+$ contains some more information. In fact, in this case it is natural to give up the constraint (\[Constr5\]) and to assume that the frame $s(\tau)$ rotates together with the particle with an angular velocity $$\boldsymbol{\omega} = (b^0)^{-1} \mathbf{b}'$$ with respect to a Fermi-Walker transported frame. It has been shown in ref. [@Toller2] that from our condition one gets the upper bound $$\label{Omega}
\|\boldsymbol{\omega}\|^2 + \|\mathbf{a}\|^2 + 2 \| \mathbf{a} \times \boldsymbol{\omega} \| \leq a_M^2.$$ The set defined by this inequality can be considered as the intersection of the cone $\mathcal{T}^+$ with the plane defined by eq. (\[Constr4\]). It is convex and compact, but not Lorentz invariant, as it deals with quantities measured in the rest frame.
In ref. [@Toller2] one also finds a justification of this formula in terms of a model that considers a spherical rigid body with radius $a_M^{-1}$, requiring that all its points have a velocity smaller than $c$. This formula implies, for instance, an upper bound to the precession angular velocity of a particle with a magnetic moment in a magnetic field.
In refs. [@SP; @Schuller2] the Born-Infeld kinematics is also treated from the point of view of the bundle of frames. We cannot give here a consistent summary of this development, which is essentially based on the pseudo-complex geometry and on the Born-Infeld dynamics, but some results can be expressed with the notations introduced above and it is interesting to compare them with the point of view described in the present section.
The relevant formula is $$b^i{}_k = a_M b^0 [\tanh \epsilon]^i{}_k,$$ where the matrix $\epsilon$ has the antisymmetry property $$\eta_{ij} \epsilon^j{}_k + \eta_{kj} \epsilon^j{}_i = 0,$$ namely it represents an element of the Lorentz Lie algebra $o(1, 3)$. It can be expressed in terms of the forces acting on the particle.
The function $\tanh$ is defined by a power series of matrices and, as in the real numeric case, its range is not the whole vector space $o(1, 3)$. The inequalities that describe this range can be calculated by choosing a suitable Lorentz frame [@SP; @Schuller2] and one obtains $$\|\mathbf{a}\|^2 - \|\boldsymbol{\omega}\|^2 +
a_M^{-2} (\mathbf{a} \cdot \boldsymbol{\omega})^2 \leq a_M^2.$$ This inequality is essentially different from eq. (\[Omega\]), in particular it does not define a convex set, it allows any value of the angular velocity $\boldsymbol{\omega}$ and it is Lorentz invariant. If one is dealing with a point particle, one can put $\boldsymbol{\omega} = 0$ and the two formulas coincide.
Feasible transformations.
=========================
The treatment of the preceding section, even after the introduction of rotating particles, still maintains the constraint (\[Constr4\]), which does not permit us to give a physical meaning to all the vectors belonging to $\mathcal{T}^+$. One can overcome this obstacle by adopting a different interpretation, which is the one originally proposed in ref. [@Toller1] and is based on the ideas discussed in ref. [@Toller3].
The starting point is the remark that a local reference frame is defined by some material object and any physical procedure, in order to be localized in space and time, must refer to some of these objects. In particular, a procedure which has the aim to build a new (material) reference frame starting from a pre-existent one is called a [*transformation*]{}.
A transformation is described, at least approximately, by an element of the Poincaré group, but only some elements correspond to [*feasible*]{} transformations, because, as pointed out in ref. [@Toller3], it takes some time to translate, accelerate or rotate a (material) reference frame. The infinitesimal feasible transformations are described by elements of the cone $\mathcal{T}^+ \subset \mathcal{T}$ described in the preceding section. The assumption that $\mathcal{T}^+$ is a cone means that the composition of two feasible infinitesimal transformations is again feasible.
This point of view can also be applied to the space-time-velocity manifold $\mathcal{V}$ and it permits a physical interpretation of tangent vectors that do not respect the constraint (\[Constr2\]). A point of $\mathcal{V}$ represents the position of the origin and the time-like four-vector $e_0 = u$ of a tetrad, without giving any information about the three other spacelike four-vectors. However, $u$ is not interpreted as the four-velocity of a particle and the constraint (\[Constr2\]) is not necessary. One may say that a point of $\mathcal{V}$ represents a class of reference frames which are (partially) physically defined by a spherically symmetric material object.
In this way we have defined a mapping $\mathcal{S} \to \mathcal{V}$. Points of $\mathcal{S}$ which differ by a rotation correspond to the same point of $\mathcal{V}$ and we can represent $\mathcal{V}$ as a quotient space $$\label{Quotient}
\mathcal{V} = \mathcal{S}/SO(3).$$ This formula provides a bridge between the formalisms of the sections 2, 3 and the one of section 4.
The corresponding linear mapping between a tangent space of $\mathcal{S}$, identified with $\mathcal{T}$, and a tangent space of $\mathcal{V}$ introduces in the latter tangent space the seven coordinates $b^0, \mathbf{b}, \mathbf{b}''$ (defined up to a rotation). This mapping transforms the cone $\mathcal{T}^+$ into a cone tangent to $\mathcal{V}$. It is interestig to compare this cone with the cones (\[Cond3\]) and (\[Cond7\]) defined in the section 2 and 3, which, with the new notation, take, respectively, the form $$\label{Cond9}
\|\mathbf{b}\|^2 + \|\mathbf{b}''\|^2 \leq (b^0)^2,
\qquad b^0 \geq 0,$$ $$\label{Cond10}
\|\mathbf{b}\|^2 + 2 |\mathbf{b} \cdot \mathbf{b}''| + \|\mathbf{b}''\|^2 \leq (b^0)^2,
\qquad b^0 \geq 0.$$
We start with another description of $\mathcal{T}^+$ as the cone generated by the elements of the form [@Toller1; @Toller2] $$\label{Extr}
\|\mathbf{b}\| = \|\mathbf{b}''\| = b^0, \qquad \mathbf{b} \cdot \mathbf{b}'' = 0,$$ $$\mathbf{b}' = (b^0)^{-1} \mathbf{b}'' \times \mathbf{b}.$$ These equations define the extremal elements of $\mathcal{T}^+$. Their projections on a tangent space of $\mathcal{V}$ are defined by eq. (\[Extr\]) alone and they generate the projection of the cone $\mathcal{T}^+$ in this space. The elements defined by eq. (\[Extr\]) satisfy neither eq. (\[Cond9\]) nor eq. (\[Cond10\]) and they generate a cone different from the cones defined by these equations.
Conclusions and outlook.
========================
In sections 2, 3, and 4 we have discussed three different geometric descriptions of the maximal acceleration hypothesis. The first and the second are based on the relativistic phase space $\mathcal{W}$ and have, respectively, the symmetry groups $O(2, 6) \times \mathbf{R}^*$ and $C \times_s (O^{\uparrow}(1, 3) \times \mathbf{R}^* \times O^{\uparrow}(1, 3) \times \mathbf{R}^*)$, while the third is based on the bundle $\mathcal{S}$ of the Lorentz frames and has the symmetry group $GL(4, \mathbf{R})$, acting linearly on Majorana spinors (not on the space-time coordinates). The third formalism also allows a treatment of the rotational degrees of freedom and implies an upper bound to the angular velocity.
Note that the above mentioned symmetry groups concern some aspects of particle kinematics, while other aspects of kinematics and dynamics have lower symmetry properties unless they are modified. Since the action of these groups mixes the space-time coordinates with other variables, the important concept of space-time coincidence of two events looses the absolute character it has in relativity theory.
In order to compare the three formalisms, one has to introduce the space-time-velocity manifold $\mathcal{V}$, obtained from $\mathcal{W}$ by imposing the “on shell” constraint (\[Constr1\]), or from the space $\mathcal{S}$ by identifying the frames that differ for a rotation, namely by performing the quotient (\[Quotient\]). We have seen that the three approaches define three different cones in the tangent spaces of $\mathcal{V}$.
However, the variables that describe a point particle are subject to additional constraints and when these constraints are taken into account the three formalisms are equivalent. In order to give a physical meaning to the differences we have found, we have to go beyond the particle interpretation. An interpretation in terms of local reference frames and infinitesimal transformations is summarized in section 5. From this wider point of view, the three geometries are physically different. We have also seen that other differences appear when one considers rotating particles.
Since experimental tests look rather difficult, the choice between these kinematic descriptions depends on the success of the attempts to find an underlying dynamical scheme in which the limitations to the acceleration appear naturally and not as an extraneous requirement superimposed to a preceding theory. Some general remarks will be given in a forthcoming paper.
Finally, we show that, if the formalism of section 3 is interpreted as an upper bound $f_M$ to the force (instead of the acceleration), it is compatible with the formalism of section 4 and one can apply both the ideas at the same time. This can be obtained by considering the cotangent bundle $T^*\mathcal{S}$ of the manifold $\mathcal{S}$ and introducing in each cotangent space the coordinates $p_k$ and $p_{ik} = - p_{ki}$, which represent, respectively, the four-momentum and the relativistic angular momentum, as it is discussed in detail in refs. [@Toller4; @Vanzo].
In order to simplify the notation, it is convenient to introduce the greek indices $\alpha, \beta$, which take the values $0, 1 ,\ldots, 9$ and to indicate the quantities introduced above by $p_{\alpha}$. In a similar way we indicate the quantities $b^i$ and $b^{ik}$ by $b^{\alpha}$. There is no $GL(4, \mathbf{R})$ invariant metric in the ten-dimensional space $\mathcal{T}$, which permits one to raise or lower the greek indices. However, as we shall discuss elsewhere, one can choose a subgroup of $GL(4, \mathbf{R})$ isomorphic to the symplectic group $Sp(4, \mathbf{R})$ and locally isomorphic to the anti-de Sitter group $SO(2, 3)$, which is still sufficient to fix the value of $a_M$ and admits an invariant metric tensor. By means of this tensor one can define the quantities $p^{\alpha}$ that transform under $Sp(4, \mathbf{R})$ in the same way as $b^{\alpha}$. We indicate by $\dot p^{\alpha}$ the derivative with respect to the parameter $\tau$ which appears in eq. (\[Deriv\]).
The ideas of sections 3 and 4 can be condensated in the inequalities $$\label{Cond11}
(b_+^{\alpha}) = (b^{\alpha} + f_M^{-1} \dot p^{\alpha}) \in \mathcal{T}^+, \qquad
(b_-^{\alpha}) = (b^{\alpha} - f_M^{-1} \dot p^{\alpha}) \in \mathcal{T}^+$$ which generalize eq. (\[Cond6\]). They define in the tangent spaces of $T^*\mathcal{S}$ a cone symmetric with respect to the group $GL(4, \mathbf{R}) \times GL(4, \mathbf{R})$, though only the subgroup $Sp(4, \mathbf{R}) \times Sp(4, \mathbf{R})$ is physically relevant. This symmetry group contains the subgroup $SL(2, \mathbf{C)} \times SL(2, \mathbf{C)}$, which is the universal covering of the group $SO^{\uparrow}(1, 3) \times SO^{\uparrow}(1, 3)$ introduced in section 3, and also the diagonal subgroup $GL(4, \mathbf{R})$ introduced in section 4.
The cone (\[Cond11\]) describes upper bounds to velocity, angular velocity, acceleration, power, force and torque. The corresponding geometry contains the fundamental constants $c$, $a_M$ and $f_M$ and therefore it fixes the fundamental scales for all the physical quantities, in particular for the action. A classical (non quantum) treatment is justified only if $$f_M a_M^{-2} c^3 \gg \hbar.$$
Characterization of the cone $V^+$.
===================================
As an introduction to the less trivial proof of the next appendix, we show that the conditions a) and b) of section 1 define the cone $V^+$ up to the choice of the time and length units and up to a time inversion.
First we remark that the intersection of the cone and the three-dimensional plane $x^0 = \tau$ is a closed convex rotation invariant set, namely the whole plane, a closed ball, a point or the empty set. The whole plane must be excluded, because it contains straight lines. Since the cone cannot contain the straight line $\mathbf{x} = 0$, the intersection must be empty for $\tau < 0$ or for $\tau > 0$. The second possibility is reduced to the first one by means of a time inversion.
The intersection with the plane $x^0 = 1$ cannot be reduced to a point, otherwise the cone is reduced to a half line, which has an empty interior. We choose the unit of lenght in such a way that it is a ball with unit radius. The cone is the union of the half lines starting from the origin that intersect this ball.
Characterization of the cone $W^+$.
===================================
In this appendix we show that the conditions a), b) and c) of section 3 define the cone $W^+$ described by eq. (\[Cond6\]) with a suitable choice of the parameter $a_M$ and up to a time inversion.
To indicate an element of $W^+$, we use the simpler notation $(x, u)$ (instead of $(dx, du)$), where $x$ and $u$ are four-vectors. From the assumption c) and the convexity property we see that if $(x, u) \in W^+$ we have also $(x, -u) \in W^+$ and $(x, 0) \in W^+$. The last condition means that $x$ belongs to a Lorentz-invariant closed cone in the four-vector space, namely $V^+$ or $-V^+$. Then, possibly after a time inversion, we have $x \in V^+$.
We choose $k \in L^+$, namely a light-like four-vector $k$ with $k^0 > 0$ and we consider the equations $$y^0 = k \cdot x \geq 0, \qquad y^1 = k \cdot u,$$ which define a linear mapping $W \to \mathbf{R}^2$. The image $I$ of this mapping is a wedge in the two-dimensional half-plane $y^0 \geq 0$, invariant under the transformation $y^1 \to - y^1$. Any pair of four-vectors $k, k' \in L^+$ are connected by an orhochronous Lorentz transformation. Since the cone $W^+$ is invariant with respect to these transformations, we see that the image $I$ does not depend on the particular choice of $k$.
We indicate by $(x', u')$ the element obtained from $(x, u) \in W^+$ by means of a Lorentz boost with rapidity $\zeta$ in the direction of $-\mathbf{k}$, the multiplication by the factor $2\exp(-\zeta)$ and the limit $\zeta \to +\infty$. Since the cone $W^+$ is Lorentz invariant and closed, $(x', u')$ belongs to $W^+$. In a suitable reference frame we have $$k = (1, 0, 0, -1), \qquad y^0 = x^0 + x^3, \qquad y^1 = u^0 + u^3,$$ $$\begin{aligned}
&x' = \lim_{\zeta \to +\infty}2\exp(-\zeta)(x^0 \cosh\zeta + x^3 \sinh\zeta, x^1, x^2, x^0\sinh\zeta + x^3 \cosh\zeta) =&
\nonumber \\
&= (x^0 + x^3, 0, 0, x^0 + x^3) = y^0 h,&\end{aligned}$$ $$\begin{aligned}
&u' = \lim_{\zeta \to +\infty}2\exp(-\zeta)(u^0 \cosh\zeta + u^3 \sinh\zeta, u^1, u^2, u^0\sinh\zeta + u^3 \cosh\zeta) =&
\nonumber \\
&= (u^0 + u^3, 0, 0, u^0 + u^3) = y^1 h,&\end{aligned}$$ where $$h = (1, 0, 0, 1).$$
We have seen that $(y^0, y^1) \in I$, if and only if $(y^0 h, y^1 h) \in W^+$. We have proven this property for a particular choice of the four-vector $h$, but it follows from the Lorentz symmetry that it holds for any $h \in L^+$. It follows that $I$, being the inverse image of $W^+$ under the continuous mapping $(y^0, y^1) \to (y^0 h, y^1 h)$, is closed.
If $I$ is the whole closed half plane, one can easily see that $W^+$ contains the plane containing the elements $(0, u)$, in contradiction with our assumptions. Then, $I$ is a closed cone defined by $$|y^1| \leq a_M y^0,$$ where $a_M$ is a positive number. This means that $(x, u) \in W^+$ implies $$k \cdot (a_M x \pm u) \geq 0$$ for all the four-vectors $k \in L^+$. This condition is equivalent to eq. (\[Cond6\]). On the other hand, $W^+$ contains all the elements of the kind $(h, \pm a_M h)$ with $h \in L^+$. In other words, it contains all the elements with $x_+ = 2h, \,\, x_- = 0$ and with $x_+ = 0, \,\, x_- = 2h$ and by convexity all the elements which satisfy the condition (\[Cond6\]).
[999]{}
E. R. Caianiello: [*Is There a Maximal Acceleration?*]{} Lett. Nuovo Cimento [**32**]{} (1981) 65.
M. Toller: [*Symmetry and Feasibility of Infinitesimal Transformations.*]{} Nuovo Cimento [**B 64**]{} (1981) 471.
E. R. Caianiello, S. De Filippo, G. Marmo and G. Vilasi: [*Remarks on the Maximal Acceleration Hypothesis.*]{} Lett. Nuovo Cimento [**34**]{} (1982) 112.
H. E. Brandt: [*Maximal Proper Acceleration Relative to Vacuum.*]{} Lett. Nuovo Cimento [**38**]{} (1983) 522.
H. E. Brandt: [*The Maximal Acceleration Group.*]{} Group Theoretical Methods in Physics, ed. W. W. Zachary, Singapore (1984).
G. Scarpetta: [*Relativistic Kinematics with Caianiello’s Maximal Proper Acceleration.*]{} Lett. Nuovo Cimento [**41**]{} (1984) 51.
E. R. Caianiello: [*Maximal Acceleration as a Consequence of Heisenberg’s Uncertainty Relations.*]{} Lett. Nuovo Cimento [**41**]{} (1984) 370.
E. R. Caianiello, and G. Landi: [*Maximal Acceleration and Sakharov’s Limiting Temperature.*]{} Lett. Nuovo Cimento [**42**]{} (1985) 70.
C. S. Sharma and S. Srirankanathan: [*On Caianiello’s Maximal Acceleration.*]{} Lett. Nuovo Cimento [**44**]{} (1985) 275.
W. Guz, and G. Scarpetta: [*Special Relativity with the Maximal Proper Acceleration.*]{} in “Quantum Field Theory”, F. Mancini editor (1986).
H. E. Brandt: [*Maximal Acceleration Invariant Phase Space.*]{} in “The Physics of Phase Space”, Lecture Notes in Physics, Y. S. Kim and W. W. Zachary editors, Springer-Verlag (1987).
H. E. Brandt: [*Maximal Proper Acceleration and the Structure of Spacetime.*]{} Found. Phys. Lett. [**2**]{} (1989) 39.
W. R. Wood, G. Papini and Y. Q. Cai: [*Maximal Acceleration and the Time-Energy Uncertainty Relation.*]{} Nuovo Cimento [**B 104**]{} (1989) 361.
W. R. Wood, G. Papini and Y. Q. Cai: [*Conformal Transformations and Maximal Acceleration.*]{} Nuovo Cimento [**B 104**]{} (1989) 653.
M. Toller: [*Maximal Acceleration, Maximal Angular Velocity and Causal Influence.*]{} Inter. Journ. Theor. Phys. [**29**]{} (1990) 963.
H. E. Brandt: [*Structure of Spacetime Tangent Bundle.*]{} Found. Phys. Lett. [**4**]{} (1991).
V. V. Nesterenko, A. Feoli and G. Scarpetta: [*Complete Integrability for Lagrangians Dependent on Acceleration in a Spacetime of Constant Curvature.*]{} Class. Quantum Grav. [**13**]{} (1996) 1201.
S. Kalyana Rama: [*Classical Velocity in kappa-deformed Poincaré Algebra and a Maximum Acceleration.*]{} Mod. Phys. Lett. [**A 18**]{} (2003) 527, hep-th/0209129.
C. Castro: [*On Maximal-Acceleration, Strings, and the Group of Minimal Planck-Area Relativity Theory.*]{} hep-th/0211053.
G. Papini: [*Revisiting Caianiello’s Maximal Acceleration.*]{} Nuovo Cimento [**117 B**]{} (2003) 1325, quant-ph/0301142.
F. P. Schuller: [*Born-Infeld Kinematics.*]{} Annals Phys. [**299**]{} (2002) 174, hep-th/0203079.
F. P. Schuller: [*Born-Infeld Kinematics and Correction to the Thomas Precession.*]{} Phys. Lett. [**B 540**]{} (2002) 119, hep-th/0207047.
F. P. Schuller, M. N. R. Wohlfarth and T. W. Grimm: [*Pauli-Villars Regularisation and Born-Infeld Kinematics.*]{} Class. Quantum Grav. [**20**]{} (2003) 4269, hep-th/0211264.
F. P. Schuller and H. Pfeiffer: [*Invariant Length Scale in Relativistic Kinematics – Lessons from Dirichlet Branes.*]{} Phys. Lett. [**B 578**]{} (2003) 402, hep-th/0307247.
F. P. Schuller: [*Almost Product Manifolds as the Low Energy Geometry of Dirichlet Branes.*]{} European Physical Journal [**C**]{} (2004) , gr-qc/0310096.
M. Born: [*A Suggestion for Unifying Quantum Theory and Relativity.*]{} Proc. Roy. Soc. Lond. [**A 165**]{} (1938) 291.
M. Born: [*Reciprocity Theory of Elementary Particles.*]{} Rev. Mod. Phys. [**21**]{} (1949) 463.
P. A. M. Dirac: [*Lectures on Quantum Mechanics.*]{} Belfer Graduate School of Science, New York (1964).
J.-M. Souriau: [*Structure des systémes dynamiques.*]{} Dunod, Paris (1970).
S. G. Low: [*$U(3, 1)$ Transformations with Invariant Symplectic and Orthogonal Metrics.*]{} Nuovo Cimento [**108 B**]{} (1993) 841.
S. G. Low: [*Canonically Relativistic Quantum Mechanics: Representations of the Unitary Semi-direct Heisenberg Group $U(1, 3) \otimes_s H(1, 3)$.*]{} J. Math. Phys. [**38**]{} (1997) 2197.
S. G. Low: [*Representations of the Canonical Group (the Semi-direct Product of the Unitary and Weyl-Heisenberg Group), Acting as a Dynamical Group on Noncommutative Extended Phase Space.*]{} J. Phys. [**A 35**]{} (2002) 5711, math-ph/0101024.
M. Born and L. Infeld: [*Foundations of the New Field Theory.*]{} Proc. Roy. Soc. Lond. [**A 144**]{} (1934) 425.
H. P. Künzle: [*Canonical Dynamics of Spinning Particles in Gravitational and Electromagnetic Fields.*]{} J. Math. Phys. [**13**]{} (1972) 739.
M. Toller: [*An Operational Analysis of the Space-Time Structure.*]{} Nuovo Cimento [**B 40**]{} (1977) 27.
M. Toller: [*Extended Test Particles in Geometric Fields.*]{} J. Math. Phys. [**24**]{} (1983) 613.
L. Vanzo: [*Presymplectic Dynamics of a Pole-Dipole Particle in External Geometric Fields.*]{} Nuovo Cimento [**B 72**]{} (1982) 122.
[^1]: e-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'F. Combes'
- 'A. Moiseev'
- 'V. Reshetnikov'
date: 'Received 2013/ Accepted 2013'
title: 'Molecular content of polar-ring galaxies [^1]'
---
Introduction
============
Polar-ring galaxies (PRGs, see Fig. \[fig:sample\]) are peculiar objects composed of a central component (usually an early-type galaxy) surrounded by an outer ring or disk, made up of gas, stars, and dust, which orbits nearly perpendicular to the plane of the gas-poor central galaxy (Whitmore 1990). For some well-known systems studied with high spatial resolution, it was possible to show that the polar system is in fact a polar disk, more than a polar ring (e.g. Iodice 2002b). However, polar disks and rings are both gathered into the same PRG class.
Measurements of their kinematics can therefore give some insight in the 3D-shape of their dark matter, which can be generalised to the progenitor spiral galaxies, provided that their formation mechanism is known.
We note that many of the best cases of PRGs have a relatively massive polar component that cannot be treated as simple test particles, but self-gravity must be taken into account.
Formation scenarios for polar-ring galaxies
-------------------------------------------
From dynamical arguments, the two misaligned systems cannot be formed simultaneously, as confirmed by the younger ages of the polar rings/disks (Iodice 2002a, b), and the fact that most of the gas of the system is in the polar disk (van Driel 2000, 2002). At least three formation mechanisms have been discussed in the literature, the first two involving galaxy interactions:
1\) the accretion scenario, where two interacting galaxies exchange mass, as invoked by Schweizer (1983) and simulated by Reshetnikov & Sotnikova (1997);
2\) the merging scenario, or the head-on collision of two orthogonal spiral galaxies, first studied by Bekki (1997, 1998). Bournaud & Combes (2003) have shown through simulations that statistically the first scenario is more frequent, and more likely to represent observations;
3\) the cosmic formation scenario, where the PRGs form through the misaligned accretion of gas from cosmic filaments (Maccio 2006; Brook 2008). The gas of the PR is then of lower metallicity than in the first scenarios.
The fact that some PRGs are polar disks more than polar rings supports the formation through cosmic accretion, as shown by Spavone (2010). Also, the low metallicity, and flat abundance gradients could favour this mechanism (Spavone 2011), while the presence of a true ring supports the tidal accretion.
Snaith (2012) have developed in more detail the formation scenario proposed by Brook (2008). In their simulations, the polar disk is progressively formed out of gas and dark matter infalling from a cosmic filament, after the last major merger has re-oriented the old system in a perpendicular direction (and formed the host). This implies that the dark matter is aligned with the polar system.
The dark matter issue
---------------------
The 3D-shape of dark matter halos has been estimated in many objects (e.g. Combes 2002), and has recently been investigated through edge-on galaxies, allowing us to study the gas flaring of the disk (e.g. O’Brien 2010, Peters 2013), through gravitational lensing (van Uitert 2012), or through satellite disruption (Law 2009). There is a large scatter in the flattening derived, i.e. an axis ratio between 0.2 and 1. Polar-ring galaxies are privileged systems for this study. Since the ring is gas rich and rotates around the pole, it is a probe of the gravitational potential in the third dimension, which is normally inaccessible in normal spirals. It is therefore a unique tool for determining the 3D-shape and consequently constraining the nature of dark matter. From the shape of dark haloes in PRGs, we should be able to deduce the shape of dark haloes in normal galaxies, knowing the formation mechanisms of PRGs.
In all previous studies (e.g. Whitmore 1987, Sackett & Sparke 1990, Sackett 1994, Reshetnikov & Combes 1994, Combes & Arnaboldi 1996, Iodice 2003) the common conclusion is that PRGs are indeed embedded in a dark halo. However, the solutions for the 3D-shape differ, according to models and accuracy of data: the dark halo is almost spherical for Whitmore (1987), flattened along the equatorial plane of the host galaxy (Sackett 1994), or flattened along the polar-ring plane (Combes & Arnaboldi 1996). This last solution is supported for a large number of PRGs by Iodice (2003), through a study of the Tully-Fisher (TF) diagram for PRGs.
The position of the PRGs in the TF diagram is very peculiar, and does not fit the sequence of normal disk galaxies (Iodice 2003, Reshetnikov 2004). The rotational velocity is obtained through HI-21cm measurement from the gas in the edge-on polar disks. Most of the PRGs have higher than normal rotational velocities, for their luminosities. This is not expected if the dark halo is spherical or flattened to the equatorial plane of the host, because then the observed velocity corresponds to the apocenter of the excentric polar orbit, and is lower than the velocity observed in the equatorial plane.
Since the contrary is observed, this must be due to a flattening of the dark matter towards the polar plane. This important suggestion must be confirmed by a much larger statistics, and therefore we want to enlarge significantly the number of objects that can be considered as PRGs. Since this implies selecting objects that are more distant because the HI-21cm line is not as easy to detect at high redshift, the CO line then becomes the preferred tracer of the gas. In the present sample, our largest redshift is 0.078, but in the future with ALMA, the CO-TF will be a unique tool.
Molecular gas in PRGs
---------------------
The molecular content of PRGs is poorly known. The first CO detection in such an object was in NGC 660 (Combes 1992). Watson (1994) then detected CO(2-1) in the polar rings of NGC 2685 and NGC 4650A, and later Schinnerer & Scoville (2002) made an interferometric map of the Spindle galaxy NGC 2685. Van Driel (1995) found abundant CO emission in the inclined ring of NGC 660, and Crocker (2008) in the center of NGC 2768. Galletta (1997) observed ten PRGs in CO and found molecular masses much larger than those in early-type galaxies and also in dwarf galaxies, suggesting that PRGs cannot get their gas in a dwarf accretion only. All these are only a few cases, and more observations are required to know better the molecular content of polar-ring galaxies and to better understand their formation scenarios.
The sample is described in Sect. \[sample\] and the observations in Sect. \[obs\]. Results are presented in Sect. \[res\] and discussed in Sect. \[disc\].
The sample {#sample}
==========
Recently, Moiseev (2011) have built a new catalogue of PRGs, significantly increasing the number of known candidate PRGs. The catalogue is based on the results of the original Galaxy Zoo project, where nearly a million galaxies from the Sloan Survey (SDSS) were classified. This results in the Polar Ring Catalog (SPRC) of 275 objects, in which 70 galaxies are classified as the best PRGs, and 115 as good PRG candidates. Among the 70, there are 15 kinematically confirmed PRGs. For our search, we have selected the latter, and among the best candidates, the brightest in terms of r-magnitude (r$<$ 15.5) (cf. Table \[tab:sample\]).
The main goal of the present work is to obtain information on the molecular content of this sample of 21 PRGs. The global CO detection allows us to locate the object in the Tully-Fisher diagram, and to have a first information in the geometry of the dark halo. Future interferometric work on the detected PRGs will provide high resolution maps, in order to be able to compare results with numerical models.
In this article, we adopt a standard flat cosmological model, with $\Lambda$ = 0.73 and a Hubble constant of 71kms$^{-1}$Mpc$^{-1}$ (Hinshaw 2009).
------------- ---------------- ------------- ------------- --------- --------- --------- -------------- --------- ------------- -------------
SPRC Name R.A. Dec. $cz$ g r i PA beam$_{10}$ beam$_{21}$
No.$^{(1)}$ \[km/s\] \[mag\] \[mag\] \[mag\] \[$^\circ$\] \[kpc\] \[kpc\]
7C 07 52 34.32 +29 20 49.7 18032 17.84 16.95 16.85 48 27.9 14.0
10C 08 20 38.19 +15 36 59.8 12736 17.16 16.29 15.79 39 19.8 9.9
13 09 14 53.66 +49 38 24.0 9521 16.10 15.37 15.06 102 14.8 7.4
14C CGCG 121-053 09 18 15.97 +20 22 05.3 9548 15.59 14.73 14.48 35 14.8 7.4
15 09 36 34.63 +21 13 57.8 10281 15.31 14.43 14.14 32 16.0 8.0
17 09 59 11.85 +16 28 41.5 7914 15.81 14.92 14.59 14 12.3 6.2
24 11 16 25.11 +56 50 17.0 14133 15.85 14.98 14.63 160 21.9 11.0
29 11 53 33.56 +47 19 07.3 14208 16.30 15.25 15.06 120 22.1 11.0
31 12 17 11.51 +31 30 37.8 14913 16.21 15.08 14.93 175 23.1 11.6
33C NGC 4262 12 19 30.57 +14 52 39.5 1358 12.22 11.24 10.98 29 21.1 10.6
39C 13 08 16.92 +45 22 35.2 8792 17.01 16.01 15.75 59 13.7 6.8
42 UGC 08634 13 39 04.59 +02 09 49.5 7041 15.75 15.04 14.85 60 11.0 5.5
47 13 59 41.70 +25 00 46.1 9370 15.62 14.49 14.28 5 14.6 7.3
48 14 14 20.82 +27 28 04.4 16788 16.30 15.04 15.10 120 26.0 13.0
52 KUG 1416+257 14 18 25.60 +25 30 06.7 4450 15.70 15.07 14.50 145 6.9 3.5
56 MCG +06-33-026 15 11 14.09 +37 02 37.7 16499 15.69 14.81 14.49 80 25.6 12.8
60C 15 47 24.32 +38 55 50.4 23519 17.71 17.29 17.02 56 36.3 18.1
61 15 49 54.81 +09 49 43.1 13753 15.78 14.81 14.50 140 21.4 10.7
67C CGCG 225-097 17 17 44.13 +40 41 52.0 8325 15.28 14.25 14.01 165 12.9 6.5
69C II Zw 092 20 48 05.67 +00 04 07.8 7396 16.22 15.36 14.93 34 11.6 5.8
260C CGCG 068-056 11 45 30.25 +09 43 44.8 6399 15.31 14.48 14.23 18 10.0 5.0
------------- ---------------- ------------- ------------- --------- --------- --------- -------------- --------- ------------- -------------
\
$^{(1)}$The numbers are followed by C when kinematically confirmed.\
PA is the position angle of the polar ring (with uncertainty $\pm$5$^\circ$).\
The kinematics were obtained for SPRC-7 by Brosch (2010), for SPRC-33 by Bettoni (2010), for SPRC-67 by Merkulova (2012), for SPRC-10, 14, 39, 60, and 69 by Moiseev (2011), for SPRC-260, by Khoperskov (2012).\
Magnitudes are from SDSS.
Observations {#obs}
============
The observations were carried out with the IRAM 30m telescope at Pico Veleta, Spain, from December 2011 to January 2012. All sources were observed simultaneously in CO(1-0) and CO(2-1) lines, with the 3mm and 1mm receivers in parallel.
The broadband EMIR receivers were tuned in single sideband mode, with a total bandwidth of 4 GHz per polarization. This covers a velocity range of $\sim$ 10,400 at 2.6mm and $\sim$ 5,200 at 1.3mm. The observations were carried out in wobbler switching mode, with reference positions offset by $2\arcmin$ in azimuth. Several backends were used in parallel, the WILMA autocorrelator with $2$ MHz channel width, covering 4$\times$4GHz, and the $4$ MHz filterbanks, covering 2$\times$4GHz.
We spent on average two hours on each galaxy, and reached a noise level between 0.6 and 1.6 mK (antenna temperature), smoothed over $30$ km s$^{-1}$ channels for all sources. Pointing measurements were carried out every two hours on continuum sources and the derived pointing accuracy was 3$''$ rms. The temperature scale is then transformed from antenna temperature $T_{\rm A}^*$ to main beam temperature $T_{\rm mb}$, by multiplying by 1.17 at 3mm and 1.46 at 1.3mm. To convert the signals to fluxes, we use $S/T_{\rm mb}$ = 5.0 Jy/K for all bands. At 2.6mm and 1mm, the telescope half-power beam width is 23$''$ and 12$''$ respectively. The data were reduced with the CLASS/GILDAS software, and the spectra were smoothed so that each line covers about ten channels in the plots. [^2]
----- --------------- ------------- ------------------- ---------------- ------------------ -----
Area V $\Delta$V$^{(1)}$ T$_{mb}^{(2)}$ M(H$_2$)$^{(3)}$ M\*
No. K km/s km/s km/s mK
14C 0.81$\pm$ 0.1 5$\pm$ 11 165$\pm$ 24 4.6 8.4 23
1.04$\pm$ 0.1 -26$\pm$ 5 127$\pm$ 12 7.7
69C 0.71$\pm$ 0.1 6$\pm$ 15 193$\pm$ 40 3.5 4.5 11
0.68$\pm$ 0.1 -17$\pm$ 10 128$\pm$ 23 5.0
42 0.91$\pm$ 0.1 -22$\pm$ 16 231$\pm$ 34 3.6 5.2 7
0.82$\pm$ 0.2 -24$\pm$ 14 131$\pm$ 36 5.8
47 0.73$\pm$ 0.1 -59$\pm$ 61 550$\pm$ 110 1.3 7.3 40
0.72$\pm$ 0.2 74$\pm$ 57 479$\pm$ 95 1.5
48 0.65$\pm$ .05 21$\pm$ 17 380$\pm$ 50 1.7 21.5 70
0.72$\pm$ 0.2 142$\pm$ 67 450$\pm$ 143 1.6
----- --------------- ------------- ------------------- ---------------- ------------------ -----
: Molecular data and stellar mass of the five detected PRGs. For each object, the first line displays the CO(1-0) and the second line the CO(2-1) results.[]{data-label="tab:det"}
Results of the Gaussian fits
$^{(1)}$ FWHM
$^{(2)}$ Peak brightness temperature
$^{(3)}$ obtained with the standard MW conversion ratio
![Tully-Fisher relation for our detected PRGs (red stars), compared with other PRGs (black triangles, Whitmore 1990, van Driel 2002, Iodice 2003), and normal spirals (black dots, Giovanelli 1997). The I absolute magnitudes, with uniform photometry and distance corrections, are plotted vs. observed rotation velocities (full width at 20% power), in the polar rings. []{data-label="fig:TF"}](TF.ps){width="8cm"}
[l c c c c c ]{} SPRC &Line&$\nu_{\rm obs}$& rms & L’$_{\rm CO}$/10$^{9}$& M(H$_2$)\
No. & & \[GHz\] & \[mK\] & \[K pc$^2$\] & \[10$^9$\]\
7C & CO(10) & 108.7 & 0.8 & 1.16 & 5.3\
& CO(21) & 217.4 & 1.6 & 0.72 & 3.3\
10C & CO(10) & 110.6 & 0.6 & 0.43 & 2.0\
& CO(21) & 221.1 & 1.1 & 0.25 & 1.1\
13 & CO(10) & 111.7 & 0.9 & 0.36 & 1.6\
& CO(21) & 223.4 & 1.2 & 0.15 & 0.7\
15 & CO(10) & 111.4 & 0.7 & 0.33 & 1.5\
& CO(21) & 222.9 & 0.8 & 0.12 & 0.5\
17 & CO(10) & 112.3 & 0.7 & 0.19 & 0.9\
& CO(21) & 224.6 & 0.7 & 0.06 & 0.3\
24 & CO(10) & 110.1 & 0.8 & 0.71 & 3.2\
& CO(21) & 220.1 & 1.4 & 0.38 & 1.8\
29 & CO(10) & 110.0 & 0.7 & 0.63 & 2.9\
& CO(21) & 220.1 & 1.1 & 0.31 & 1.4\
31 & CO(10) & 109.8 & 0.7 & 0.69 & 3.2\
& CO(21) & 219.6 & 1.5 & 0.46 & 2.1\
33C & CO(10) & 114.7 & 1.1 & 0.90 & 4.1\
& CO(21) & 229.5 & 0.9 & 0.23 & 1.0\
39C & CO(10) & 111.9 & 0.7 & 0.24 & 1.1\
& CO(21) & 223.9 & 0.7 & 0.07 & 0.3\
52 & CO(10) & 113.6 & 0.9 & 0.08 & 0.3\
& CO(21) & 227.2 & 1.1 & 0.03 & 0.1\
56 & CO(10) & 109.2 & 0.9 & 1.09 & 5.0\
& CO(21) & 218.5 & 1.3 & 0.49 & 2.2\
60C & CO(10) & 106.9 & 0.6 & 1.49 & 6.9\
& CO(21) & 213.8 & 0.9 & 0.70 & 3.2\
61 & CO(10) & 110.2 & 0.6 & 0.50 & 2.3\
& CO(21) & 220.4 & 1.1 & 0.29 & 1.3\
67C & CO(10) & 112.1 & 0.7 & 0.21 & 1.0\
& CO(21) & 224.3 & 1.0 & 0.09 & 0.4\
260C & CO(10) & 112.9 & 1.0 & 0.18 & 0.8\
& CO(21) & 225.7 & 0.9 & 0.05 & 0.2\
The rms are in T$_{\rm A}^*$ in channels of 30 .
The upper limits in L’$_{\rm CO}$ and M() are at 3$\sigma$ with an assumed $\Delta$V = 300 .
Results {#res}
=======
CO detection in PRGs {#COdet}
--------------------
Figures \[fig:spec1\] and \[fig:spec2\] display the CO-detected sources, in both CO lines. Two of the five detections involved kinematically confirmed PRGs. Table \[tab:det\] reports all line parameters for the detections, and the upper limits for the non-detections are reported in Table \[tab:uplim\]. Integrated signals and velocity widths have been computed from Gaussian fits. These also give the central velocities, with respect to the optical redshift of Table \[tab:sample\]. The upper limits are computed at 3$\sigma$, assuming a common line width of 300and getting the rms of the signal over 300.
The detection rate of 24$\pm$9% is comparable to that found for early-type galaxies, for which the gas content is believed to be due to recent accretion, as was found in the SAURON and ATLAS$^{\rm 3D}$ samples (e.g. Combes 2007, Young 2011).
One problem in the interpretations of the data is the size of the CO beams, which are sometimes smaller than the total extent of the polar rings. Table \[tab:sample\] lists the values of the two beams in kpc on the major axis of the galaxies. For the CO(1-0) beam, two detected galaxies, SPRC-42 and SPRC-47, are particularly in this case, where the H$_2$ mass derived in Table \[tab:det\] might only be a lower limit. The shape of the CO(1-0) and CO(2-1) velocity profiles are not the same, since both lines resolve the polar ring differently. Among the upper limits, four polar rings are also in this case, SPRC-33, 52, 67 and 69. A full map would be required to settle the issue.
In all the detected galaxies, we interpret the CO emission as coming from the polar-ring system, since it is always the bluer one (cf. Fig. \[fig:sample\]). All five of the detected polar rings are almost edge-on, which minimizes the errors done in correcting for inclination the observed linewidths, to obtain indicative rotational velocities (cf. Fig. \[fig:TF\]).
The line widths detected are in average 304 Full Width at Half Maximum (FWHM). Two of the PRGs, P47 and P48, clearly show double-horn profiles, indicative of rotating disks or rings. For two of the detected galaxies, the rotational velocity has also been estimated in the ionized gas (Moiseev 2011). For SPRC-14 the ionized gas data in projection on the sky has a symmetrized maximal line-of-sight rotation velocity of 160 km/s, in good agreement with the CO(1-0) value of 165$\pm$24 km/s. The polar ring is, however, asymmetric, and with an extension on the NE side, where the velocity reaches up to Vmax= 262$\pm$11 km/s (H$\alpha$) and 268$\pm$5 (NII).
For SPRC-69 the ionized gas velocity field yields Vmax=178$\pm$3 km/s, (172 after projecting with the inclination of 76$^\circ$ of the polar ring) a value also in agreement with the CO(1-0) value 193$\pm$40 km/s.
These comparisons show that we are not underestimating too much the maximum velocities of the polar rings, even though the CO beam might not encompass all the optical extent of it. This might come from the fact that the maximum velocity is reached already at small radii, especially for the molecular gas, which is more concentrated towards the center. We have plotted our detected objects in the Tully-Fisher diagram of Fig. \[fig:TF\], in comparison with normal spirals from Giovanelli (1997) and other PRGs from Iodice (2003). We note that we have corrected the velocity for AM 2020-504, which is erroneous in Iodice , replacing it with the value from Whitmore (1990) and van Driel (2002). We have converted the $i$ to $I$ magnitudes by the formula $$\begin{aligned}
\nonumber
I-i &=& -0.0307(r-i)^3 +0.1163(r-i)^2 \\ & &
\nonumber
-0.3341(r-i)-0.3584\end{aligned}$$ from Ivezic (2007). The CO-detected PRGs show the same tendency to lie to the right of the main relation, i.e. their velocities are too high. Since the velocities determined from CO are lower limits (because of the restricted beam), this result is robust.
CO luminosity and mass {#CO-mass}
----------------------
We have simultaneously observed the two first lines of the CO rotational ladder, and it is interesting to compare them, to have an idea of the excitation of the gas. Therefore, we compute L’$_{CO}$, the special unit CO luminosity, through integrating the CO intensity over the velocity profile. This luminosity, expressed in units of K pc$^2$, will give the same value irrespective of J, if the CO lines are saturated and have the same brightness temperature.
This CO luminosity is given by $$L'_{CO} =
23.5 I_{CO} \Omega_B {{D_L^2}\over {(1+z)^3}} \hskip6pt \rm{K\hskip3pt
km \hskip3pt s^{-1}\hskip3pt pc^2}$$, where $I_{CO}$ is the intensity in K , $\Omega_B$ the area of the main beam in square arcseconds, and $D_L$ the luminosity distance in Mpc.
In most cases (15 out of the 21 objects in the sample) the beam in CO(1-0) is large enough that it is likely to encompass all the emission of the polar-ring system; however, it is not the same in CO(2-1), and only the CO(1-0)-derived H$_2$ masses should be trusted. For the detected objects, the integrated intensities are always comparable between the CO(1-0) and CO(2-1) lines, as can be seen in Table \[tab:det\]. For point sources, with saturated and thermalized CO lines, the ratio should be as high as 4, in favour of the CO(2-1). The fact that intensities are comparable can be interpreted either in terms of an extended emission, or a sub-thermal excitation, or both.
We have computed the molecular mass from the CO(1-0) flux, using M$_{\rm H_2} = \alpha$ L’$_{\rm CO}$, with $\alpha=4.6$ M$_\odot$ (K pc$^2$)$^{-1}$, the standard factor for nearby quiescent galaxies like the Milky Way. The molecular gas masses are listed in Table \[tab:det\] and the upper-limits in Table \[tab:uplim\].
The average CO luminosity for the five galaxies detected is L’$_{\rm CO}$ = 2.0 10$^{9}$ K pc$^2$, corresponding to an average mass of 9.4 10$^{9}$ . These gas masses are relatively high, and we now compare them to stellar masses for each system.
Stellar mass and gas fraction {#GFR}
-----------------------------
We compute stellar masses from observed optical (SDSS) and near infrared (2MASS) magnitudes, using calibrated relations between mass-to-light ratios and colours (see e.g. Bell 2003). The multi-wavelength luminosities were K-corrected according to the colours (cf. Chilingarian 2010). Stellar masses are displayed in Table \[tab:det\]. The gas fractions derived from these stellar masses show large variations, between 15% and 43%, with an average of 27%. These gas fractions are quite high, compared to normal spirals at the same redshifts. The selection of bright polar-ring galaxies at these distances means therefore the selection of galaxies that have just accreted a large amount of gas mass.
Summary and discussion {#disc}
======================
We have presented our CO survey in 21 polar-ring galaxies, observed with the IRAM-30m telescope. Five galaxies were detected, and among them two kinematically confirmed PRGs. The detection rate of 24$\pm$9% is comparable to early-type galaxies, where the gas is thought to have been recently accreted. The first two CO lines were observed, and the L’(CO) luminosities in CO(2-1) are lower than in CO(1-0), indicating either a subthermal gas, and/or a gas extent larger than the $\sim$ 12” CO(2-1) beam. Assuming a standard CO-to- conversion factor, the average molecular gas mass is found to be 9.4 10$^{9}$ . The average ratio between gas and stellar mass is 0.4, or the average gas fraction is 27%. This high fraction means that bright polar-ring galaxies have just accreted a large amount of gas.
We interpret our CO detections as coming from the polar-ring systems in the detected galaxies, since it is the bluer and younger component. Deriving the rotational velocity of the gas from the CO profile, we have placed our observed galaxies in the Tully-Fisher diagram, allowing us to compare them with the control sample of normal spirals. The new detected objects confirm the offset position of PRGs already noticed by Iodice (2003).
In Fig. \[fig:TF\], one of the PRGs falls in the expected range for a spherical potential (SPRC-14), two are mildly offset (SPRC-69 and SPRC-42), and two are very far from the usual relation, with very broad velocities (SPRC-48 and SPRC-47). This large scatter might indicate different formation mechanisms for the systems, and at least four out of the five should have their dark matter halo aligned with the polar plane. The position of the five PRGs in the TF diagram does not appear related to the mass ratio between the polar disk/ring and the host. The galaxies SPRC-42 and SPRC-69 have polar structures brighter than their hosts, while SPRC-14 and SPRC-48 are comparable, and SPRC-47 is weaker than its host. This was also the case in the previous PRGs (Iodice 2003). There is no obvious relation with the PR spatial extent either.
The two most extreme cases (SPRC-47 and SPRC-48) must have a very flattened dark-matter system, as flat as a disk. This result is however uncertain, since the mass distribution is not yet known, and the CO beam does not cover the whole polar disk: the velocity could be higher in the center, and the flat portion of the rotation curve be lower. A full CO map at high spatial resolution is required to conclude.
We warmly thank the referee for constructive comments and suggestions. The IRAM staff is gratefully acknowledged for their help in the data acquisition. F.C. acknowledges the European Research Council for the Advanced Grant Program Number 267399-Momentum. A.M. is also grateful to the ‘Dynasty’ Foundation and RFBR grant 13-02-00416. V.R. acknowledges partial financial support from the RFBR grant 11-02-00471−a. We made use of the NASA/IPAC Extragalactic Database (NED), and of the HyperLeda database.
Bekki K. 1997, ApJ, 490, L37 Bekki K. 1998, ApJ, 499, 635 Bell E., McIntosh, D. H., Katz, N., Weinberg, M. D.: 2003, ApJS 149, 289 Bettoni D., Buson L. M. & Galletta G. 2010, A&A, 519, 72 Bournaud F., Combes F. 2003, A&A, 401, 817 Brook C. B., Governato F., Quinn T. 2008, ApJ, 689, 678 Brosch, N., Kniazev, A., Moiseev, A., Pustilnik, S. A.: 2010, MNRAS 401, 2067 Chilingarian, I. V., Melchior, A.-L., Zolotukhin, I.: 2010, MNRAS 405, 1409 Combes F., Braine J., Casoli F., Gerin M., van Driel W.:1992, A&A 259, L65 Combes F., Arnaboldi M. 1996, A&A, 305, 763 Combes F.: 2002, NewAR 46, 755 (astro-ph/0206126) Combes F., Young L.M., Bureau M.: 2007 MNRAS 377, 1795 Crocker, A. F., Bureau, M., Young, L. M., Combes, F.: 2008, MNRAS 386, 1811 Galletta G., Sage L.J., Sparke L.S.: 1997, MNRAS 284, 773 Giovanelli R., Haynes M.P., Herter T. 1997, AJ 113, 22 Iodice, E., Arnaboldi, M., Sparke, L. S. : 2002a, A&A 391, 103 & 117 Iodice, E., Arnaboldi, M., de Lucia G. : 2002b, AJ 123, 195 Iodice E., Arnaboldi, M., Bournaud, F. : 2003, ApJ 585, 730 Ivezic, Z., Smith, J. A., Miknaitis, G. 2007, ASP-Conf 364, 165 Hinshaw G., Weiland J.L., Hill R.S. : 2009, ApJS 180, 225 Khoperskov S.A., Moiseev A.V., Khoperskov A.V.: (2012), EWASS in press Law, D.R., Majewski, S.R., Johnston, K.V.: 2009, ApJ 703, L67 Lüghausen F. , Famaey B., Kroupa P. : 2013, MNRAS, in press Maccio A. V., Moore B., Stadel J.: 2006, ApJ, 636, L25 Merkulova O.A.,, Karataeva G. M., Yakovleva V.A., Burenkov A.N 2012, AstBu 67, 374 Moiseev A., Smirnova K.I., Smirnova A.A., Reshetnikov V.: 2011, MNRAS 418, 244 O’Brien, J. C., Freeman, K. C., van der Kruit, P. C. 2010, A&A 515, A62 Peters, S.P.C., van der Kruit, P.C., Allen R.J., Freeman, K.C.: 2013, A&A, sub Reshetnikov V., Combes F. 1994, A&A, 291, 57 Reshetnikov V., Sotnikova N. 1997, A&A, 325, 933 Reshetnikov V., 2004, A&A 416, 889 Sackett P.D., Sparke L. 1990, ApJ, 361, 408 Sackett P.D. 1994, ApJ, 436, 629 Schinnerer, E., Scoville, N.: 2002, ApJ 577, L103 Schweizer F., Whitmore B.C., Rubin V.C. 1983, AJ, 88, 909 Snaith, O. N., Gibson, B. K., Brook, C. B. 2012, MNRAS 425, 1967 Spavone, M., Iodice, E., Arnaboldi, M. 2010, ApJ 714, 1081 Spavone, M., Iodice, E., Arnaboldi, M. 2011, A&A 531, A21 van Driel W., Arnaboldi M., Combes F., Sparke L.: 2000, A&AS 141, 385 van Driel W., Combes F., Arnaboldi M., Sparke L.: 2002, A&A 386, 140 van Driel, W., Combes, F., Casoli, F. : 1995, AJ 109, 942 van Uitert, E., Hoekstra, H., Schrabback, T. 2012, A&A, 545, A71 Watson, D. M., Guptill, M. T., & Buchholz, L. M. 1994, ApJ, 420, L21 Whitmore B.C., McElroy D.B., Schweizer F. 1987, ApJ, 314, 439 Whitmore B.C., Lucas R.A., McElroy D.B. 1990, AJ, 100, 1489 Young L.M., Bureau M., Davis T.A. : 2011 MNRAS 414, 940
[^1]: Based on observations carried out with the IRAM 30m telescope. IRAM is supported by INSU/CNRS (France), MPG (Germany), and IGN (Spain)
[^2]: Spectra of detections are available in electronic form at the CDS via anonymous ftp to cdsarc.u-strasbg.fr (130.79.128.5) or via http://cdsweb.u-strasbg.fr/cgi-bin/qcat?J/A+A/
| {
"pile_set_name": "ArXiv"
} |
Dispersive imaging with off-resonant light is an important technique for observing Bose-Einstein condensates [@andr96; @kett99var]. Compared to absorption imaging it causes much less heating, and hence, allows the recording of non-destructive real-time “movies” of the dynamics of Bose-Einstein condensates [@andr97prop]. We observed that a limitation of dispersive imaging comes from residual absorption or Rayleigh scattering. The momentum transfer to the condensate atoms depletes the condensate and heats the cloud due to the transferred recoil energy [@andr96; @kett99var].
In contrast, a recent paper [@leon99] emphasized that the limit of dispersive imaging is not residual absorption, but a different form of quantum backaction of the probe light which was determined with a new approach to quantum-optical propagation. This note points out that these conclusions are incorrect, and that Rayleigh scattering is the dominant quantum backaction of dispersive imaging.
First, the absorption rate cannot be completely suppressed by imaging with far-detuned light. For a desired signal-to-noise ratio, a further detuning has to be compensated by higher laser intensity in such a way that the rate of far-wing absorption is constant [@kett99var]. The absorption rate per atom is simply the Rayleigh scattering rate $ \gamma_s=\Gamma f_{\rm exc}$, where $\Gamma$ is the natural linewidth and the excited state fraction $f_{\rm exc} = (\omega_R/2\Delta)^2$ is given by the Rabi frequency $\omega_R $ of the probe light and its detuning $\Delta$. The recoil due to the scattering of photons knocks atoms out of the condensate and depletes it with a rate $\gamma_s$.
Leonhardt et al. [@leon99] derived an expression for the depletion of the condensate $\gamma_L$ (their Eq. (62)). The rate $\gamma_L$ turns out to be proportional to the absorption rate $\gamma_s$ but is smaller by a factor of (3/16) [@leon-pc]. This indicates that the calculated backaction is related to Rayleigh scattering. It seems that it is just Rayleigh scattering with the smaller prefactor caused by approximations of the theory. Therefore, the statement by the authors that their result is *qualitatively* different from Rayleigh scattering is inconsistent with their results.
Another major result of Ref. [@leon99] is that the phase diffusion rate is always smaller than the depletion rate. Our experiments [@andr96; @andr97prop] were not sensitive to perturbations of the phase, and we didn’t estimate this effect.
In conclusion, residual absorption or Rayleigh scattering is the dominant perturbation of dispersive imaging, and this process is the dominant quantum backaction of the probe light on the Bose-Einstein condensate.
[1]{}
M. Andrews, M.-O. Mewes, N. van Druten, D. Durfee, D. Kurn, and W. Ketterle, Science [**273**]{}, 84 (1996).
W. Ketterle, D. Durfee, and D. Stamper-Kurn, in [*Bose-Einstein condensation in atomic gases*]{}, [*Proceedings of the International School of Physics Enrico Fermi, Course CXL*]{}, edited by M. Inguscio, S. Stringari, and C. Wieman (IOS Press, Amsterdam, 1999), pp. 67–176.
M. Andrews, D. Kurn, H.-J. Miesner, D. Durfee, C. Townsend, S. Inouye, and W. Ketterle, Phys. Rev. Lett. [**79**]{}, 553 (1997).
U. Leonhardt, T. Kiss, and P. Piwnicki, Eur. Phys. J. D [**7**]{}, 413 (1999).
U. Leonhardt, T. Kiss, and P. Piwnicki, reply to this comment.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Systems composed of large numbers of interacting agents often admit an effective coarse-grained description in terms of a multidimensional stochastic dynamical system, driven by small-amplitude intrinsic noise. In applications to biological, ecological, chemical and social dynamics it is common for these models to posses quantities that are approximately conserved on short timescales, in which case system trajectories are observed to remain close to some lower-dimensional subspace. Here, we derive explicit and general formulae for a reduced-dimension description of such processes that is exact in the limit of small noise and well-separated slow and fast dynamics. The Michaelis-Menten law of enzyme-catalyzed reactions, and the link between the Lotka-Voltera and Wright-Fisher processes are explored as a simple worked examples. Extensions of the method are presented for infinite dimensional systems and processes coupled to non-Gaussian noise sources.'
author:
- 'Todd L. Parsons'
- Tim Rogers
bibliography:
- 'DRvTS.bib'
title: 'Dimension reduction for stochastic dynamical systems forced onto a manifold by large drift: a constructive approach with examples from theoretical biology'
---
Introduction
============
One important reason for the observed “unreasonable effectiveness” of mathematical modelling in describing natural phenomena [@Wigner1960] is the gigantic separation of scales apparent in physical systems. This convenient property allows a process of interest to be treated separately from what is happening on much slower or faster time scales, or on much larger or smaller spatial scales. The result is that mathematical models can be simple to state and, having few degrees of freedom, are more likely to be solvable. Sadly, other fields to which mathematics is applied — biology, ecology, finance, social science, to name a few — do not enjoy this separation, and consequently modellers in these areas have not yet come close to the same levels of predictive success achieved in physics. In writing a successful mathematical model, one faces two intertwined problems: the ability of a model to capture the essential behaviours of the system of interest, and our ability to ‘solve’ the model to extract quantitative predictions and qualitative understanding. Biological systems, for example, often comprise feedbacks over several scales, meaning that more degrees of freedom must be included in a model to achieve an acceptable representation of reality. This has negative consequences for the analysis, however, as higher-dimensional systems tend to be harder to work with, both analytically and in simulations.
These concerns are particularly relevant for the study of emergent phenomena arising from the interactions of large numbers of individual elements (typically particles, agents, or organisms). Important examples include the modelling of population dynamics, evolution, and epidemic spread. Significant progress has been made in this area through the application of so-called system size expansion techniques (e.g. [@vanKampen92; @Kurtz78]). These methods provide a coarse-grained description of an interacting system, typically in terms of a stochastic process driven by intrinsic noise (i.e. with amplitude depending on the state of the system, and scaling inversely with system size), which is precise in the limit of large system size. Although very much less complicated than the original system, these coarse-grained models are still likely to resist exact solution, being non-linear multi-dimensional stochastic dynamical systems. Previously, important results have been obtained though analysis of linear perturbations around fixed points, however, to more fully characterise the phenomena possible in these models requires a fully non-linear treatment.
Working in this direction, several groups have independently found and exploited a natural separation of scales emerging in certain models of interacting populations [@Parsons2007b; @Parsons2008; @Durrett2009; @Parsons2010; @Parsons2012; @Rogers2012a; @Rogers2012b; @Doering2012; @Constable2013; @Constable2014; @Pigolotti2014; @Constable2015; @Chotibut2015]. Loosely speaking, it is often the case that the total size of a population varies much more rapidly than its composition, as is evident in the disparate timescales of ecology and evolution. In dynamical systems terminology, it is observed that trajectories remain in the neighbourhood of a lower-dimensonal manifold; a subspace of the system state space in which the total size of the population is a function of its composition. Intrinsic noise drives small perturbations from this manifold, which are quickly suppressed by a large deterministic drift back (see, for example, the trajectories of Michaelis-Menten dynamics in Fig. \[MMfig\]). The works cited above pursue various related approximation strategies, allowing for a simplified, often solvable, effective model to be derived describing motion along the lower-dimensional manifold.
![**Thin Red:** Simulation of a single stochastic trajectory of an SDE of the type (\[sdex\]), with $\bm{f}$, $\bm{h}$ and $\bm{G}$ corresponding to the Michaelis-Menten model (\[MMfh\], \[MMG\]). **Thick Blue:** The slow manifold for this system, which the stochastic trajectory stays close to after the fast initial transient carrying it away from the initial condition $(1,0)$. **Dashed Black:** The flow field of the outer drift term $\bm{f}$, to which the fast motion is approximately parallel.[]{data-label="MMfig"}](mm_trajectory.pdf){width="80.00000%"}
This is a kind of timescale separation that cannot be put by hand into a model as there is a complex feedback between the fast and slow degrees of freedom, which must be carefully computed. Mathematically, this is not straightforward, and a host of different strategies have previously been employed to approximate the effects of this feedback between scales. A rigorous treatment of this situation can in fact be found in the theorems proved in [@Katzenberger91] and [@Funaki1993], however, the results of these works are difficult to apply in practice as they lack explicit formulations of certain key quantities. Our purpose here is to synthesise the various methods of the above mentioned authors with the rigorous theory of Katzenberger, and distil this into a single robust, systematic and provably correct procedure for timescale separation in stochastic dynamical systems with intrinsic noise, which we believe will be of considerable general use. The main results are contained in the short Section \[S2\], where we describe a map from a high-dimensional system of equations (\[sdex\]) to a lower-dimensional one (\[sdez\]), via explicit formulae that are summarised in Table \[tab1\].
**Case** **Procedure**
-------------------------------------------------------- ---------------------------------------------------------------
Outer system ($\varepsilon=\mu=0$) is solvable $\quad$ Use equations (\[ivp\], \[defpi\]) and (\[defPQ\], \[drift\])
Manifold is one-dimensional Use equations (\[drift\]) and (\[P1D\], \[Q1D\])
Manifold has co-dimensional one Use equations (\[Pcd1\]) and (\[Qcd1\])
Manifold is $m$-dimensional Use equations (\[drift\]) and (\[diagJ\] – \[QmD\])
: Quick reference table of equations applying to different cases of slow-manifold reduction.[]{data-label="tab1"}
The general setting for our calculations will be models expressed as coupled stochastic differential equations, with some small parameter $\varepsilon$ that controls the separation of timescales. As stated above, we are particularly interested in systems that are derived from a complex underlying interacting process, so that the noise terms are intrinsic, representing the cumulative random effects of many interactions. Importantly in this case the noise, although non-negligible in its effect, is typically small in amplitude[^1]. Three example application areas are: (i) chemical and biological reaction networks, *e.g.* gene regulation in a cell t, where small copy numbers imply noise but homoeostasis suggests timescale separation [@Ball2006], (ii) evolutionary models where noise arising from demographic fluctuations can alter the course of selection (references above), (iii) dynamical networks which are naturally extremely high-dimensional systems in need of low-dimensional proxies [@Rogers2013]. It is worth pointing out that a plethora of different timescale separation techniques exist in the literature, and the most useful choice of method depends on the system in question. In Appendix \[lit\] we give a very brief history of notable developments in stochastic timescale separation.
In addition to the general theory outlined in Section \[S2\], we present in Section \[WE\] some worked examples and variations to the method. As a prototypical example of the standard method, we work through the derivation of a stochastic form of the Michaelis-Menten law for enzyme-catalysed reactions. The second example demonstrates the extension of the methods to an infinite dimensional setting. Katzenburger’s theorem is discussed in detail in Appendix \[Katz\], as well as the generalisation of the noise sources (for example to jump processes) and alternative characterisations of the diffusion process on the manifold.
Reduced model description {#S2}
=========================
We consider Langevin stochastic differential equations of the general type $$\frac{d\bm{x}}{dt}=\bm{f}(\bm{x})+\varepsilon\bm{h}(\bm{x})+\sqrt{\mu}\,\bm{G}(\bm{x})\,\bm{\eta}(t)\,,
\label{sdex}$$ where the state variable $\bm{x}=(x_1,\ldots,x_d)^T$ is an $d$-dimensional vector, and there are $s$ independent Itô white noise sources $\bm{\eta}(t)=(\eta_1(t)\,,\ldots,\eta_s(t))^T$. The vector-valued functions $\bm{f}:\mathbb{R}^d\to\mathbb{R}^d$ and $\bm{h}:\mathbb{R}^d\to\mathbb{R}^d$ are the ‘outer’ and ‘inner’ parts of the drift respectively, and the matrix valued function $\bm{G}:\mathbb{R}^d\to\mathbb{R}^{d,s}$ specifies the coupling of state variables to noise sources. We assume throughout that $\bm{f}$ is twice differentiable, but place no constraints on the other functions. The parameters $\varepsilon$ and $\mu$ determine the separation of timescales and the strength of the noise, respectively.
We do not assume an *a priori* separation into slow and fast variables, as is common in the literature, as in the applications that motivate us, an appropriate change of variables is frequently neither evident nor analytically tractable (although see [@Parsons2007b]), and our method does not require that they be known.
We are interested in the case when $\varepsilon$ and $\mu$ are small and $\bm{f}$ possesses a $m$-dimensional manifold of equilibria $\Gamma\subset\mathbb{R}^d$ such that $\bm{f}(\bm{\tilde{x}})=\bm{0}$ for all $\bm{\tilde{x}}\in\Gamma$. We assume that this manifold is unique, connected, and globally attractive (*i.e.* it is a globally unique, normally hyperbolic slow manifold see *e.g* [@Berglund2006]); then we expect solutions of (\[sdex\]) to rapidly approach and remain very close to $\Gamma$. In fact, it has been rigourously proved by Katzenberger [@Katzenberger91] that the trajectories of $\bm{x}\in\mathbb{R}^d$ converge those of a stochastic variable $\bm{\tilde{x}}\in\Gamma$ with dynamics $$\frac{d\bm{\tilde{x}}}{dt}=\varepsilon\bm{P}(\bm{\tilde{x}})\bm{h}(\bm{\tilde{x}})+\mu\bm{g}(\bm{\tilde{x}})+ \sqrt{\mu}\bm{P}(\bm{\tilde{x}})\bm{G}(\bm{\tilde{x}})\bm{\eta}(t)\,,
\label{sdez}$$ where $\bm{P}$ is a certain projection matrix derived from $\bm{f}$, and $\bm{g}$ is a new contribution to the drift arising from the way in which fluctuations away from the manifold are suppressed; as our examples illustrate, unlike the deterministic situation, it is not sufficient to simply restrict to $\Gamma$ to obtain the slow dynamics. Our purpose here is to derive explicit expressions for $\bm{P}$ and $\bm{g}$. Readers with a specific problem in mind may wish to jump straight to the appropriate result, which can be found by referring to Table \[tab1\].
Before we proceed with our main task, we give a brief sketch of the derivation of (\[sdez\]). Examining (\[sdex\]) when $\varepsilon$ and $\mu$ are small, one might imagine a picture in which the state of the system is quickly carried onto the manifold by the fast outer drift term $\bm{f}$. Following this fast initial transient, it may then receive multiple stochastic ‘kicks’ carrying it away from the manifold, each time only to return again via the paths described by $\bm{f}$. See Figure \[MMfig\] for an illustrative example. This intuition can be made concrete by considering the flow map of the outer system. Let $\bm{x}$ be a point in the state space and consider the deterministic initial value problem $$\begin{cases}\displaystyle\frac{d\bm{\xi}_{\bm{x}}}{dt}=\bm{f}(\bm{\xi}_{\bm{x}})\\ \bm{\xi}_{\bm{x}}(0)=\bm{x}\,.\end{cases}
\label{ivp}$$ Since the center manifold is globally attractive, all trajectories lead eventually to $\Gamma$ and we may thus define a function $\bm{\pi}:\mathbb{R}^d\to\Gamma$ giving the endpoint of the deterministic trajectories $$\bm{\pi}(\bm{x})=\lim_{t\to\infty}\bm{\xi}_{\bm{x}}(t)\,,
\label{defpi}$$ where $\bm{\xi}_{\bm{x}}$ is the solution of (\[ivp\]).
If we take the point $\bm{x}$ to be the current location of the random variable governed by equation (\[sdex\]), then $\bm{\pi}(\bm{x})$ defines another random variable that tracks the motion of $\bm{x}$ but is constrained to the manifold. Application of Itô’s formula [@Ito1974] gives the Langevin equations for each spatial coordinate: $$\begin{split}
\frac{d}{dt}\pi_i(\bm{x})&=\sum_j \frac{\partial\pi_i}{\partial x_j} \frac{dx_j}{dt} + \frac{\mu}{2}\sum_{s,j,k} G_{js}(\bm{x})G_{ks}(\bm{x})\frac{\partial^2\pi_i}{\partial x_j\partial x_k}\\
&=\sum_j \frac{\partial\pi_i}{\partial x_j}\Big(f_j(\bm{x})+\varepsilon h_j(\bm{x})\Big) +\frac{\mu}{2}\sum_{s,j,k} G_{js}(\bm{x})G_{ks}(\bm{x})\frac{\partial^2\pi_i}{\partial x_j\partial x_k} +\sqrt{\mu} \sum_{s,j}G_{js}(\bm{x})\frac{\partial\pi_i}{\partial x_j}\eta_s(t)\\&
=\varepsilon\sum_j \frac{\partial\pi_i}{\partial x_j} h_j(\bm{x}) +\frac{\mu}{2}\sum_{s,j,k} G_{js}(\bm{x})G_{ks}(\bm{x})\frac{\partial^2\pi_i}{\partial x_j\partial x_k} +\sqrt{\mu} \sum_{s,j}G_{js}(\bm{x})\frac{\partial\pi_i}{\partial x_j}\eta_s(t)
\,.
\label{sdezx}
\end{split}$$ where the last equality comes from the observation that $\bm{\pi}(\bm{\xi}_{\bm{x}}(t)) = \bm{\pi}(\bm{x})$ for all $t$, and thus $$0 = \frac{d}{dt}\Big\vert_{t = 0} \pi_{i}(\bm{\xi}_{\bm{x}}(t))
= \sum_j \frac{\partial\pi_i}{\partial x_j} \frac{d\xi_{j}}{dt}\Big\vert_{t = 0}
= \sum_j \frac{\partial\pi_i}{\partial x_j} f_j(\bm{x})\,.$$ Unfortunately equation is not closed since it relies on full knowledge of the random variable $\bm{x}$. However, if we believe that $\bm{x}$ remains very close to $\Gamma$ (as is the case when $\varepsilon$ and $\mu$ are small) then we might be motivated to consider a new random variable $\bm{\tilde{x}}\in\Gamma$ which we assume is a close approximation to both $\bm{x}$ and $\bm{\pi}(\bm{x})$. Substituting $\bm{\tilde{x}}$ for both these quantities in , we obtain the closed expression $$\frac{dz_i}{dt}=\varepsilon \sum_j P_{ij}(\bm{\tilde{x}}) h_j(\bm{\tilde{x}}) +\frac{\mu}{2} \sum_{s,j,k}G_{js}(\bm{\tilde{x}})G_{ks}(\bm{\tilde{x}})Q_{ijk}(\bm{\tilde{x}}) +\sqrt{\mu} \sum_{s,j}P_{ij}(\bm{\tilde{x}})G_{js}(\bm{\tilde{x}})\eta_s(t)\,,
\label{sdez2}$$ where $\bm{P}$ is a matrix and $\bm{Q}$ an array defined by $$P_{ij}(\bm{\tilde{x}})=\left.\frac{\partial}{\partial x_j}\pi_i(\bm{x})\right|_{\bm{x}=\bm{\tilde{x}}}\,,\quad Q_{ijk}(\bm{\tilde{x}})=\left.\frac{\partial^2}{\partial x_j\partial x_k}\pi_i(\bm{x})\right|_{\bm{x}=\bm{\tilde{x}}}\,.
\label{defPQ}$$ Equivalently we may rewrite (\[sdez2\]) as equation (\[sdez\]), where the additional drift term is $$\label{drift}
\bm{g}(\bm{\tilde{x}})=\frac{1}{2}\sum_{s,j,k} G_{js}(\bm{\tilde{x}})G_{ks}(\bm{\tilde{x}})Q_{ijk}(\bm{\tilde{x}}).$$
![**Left:** Here, the variation in the angle between the fast (dashed) and slow (solid) subspaces creates a bias in the location of the return to the manifold of a perturbation away from it; an upward perturbation returns quite close on the left of the origin, but an equally likely downward perturbation is carried far to the right. **Centre:** The same effect can occur as a result of curvature of the manifold. In this figure the flow fields are parallel, but the manifold curves, resulting in the same rightwards bias in the projected system. **Right:** Curvature of the flow field may also induce bias, even when the angle of intersection is constant.[]{data-label="ffd_fig"}](Q_decomp.pdf){width="\textwidth"}
The projection matrix $\bm{P}(\bm{\tilde{x}})$ is entirely determined by the first order terms of this expansion, and typically it can be straightforwardly reconstructed from knowledge of the eigenvectors of the Jacobian matrix of $\bm{f}$. The calculation of the noise-induced drift term is more complicated, having contributions from three possible sources: variation of the alignment of the flow field, curvature of the manifold, and curvature of the flow field. Each of these mechanisms can induce a bias in the direction of flow of the reduced dimension system, as illustrated in Figure \[ffd\_fig\].
In the following subsections we will present explicit procedures for computing $\bm{P}$ and $\bm{Q}$.
One-dimensional manifolds
-------------------------
The simplest case to treat is that of a one-dimensional manifold, as the second-order perturbation expansion is explicitly solvable. Suppose that the slow manifold $\Gamma$ is a curve parameterised by the first spatial co-ordinate of the system[^2]. That is, there exists function $\bm{\gamma}$ such that $$\bm{x}\in\Gamma\quad\Leftrightarrow\quad \bm{x}=\bm{\gamma}(x_1)\,.$$ In this case the dynamics of the reduced system $\bm{\tilde{x}}$ defined in (\[sdez\]) are determined entirely by the first component, so we need only to compute the partial derivatives of $\pi_1$. For ease of notation we will drop the subscript $1$ from now on, writing $\tilde x:=\tilde x_1$ as well as $P_{j}:=P_{1j}(\bm{\gamma}(\tilde x_1))$ and $Q_{jk}:=Q_{1jk}(\bm{\gamma}(\tilde x_1))$.
Consider a point $\bm{x}=\bm{\gamma}(\tilde x)$ on the manifold. To obtain expressions for $P_{j}$ and $Q_{jk}$ we undertake a second-order perturbation theory. Since $\bm{x}\in\Gamma$ is a point on the manifold we have by definition that it is unchanged by the action of the outer flow field, so $\bm{\pi}(\bm{x})=\bm{\gamma}(\tilde x)$. We make a small perturbation $\bm{x}\mapsto \bm{x}+\bm{\Delta x}$, and ask what perturbation $\tilde x\mapsto \tilde x+\Delta \tilde x$ is required so that $\bm{\pi}(\bm{x}+\bm{\Delta x})=\bm{\gamma}(\tilde x+\Delta \tilde x)$. See Figure \[1Dfig\] for an illustration.
![Illustration of the perturbation calculation for a 1D manifold. []{data-label="1Dfig"}](1D2.pdf){width="80.00000%"}
Becuase we set the problem up so that $\pi_1(\bm{x})=\tilde x$, Taylor expanson gives $$\Delta \tilde x=\sum_j P_{j} \Delta x_j+\frac{1}{2}\sum_{j,k}Q_{jk}\Delta x_j\Delta x_k+\ldots$$ Near the point $\bm{x}\in \Gamma$ we can approximate the action of $\bm{\pi}$ by constructing the quadratic expansion of the preimage. Specifically, it can be shown that in the neighbourhood of $\bm{x}$, the collection of nearby points that would be mapped to $\bm{x}$ by $\bm{\pi}$ (i.e. the set $\{\bm{y}\,:\,\bm{\pi}(\bm{y})=\bm{x}\}$) is approximated to second order by the set of points $\bm{y}$ such that $$\bm{v}(\tilde x)^T(\bm{y}-\bm{\gamma}(\tilde x))+(\bm{y}-\bm{\gamma}(\tilde x))^T\Theta(\tilde x)(\bm{y}-\bm{\gamma}(\tilde x))=0\,,
\label{quadratic}$$ where $\bm{v}(\tilde{x})$ is a perpendicular vector to the flow field near $\bm{x}=\bm{\gamma}(\tilde x)$ and $\Theta(\tilde{x})$ a matrix describing the curvature of the flow field near the same point. In Appendix \[Theta\] we give an explicit derivation of these quantities from $\bm{f}$; for now we assume they are known. Recall that we are seeking the perturbation $\Delta\tilde{x}$ such that $\bm{\pi}(\bm{x}+\bm{\Delta x})=\bm{\gamma}(\tilde x+\Delta \tilde x)$, to second order. We make the following Taylor expansions of various orders: $$\begin{split}
\Big[\bm{x}+\Delta\bm{x}-\bm{\gamma}(\tilde x+\Delta \tilde x)\Big]_\ell &= \Delta x_\ell- \gamma_\ell'\Delta\tilde x-\frac{1}{2}\gamma_\ell''(\Delta \tilde x)^2 +\ldots\\
&= \sum_k (\delta_{k,\ell}- \gamma_\ell'P_{k}) \Delta x_k -\frac{1}{2}\sum_{j,k} \left(\gamma_\ell' Q_{jk} + \gamma_\ell'' P_{j}P_k\right)\Delta x_j\Delta x_k+\mathcal{O}(\bm{\Delta x}^3)\\
\Big[\bm{v}(\tilde x+\Delta \tilde x)\Big]_\ell &=v_\ell+v'_\ell\Delta\tilde x+\ldots =v_\ell+v'_\ell\sum_j P_{j} \Delta x_j+\mathcal{O}(\bm{\Delta x}^2)\\
\Big[\Theta(\tilde x+\Delta \tilde x)\Big]_{jk} &= \Theta_{jk}+\mathcal{O}(\bm{\Delta x})\,.
\end{split}$$ Here we drop the argument $\tilde x$ from $\gamma_\ell'$, $\gamma_\ell''$, $v_\ell$, $v_\ell'$ and $\Theta_{jk}$ to avoid clutter. Following (\[quadratic\]), the requirement that $\bm{\pi}(\bm{x}+\bm{\Delta x})=\bm{\gamma}(\tilde x+\Delta \tilde x)$ to second order becomes $$\begin{split}
0&=\bm{v}(\tilde x+\Delta \tilde x)^T\Big(\bm{x}+\bm{\Delta x}-\bm{\gamma}(\tilde x+\Delta \tilde x)\Big)\\
&\quad+\Big(\bm{x}+\bm{\Delta x}-\bm{\gamma}(\tilde x+\Delta \tilde x)\Big)^T\Theta(\tilde x+\Delta \tilde x)\Big(\bm{x}+\bm{\Delta x}-\bm{\gamma}(\tilde x+\Delta \tilde x)\Big)+\mathcal{O}(\bm{\Delta x}^3)\\
&=\sum_\ell \Big[v_\ell+v'_\ell\sum_j P_{j} \Delta x_j\Big]\Big[\sum_k (\delta_{k,\ell}- \gamma_\ell'P_{k}) \Delta x_k -\frac{1}{2}\sum_{j,k} \left(\gamma_\ell' Q_{jk} + \gamma_\ell'' P_{j}P_k\right)\Delta x_j\Delta x_k\Big]\\
&\quad+\sum_{j,k}\Theta_{jk}\Delta x_j\Delta x_k+\mathcal{O}(\bm{\Delta x}^3)\\
&=\sum_k\Big\{\sum_\ell v_\ell(\delta_{k,\ell}- \gamma_\ell'P_{k}) \Big\}\Delta x_k\\
&\quad +\frac{1}{2}\sum_{j,k} \Big\{ \sum_\ell v_\ell' (\delta_{k,\ell}+\delta_{j,\ell}- 2\gamma_\ell'P_{k})P_{j}-\sum_\ell v_\ell\left(\gamma_\ell' Q_{jk} + \gamma_\ell'' P_{j}P_k\right)+2\Theta_{jk}\Big\}\Delta x_j\Delta x_k \\&\qquad+\mathcal{O}(\bm{\Delta x}^3)\,.
\end{split}$$ Since the perturbation $\bm{\Delta x}$ was abritrary, we require each term in curly brackets above to be equal to zero. From the first order terms we conclude that $$\label{P1D}
P_k=\frac{v_k}{\sum_\ell v_\ell\gamma_\ell'}\,,$$ and from the second order that $$\label{Q1D}
Q_{jk}=\frac{1}{\sum_\ell v_\ell\gamma_\ell'}\left( v_k'P_j+v_j'P_k+2\Theta_{jk} - \sum_\ell (2v_\ell' \gamma_\ell'+v_\ell\gamma_\ell'' ) P_{j}P_k\right).$$ Written this way, the separate contributions from variation of the flow field (terms involving $\bm{v'}$), curvature of flow field ($\Theta$), and curvature of the manifold (the $\bm{\gamma''}$ term) are clearly visible.
In higher dimensions, the above perturbation expansion is less useful, as it produces a larger system of equations which lacks an explicit solution. A different line of attack is necessary.
General case
------------
If the linearisation of the flow field $\bm{\phi}_t$ is known in the neighbourhood of the manifold then $\bm{P}$ can be reconstructed easily. Specifically, around a point $\bm{z}\in\Gamma$ the state space $\mathbb{R}^d$ can be decomposed into a product of ‘slow’ and ‘fast’ subspaces of dimension $m$ and $d-m$, respectively. The slow subspace is the tangent plane to the manifold at the given point; a perturbation in one of these directions is unaffected by the action of $\bm{f}$. Conversely, the fast subspace comprises perturbation directions that collapse quickly back to the manifold. The projection matrix $\bm{P}(\bm{z})$ acts as the identity on the slow subspace and as zero on the fast subspace.
Unfortunately, no such simple formulation is available for $\bm{Q}(\bm{z})$ in general. This problem was explored in [@Parsons2012], where the following method was developed. This result is explained fully in Appendix \[Q\], for now we simply present the computational steps.
1. Compute the Jacobian $\bm{J}$ and diagonalize it, writing $$\bm{J}=\bm{W}\bm{\Lambda}\bm{W}^{-1}\,.
\label{diagJ}$$ where $\bm{W}=(\bm{w}_1\cdots\bm{w}_n)$ is a matrix of eigenvectors forming a basis of $\mathbb{R}^d$, with the $m$ slow directions written first. $\bm{\Lambda}$ is a diagonal matrix of eigenvalues with $\lambda_1=\cdots=\lambda_m=0$ and $\textrm{Re}(\lambda_{m+1}),\ldots,\textrm{Re}(\lambda_n)<0$. Also compute the pseudo-inverse $$\bm{J}^+=\bm{W}\bm{\Lambda}^+\bm{W}^{-1}\,,$$ where $\bm{\Lambda}^+$ is the diagonal matrix with eigenvalues $\lambda^+_1=\cdots=\lambda^+_n$, where $$\lambda^+=\begin{cases}0\quad&\text{if}\quad \lambda=0\\1/\lambda&\text{if}\quad \lambda\neq 0\,.\end{cases}$$
2. For each $i$, compute the Hessian $\bm{H}_i$ defined by $$H_{ijk}=\frac{\partial f_i(\bm{x})}{\partial x_j\partial x_k}\Bigg|_{\bm{x}=\bm{z}}\,.$$ Then find the (matrix-valued) solution $\bm{X}_i$ of the Lyapunov equation $$\label{lyapunov}
\bm{J}^T\bm{X}_i+\bm{X}_i\bm{J}=-\bm{H}_{i}\,.$$ NB: this is a linear problem that is straightforwardly solved [@Bartels1972].
3. Finally, the projection matrix is given by $$\bm{P}=\bm{I}-\bm{J}^+\bm{J}\,.
\label{PmD}$$ and for $\bm{Q}$ we have $$Q_{ijk}=\sum_l -J^+_{il}[\bm{P^T}\bm{H}_l\bm{P}]_{jk}+P_{il}[\bm{X}_l-\bm{J}^{+T}\bm{H}_l\bm{P}-\bm{P}^T\bm{H}_l\bm{J}^+]_{jk}\,.
\label{QmD}$$
Co-dimension one manifolds
--------------------------
We now use the results of the previous section to obtain explicit expressions for the derivatives in the case when $\Gamma$ is a $(d-1)$-dimensional manifold. In this case, in a small neighbourhood around any point $\bm{z}\in\Gamma$, the flow field can be decomposed as $\bm{f} = \phi\,\bm{r}$, into a scalar part $\phi: \mathbb{R}^{d} \to \mathbb{R}$ that vanishes on $\Gamma$, and a non-vanishing vector part $\bm{r} : \mathbb{R}^{d} \to \mathbb{R}^{d}$. Using this decomposition we compute an expression for the Jacobian around a point: $$\bm{J}(\bm{x}) = \phi(\bm{x}) \frac{\partial \bm{r}}{\partial \bm{x}}
+ \bm{r}(\bm{x})\nabla \phi(\bm{x})^{T} .$$ In particular, evaluated exactly at the point $\bm{z}$ on the manifold we have $\bm{J} = \bm{r} \nabla\phi^{T}$. Meaning that $\bm{r}$ is, up to scalar multiple, the unique eigenvector corresponding to $$\lambda = \nabla\phi^{T} \bm{r}\,,$$ which is the sole non-zero eigenvalue of $\bm{J}$. Note that here and hereafter we drop the arugment $\bm{z}$ to avoid notational clutter. As the Jacobian is given by the outer product of vectors $\bm{r}$ and $\nabla\phi$, it is straightforward to check that the pseudo-inverse may be written as $$\bm{J}^+=\frac{1}{\lambda^2}\bm{J}^T\,.$$ We conclude from that $$\label{Pcd1}
\bm{P}=\bm{I}-\bm{J}^+\bm{J}=\bm{I}-\frac{\bm{J}^T\bm{J}}{\lambda^2}\,.$$ To determine $\bm{Q}$, it thus remains to solve , $$\bm{J}^T\bm{X}_i+\bm{X}_i\bm{J}=-\bm{H}_{i}\,,$$ for $\bm{H}$ and insert into . As developed in Appexdix \[Q\], equation , the solution to this Lyapunov equation can be expressed by the exponential integral $$\bm{X}_i
= \int_{0}^{\infty} \Big(e^{s\bm{J}}-\bm{P}\Big)^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}} \Big(e^{s\bm{J}}-\bm{P}\Big)\,ds.$$ For an arbitrary vector $\bm{Y}$, we have $\bm{J} \bm{Y} = \bm{r} \nabla\phi^{T}\bm{Y}$, so that $$\bm{J}^{2} \bm{Y}
= \bm{r} \nabla\phi^{T}\bm{r} \nabla\phi^{T}\bm{Y}
= \lambda (\nabla\phi^{T}\bm{Y}) \bm{r},$$ $\bm{J}^{n} \bm{Y} = \lambda^{n-1} (\nabla\phi^{T}\bm{Y}) \bm{r}$, and $$\begin{aligned}
e^{s\bm{J}} \bm{Y} &= \sum_{n=0}^{\infty} \frac{s^{n}}{n!} \bm{J}^{n} \bm{Y}\\
&= \bm{Y} + (\nabla\phi^{T}\bm{Y}) \bm{r} \lambda^{n-1}
\sum_{n=1}^{\infty} \frac{s^{n}}{n!}\\
&= \frac{(\nabla\phi^{T}\bm{Y})}{\lambda}(e^{\lambda t}-1)
\bm{r}.
\end{aligned}$$ Thus, $e^{s\bm{J}}-\bm{P} \bm{Y} = \frac{(\nabla\phi^{T}\bm{Y})}{\lambda} e^{\lambda t} \bm{r}$ and, recalling that $\lambda < 0$, we have that $$\begin{aligned}
X_{ijk}
&= \bm{e}_{j}^{T} \left(\int_{0}^{\infty} \bm{e}_{i}^{T} (e^{s\bm{J}}-\bm{P})^{T}
\frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}} (e^{s\bm{J}}-\bm{P})\,ds
\right)\bm{e}_{k}\\
&= \frac{(\nabla\phi^{T}\bm{e}_{j})(\nabla\phi^{T}\bm{e}_{k})}
{\lambda^{2}}
\bm{r}^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}} \bm{r}
\int_{0}^{\infty} e^{2 \lambda t}\, dt\\
&= -\frac{\frac{\partial \phi}{\partial x_{j}}\frac{\partial \phi}{\partial x_{k}}}
{2 \lambda^{3}}
\bm{r}^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}} \bm{r}.
\end{aligned}$$ Finally, observing that $\frac{\partial^{2} f_{i}}{\partial x_{j}\partial x_{k}} =
\frac{\partial r_{i}}{\partial x_{j}}\frac{\partial \phi}{\partial x_{k}} + r_{i} \frac{\partial^{2} \phi}{\partial x_{j}\partial x_{k}}$, substituting the above into and considerable algebraic simplification yields $$\begin{gathered}
\label{Qcd1}
Q_{ijk} = \frac{1}{\lambda} \left(
[\bm{P}^{T}\frac{\partial^{2} \phi}{\partial \bm{x}^{2}}\bm{P}]_{jk}
r_{i}
- \frac{\partial \phi}{\partial x_{j}}
[\bm{P} \frac{\partial \bm{r}}{\partial \bm{x}}]_{ik}
- \frac{\partial \phi}{\partial x_{k}}
[\bm{P} \frac{\partial \bm{r}}{\partial \bm{x}}]_{ij}
\right)+ \frac{1}{\lambda^{2}} \frac{\partial \phi}{\partial x_{j}}
\frac{\partial \phi}{\partial x_{k}}
[\bm{P} \frac{\partial \bm{r}}{\partial \bm{x}} \bm{r}]_{i}.\end{gathered}$$
Worked examples {#WE}
===============
Simple example: Michaelis-Menten kinetics
-----------------------------------------
The Michaelis-Menten law is perhaps one of the most widely-applied examples of timescale separation. It is a model for the net rate of production in a chemical reaction that is catalysed by an enzyme, in which it is assumed that the process of enzyme binding and unbinding occurs very much faster than the catalytic reaction of interest. Using the notation of chemical reactions, one may write $$E+S\xrightleftharpoons[k_r]{k_f} C \xrightarrow{k_{\text{cat}}} E+P\,,
\label{MM}$$ where $E$ symbolises the enzyme, $S$ the substrate, $C$ the enzyme-substrate complex, and $P$ the product. The parameters $k_f$ and $k_r$ give the rate of binding (forward) and unbinding (reverse) of the enzyme to the substrate, while $k_{\text{cat}}$ specifies the rate of catalysis.
Assuming the reaction takes place in a domain of infinite volume, one may write rate the deterministic equations $$\begin{split}
&\frac{dS}{dt}=-k_fES+k_rC\,,\\
&\frac{dE}{dt}=-k_fES+(k_r+k_{\text{cat}})C\,,\\
&\frac{dC}{dt}=k_fES-(k_r+k_{\text{cat}})C\,,\\
&\frac{dP}{dt}=k_{\text{cat}}C\,,\\
\end{split}$$ where $S$, $C$, $P$ and $E$ now represent the *concentrations* of the various reactants. Note that this system has only two degrees of freedom due to conservation relations $E+C=E_0$ and $S+C+P=S_0$, where $E_0$ and $S_0$ are the initial concentrations of the enzyme and substrate, respectively. If $k_f,k_r\gg k_{\text{cat}}$ we might approximate the concentration of the complex $C$ by the equilibrium value it would have if $k_{\text{cat}}$ were actually zero: $$k_fES-k_rC\approx 0\quad \Rightarrow \quad C\approx E_0\frac{S}{k+S}\,,$$ where $k=k_r/k_f$. Introducing $v^\ast=k_{\text{cat}}E_0$, on the slower timescale the net production rate is then found to be $$\frac{dP}{dt}=\frac{v^\ast\, S}{k+S}\,.$$ This is the Michaelis-Menten law.
In finite volume domains chemical reactions are subject to random fluctuations arising from the discrete nature of the molecules involved. A more appropriate description in these circumstances is a stochastic differential equation, with noise terms that are derived from the instantaneous reaction rates (each possible reaction introduces its own source of noise). For the reaction described above in (\[MM\]) occurring in a domain of volume $V$, equations are derived following Kurtz [@Kurtz78][^3]: $$\begin{split}
&\frac{dS}{dt}=-k_f(E_0-C)S+k_rC-\sqrt{\frac{k_f(E_0-C)S}{V}}\,\eta_{f}(t)+\sqrt{\frac{k_rC}{V}}\,\eta_{r}(t)\,,\\
&\frac{dC}{dt}=k_f(E_0-C)S-(k_r+k_{\text{cat}})C+\sqrt{\frac{k_f(E_0-C)S}{V}}\,\eta_{f}(t)-\sqrt{\frac{k_rC}{V}}\,\eta_{r}(t)-\sqrt{\frac{k_{\text{cat}}C}{V}}\,\eta_{\text{cat}}(t)\,.\\
%&\frac{dP}{dt}=k_{\text{cat}}C+\sqrt{\frac{k_{\text{cat}}C}{V}}\,\eta_{\text{cat}}(t)\\
\end{split}$$ Following similar lines to [@Heineken1967] a dimensionless form may be found by rescaling time $t\mapsto k_fE_0t$ and introducing variables $$\bm{x}=\left(\begin{array}{c}S/S_0\\C/E_0\end{array}\right)\,,$$ and parameters $$\begin{split}
&\varepsilon=\frac{k_{\text{cat}}}{k_fE_0}>0\,,\quad\mu=\frac{1}{S_0V}\,,\quad\alpha=\frac{k_r}{k_fS_0}>0 \,,\quad\beta=\frac{S_0}{E_0}>0\,.\\
\end{split}$$ The result is a system of exactly the form of equation (\[sdex\]): $$\frac{d\bm{x}}{dt}=\bm{f}(\bm{x})+\varepsilon\bm{h}(\bm{x})+\sqrt{\mu}\,\bm{G}(\bm{x})\,\bm{\eta}(t)\,,$$ where $$\begin{split}
&\bm{f}(\bm{x})=\left(\begin{array}{c}-x_1+(x_1 +\alpha)x_2\\\beta(x_1-(x_1 +\alpha)x_2) \end{array}\right)\,,\qquad\bm{h}(\bm{x})=\left(\begin{array}{c}0\\-x_2 \end{array}\right)\,,\\
\end{split}
\label{MMfh}$$ and $$\begin{split}
&\bm{G}(\bm{x})=\left(\begin{array}{ccc}-\sqrt{(1-x_2)x_1}&\sqrt{\alpha x_2}&0\\\beta\sqrt{(1-x_2)x_1}&-\beta\sqrt{\alpha x_2}&-\sqrt{\varepsilon \beta x_2} \end{array}\right)\,,\qquad\bm{\eta}(t)=\left(\begin{array}{c}\eta_{f}(t)\\\eta_r(t)\\\eta_{\text{cat}}(t)\end{array}\right)\,.\\
\end{split}
\label{MMG}$$ The slow manifold in this case is the curve $x_1-x_2(x_1+\alpha)=0$, along which $\bm{f}(\bm{x})=\bm{0}$. See Figure \[MMfig\] for an illustration.
Let us take $z=x_1$ as the slow variable and proceed to calculate a reduced system in terms of $z$ only. As the manifold is one-dimensional, we are able to simply follow the procedure laid out above. We begin by writing down the formula for the slow manifold and its $z$ derivatives: $$\bm{\gamma}(z)=\left(\begin{array}{c}z\\\frac{z}{z+\alpha}\end{array}\right)\qquad \bm{\gamma}'(z)=\left(\begin{array}{c}1\\\frac{\alpha}{(z+\alpha)^2}\end{array}\right)\qquad\bm{\gamma}''(z)=\left(\begin{array}{c}0\\\frac{-2\alpha}{(z+\alpha)^3}\end{array}\right)\,.$$ Next, we find the Jacobian matrix on the manifold $$\bm{J}(\bm{x})=\left(\begin{array}{cc}x_2-1&x_1+\alpha\\\beta(1-x_2)&-\beta(x_1+\alpha) \end{array}\right)\quad\Rightarrow\quad \bm{J}(z)=\left(\begin{array}{cc}\frac{z}{z+\alpha}-1&z+\alpha\\\beta(1-\frac{z}{z+\alpha})&-\beta(z+\alpha) \end{array}\right)\,.$$ Diagonalizing $\bm{J}(z)$ we find the left eigenvector $\bm{v}(z)$ corresponding to the eigenvalue zero, and its $z$ derivative: $$\bm{v}(z)=\left(\begin{array}{c}\beta\\1\end{array}\right)\qquad \bm{v}'(z)=\left(\begin{array}{c}0\\0\end{array}\right)\,.\\$$ Following equation (\[P1D\]) we obtain $$\bm{P}_{1\ast}(z)=\frac{(z+\alpha)^2}{\alpha+\beta(z+\alpha)^2}\left(\begin{array}{cc}\beta & 1\end{array}\right)\,,$$ and from equation (\[Q1D\]) $$\bm{Q}_{1\ast\ast}(z)=2\alpha\left(\frac{(z+\alpha)}{\alpha+\beta(z+\alpha)^2}\right)^3\left(\begin{array}{cc}\beta^2 & \beta\\ \beta&1\end{array}\right)\,.$$ Plugging these results into the general formula (\[sdez2\]) gives the reduced model $$\frac{dz}{dt}=-\varepsilon\frac{z(z+\alpha)}{\alpha+\beta(z+\alpha)^2}+\varepsilon\mu\frac{\alpha\beta z(z+\alpha)^2 }{(\alpha+\beta(z+\alpha)^2)^3}-\frac{(z+\alpha)^2}{\alpha+\beta(z+\alpha)^2}\sqrt{\varepsilon\mu\frac{\beta z}{z+\alpha}}\,\eta_{\text{cat}}(t)\,.
\label{MMz}$$
![Trajectories of $x_1$, $z$ and $\pi_1(\bm{x})$ from a single stochastic simulation of the Michaelis-Menten model. Note that the reduced dimension model for $z$ given by equation (\[MMz\]) captures the dynamics of the full system under the projection $\bm{\pi}$ (hence the extremely close agreement between the solid and dashed black lines above). The original coordinate $x_1$ is subject to additional noise in the kernel of the projection. []{data-label="MMfig2"}](xzpi.pdf){width="60.00000%"}
Notice that there is a positive noise-induced drift term, meaning that the rate of decrease of $z$ is slowed by the noise. Figure \[MMfig2\] shows the dynamics of $z$ compared with those of $x_1$ in the full system for a single realization of the noise.
At first sight equation (\[MMz\]) is considerably more complex than the traditional Michaelis-Menten law, however, carefully transforming back to the original coordinates[^4], one finds the simple result $$\frac{dP}{dt}=\frac{v^\ast S}{k+S}+\sqrt{\frac{v^\ast S}{V(k+S)}}\,\eta_{\text{cat}}(t)\,.$$
Co-dimension one: the Wright-Fisher diffusion as a limit of a near-neutral stochastic Lotka-Volterra process
------------------------------------------------------------------------------------------------------------
Consider a well mixed-population of $d$ interacting species in an environment of carrying capacity $K$: there are $K$ “slots” in the environment that at most one individual may occupy. Let $X_{i}$ denote the number of individuals of species $i$, and suppose that each individual of species $i$ gives birth at rate $b_{i}$ and dies at rate $d_{i}$. Further, suppose that the offspring is only viable if it lands in an empty patch, or if it lands in an occupied patch and out-competes the resident; say that an individual of type $i$ successfully displaces a resident of type $j$ with probability $c_{ij}$. Then, there are three types of events:
(i) $X_{i}$ increases by 1 at rate $b_{i}X_{i} \left(1-\frac{\sum_{j} X_{j}}{K}\right)$,
(ii) $X_{i}$ decreases by 1 at rate $d_{i}X_{i}$, or,
(iii) $X_{i}$ increases by 1 and $X_{j}$ decreases by 1 at rate $b_{i} X_{i} \left(\frac{c_{ij}X_{j}}{K}\right)$.
This gives a stochastic model of a population with density-dependent competition; *n.b.* the total population size is *not* fixed at $K$, but is rather allowed to fluctuate stochastically with an upper bound of $K$, as we allow the possibility of empty slots in the environment.
Let $x_{i}(t)$ denote the density of species $i$ (*i.e.* $\frac{X_{i}(t)}{K}$). As in the previous section, this system may be approximated by a system of stochastic differential equations, $$\begin{gathered}
\frac{dx_{i}}{dt} = \left((b_{i}-d_{i})- \sum_{j} (b_{i} - b_{i} c_{ij}+b_{j}c_{ji})x_{j}\right)x_{i}\\
+ \sqrt{\frac{b_{i} x_{i}(1-\sum_{j} x_{j})}{K}} \eta_{b,i}(t)
- \sqrt{\frac{d_{i} x_{i}}{K}} \eta_{d,i}(t)t
+ \sum_{j} \sqrt{\frac{b_{i} c_{ij} x_{i} x_{j}}{K}} \eta_{i,j}(t)
- \sum_{j} \sqrt{\frac{b_{j} c_{ji} x_{i} x_{j}}{K}} \eta_{j,i}(t).\end{gathered}$$ We will be interested in large carrying capacity limits as $K \to \infty$, so here, $\mu = \frac{1}{K}$. To explore the link between population genetics and population dynamics, we will further postulate that there exist values $\epsilon_{1},\ldots,\epsilon_{d}$ so that $$b_{i} = b\left(1 + \frac{\epsilon_{i}}{K}\right), \quad
d_{i} = d + \frac{\eta_{i}}{K},
\quad \text{and} \quad c_{ij} = c + \frac{a_{ij}}{K},$$ for all $i,j$; this corresponds to the weak selection hypothesis [@Ewens1979]: all species are competitively equivalent and differ in their demographic rates by terms of $O\left(\frac{1}{K}\right)$. Then, $$\begin{gathered}
f_{i}(\bm{x}) = x_{i}\left((b-d) - b \sum_{j=1}^{K} x_j\right),\\
\varepsilon = \frac{1}{K},\\
\intertext{and}
h_{i}(\bm{x}) = x_{i} \left(\left(b \epsilon_{i} - d \eta_{i}\right) -
b \sum_{j} \left( (1-c)\epsilon_{i} - c \epsilon_{j} - a_{ij} + a_{ij} \right) x_{j} \right)\end{gathered}$$ Under these assumptions, $\text{rank}(\bm{A} \vert \bm{b}) = 1$, $$\Gamma = \left\{\bm{x} \in \mathbb{R}^{d} : \sum_{j=1}^{K} x_{j} = 1-\frac{d}{b}\right\}$$ and for $\bm{x} \in \Gamma$, the derivatives and simplify to $$P_{ij}(\bm{x}) = \delta_{ij} - \frac{x_{i}}{1-\frac{d}{b}} \quad \text{and} \quad
Q_{ijk}(\bm{x})
= - \frac{1}{1-\frac{d}{b}} \left(\delta_{ij} + \delta_{ik} - \frac{2 x_{i}}{1-\frac{d}{b}}\right),$$ whereas for $\bm{x} \in \Gamma$, $$h_{i}(\bm{x}) = x_{i} \left(d(\epsilon_{i} - \eta_{i}) + c \sum_{j} (\epsilon_{i} - \epsilon_{j}) x_{j}
+ b \sum_{j} (a_{ij}-a_{ji})x_{j}\right).$$ A straightforward if lengthy calculation shows that $\bm{g}(\bm{x}) = O\left(\frac{1}{K^{2}}\right)$.
Substituting into our general formula then gives $$\begin{gathered}
\frac{d\tilde{x}_{i}}{dt}
= \frac{1}{K}
\left(h_{i}(\tilde{\bm{x}}) - \frac{\tilde{x}_{i}}{1-\frac{d}{b}} \sum_{j} h_{j}(\tilde{\bm{x}}) \right)\\
+ \sum_{j} (\delta_{ij} - \frac{\tilde{x}_{i}}{1-\frac{d}{b}})
\left(\sqrt{\frac{d \tilde{x}_{j}}{K}} (\eta_{b,j}(t) - \eta_{d,j}(t))
+ \sum_{k} \sqrt{\frac{b c \tilde{x}_{j} \tilde{x}_{k}}{K}} (\eta_{j,k}(t) - \eta_{k,j}(t))\right),\end{gathered}$$ or, changing variables to $p_{i} = \frac{\tilde{x}_{i}}{1-\frac{d}{b}}$, so $p_{i}$ is the proportion of species $i$, $$\begin{gathered}
\frac{dp_{i}}{dt}
= \frac{1}{K}
p_{i} \left(s_{i}(\bm{p}) - \sum_{j} s_{j}(\bm{p}) p_{j} \right)\\
+ \sum_{j} (\delta_{ij} - p_{i})
\left(\frac{1}{\sqrt{1-\frac{d}{b}}} \sqrt{\frac{d p_{j}}{K}} (\eta_{b,j}(t) - \eta_{d,j}(t))
+ \sum_{k} \sqrt{\frac{b c p_{j} p_{k}}{K}} (\eta_{j,k}(t) - \eta_{k,j}(t))\right),\end{gathered}$$ where $$s_{i}(\bm{p}) = d(\epsilon_{i} - \eta_{i})
+ c\left(1 - \frac{d}{b}\right) \sum_{j} (\epsilon_{i} - \epsilon_{j}) p_{j}
+ (b - d) \sum_{j} (a_{ij}-a_{ji})p_{j}.$$
The corresponding Fokker-Planck equation for the density $f(\bm{p},t)$ is then $$\frac{\partial f}{\partial t} = -\frac{1}{K} \frac{\partial}{\partial p_{i}} \left[
p_{i} \left(s_{i}(\bm{p}) - \sum_{j} s_{j}(\bm{p}) p_{j} \right) f\right]
+ \frac{1}{2} \frac{{2\left(bc + \frac{d}{1-\frac{d}{b}}\right)}}{K}
\frac{\partial^{2}}{\partial p_{i} \partial p_{j}}\left[p_{i}(\delta_{ij}-p_{j})f\right]$$ which we recognise as the equation for the Wright-Fisher diffusion, where the (frequency dependent) selection coefficient is $\frac{s_{i}(\bm{p})}{K}$ and the effective population size is $N_{e} = \frac{\left(1-\frac{d}{b}\right)K}{2\left(c(b-d) + d\right)}$; $\left(1-\frac{d}{b}\right)K$ is the population size at the deterministic equilibrium, whereas the other terms reflect variance in the total population size. This gives an alternate derivation of the results presented in [@Parsons2007b; @Parsons2008; @Parsons2010; @Constable2015].
Continuous degrees of freedom: example of competition-limited diffusion
-----------------------------------------------------------------------
The methods of Section \[S2\] can readily be extended to infinite dimensional settings. Two recent examples come from work exploring the role of stochasticity in spatial ecological models [@Rogers2012a; @Pigolotti2014]. Here we work through a simple illustrative example of diffusing particles coupled by a competitive birth-death interaction; we will show that this competition acts to limit the speed of diffusion of the population. Interested readers are referred to [@Etheridge1991], where the continuum limit of this example has been studied in considerable depth.
Consider the following stochastic process. At time $t$ there are $N_t$ individual particles wandering in a one-dimensional space, each following their own Brownian motion with diffusion constant $D=\sqrt{2\varepsilon}$. With rate one, each particle may independently ‘reproduce’, creating a daughter particle that initially shares the location of the parent, but thereafter moves independently. Particles ‘die’ with rate proportional to their total number; specifically, the death rate for each particle is $\mu(N_t-1)$. We assume the constants $\mu$ and $\varepsilon$ are small, but of the same order.
Since the location of the particles does not influence the birth or death rates, it is easy to see that the total number of particles follows a logistic growth law, quickly reaching an equilibrium $N_t\approx \mu^{-1}$. The total population size remains at this level while the spatial distribution of particles evolves slowly over a much longer timescale. We are interested in the long-term behaviour of the distribution of particle locations. Introduce the population density $$u(x)=\mu\sum_{n=1}^{N_t}\delta\left(x-X_n(t)\right)\,,$$ where $X_n$ is the location of particle $n$ at time $t$, $\delta$ is the Dirac delta function, and we suppress the dependence of $u$ on $t$ to reduce clutter. Simulations suggest that the competitive interaction of the particles limits the extent to which they are able to diffuse away from each other (Figure \[clbm\_fig\], left panel). This observation can be made quantitative by computing the mean square distance between pairs of particles, $$\Delta[u]:=\mu^2\sum_{n,m}(X_n(t)-X_m(t))^2=\iint (x-y)^2u(x)u(y)\,dx\,dy\,.$$ The right panel of Figure \[clbm\_fig\] shows the time evolution of $\Delta$ for the population, compared to the growth $\Delta \sim t$ observed for independent diffusing particles. The solid lines show our theoretical prediction for this phenomenon, which we will now derive using timescale separation.
![Simulation of competition-limited diffusion (dark red), contrasted with a collection of $N$ independent Brownian particles (light purple). The left panel shows the particle trajectories, on the right is shown the mean square distance between pairs of particles. []{data-label="clbm_fig"}](clbm.pdf){width="80.00000%"}
Following a system-size expansion [@McKane2014], we find that the time-evolution of $u(x)$ is described to close approximation by the stochastic partial differential equation (SPDE) $$\frac{\partial}{\partial t}u(x)=\varepsilon\frac{\partial^2}{\partial x^2}u(x)+u(x)\left(1-\int u(y)\,dy\right)+\sqrt{\mu\, u(x)\left(1+\int u(y)\,dy\right)}\,\eta(x,t)\,,
\label{spdeu}$$ where $\eta(x,t)$ is spatio-temporal white noise and the integrals run over the real line.
Equation (\[spdeu\]) has the same essential structure as our basic object of interest (\[sdex\]). If we identify $$\begin{split}
&f[u](x)=u(x)\left(1-\int u(y)\,dy\right)\\
&h[u](x)=\frac{\partial^2}{\partial x^2}u(x)\\
&G[u](x,s)=\delta(x-s)\sqrt{u(x)\left(1+\int u(y)\,dy\right)}\,,
\end{split}
\label{fhG_ex}$$ then (\[spde\]) becomes $$\frac{\partial }{\partial t}u(x)=f[u](x)+\varepsilon h[u](x)+\sqrt{\mu}\int G[u](x,s)\eta(s,t)\,ds\,.
\label{spde}$$ The integral here is the analogue of the matrix-vector multiplication $\bm{G}(\bm{x})\bm{\eta}(t)$ appearing in (\[sdex\]). The delta function appearing in $G[u]$ means that the noise in our example is spatially uncorrelated; this may not hold for other models.
In this section we will show how the timescale separation techniques discussed above may also be applied to equations of the form (\[spde\]). First, an important caveat: the irregular nature of spatio-temporal noise creates enormous mathematical complications in the rigorous analysis of SPDEs, we refer interested readers to the Fields-medal winning work [@Hairer2014]. In what follows we will turn a blind eye to deeper questions concerning the nature of the solution space and simply apply the techniques developed in the previous sections.
First we examine the outer part $\partial u/\partial t=f[u]$. In our example, the PDE $$\frac{\partial}{\partial t}u(x)=u(x)\left(1-\int u(y)\,dy\right)\,,$$ is straightforward to solve: $$u(x,t)=\frac{u(x,0)\,e^t}{1+(e^t-1)\int u(y,0)\,dy}\,,$$ which describes the fast relaxation of $u$ to a state in which it has total mass one. In this infinite-dimensional setting, the map that describes the long-time limit of the outer solution (previously defined in (\[defpi\])) is an operator $\pi$, whose action is specified by $$\pi[u](x)=\frac{u(x)}{\int u(y)\,dy}\,.$$ We suppose that there exists a suitable space of functions $\mathcal{U}$ describing possible solutions of (\[spde\]). Exactly what kind of space is a deep question beyond our present focus. The analogue of the slow manifold is the subspace $\mathcal{V}\subset\mathcal{U}$ containing functions $v$ satisfying $f[v]=0$, or equivalently for our example, $\int v(y)\,dy = 1$. We aim to derive an equation describing slow stochastic evolution in $\mathcal{V}$ that well-approximates the behaviour of solutions to the full system (\[spde\]).
Where previous calculations involved partial differentiation, we now apply a functional derivative. In analogue to the definitions in (\[defPQ\]) we introduce $$P[v](x,y)=\frac{\delta }{\delta u(y)}\pi[u](x)\Bigg|_{u=v}\,,\quad Q[v](x,y,z)=\frac{\delta^2 }{\delta u(y)\delta u(z)}\pi[u](x)\Bigg|_{u=v}\,.$$ The reduced system may then be written down: $$\begin{split}
\frac{\partial}{\partial t}v(x)=&\int P[v](x,y)\left[\varepsilon h(y)\,dy +\sqrt{\mu}\int G[v](y,s)\eta(s,t)\,ds\right]\\
&+\frac{\mu}{2}\iiint G[v](y,s)G[v](z,s)Q[v](x,y,z)\,dy\,dz\,ds\,.
\end{split}
\label{spde_re}$$
For the example at hand we compute $$\begin{split}
&\frac{\delta }{\delta u(y)}\pi[u](x)=\frac{\delta(x-y)}{\int u(z)\,dz}-\frac{u(x)}{\left(\int u(z)\,dz\right)^2}\,,\\
&\frac{\delta^2 }{\delta u(y)^2}\pi[u](x)=\frac{2u(x)}{\left(\int u(z)\,dz\right)^3}-\frac{2\delta(x-y)}{\left(\int u(z)\,dz\right)^2}\,,
\end{split}$$ and thus $$P[v](x,y)=\delta(x-y)-v(x)\,,\quad Q[v](x,y,y)=2v(x)-2\delta(x-y)\,.
\label{PQ_ex}$$ Note that we only need the $z=y$ parts of $Q[v](x,y,z)$ because of the delta function in $G$. Plugging (\[fhG\_ex\]) and (\[PQ\_ex\]) into (\[spde\_re\]), we obtain the reduced model $$\frac{\partial}{\partial t}v(x)=\varepsilon \frac{\partial^2}{\partial x^2}v(x)+\sqrt{2\mu}\int \big[\delta(x-y)-v(x)\big]\sqrt{v(y)}\eta(y,t)\,dy\,.
\label{spdev}$$
Comparing (\[spdev\]) to the original equation (\[spdeu\]) we see two main differences: the non-linearity in the drift has vanished, but the noise is now spatially coupled.
To compute a prediction for the mean squared distance between particles, it is simpler to work in Fourier space. Introducing $\tilde{v}(k)=\int e^{-2\pi i k x}v(x)\,dx$, we note first that $$\mathbb{E}\Delta[v]=\iint z^2e^{2\pi i k z}\,\mathbb{E}|\tilde{v}(k)|^2\,dk\,dz=-\frac{1}{4\pi^2}\frac{\partial^2}{\partial k^2}\mathbb{E}|\tilde{v}(k)|^2\Bigg|_{k=0}
\label{EDv}$$ Translating (\[spdev\]) to Fourier space we find $$\frac{\partial}{\partial t}\tilde{v}(k)=-4\varepsilon\pi^2k^2\tilde{v}(k)+\sqrt{2\mu}\int \widetilde{G}[\tilde{v}](k,x)\,\eta(x,t)\,dx\,,$$ where $$\widetilde{G}[\tilde{v}](k,x)=\left(e^{-2\pi ikx}-\tilde{v}(k)\right)\sqrt{\int\,e^{2\pi i \ell x}\tilde{v}(\ell)d\ell}\,.$$ In mean, this process behaves exactly as a straightforward diffusion: $$\frac{d}{dt}\mathbb{E}[\tilde{v}(k)]=-4\varepsilon\pi^2k^2\,\mathbb{E}[\tilde{v}(k)]\,.$$ However, the noise introduces a correction to the variance following Itô’s formula. Specifically, $$\begin{split}
\frac{d}{dt}\mathbb{E}|\tilde{v}(k)|^2&=-8\varepsilon\pi^2k^2\mathbb{E}|\tilde{v}(k)|^2+\frac{1}{2}\iiint \widetilde{G}[\tilde{v}](\ell,x)\widetilde{G}[\tilde{v}](m,x)\frac{\delta^2|\tilde{v}(k)|^2}{\delta \tilde{v}(\ell)\delta \tilde{v}(m)} \,dx\,d\ell\,dm\\
&=-8\varepsilon\pi^2k^2\mathbb{E}|\tilde{v}(k)|^2+2\mu(1-\mathbb{E}|\tilde{v}(k)|^2)\,.
\label{dEv2}
\end{split}$$ Solving (\[dEv2\]) and plugging into (\[EDv\]) gives the prediction $$\mathbb{E}\Delta[v]=\frac{2\varepsilon}{\mu}\left(1-e^{-2\mu t}\right)\,.$$ This result is shown as the dark red curve in Figure \[clbm\_fig\]. In particular, notice that whilst the mean square distance between diffusing particles grows indefinitely, in the competition coupled process it attains a finite limit $2\varepsilon/\mu$.
Discussion
==========
The purpose of this article has been to show the derivation and application of a systematic computational framework for dimension reduction in stochastic dynamical systems that exhibit a separation of timescales via a globally stable normal hyperbolic slow manifold *i.e.* in the limit of small noise the limiting deterministic dynamical system defined by $\bm{f}$ possesses a single, connected and globally attractive manifold of fixed points. The method is exact in the limit of small noise and well-separated slow and fast dynamics, and experimentally found to be valid as an approximation scheme over a sensible parameter range. We have also presented extensions of the method for infinite dimensional systems and processes coupled to general noise sources.
In some applications more general scenarios may occur, we now briefly discuss two of interest. Some models may exhibit more than one connected manifold of equilibria or dynamic bifurcations, , points where the critical manifold ceases to be normally hyperbolic [@Berglund2006]; in this case the theory developed here will apply locally to trajectories in the basin of attraction of each manifold individually, but further analysis will be necessary to describe the statistics of noise-driven transitions between manifolds. A possibly more exciting direction for further research is the analysis of noisy behaviour around more general attractors such as limit cycles, limit tori and strange attractors. In the case of limit cycles some work exists on stochastic extensions to Floquet theory [@Boland2009], however, this is a linear description that cannot capture any bias analogous to the noise-induced drift in the slow manifold setting.
Finally, it is worth returning to discuss the motivation for this work. As mentioned earlier, variations of the work of Katzenberger have been independently rediscovered by several groups in recent years, almost all of whom have been interested in questions about the role of noise in ecology and evolution. Historically, many theoretical results in this field have been derived from models that assume for convenience a fixed population size. In the deterministic limit this assumption is not important, but we are now beginning to realise that the inclusion of noise can induce radically different and sometimes unexpected behaviour. Mathematically, this is a consequence of the noise-induced drift term $\bm{g}$ that appears in our equation (\[sdez\]), and more generally of the seemingly endless capacity of Itô’s lemma to cause surprise. There have been some tentative explorations of the possible evolutionary and ecological consequences of these effects [@Parsons2010; @Rossberg2013; @Constable2014b], but much more is yet to be discovered.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank George Constable, Alan McKane and Christopher Quince for useful discussions and Tom Kurtz for bringing [@Katzenberger91] to our attention. Work on this paper began when both TLP and TR were visitors at the Isaac Newton Institute for Mathematical Sciences programme *Understanding Microbial Communities; Function, Structure and Dynamics*.
Funding statement {#funding-statement .unnumbered}
=================
TLP is supported by the CNRS, and TR by the Royal Society.
Related literature {#lit}
==================
As mentioned in the introduction, the mathematical and theoretical physics literatures contain a multitude of techniques for separation of timescales, many of which have been extremely well-studied. In this appendix we present a brief overview of some of the historical developments that we consider to be more relevant to the class of systems we are interested in.
The canonical example of timescale separation in physics is Brownian motion; stochastic interactions with the water molecules cause changes to the *velocity* of the pollen grain on a fast timescale, which have a cumulative effect of perturbing the *position* on a slower timescale. The process of moving from a description of the particle’s motion in terms of position and velocity to one concerned just with position (thus reducing the dimension of the model from two to one) is known as *adiabatic elimination*[^5]. Following the introductory discussion in [@Gardiner1985], one may write equations of motion for the pollen grain of the form of Langevin equations such as $$\frac{dx}{dt}=v\,,\quad m\frac{dv}{dt}=-v+\sqrt{\mu}\,\eta(t)\,,$$ where $m$ is the mass, $\mu$ a constant derived from the temperature of the water bath, and $\eta(t)$ is Gaussian white noise. If the mass $m$ is very small, we might approximate $m\approx0$ and thus $v\approx \sqrt{\mu}\,\eta(t)$. From this we derive the reduced description $$\frac{dx}{dt}=\sqrt{\mu}\,\eta(t)\,.$$ A reader encountering arguments of this type for the first time is likely to be suspicious, and rightly so. Although provably exact for the simple case described here, taking limits in a brusque fashion like this is generally inadvisable when dealing with stochastic dynamical systems. Nonetheless, the method is quite powerful and the same basic principle has been developed to various levels of rigour and generality by many authors, most notably Haken [@Haken1982].
Although conceptually appealing, the Langevin-style description of a stochastic system in terms of dynamical equations with noise hides some important complications. As any introductory text will warn, when writing such an equation we must specify the sense in which we are to understand the noise term $\eta(t)$. This choice then impacts the behaviour of the system under certain limits or changes of variables, potentially complicating the process of timescale separation. An alternative description of a stochastic system that avoids this ambiguity is in terms of a PDE describing the time evolution of the probability density, known as the Fokker-Planck Equation (also, the Kolmogorov forward equation). Timescale separation in this setting amounts to integrating out one or more degrees of freedom from the PDE to reduce its dimension. Physicists might understand this process through its natural analogue in quantum mechanics, the Born-Oppenheimer approximation [@Born1927]. In most applications the Fokker-Planck equation will not be exactly separable, necessitating the application by hand of a carefully chosen projection operator, an approach going back to the work of Zwanzig [@Zwanzig1960].
Existing in parallel with the development of adiabatic elimination techniques in physics is a separate body of literature concerned with probabilistic models of biological processes that also exhibit a separation of timescales. This thread begins with early work on the convergence of Markov chains — particularly those appearing in genetics — to diffusion processes [@Feller1951; @Trotter1958]. Analyzing the Wright-Fisher model of allele frequencies in a population of large size $N$, Feller used the standard convention of measuring the population in units of $N$ individuals (so one individual has weight $\frac{1}{N}$), but rather than truncating the master equation by discarding terms of order $\frac{1}{N^{2}}$ and greater, he also rescaled to a “slow time” so that $\Delta t = \frac{1}{N}$, which led to a well defined limit as $N \to \infty$, a procedure proved rigorously by Trotter.
Mathematical population genetics also provides an important early example of a stochastic process explicitly considered at multiple timescales. In [@Ethier1980; @Nagylaki1980], models of diploid populations with non-overlapping generations and geographic structure were analysed by separating a fast timescale on which, for example, the genotype frequencies would rapidly equilibrate to Hardy-Weinberg proportions, from a slow timescale over which allele frequencies would vary due to mutation and genetic drift. Their approach, obtained by an extension of the semi-group methods of [@Trotter1958], requires the explicit characterization of infinite-dimensional spectral projection operators, and an explicit separation of the process into fast and slow variables, the former of which must be effectively constant on the slow-timescale. These requirements somewhat limit the generality of the approach, but it does have the virtue of being applicable to Markov processes other than Brownian motion, such as measure-valued processes [@Harris2015].
Slightly over a decade later, two very general papers appeared that used similar approaches based in stochastic differential equations and classical, deterministic formulations of time-scale separation. The first, [@Katzenberger91], considered processes taking values in $\mathbb{R}^{d}$, but with a very general approach to stochastic noise, so that the results can be applied both to diffusion processes and discrete Markov processes (see the discussion in section \[Katz\]), whilst the latter [@Funaki1993], allows the process to be defined on a general Riemann manifold $M$, but requires the noise to be the canonical Brownian motion on the manifold. In both, there is a fast timescale on which the process essentially behaves like a deterministic dynamical system, with trajectories that approach a lower-dimensional “slow manifold” $\Gamma$. On the slow time-scale, the process is asymptotic to a diffusion process that is confined to the manifold $\Gamma$. Katzenberger characterized the diffusion on the slow manifold via stochastic differential equations and a function $\bm{\pi}$ (defined formally below), which, given an initial point $\bm{x}$, gives the point in $\Gamma$ to which it will be carried by the fast dynamics, whereas [@Funaki1993] used the backward equation to describe the diffusion.
Unfortunately, in spite of their generality, [@Katzenberger91] and [@Funaki1993] have been frequently overlooked, as was the follow up [@Funaki1995], where the problem of dimension reduction in infinite dimensions was considered for an SPDE of Ginzburg-Landau type. Indeed, special cases have been subsequently rediscovered. This is perhaps because, whilst these papers have the virtue of providing rigorous proofs, they are not necessarily useful for the practitioner: [@Katzenberger91] supposes the function $\bm{\pi}$ is given, whereas in applications it is frequently difficult or impossible to obtain in closed form, as the fast system may not be solvable. [@Funaki1993] gives a closed expression for the diffusion equation, though one that, except in the case when the fast system is a gradient flow, requires extensive calculations with Fermi coordinates (like Katzenberger’s $\bm{\pi}$, these are often impossible to obtain in closed form) and only applies when the noise is Brownian motion.
Katzenberger’s Theorem {#Katz}
======================
Above we developed our results in the context of Itô SDEs, however, [@Katzenberger91] proved a more general result that allows us to consider a much broader class of noise processes: semimartingales. Semimartingales are the most general class of stochastic processes for which one may define a stochastic integral and stochastic differential equations (Brownian motion is included as a special case). Suitably adapted, most of the familiar results for SDEs and white-noise integrals, including Itô’s formula, remain true in the more general setting [@Protter04].
To define a semimartingale, we must first make a few auxiliary definitions. A Markov process $M(t)$ is a *martingale* if $$\mathbb{E}\left[M(t) \middle\vert M(s) \right] = M(s).$$ A random variable $\tau$ taking values in $[0,\infty)$ is a *stopping time* if one can determine if $\tau < t$ without knowledge of the future beyond $t$; an example of a stopping time is the first time a diffusion started from 0 exits an interval $[-a,a]$. $M(t)$ is a *local martingale* if there is a sequence of stopping times $\tau_{n} \to \infty$ such that $M(\min\{t,\tau_{n}\})$ is a martingale for each $n$.
A function is *càdlàg* if it is continuous from the right and has left-hand limits at every point.
The *total variation* of a function $f$ on an interval $[a,b]$ is $$V^{a}_{b}(f) = \min_{\{t_{i}\}} \sum_{i} |f(t_{i+1})-f(t_{i})|,$$ where the minimum is over all partitions $a=t_{0} < t_{1} < \cdots < t_{n} = b$ of $[a,b]$. A stochastic process $A(t)$ is of *finite variation* if it is càdlàg and has finite total variation on all intervals $[a,b]$ (note that $A(t)$ is allowed to have jump discontinuities).
Finally, $Z(t)$ is a semimartingale if it may be written as the sum of a local martingale and a finite variation process, $$Z(t) = M(t) + A(t).$$ Diffusion processes are the prototypical example of semimartingales, but the class is much broader, and includes processes with jumps, such as Lévy processes; *e.g.* if $N(t)$ is a Poisson process, then $M(t) = N(t) - t$ is a local martingale and $A(t) = t$ is of finite variation, so $N(t)$ is a semimartingale. Integration with respect to a semimartingale is defined analogously to the Stieltjes integral, except that we require the approximating sum to converge in probability, and, as with the Itô integral, the integrand is always evaluated at the left endpoint of each interval in the partition.
More generally, we can define vector and matrix valued martingales, local martingales, finite variation processes and semimartingales, $\bm{M}(t)$, $\bm{A}(t)$, and $\bm{Z}(t)$, by requiring the components, $M_{i}(t)$ *etc.*, have the corresponding property.
We can now formulate Katzenberger’s result. Let
1. $\bm{Z}_{n}(t)$ be a convergent sequence of vector valued semimartingales such that the jumps $\Delta \bm{Z}_{n}(t) \to 0$ as $n \to \infty$,
2. $A_{n}(t)$ be a sequence of non-decreasing finite variation processes such that $\Delta A_{n}(t) \to 0$, and $$\int_{a}^{b} dA_{n}(s) = A_{n}(b) - A_{n}(a) \to \infty;$$ Katzenberger notes that most frequently in applications, $A_{n}(t) = \alpha_{n} t$ for some sequence $\alpha_{n} \to \infty$ (*n.b.*, in this formulation, this explosion in $A_{n}(s)$ corresponds to the drift becoming infinitely strong, rather than the noise infinitely weak, as in . The two are equivalent, if one changes the timescale accordingly; recall we had $$\frac{d\bm{x}}{dt}=\bm{f}(\bm{x})+\varepsilon\bm{h}(\bm{x})+\sqrt{\mu}\,\bm{G}(\bm{x})\,\bm{\eta}(t).$$ If instead, we consider the process $\tilde{\bm{x}}(t) = \bm{x}(\mu t)$, we get $$\frac{d\tilde{\bm{x}}}{dt}= \frac{1}{\mu} \bm{f}(\tilde{\bm{x}})+\frac{\varepsilon}{\mu}\bm{h}(\tilde{\bm{x}})+\,\bm{G}(\tilde{\bm{x}})\,\bm{\eta}(t),$$ with a drift that blows up as $\mu \to 0$).
3. $\bm{f}$ and $\Gamma$ be as before,
4. $\bm{G}_{n}(\bm{x})$ be a sequence of matrix-valued functions converging to a limit $\bm{G}(\bm{x})$, and
5. $\bm{x}_{n}(t)$ be a sequence of stochastic processes satisfying the (semimartingale) SDE $$\ d\bm{x}_{n} = \bm{f}(\bm{x}_{n})\, dA_{n} + \bm{G}_{n}(\bm{x}_{n})\, d\bm{Z}_{n}.$$
Then, as before, subject to a few technical considerations, as $n \to \infty$, $\bm{x}_{n}$ converges to a diffusion process on $\Gamma$ satisfying $$\label{KSDE}
\frac{d\bm{z}}{dt}=\bm{g}(\bm{z})+ \bm{P}(\bm{z})\bm{G}(\bm{z})\bm{\eta}(t)\,,$$ where $\bm{g}$ is as in equation and $\bm{\eta}$ is white noise.
Some care is required in understanding the sense of convergence in [@Katzenberger91]; if $\bm{x}_{n}(0)$ converges weakly to $\bm{z} \in \Gamma$ in $\mathbb{R}^{d}$ (*i.e.*, for all continuous functions $F:\mathbb{R}^{d} \to \mathbb{R}$, $\mathbb{E}[F(\bm{x}_{n}(0))] \to \mathbb{E}[F(\bm{z})]$) then $\bm{x}_{n}(t)$ converges weakly to $\bm{z}(t)$ in the space of càdlàg functions: $$\mathbb{E}[F(\bm{x}_{n}(t))] \to \mathbb{E}[F(\bm{z}(t))]$$ for all continuous functions $F$ from the space of càdlàg functions on $[0,\infty)$ to $\mathbb{R}$, see [@Billingsley68; @Ethier+Kurtz86] for a definition of the topology on càdlàg functions and results on weak convergence. When $\bm{x}_{n}(0)$ converges to a limit $\bm{x}$ that is not in $\Gamma$, additional care is required: in this case, the process will jump instantaneously from $\bm{x}$ to $\bm{\pi}(\bm{x}) \in \Gamma$, which is not compatible with convergence in the weak topology on càdlàg functions. However, if one considers $$\hat{\bm{x}}_{n}(t) = \bm{x}_{n}(t) - \bm{\xi}(nt) + \bm{\pi}(\bm{x}),$$ (recall, $\bm{\xi}(t)$ is the solution to the outer system, ) then $\hat{\bm{x}}_{n}(0) \to \bm{\pi}(\bm{x}) \in \Gamma$ and $\hat{\bm{x}}_{n}(t)$ converges weakly to the diffusion $\bm{z}(t)$ on $\Gamma$ as before; intuitively $\hat{\bm{x}}_{n}(t)$ is obtained by removing the initial transient phase when $\bm{x}_{n}(t)$ follows the trajectories of the outer system, and starting the process instead from the endpoint of that trajectory, $\bm{\pi}(\bm{x})$ (see Figure \[xi\_fig\]).
![Illustration of $\hat{x_1}$ on the fast timescale for a prototypical stochastic dynamical system with a slow manifold $\Gamma=\{\bm{x}\,:\,x_1=1\}$. In the slow timescale the initial transit to the manifold is compressed into an instantaneous jump at $t=0$.[]{data-label="xi_fig"}](xi.pdf){width="80.00000%"}
Density dependent population processes
--------------------------------------
While Katzenberger’s result might seem unnecessarily abstract, it allows one to apply the same slow-manifold reduction to a number of individual-based, discrete stochastic processes that include a number of well-known examples from applications. In [@Kurtz70; @Kurtz71; @Kurtz78; @Kurtz81], Kurtz introduced and studied what he called density dependent population processes. While his original motivation was chemical reaction networks, the class also includes many examples of interest in biology and epidemiology.
A sequence of Markov processes $\bm{x}_{n}(t)$ is a *density dependent population process* if $\bm{x}_{n}$ takes values in $\frac{1}{n} \mathbb{Z}^{d}$, and, if $q^{(n)}_{\bm{x},\bm{y}}$ is the jump rate between $\bm{x},\bm{y} \in \frac{1}{n} \mathbb{Z}^{d}$, then $$q^{(n)}_{\bm{x},\bm{y}} = n \lambda_{n(\bm{y}-\bm{x})}(\bm{x})$$ for some non-negative function $\lambda_{\bm{l}}(\bm{x})$ on $\mathbb{R}^{d}$, where $\bm{l} = n(\bm{y}-\bm{x}) \in \mathbb{Z}^{d}$. More generally, one can consider the case of functions $\lambda^{(n)}_{\bm{l}}(\bm{x})$ that depend on $n$, provided $\lambda^{(n)}_{\bm{l}}(\bm{x})$ converges to a limit $\lambda_{\bm{l}}(\bm{x})$ sufficiently quickly as $n \to \infty$; see [@Pollett1990].
The parameter $n$ corresponds to the “system size” in [@vanKampen92], and can be interpreted differently according to the context, as *e.g.* total population size, area, or volume. For example, consider the stochastic logistic process $X_{n}(t)$ with birth and death rates $$Q^{(n)}_{X,X+1} = \beta X\left(1 - \frac{X}{n}\right) \qquad Q^{(n)}_{X,X-1} = \delta X.$$ Here, $n$ plays the role of the carrying capacity in the deterministic logistic equation, *i.e.* the number of individuals the environment can support: individuals have an intrinsic per-capita birth rate $\beta$, but the offspring will only survive if it arrives in an unoccupied spot in the habitat. Nondimensionalising, we might consider instead the process $x_{n}(t) = \frac{1}{n} X_{n}(t)$, with rates $$q^{(n)}_{x,x+\frac{1}{n}} = n \beta x(1-x) \qquad q^{(n)}_{x,x-\frac{1}{n}} = n \delta x.$$ The latter is an example of a density-dependent population process, with $$\lambda_{1}(x) = \beta x (1-x) \qquad \lambda_{-1}(x) = \delta x.$$
In [@Kurtz70], Kurtz shows that provided $$\sum_{\bm{l} \in \mathbb{Z}^{d}} \|\bm{l}\| \sup_{\bm{x} \in \mathcal{K}} \lambda_{\bm{l}}(\bm{x}) < \infty$$ for all closed and bounded sets $\mathcal{K}$, then if $$\bm{f}(\bm{x}) = \sum_{\bm{l} \in \mathbb{Z}^{d}} \bm{l} \lambda_{\bm{l}}(\bm{x})$$ is differentiable and $\bm{x}_{n}(t) \to \bm{x}_{0}$, then for any fixed $T > 0$, $$\lim_{n \to \infty} \sup_{t \leq T} |\bm{x}_{n}(t)- \bm{x}(t)| = 0,$$ where $\bm{x}(t)$ is the solution of $\frac{d\bm{x}}{dt} = \bm{f}(\bm{x})$ with $\bm{x}(0) = \bm{x}_{0}$.
If one assumes that $\lambda_{\bm{l}}(\bm{x})$ is non-zero for only finitely many transitions, say $\bm{l}_{1},\ldots,\bm{l}_{s}$, then, letting $\bm{G}(\bm{x})$ be the matrix with $i$^th^ column $\bm{l}_{i} \sqrt{\lambda_{\bm{l}_{i}}(\bm{x})}$, $\bm{\eta}(t)$ be an $s$-dimensional Itô white noise, and $\bm{z}_{n}(t)$ be the solution of $$\frac{d\bm{z}_{n}}{dt} = \bm{f}(\bm{z}_{n}) + \frac{1}{\sqrt{n}} \bm{G}(\bm{z}_{n}) \bm{\eta}(t),$$ then for any fixed $T > 0$, there exists a constant $C_{T}$ such that $$\lim_{n \to \infty} \mathbb{P}\left(\sup_{t \leq T} |\bm{x}_{n}(t) - \bm{z}_{n}(t)|
> \frac{C_{T} \log{n}}{n}\right) = 0.$$
In our current setting, if $\bm{f}(\bm{x})$ is twice continuously differentiable and once again has a globally attractive $m$-dimensional manifold of equilibria $\Gamma$, then the process $\bm{z}_{n}(t) = \bm{x}_{n}(nt)$ satisfies the conditions of [@Katzenberger91], so that as $n \to \infty$, $\bm{z}_{n}(t)$ converges to a diffusion $\bm{z}(t)$ satisfying equation for $\bm{f}$ and $\bm{G}$ defined as above. This result was applied to the study population genetic and epidemiological models in [@Parsons2010; @Parsons2012].
Local representations of one-dimensional manifolds {#Theta}
==================================================
In this section, we will discuss how one may obtain a parameterisation $\bm{\gamma}$ of a one-dimensional slow manifold $\Gamma$ and compute the quadratic expansion of the flow field (i.e. the quantities $\bm{v}$ and $\bm{\Theta}$) in the neighbourhood of a point $\bm{x} \in \Gamma$.
We start by fixing a basis of generalised eigenvectors of the Jacobian at $\bm{x}_{0}$, $\bm{J}(\bm{x}_{0})$, say $\bm{w}_{1},\ldots,\bm{w}_{n}$, and letting $\bm{W}$ be the corresponding change of basis matrix with the $\bm{w}_{i}$ as columns. Let $\bm{w}_{1}$ to be the eigenvector corresponding to the eigenvalue 0 (which we take to be unique up to scalar multiplication). Then, $$\label{WJW}
\bm{W}^{-1}\bm{J}(\bm{x}_{0})\bm{W} = \begin{bmatrix} 0 & \\ & \bm{J}_{\bm{2}} \end{bmatrix},$$ where $\bm{J}_{2}$ is a block-diagonal matrix, with each block acting invariantly on one of the eigenspaces corresponding to the non-zero eigenvalues.
We introduce a new coordinate system $$\bm{z} = \bm{W}^{-1} \left(\bm{x} - \bm{x}_{0}\right).$$ In this coordinate system, we will construct a parameterisation $\bm{\gamma}(z_{1})$ of $\Gamma$ such that that $\bm{x}_{0} =\bm{\gamma}(0)$.
In the new coordinate system, the dynamics are then given by $\frac{d\bm{z}}{dt} = \hat{\bm{f}}(\bm{z})$, where $$\hat{\bm{f}}(\bm{z}) = \bm{W}^{-1}\bm{f}\left(\bm{x}_{0} + \bm{W}\bm{z}\right)$$ (thus, the Jacobian of $\hat{\bm{f}}$ at $\bm{0}$, say $\hat{\bm{J}}$, is $\bm{W}^{-1}\bm{J}(\bm{x}_{0})\bm{W}$). Setting $\bm{z}_{\bm{2}}=(z_2\,\ldots,z_d)$, we may write this as $$\label{NormalODE}
\begin{aligned}
\frac{dz_1}{dt} &= \varphi_1(z_1,\bm{z}_{\bm{2}})\\
\frac{d\bm{z}_{\bm{2}}}{dt} &= \bm{J}_{\bm{2}}\bm{z}_{\bm{2}} + \bm{\varphi}_{\bm{2}}(z,\bm{z}_{\bm{2}})\,.
\end{aligned}$$ where $\bm{\varphi}_{\bm{2}}(z_1,\bm{z}_{\bm{2}}) = (\varphi_{2},\ldots,\varphi_{d})$ is quadratic. We may thus Taylor expand $\varphi_{i}(\bm{z})$ about $\bm{0}$ as $$\varphi_{i}(\bm{z}) = \sum_{j, k = 1}^{d} c_{ijk} z_{j} z_{k} + O\left(|\bm{z}|^{3}\right).$$
Computing $\bm{\gamma}$ or $\bm{\Theta}$, is essentially the task of characterizing the centre and stable manifolds at $\bm{x}_{0}$ respectively. The centre manifold theorem (we follow the treatment in [@Glendinning94]) tells us that at $\bm{x}_{0}$ the centre manifold is tangent to $\bm{w}_{1} $, whereas the stable manifold is tangent to the space spanned by $\bm{w}_{2},\ldots,\bm{w}_{d}$. Moreover, we may locally represent each manifold as the graph of a function over the tangent space. In particular, in the new coordinate system, there exists a function $$\bm{\gamma}_{\bm{2}}(z_{1}) = (\gamma_{2}(z_{1}),\ldots,\gamma_{d}(z_{1}))$$ such that $\bm{\gamma}(z) = (z,\bm{\gamma}_{\bm{2}}(z))^T$ is a point on $\Gamma$ for all $z_{1}$ sufficiently close to 0, and a function $\vartheta(\bm{z}_{\bm{2}})$ such that $(\bm{z}_{\bm{2}},\vartheta(\bm{z}_{\bm{2}}))$ is a point in the stable manifold near $\bm{x}_{0}$ for $\bm{z}_{\bm{2}}$ sufficiently close to $\bm{0}$. We will demonstrate the calculation of $\bm{\gamma}_{\bm{2}}(z_{1})$ below; the calculation of $\vartheta(\bm{z}_{\bm{2}})$ is similar, so we will simply give the result. Finally, we will show how one obtains $\bm{\Theta}$ from $\vartheta(\bm{z}_{\bm{2}})$.
To begin, we observe that in our new coordinate system $\bm{x}_{0}$ is the origin and $\Gamma$ is tangent to the $z_{1}$ axis (*i.e.* the span of $\bm{w}_{1}$), so we must have $\bm{\gamma}_{\bm{2}}'(0) = \frac{d\bm{\gamma}_{\bm{2}}}{dz_{1}} = \bm{0}$. We thus look for $\bm{\gamma}_{\bm{2}}(z_{1})$ of the form $$\gamma_{i}(z_{1}) = a_{i} z_{1}^{2} + O\left(z_{1}^{3}\right).$$ (as we shall only be interested in the first and second order derivatives of $\bm{\gamma}$ at $\bm{x}_{0}$ – *i.e.* at $z_{1} = 0$ – this is adequate for our purposes).
Substituting into , for points on $\Gamma$ we have $$\begin{aligned}
\frac{dz_{1}}{dt} &= \varphi_{1}(z_{1},\bm{\gamma}_{\bm{2}}(z_{1}))\\
\frac{d}{dt}\bm{\gamma}_{\bm{2}}(z_{1}) &= \bm{J}_{\bm{2}}\bm{\gamma}_{\bm{2}}(z_{1}) + \bm{\varphi}_{\bm{2}}(z_{1},\bm{\gamma}_{\bm{2}}(z_{1})),
\end{aligned}$$ or, expanding the latter using the chain rule, $$\varphi_{1}(z_{1},\bm{\gamma}_{\bm{2}}(z_{1})) \frac{d\bm{\gamma}_{\bm{2}}}{dz_{1}}
= \bm{J}_{\bm{2}}\bm{\gamma}_{\bm{2}}(z_{1}) + \bm{\varphi}_{\bm{2}}(z_{1},\bm{\gamma}_{\bm{2}}(z_{1})).$$ Substituting our series expressions for the $\varphi_{i}$ and $h_{i}$, to lowest order this gives us $$2 c_{i11} a_{i} z_{1}^{3} + O\left(z_{1}^{4}\right)
= \left(\sum_{j=2}^{d} \hat{J}_{ij} a_{j} + c_{i11}\right) z_{1}^{2} + O\left(z_{1}^{3}\right)$$ *i.e.* we may obtain the quantities $a_{i}$, $i=2,\ldots,d$ by solving the system of equations $$\sum_{j=2}^{d} \hat{J}_{ij} a_{j} = - c_{i11},\quad i = 2,\ldots,d.$$ Noting that $(a_{2},\ldots,a_{d})^{T} = \frac{1}{2} \frac{d^{2}\bm{\gamma}_{\bm{2}}}{dz_{1}^{2}}(0)$ whereas $(c_{111},\ldots,c_{d11})^{T} = \frac{1}{2} \frac{\partial^{2}\bm{\varphi}_{\bm{2}}}{\partial z_{1}^{2}}(\bm{0})$, we can solve the previous equation as $$\frac{d^{2}\bm{\gamma}_{\bm{2}}}{dz_{1}^{2}}(0)
= -\hat{\bm{J}}^{-1} \frac{\partial^{2}\bm{\varphi}_{\bm{2}}}{\partial z_{1}^{2}}(\bm{0}).
$$ To return to our original functions as expressed in the original coordinate system, we first observe that for $i = 2,\ldots,d$, $$\frac{d^{2}\varphi_{i}}{dz_{1}^{2}}(\bm{0})
= \frac{\partial^{2}\hat{f}_{i}}{\partial z_{1}^{2}}(\bm{0}),$$ whereas $$\frac{\partial^{2}\hat{\bm{f}}}{\partial z_{1}^{2}}(\bm{0})
= \bm{W}^{-1} \sum_{j,k =1}^{d} \frac{\partial^{2} \bm{f}}{\partial x_{j} \partial x_{k}}(\bm{x}_{0}) W_{j1} W_{k1}.$$ In particular, recalling , we see that for $i = 2,\ldots,d$, $\frac{d^{2}\gamma_{i}}{dz_{1}^{2}}(0)$ agrees with the $i$^th^ entry of $$- \bm{W}^{-1}\bm{J}^{+} \sum_{j,k =1}^{d}
\frac{\partial^{2} \bm{f}}{\partial x_{j} \partial x_{k}}(\bm{x}_{0}) W_{j1} W_{k1},$$ *i.e.* $$- \left[\bm{W}^{T}
\frac{\partial^{2} \left[\bm{W}^{-1}\bm{J}^{+} \bm{f}\right]_{i}}{\partial \bm{x}^{2}}(\bm{x}_{0})
\bm{W}\right]_{11},$$ where, as before, $\bm{J}^{+}$ is the pseudo-inverse of $\bm{J}$, which is defined by $\bm{J}^{-1}$ on the image of $\bm{J}$ and is $\bm{0}$ on the kernel of $\bm{J}$.
Thus, $$\begin{gathered}
\bm{\gamma}(0) = \bm{x}_{0}\\
\bm{\gamma}'(0) = \bm{w}_{1}\\
\bm{\gamma}''(0) = \sum_{i=2}^{d} \frac{d^{2}\gamma_{i}}{dz_{1}^{2}}(0) \bm{w}_{i}
= - \sum_{i=2}^{d} \left[\bm{W}^{T} \frac{\partial^{2}
\left[\bm{W}^{-1}\bm{J}^{+} \bm{f}\right]_{i}}{\partial \bm{x}^{2}}(\bm{x}_{0})\bm{W}\right]_{11}\bm{w}_{i}
$$ and $$\bm{\gamma}(z_1) = \bm{x}_{0} + z_{1}\bm{w}_{1}
- \frac{1}{2} z_{1}^{2} \sum_{i=2}^{d} \left[\bm{W}^{T} \frac{\partial^{2}
\left[\bm{W}^{-1}\bm{J}^{+} \bm{f}\right]_{i}}{\partial \bm{x}^{2}}(\bm{x}_{0})\bm{W}\right]_{11}\bm{w}_{i}
+ O\left(z_{1}^{3}\right)$$ is the desired parametrization of $\Gamma$ in the $\bm{z}$ coodinates.
Proceeding similarly, we find that $$\vartheta(\bm{z}_{\bm{2}}) = \bm{z}_{\bm{2}}^{T} (\hat{\bm{J}}^{T})^{-1}
\frac{\partial^{2} \hat{f}_{1}}{\partial \bm{z}_{\bm{2}}^{2}}(\bm{0}) \bm{z}_{\bm{2}}$$ and $(\hat{\bm{J}}^{T})^{-1} \frac{\partial^{2} \hat{f}_{1}}{\partial \bm{z}_{\bm{2}}^{2}}(\bm{0})$ and $$\bm{W}^{T} \left(\bm{J}^{T}\right)^{+}
\frac{\partial^{2} \left[\bm{W}^{-1}\bm{f}\right]_{1}}{\partial \bm{x}^{2}}(\bm{x}_{0})
\bm{W}$$ have equal $j,k$^th^ entry for all $j,k=2,\ldots,d$ (the first row of the latter is zero, but the first column need not be). Thus, if we set $$\bm{\Theta(\bm{x}_{0})} = \bm{P}(\bm{x}_{0}) \bm{W}^{T} \left(\bm{J}^{T}\right)^{+}
\frac{\partial^{2} \left[\bm{W}^{-1}\bm{f}\right]_{1}}{\partial \bm{x}^{2}}(\bm{x}_{0})
\bm{W},$$ then the stable manifold at $\bm{x}_{0}$ is thus the set of all points $\bm{z}$ such that $$z_{1} = \bm{z}_{\bm{2}}^{T} \bm{\Theta(\bm{x}_{0})} \bm{z}_{\bm{2}}.$$ Now, if we choose $\bm{v}(\bm{x}_{0})$ so that $$\bm{v}(\bm{x}_{0})^{T} \bm{w}_{i} = \begin{cases}
1 & \text{if $i =1$, and}\\
0 & \text{otherwise,}
\end{cases}$$ then for a point $\bm{x} = \bm{x}_{0} + \Delta \bm{x}$, $z_{1} = \bm{v}(\bm{x}_{0})^{T} \Delta \bm{x}$, whereas $$\bm{z}_{\bm{2}} = \Delta \bm{x} - \left(\bm{v}(\bm{x}_{0})^{T} \Delta \bm{x}\right) \bm{w}_{1},$$ so that $\bm{x}$ is in the stable manifold at $\bm{x}_{0}$ (to lowest order in $\Delta \bm{x}$) provided $$\bm{v}(\bm{x}_{0})^{T} \Delta \bm{x}
- \left(\Delta \bm{x} - \left(\bm{v}(\bm{x}_{0})^{T} \Delta \bm{x}\right) \bm{w}_{1}\right)^{T}
\Theta(\bm{x}_{0})
\left(\Delta \bm{x} - \left(\bm{v}(\bm{x}_{0})^{T} \Delta \bm{x}\right) \bm{w}_{1}\right) = 0,$$ or, rearranging, $$\bm{v}(\bm{x}_{0})^{T} \Delta \bm{x}
- \Delta \bm{x}^{T}\left(\bm{I} - \bm{w}_{1}\bm{v}(\bm{x}_{0})^{T}\right)^{T}
\Theta(\bm{x}_{0}) \left(\bm{I} - \bm{w}_{1}\bm{v}(\bm{x}_{0})^{T}\right) \Delta \bm{x}
= 0.$$
Derivation of general case {#Q}
==========================
First we examine the projection matrix $\bm{P}$. Consider the outer system $$\frac{d\bm{\xi}}{dt}=\bm{f}(\bm{\xi})\,,\quad \bm{\xi}(0)=\bm{x}\,,$$ where $\bm{x}$ lies close to a point $\bm{z}$ on the manifold. Varying the initial conditions yields $$\label{odedxidx}
\frac{d}{dt}\frac{\partial\xi_i}{\partial x_j}=\frac{\partial}{\partial x_j}f_i(\bm{\xi})=\sum_k \frac{\partial\xi_k}{\partial x_j}\frac{\partial }{\partial \xi_k}f_i(\bm{\xi})=\sum_kJ_{ik}(\bm{\xi})\frac{\partial\xi_k}{\partial x_j}$$ *i.e.* $$\frac{d}{dt}\frac{\partial \bm{\xi}}{\partial \bm{x}}
= \bm{J}(\bm{\xi}(t))\frac{\partial \bm{\xi}}{\partial \bm{x}}$$ where $\bm{J}$ is the Jacobian matrix of $\bm{f}$. Now, since $\bm{\xi}(0)=\bm{x}$, $$\frac{\partial \bm{\xi}}{\partial \bm{x}}(0,\bm{x}) = \bm{I}$$ and thus this variational equation has solution $$\frac{\partial\xi_i}{\partial x_j} = \bm{\Pi}(0,t)$$ where $\bm{\Pi}(s,t)$ is the fundamental matrix solving $$\frac{d}{dt} \bm{\Pi}(s,t) = \bm{J}(\bm{\xi}(\bm{x},t))\bm{\Pi}(s,t), \quad \bm{\Pi}(s,s) = I.$$ When $\bm{x}$ is taken to be $\bm{z} \in \Gamma$, since $\bm{\xi}(\bm{z},t) = \bm{z}$ for all $\bm{z} \in \Gamma$, we have $$\bm{\Pi}(s,t) = e^{(t-s)\bm{J}(\bm{z})}.$$ so, in this case, $\frac{\partial \bm{\pi}}{\partial \bm{x}}(\bm{z},t) = e^{t \bm{J}(\bm{z})}$ (*i.e.* informally, $\frac{d}{dt}\frac{\partial \bm{\xi}}{\partial \bm{x}} \approx \bm{J}(\bm{z})\frac{\partial \bm{\xi}}{\partial \bm{x}}$. Under this approximation the equation is linear and admits the solution $\frac{\partial\bm{\xi}}{\partial \bm{x}}=e^{t\bm{J}(\bm{z})}$).
From the definitions (\[defpi\]) and (\[defPQ\]) we recover $\bm{P}$ by taking the limit of large $t$, $$\bm{P}(\bm{z})=\lim_{t\to\infty}e^{t\bm{J}(\bm{z})}\,.$$ To compute the limit we consider the action of $e^{t\bm{J}(\bm{z})}$ on an eigenvector of the Jacobian[^6]. If $\bm{u}_i$ is tangent to the manifold then the corresponding eigenvalue $\lambda_i$ is zero and so $e^{t\lambda_i}=1$ and $\bm{P}(\bm{z})$ leaves $\bm{u}_i$ unchanged. Alternatively, if $\bm{u}_i$ corresponds to a direction of fast collapse then its eigenvalue is negative and $e^{t\lambda_i}\to0$, so $\bm{u}_i$ is annihilated by $\bm{P}(\bm{z})$.
Let $\bm{U}=(\bm{u}_1,\ldots,\bm{u}_m)$ be a basis of the tangent plane to the manifold at $\bm{z}$ (the slow subspace) and let $\bm{V}=(\bm{v}_1,\ldots,\bm{v}_{m})$ a basis of the orthogonal complement of the fast subspace. Then we may write $$\bm{P}(\bm{z})=\bm{U}(\bm{V}^T\bm{U})^{-1}\bm{V}^T\,.
\label{Puv}$$ In the above we assumed that the tangent plane to the manifold was precisely the kernel of the Jacobian, in which case $\bm{U}$ would be the first $m$ columns of the right eigenvector matrix, and $\bm{V}^T$ the bottom $d-m$ rows of the left eigenvector matrix. This may not hold if the manifold is not hyperbolic (for example if $\bm{f}$ has a component like $-x_i^3$, which is stable but not linearly so), however, equation (\[Puv\]) remains true for all flow fields, provided we somehow have access to bases $\bm{U}$ and $\bm{V}$.
Let us move on to calculate $\bm{Q}$. We start by obtaining some simple identities: first note that by the definition of $\bm{\pi}$, we have $f_{i}(\bm{\pi}(\bm{x})) = 0$ for all $\bm{x}$. Differentiating this, we obtain $$\label{A1}
\sum_{m}\frac{\partial f_{i}}{\partial x_{m}}(\bm{\pi}(\bm{x})) \frac{\partial \pi_m}{\partial x_j} = 0,$$ or, in matrix form, $\bm{J}(\bm{\pi}(\bm{x})) \frac{\partial \bm{\pi}}{\partial \bm{x}} = \bm{0}$. Replacing $\bm{x}$ by $\bm{z} \in \Gamma$, and recalling that $ \frac{\partial \bm{\pi}}{\partial \bm{x}}(\bm{z}) =\bm{P}(\bm{z})$, we have $$\bm{J}(\bm{z})\bm{P}(\bm{z}) = \bm{0},$$ *i.e.* $\bm{J}(\bm{z})$ annihilates all the slow directions, as we have already observed. Differentiating , we obtain $$\sum_{m,n}\frac{\partial^{2} f_{i}}{\partial x_{m}\partial x_{n}}(\bm{\pi}(\bm{x})) \frac{\partial \pi_m}{\partial x_j} \frac{\partial \pi_n}{\partial x_k} + \sum_{m}\frac{\partial f_{i}}{\partial x_{m}}(\bm{\pi}(\bm{x})) \frac{\partial^{2} \pi_m}{\partial x_j \partial x_k} = 0,$$ which we can write in vector form as $$\label{A2}
\bm{\mathcal{H}}_{jk}\left(\frac{\partial \bm{\pi}}{\partial \bm{x}}\right) + \bm{J}(\bm{\pi}(\bm{x})) \frac{\partial^{2} \bm{\pi}}{\partial x_j \partial x_k} = 0,$$ where, for any $n \times n$-matrix $A$, $\bm{\mathcal{H}}_{jk}(\bm{A})$ is the vector with $i$^th^ entry $$\mathcal{H}_{ijk}(\bm{A}) = \bm{e}_{j}^{T}\bm{A}^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}\bm{A}\bm{e}_k,$$ where $\bm{e}_j$ is the $j$^th^ standard basis vector, and we have written $\frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}$ for the Hessian matrix with $j,k$^th^ entry $\frac{\partial^{2} f_{i}}{\partial x_j \partial x_k}$. *i.e.*, since $\frac{\partial \bm{\pi}}{\partial x_j} = \frac{\partial \bm{\pi}}{\partial \bm{x}} \bm{e}_j$, $$\mathcal{H}_{ijk}\left(\frac{\partial \bm{\pi}}{\partial \bm{x}}\right)
= \left(\frac{\partial \bm{\pi}}{\partial x_j}\right)^T \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}\frac{\partial \bm{\pi}}{\partial x_k}
= \sum_{m,n}\frac{\partial^{2} f_{i}}{\partial x_{m}\partial x_{n}}(\bm{\pi}(\bm{x})) \frac{\partial \pi_m}{\partial x_j} \frac{\partial \pi_m}{\partial x_k}.$$
Now, recalling that at $\bm{z} \in \Gamma$, $\bm{\pi}(\bm{z}) = \bm{z}$, $\frac{\partial \bm{\pi}}{\partial \bm{x}} = \bm{P}(\bm{z})$, and $\frac{\partial^{2} \pi_i}{\partial x_j \partial x_k}(\bm{z}) = Q_{ijk}(\bm{z})$, we can write as $$\label{A3}
\bm{J}(\bm{z}) \bm{Q}_{jk}(\bm{z}) = - \bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z})),$$ where we continue with the convention that $\bm{Q}_{jk}(\bm{z})$ is the vector with $i$^th^ entry $Q_{ijk}(\bm{z})$. Applying $\bm{P}(\bm{z})$ to both sides of gives $$\bm{P}(\bm{z})\bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z})) = \bm{0}$$ so we see $\bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z}))$ is entirely contained in the eigenspace of fast directions. Notice that restricted to the fast subspace, $\bm{J}(\bm{z})$ is a full-rank operator, so that, *regarded as an operator on the fast subspace*, has a unique solution, which we will write as $$- \bm{J}(\bm{z})^{+} \bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z})).$$ where we recall that $\bm{J}(\bm{z})^{+}$ is the pseudo-inverse of $\bm{J}(\bm{z})$, which acts as the inverse of $\bm{J}(\bm{z})$ when restricted to the fast directions and which annihilates all vectors in the slow directions. However, regarded as an equation on all of $\mathbb{R}^{d}$, the solution to is not unique, but rather takes the form $$\bm{Q}_{jk}(\bm{z}) = - \bm{J}(\bm{z})^{+} \bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z})) + \bm{S}_{jk}(\bm{z})$$ for some vector $\bm{S}_{jk}(\bm{z})$ in the slow directions.
To obtain $\bm{S}_{jk}(\bm{z})$, we proceed as we did to obtain $\bm{P}(\bm{z})$, differentiating to obtain $$\frac{d}{dt}\frac{\partial^{2}\xi_i}{\partial x_j \partial x_k}
= \sum_l \bm{J}_{il}(\bm{\xi})\frac{\partial^{2}\xi_l}{\partial x_j \partial x_k}
+\sum_{m,n}\frac{\partial^{2} f_{i}}{\partial x_{m}\partial x_{n}}(\bm{\xi})\frac{\partial\xi_m}{\partial x_j}\frac{\partial\xi_n}{\partial x_k}$$ which again write in vector form as $$\label{QPDE}
\frac{d}{dt}\frac{\partial^{2}\bm{\xi}}{\partial x_j \partial x_k}
= \bm{J}(\bm{\xi}) \frac{\partial^{2}\bm{\xi}}{\partial x_j \partial x_k} + \bm{\mathcal{H}}_{jk}\left(\frac{\partial \bm{\xi}}{\partial \bm{x}}\right).$$ This may be formally solved by Duhamel’s principle to give $$\frac{\partial^{2}\bm{\xi}}{\partial x_j \partial x_k} =
\int_{0}^{t} \bm{\Pi}(s,t)\bm{\mathcal{H}}_{jk}\left(\frac{\partial \bm{\xi}}{\partial \bm{x}}(x,s)\right)\, ds$$ where $\bm{\Pi}(s,t)$ is the fundamental matrix from above.
As before, when $\bm{x}$ is taken to be a point $\bm{z} \in \Gamma$, since $\bm{\xi}(\bm{z},t) = \bm{z}$ for all $\bm{z} \in \Gamma$, we have $\bm{\Pi}(s,t) = e^{(t-s)\bm{J}(\bm{z})}$ and $\frac{\partial \bm{\pi}}{\partial \bm{x}}(\bm{z},t) = e^{t \bm{J}(\bm{z})}$, and the solution to simplifies to $$\int_{0}^{t} e^{(t-s) \bm{J}(\bm{z})} \bm{\mathcal{H}}_{jk}\left(e^{s\bm{J}(\bm{z})}\right)\, ds.$$ Thus, $$\bm{Q}_{jk}(\bm{z})
= \lim_{t \to \infty} \frac{\partial^{2}\bm{\xi}}{\partial x_j \partial x_k}
= \lim_{t \to \infty} \int_{0}^{t} e^{(t-s) \bm{J}(\bm{z})} \bm{\mathcal{H}}_{jk}\left(e^{s\bm{J}(\bm{z})}\right)\, ds.$$
Now, $e^{t\bm{J}(\bm{z})} \to\bm{P}(\bm{z})$ as $t \to \infty$, and $$\lim_{t \to \infty} \bm{\mathcal{H}}_{jk}\left(e^{t\bm{J}(\bm{z})}\right)
= \bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z})),$$ both of which are non-zero, so it is not immediately obvious that the integral above converges. However, the information obtained above allows us to resolve these issues. We start by observing that $$\bm{S}_{jk}(\bm{z}) =\bm{P}(\bm{z})\bm{Q}_{jk}(\bm{z}) = \lim_{t \to \infty} \int_{0}^{t}\bm{P}(\bm{z}) e^{(t-s) \bm{J}(\bm{z})} \bm{\mathcal{H}}_{jk}\left(e^{s\bm{J}(\bm{z})}\right)\, ds,$$ and, since $e^{(t-s) \bm{J}(\bm{z})}$ acts like the identity matrix on the slow directions, $\bm{P}(\bm{z}) e^{(t-s) \bm{J}(\bm{z})} =\bm{P}(\bm{z})$, so that $$\bm{S}_{jk}(\bm{z}) = \lim_{t \to \infty} \int_{0}^{t} \bm{P}(\bm{z}) \bm{\mathcal{H}}_{jk}\left(e^{s\bm{J}(\bm{z})}\right)\, ds =\bm{P}(\bm{z}) \int_{0}^{\infty} \bm{\mathcal{H}}_{jk}\left(e^{s\bm{J}(\bm{z})}\right)\, ds.$$ Moreover, we’ve already observed that $\bm{P}(\bm{z})\bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z})) = \bm{0}$, so $$\bm{S}_{jk}(\bm{z}) =\bm{P}(\bm{z}) \int_{0}^{\infty} \bm{\mathcal{H}}_{jk}\left(e^{s\bm{J}(\bm{z})}\right) - \bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z}))\, ds,$$ and we are left with evaluating the integral $$\begin{multlined}
\int_{0}^{\infty} \mathcal{H}_{ijk}\left(e^{s\bm{J}(\bm{z})}\right) - \mathcal{H}_{ijk}(\bm{P}(\bm{z}))\, ds
= \int_{0}^{\infty} \bm{e}_j^T e^{s\bm{J}(\bm{z})^{T}} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z}) e^{s\bm{J}(\bm{z})} \bm{e}_k
- \bm{e}_j^T\bm{P}(\bm{z})^T \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z})\bm{P}(\bm{z}) \bm{e}_k\\
= \bm{e}_j^T \left(\int_{0}^{\infty} e^{s\bm{J}(\bm{z})^{T}} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z}) e^{s\bm{J}(\bm{z})} -\bm{P}(\bm{z})^T \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z})\bm{P}(\bm{z})\,ds\right) \bm{e}_k.
\end{multlined}$$ Now, $$\begin{multlined}
\int_{0}^{\infty} e^{s\bm{J}(\bm{z})^{T}} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z}) e^{s\bm{J}(\bm{z})} -\bm{P}(\bm{z})^T \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z})\bm{P}(\bm{z})\,ds\\
= \int_{0}^{\infty} (e^{s\bm{J}(\bm{z})}-\bm{P}(\bm{z}))^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z}) (e^{s\bm{J}(\bm{z})}-\bm{P}(\bm{z}))\,ds\\
+ \int_{0}^{\infty} (e^{s\bm{J}(\bm{z})}-\bm{P}(\bm{z}))^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z})\bm{P}(\bm{z})\,ds
+ \int_{0}^{\infty}\bm{P}(\bm{z})^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z}) (e^{s\bm{J}(\bm{z})}-\bm{P}(\bm{z}))\,ds,
\end{multlined}$$ and, since $e^{t\bm{J}(\bm{z})}-\bm{P}(\bm{z})$ vanishes on the slow directions, and acts as $e^{t\bm{J}(\bm{z})}$ restricted to the fast directions, $$\int_{0}^{\infty} e^{s\bm{J}(\bm{z})}-\bm{P}(\bm{z})\, ds = -\bm{J}(\bm{z})^{+}$$ whereas $$\label{lyapunovsol}
\bm{X}_i(\bm{z}) = \int_{0}^{\infty} (e^{s\bm{J}(\bm{z})}-\bm{P}(\bm{z}))^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z}) (e^{s\bm{J}(\bm{z})}-\bm{P}(\bm{z}))\,ds$$ is the unique solution to the Lyapunov equation $$\bm{J}(\bm{z})^{T} \bm{X}_i(\bm{z}) + \bm{X}_i(\bm{z}) \bm{J}(\bm{z})
= - \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}(\bm{z})$$ in the fast subspace [@Bellman60]. Thus, $$\bm{S}_{jk}(\bm{z}) =\bm{P}(\bm{z}) \tilde{\bm{S}}_{jk}(\bm{z}),$$ where $$\tilde{S}_{ijk}(\bm{z}) = \bm{e}_j^T \left(\bm{X}_i(\bm{z})
- (\bm{J}(\bm{z})^{+})^{T}\frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}\bm{P}(\bm{z})
- \bm{P}(\bm{z})^{T} \frac{\partial^{2} f_{i}}{\partial \bm{x}^{2}}\bm{J}(\bm{z})^{+}\right) \bm{e}_k$$ and, finally, $$\bm{Q}_{jk}(\bm{z})
= - \bm{J}(\bm{z})^{+} \bm{\mathcal{H}}_{jk}(\bm{P}(\bm{z})) + \bm{P}(\bm{z}) \tilde{\bm{S}}_{jk}(\bm{z}).$$
[^1]: A complementary branch of theory exists dealing with the relaxation of this assumption, see [@Arnold1998; @Roberts2008] for starting points in the literature.
[^2]: Note that we have chosen this case for simplicity of presentation, and not all 1D manifolds can in fact be tackled in this way (e.g. a circular manifold would fail here). The more general case of a manifold described by an arbitrary parameterised curve is not substantially different, however, as we only ever require the local properties of the projection $\pi$, and for smooth manifolds there is always a local coordinate system in which the problem can be set up in the required format.
[^3]: In fact this step is not strictly necessary; we could choose to work directly with the process of particle numbers, as described in Appendix \[Katz\].
[^4]: From the conservation rule $S+C+P=S_0$, we deduce that on the slow manifold we have $P=S_0-S_0z-E_0z/(z+\alpha)$. Use Itô’s lemma to compute $dP/dt$ from (\[MMz\]) and finally undo the coordinate change via $t\mapsto t/k_fE_0$, $z\mapsto S/S_0$.
[^5]: Also *fast variable elimination*, *fast mode elimination*, the *quasi-static approximation*, and many other alternatives.
[^6]: To simplify the discussion we assume that $\bm{J}(\bm{z})$ is diagonalizable and that its kernel contains only the tangent plane to the manifold. Neither assumption is necessary.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Maxence Lef[è]{}vre'
- Aymeric Spiga
- 'S[é]{}bastien Lebonnois'
date: 'Accpeted in Icarus [DOI](https://doi.org/10.1016/j.icarus.2019.07.010)'
title: ' **Mesoscale modeling of Venus’ bow-shape waves** '
---
Abstract {#abstract .unnumbered}
========
The *Akatsuki* instrument LIR measured an unprecedented wave feature at the top of Venusian cloud layer. Stationary bow-shape waves of thousands of kilometers large lasting several Earth days have been observed over the main equatorial mountains. Here we use for the first time a mesoscale model of the Venus’s atmosphere with high-resolution topography and fully coupled interactive radiative transfer computations. Mountain waves resolved by the model form large-scale bow shape waves with an amplitude of about 1.5 K and a size up to several decades of latitude similar to the ones measured by the *Akatsuki* spacecraft. The maximum amplitude of the waves appears in the afternoon due to an increase of the near-surface stability. Propagating vertically the waves encounter two regions of low static stability, the mixed layer between approximately 18 and 30 km and the convective layer between 50 and 55 km. Some part of the wave energy can pass through these regions via wave tunneling. These two layers act as wave filter, especially the deep atmosphere layer. The encounter with these layers generates trapped lee waves propagating horizontally. No stationary waves is resolved at cloud top over the polar regions because of strong circumpolar transient waves, and a thicker deep atmosphere mixed layer that filters most of the mountain waves.
Intro
=====
The influence of the topography on the Venusian atmosphere, especially on the dynamics of the cloud layer, is still not fully understood. The *VeGa* balloons campaign was the first measurement of the impact of the surface on the atmospheric dynamics and demonstrated an increase of the vertical wind above Aphrodite Terra, speculated to be linked to the propagation of orographic gravity waves [@Blam86]. Numerical modeling was then used to assess the possibility for such kind of wave to emerge [@Youn87; @Youn94]. These modeling efforts showed that gravity waves generated by the topography can propagate up to the cloud layer, even in the observed conditions of strong vertical variations of zonal wind and stability. Additional observations of the interactions between the surface and the atmosphere have been made with the *Venus Express* mission [@Bert16]. Correlations between the zonal wind at the top of the cloud and the underlying topography has been evidenced by UV measurements with the Venus Monitoring Camera, and interpreted as the result of stationary gravity waves. A water minimum has also been measured with *Venus Express* instrument SPICAV at cloud top above Aphrodite Terra [@Fedo16] possibly linked to an interaction between the surface and the cloud layer.
Using the LIR instrument on board the *Akatsuki* mission, [@Fuku17] discovered large-scale stationary bow-shaped oscillations at the top of the cloud above Aphrodite Terra. This bow-shape signature, extending over 60 of latitude (about 10 000 km across), was observed during 5 days. Similar signatures were also reported above the main equatorial topographic features [@Kouy17 e.g., Atla and Beta Regio]. Those signatures were interpreted as stationary orographic gravity waves. The bow-shape wave above Aphrodite Terra has been observed in the afternoon, with a maximum amplitude close to the evening terminator. The cloud-top signatures above the other Venusian mountains are visible during the afternoon, with a maximum close to the evening terminator. Cloud-top signatures associated with the underlying topography were also observed above Beta Regio with the instrument *Akatsuki*/IR2 [2.02 $\mu$m wavelength, @Sato17] as well as with *Akatsuki* UV imager above Aphrodite Terra, Atla and Beta Regio [@Kita19]. Stationary features have also been measured by the *Venus Express* instrument VIRTIS above the main topographical obstacles in the southern hemisphere in the nightside with no apparent dependence with the local time [@Pera17].
To investigate the atmospheric dynamics of those observed phenomena, two modeling approaches were adopted: one by [@Fuku17] using a very idealized (i.e. no representation of physical processes and topography) high-resolution Global Circulation Model (GCM), with an arbitrary perturbation as the source of the waves and one by [@Nava18] using a subgrid-scale paramaterization for gravity waves in a low-resolution GCM with complete physics [the IPSL Venus GCM described in @Gara18]. In the study by [@Nava18], the parametric representation of the orographic gravity waves in a GCM demonstrated the link between the topography and the bow-shaped waves. In addition, the impact of the large-scale mountain waves detected by Akatsuki on the global-scale dynamics was shown to induce, along with the thermal tide and baroclinic waves, a significant change in the rotation rate of the solid body, with variability of the order of minutes.
Despite those recent modeling studies, open questions remain on the source, propagation and impact of the large-scale gravity waves evidenced by Akatsuki [@Fuku17; @Kouy17]. Notably, [@Nava18] used a parameterization for orographic gravity waves on Venus and did not resolve the complete emission and propagation of gravity waves from the surface to Venus’ cloud top, which is not possible with current GCMs for Venus. Following the methodology described in [@Lefe18] to study the turbulence and gravity waves in the cloud layers, we propose a method combining high-resolution atmospheric dynamics with detailed physics of the Venusian atmosphere to resolve the emission and propagation of mountain waves in a realistic atmosphere. In other words, in order to address the Venusian bow-shaped mountain waves, we developed the first Venusian mesoscale model. With high-resolution topography from *Magellan* data, we can model the atmospheric dynamics of a limited area of Venus using the dynamical core of the Weather Research and Forecast [WRF, @Skam08]. Following a method used on Mars by [@Spig09], the WRF dynamical core is interfaced with the physics of the IPSL Venus GCM [@Lebo10; @Gara18] to be able to compute radiative transfer and subgrid-scale turbulent mixing coupled in real-time with the dynamical integrations.
In this paper, using our new mesoscale model for Venus, we carry out simulations in areas surrounding the main mountains of the Venusian equatorial belt at various local times, in order to understand the generation and vertical propagation of the waves and to interpret the signal detected by Akatsuki at the top of the cloud. Our new mesoscale model for Venus can potentially be used for many applications other than studying the bow-shaped mountain waves, e.g. near-surface slope winds [@Lebo18], polar meteorology [@Gara15], mesoscale structures in the vicinity of the super-rotating jet [@Hori17].
This paper is organized as follows. Our mesoscale model for Venus is described in Section 2. In Section 3, the emission of resolved gravity waves is presented. The resulting top-of-the-cloud signal above Aphrodite Terra is developed in Section 4, and above Atla Regio and Beta Region in Section 5. In Section 6, mountain waves in the polar regions are discussed. Our conclusions are summarized in Section 7.
The LMD Venus mesoscale model {#Sec:model}
=============================
Dynamical core
--------------
The dynamical core of our LMD Venus mesoscale model is based on the Advanced Research Weather-Weather Research and Forecast (hereafter referred to as WRF) terrestrial model [@Skam08]. This methodology is similar to the one adopted for the LMD Mars mesoscale model (see [@Spig09] for a reference publication of this model and [@Spig18] for the most up-to-date description). The WRF dynamical core integrates the fully compressible non-hydrostatic Navier-Stokes equations on a defined area of the planet. The conservation of the mass, momentum, and entropy is ensured by an explicitly conservative flux-form formulation of the fundamental equations [@Skam08], based on mass-coupled meteorological variables (winds and potential temperature). To ensure the stability of the model, and given the typical horizontal scales aimed at in our studies, the fundamental equations are integrated under the hydrostatic approximation, available as a runtime option in WRF.
Coupling with complete physical packages for Venus
--------------------------------------------------
The radiative heating rates, solar and IR, are calculated using the IPSL Venus GCM radiative transfer scheme [@Lebo15]. This setting is similar to the “online” mode described in [@Lefe18]. The time step ratio between the dynamical and physical integrations is set to 250, as a trade-off between computational efficiency and the requirement that the physical timestep is significantly less than the typical radiative timescale for Venus.
The version of the Venus radiative transfer model used in our mesoscale model is the same as in the version of the IPSL Venus GCM described in [@Gara18]. The infrared (IR) transfer uses [@Eyme09] net-exchange rate (NER) formalism: the exchanges of energy between the layers are computed before the dynamical simulations are carried out by separating temperature-independent coefficients from the temperature-dependent Planck functions of the different layers. These temperature-independent coefficients are then used in the mesoscale simulations to compute the infrared cooling rates of each layer. The solar rates are based on [@Haus15] computations: look-up tables of vertical profiles of the solar heating rate as a function of solar zenith angle are read, before being interpolated on the vertical grid adopted for mesoscale integrations.
The cloud model used in this study is based on the [@Haus14] and [@Haus15] models derived from retrievals carried out with *Venus Express* instruments. The cloud structure is latitude-dependent, the cloud top varies from 71 km to 62 km between the Equator and the poles [@Haus14]. This latitudinal variation of the cloud takes the form of five distinct latitude intervals: 0$^{\circ}$ to 50$^{\circ}$, 50$^{\circ}$ to 60$^{\circ}$, 60$^{\circ}$ to 70$^{\circ}$, 70$^{\circ}$ to 80$^{\circ}$ and 80$^{\circ}$ to 90$^{\circ}$. For each latitudinal intervals different NER-coefficient matrices are computed on the vertical levels of the model, ranging from the surface to roughly 100 km altitude. The cloud structure used for the calculations is fixed prior to the simulation and does not evolve with time.
Since the horizontal grid spacing for the mesoscale simulations is set to several tens of kilometer, the convective turbulence in the planetary boundary layer and the cloud layer is not resolved. Therefore, similarly to what is done in GCMs, our mesoscale model uses subgrid-scale parameterizations for turbulent mixing. As in the IPSL Venus GCM, for mixing by smaller-scale turbulent eddies, the formalism of [@Mell82] is adopted, which calculates with a prognostic equation the turbulent kinetic energy and mixing length. For mixing by larger-scale turbulent plumes [such as those resolved by Large-Eddy Simulations, see @Lefe17; @Lefe18], a simple dry convective adjustment is used to compute mixed layers in situations of convectively-unstable temperature profiles. In the physics (same as in the GCM), the dependency of the heat capacity with temperature $c_p(T)$ is taken into account, but in the dynamical core, we use a constant heat capacity as in @Lefe18, with a mean value of $c_p = 1000$ J kg$^{-1}$ K$^{-1}$ suitable for the vertical extent of our model from the surface to 100 km altitude. The initial state and the boundary conditions of the mesoscale domain, detailed in the next section, are calculated using a variable $c_p$ to ensure a realistic forcing of the mesoscale model by the large-scale dynamics. Therefore the impact of the constant value of $c_p$ is minimum in our configuration because the equations of the WRF dynamical core are formulated in potential temperature and the conversion from the dynamics to the physics (and vice versa) between potential temperature and temperature is done with a variable $c_p (T)$.
Simulation settings
-------------------
The topography used in our mesoscale model is presented in Figure \[fig:231\], it is based on *Magellan* data [@Ford92] for the majority of the surface, with *Pioneer Venus* data used to fill the blank spots wherever needed [@Pette80]. The dataset has a resolution of 8192 points along the longitude and 4096 points along the latitude. Given the detection of bow-shaped features by the *Akatsuki* spacecraft [@Kouy17], we choose three domains of interest : Aphrodite Terra, Atla Regio and Beta Regio. These three mesoscale domains has been computed with a horizontal resolution of 40 km for Aphrodite Terra, 30 km for Beta Region and 15 km for Atla Regio. Given the horizontal resolutions involved, the dynamical timestep for mesoscale integrations is set between 8 and 10 s [typical for those horizontal resolutions, see @Spig09]. The horizontal domains and resolutions have been chosen to enclose the whole latitudinal extent of the bow-shaped waves, while keeping a feasible computing time. The vertical grid is composed of 150 levels from the ground to 100 km, with a similar distribution as the LMD Venus LES model [@Lefe18].
![Topographic map of Venus used for the mesoscale simulations, from *Magellan* [@Ford92] and *Pioneer Venus* [@Pette80] datasets. The different domains used in the mesoscale simulations are indicated : Aphrodite Terra, Beta Regio and Atla Terra.[]{data-label="fig:231"}](./Figure1.png){width="10cm"}
The horizontal boundary conditions have been chosen as ‘specified’, i.e. meteorological fields are extracted from IPSL Venus GCM simulations [detailed in @Gara18] and interpolated both on the vertical grid, accounting for the refined topography of the mesoscale, and on the temporal dimension, accounting for the evolution of those fields over the low dynamical timestep in mesoscale integrations. This ensures that the planetary-scale super-rotation of the Venusian atmosphere is well represented. The mesoscale simulations presented in this study are performed using an update frequency of a 1/100 Venus day, which enables a correct representation of the large-scale variability simulated by the GCM at the boundaries of the mesoscale domain. Between the mesoscale domain and the specified boundary fields, a relaxation zone is implemented in order to allow for the development of the mesoscale circulations inside the domain, while keeping prescribed GCM fields at the boundaries [@Skam08]. In this study, the number of relaxation grid points is set to 5. An exponential function is used to reach a smooth transition between the domain and the specified boundary fields; the coefficient is set to 1, a value used for terrestrial and Martian applications that appeared suitable for the Venusian case as well. At the top of the mesoscale model, a diffusive Rayleigh damping layer is applied to avoid any spurious reflection of vertically-propagating gravity waves. The damping coefficient is set to 0.01 and the depth of the damping layer to 5 km.
The initialization of the meteorological fields for the mesoscale domain are, similarly to boundary conditions, extracted from the [@Gara18] GCM. First the fields are interpolated horizontally to the mesoscale refined-resolution grid, then interpolated vertically on the mesoscale vertical grid – accounting for high-resolution topography. To extrapolate the high-resolution features, the same methodology is used as in the LMD Mars Mesoscale Model [@Spig09].
Venus’s rotation is retrograde, but the WRF dynamical core has been built for Earth applications, and therefore assuming implicitly prograde rotation. In order to solve this issue, as is the case for the IPSL GCM runs [@Lebo10], the GCM fields are turned upside down and the meridional wind is multiplied by minus 1. At the post-processing stage, the fields are turned upside down again and the meridional wind is again multiplied by minus 1, to obtain the diagnostics enclosed in this paper.
Generation and propagation of orographic gravity waves {#Sec:gwg}
======================================================
To discuss the orographic wave generation, we focus first on Atla Regio which facilitates the visualization of the phenomena given its very sharp mountains. Figure \[fig:31\] shows the domain chosen for Atla Regio.
![Elevation map (km) of the selected domain for Atla Regio with a resolution of 15 km.[]{data-label="fig:31"}](./Figure2.png){width="10cm"}
Figure \[fig:32\] shows a vertical cross-section of the vertical wind (m s$^{-1}$) at 1 of latitude, between the surface and approximately 55 km in the beginning of afternoon. Contours represent potential temperature. For this figure \[fig:32\], and all the similar cross-sections shown in this paper, the zonal wind comes for the right side of the plot, i.e. wind is blowing westward. The Pioneer Venus probe measured horizontal wind (u and v) from the ground to 3 km between 0.1 and 1.5 m s$^{-1}$ [@Schu80]. In the GCM, the horizontal wind on the same vertical extent is between 0.1 and about 3.0 m s$^{-1}$. The horizontal wind is of the same order of magnitude as in-situ measurements although slightly overestimated.
Strong vertical winds are visible above the two main mountains, Maat Mons at -165 of longitude and Ozza Mons at -156 of longitude. The interaction between the incoming zonal flow and those sharp topographical obstacles leads to the generation of gravity waves. Near the surface, the value of the dimensionless mountain height $H_d = H \, N_B / u$ can reach value close to 1, where $H$ is the maximum mountain height, $N_B$ is Brunt-Väisälä frequency and $u$ the zonal wind, meaning that the flow is in a non-linear regime [@Durr03]. The waves propagate vertically and encounter two regions of low static stability, the mixed layer between 18 and 35 km and the convective layer between 48 and 52 km. Above, the waves propagate into the stratified layers. The vertical wavelength of the waves is around 30 km.
![Vertical cross-section of the vertical wind (m s$^{-1}$) at 1 of latitude, between the surface and approximately 55 km in the beginning of afternoon. Contours represent the potential temperature. Direction of the zonal is indicated with the black arrow.[]{data-label="fig:32"}](./Figure3.png){width="10cm"}
The impact of these two near-neutral low-stability regions on the vertical propagation of the wave can be understood via the Scorer parameter (km$^{-1}$) [@Scor49] : $l^2 = \frac{{N_B}^2}{u^2} - \frac{1}{u}\frac{d^2 u}{d z^2}$. This parameter represents the minimum vertical wavelength of propagation. Figure \[fig:33\] shows the vertical cross-section of the Scorer parameter, and its domain-averaged vertical profile, at the same location as Figure. \[fig:32\].
![Left : Instantaneous vertical cross-section of the Scorer parameter (km$^{-1}$) at 1 of latitude, between the surface and approximately 55 km in the beginning of afternoon. Contours represent the potential temperature. Right : domain-averaged vertical profile of the Scorer parameter at the same local time. Direction of the zonal is indicated with the white arrow.[]{data-label="fig:33"}](./Figure4_1.png "fig:"){width="8cm"} ![Left : Instantaneous vertical cross-section of the Scorer parameter (km$^{-1}$) at 1 of latitude, between the surface and approximately 55 km in the beginning of afternoon. Contours represent the potential temperature. Right : domain-averaged vertical profile of the Scorer parameter at the same local time. Direction of the zonal is indicated with the white arrow.[]{data-label="fig:33"}](./Figure4_2.png "fig:"){width="8cm"}
Beyond 12 km altitude, the Scorer parameter decreases strongly up to 18 km and has very small values, sometimes negative, from 18 to 35 km. Then it increases up to 42 km, and decreases until reaching the convective layer at 48 km. Inside this convective layer, the Scorer parameter is again very small. Above the convective layer, the Scorer parameter increases through the stable atmosphere. These two regions of low Scorer parameter indicate the predominance of trapped lee waves that propagate horizontally. The horizontally-propagating trapped waves are visible in Figure \[fig:32\] within layers exhibiting low Scorer parameter, between 18 and 35 km and between 48 and around 54 km. An additional vertical propagation of those trapped waves is visible at some longitudes. This phenomenon is similar to the leakage of trapped lee waves into the stratosphere on Earth [@Brow83]. On Venus, the horizontal wavelength is approximately 150 km, one order of magnitude larger to typical Earth’s trapped lee waves [@Ralp97]. Large horizontal wavelengths are known to increase the vertical leakage of trapped lee waves [@Durr15]. The horizontal wavelength of the trapped lee waves generated above Atla Regio is consistent with VEx/VIRTIS nightside measurements in the low cloud region [@Pera08], whereas the trapped lee waves generated above Aphrodite Terra and Beta Regio are larger by at least a factor of 2.
One of the main questions about the vertical propagation of bow-shaped waves on Venus (and the reason why they can be detected at high altitudes by Akatsuki) is how the waves can propagate through two (near-neutral) mixed layers. A similar configuration can be found in the Earth’s oceans where the seasonal and main thermoclines can be separated by a relatively weakly stratified region [@Ecka61]. An analogous case in the Earth atmosphere is the propagation of gravity waves from the troposphere to the upper mesosphere and lower thermosphere, tunneling through an evanescent region [@Walt01]. Gravity waves are evanescent in neutral-stability layers, in which the energy of the waves decrease exponentially with altitude. However, if the vertical extension of the neutral region is not too large, it is possible for a significant fraction of the incident wave energy to pass through the neutral stability region. By analogy with quantum mechanics, some transfer of energy is possible via wave tunneling [@Suth04] through a neutral barrier. To quantify the energy transmitted from a region with a Brunt-Väisälä frequency $N$ through a barrier of zero Brunt-Väisälä frequency, the transmission $T$ is defined as
$$T = \left[ 1 + \left( \frac{\sinh(k_x H )}{\sin2\theta} \right)^2 \right]^{-1}$$
according to @Suth04, with $k_x$ the horizontal wavenumber, $H$ the height of the barrier and $\theta$ equals to $\cos(\omega / N)$ with $\omega$ the frequency of the wave.
For the first barrier, between 18 and 35 km, the horizontal wavelength is of the order of 200 km, $H$ is equal to 17 km, $\omega$ to 3.4 10$^{-4}$ s$^{-1}$ and $N$ to 2.3 10$^{-3}$ s$^{-1}$. For the second barrier, the convective layer between 48 and 52 km, the horizontal wavelength is of the order of 200 km, $H$ is equal to 4 km, $\omega$ to 1.3 10$^{-3}$ s$^{-1}$ and $N$ to 9.0 10$^{-3}$ s$^{-1}$. With these values, the transmission $T$ for the first layer is equal to 21 % and to 84 % for the second barrier. The cloud convective layer depth is closer to 10 km [@Tell09], with that value the transmission drops to about 45 %.
We conclude that, despite the presence of the two neutral-stability layers in the atmosphere of Venus, there is a significant energy transmission through these two barriers. In other words, this tunneling effect enables the orographic gravity waves to propagate towards the top Venusian clouds, where those waves are detected by Akatsuki [@Fuku17; @Kouy17].
The mixed layer in the deep atmosphere between altitudes 18 and 35 km [@Schu80] is the thickest of the two barriers, and is the one that affects the most the vertically-propagating wave, since a fifth of the wave energy makes it through this mixed layer. Our simulations thereby provide insights into the mechanisms responsible for the propagation of the bow-shape perturbation from the surface to 35 km, left unexplained by the approach based on gravity-wave parameterization used in @Nava18. Conversely, the cloud convective layer between altitudes 48 and 52 km, thinner than the deep atmosphere mixed layer but with strong vertical plumes [@Lefe18], plays a minor role and does not affect substantially the wave vertical propagation (more than 80% of the incoming wave energy is propagating through this mixed layer). These two neutral barriers would also be an obstacle to the vertical propagation of sound waves [@Mart18] induced by putative Venusian seismic activity, possibly detectable via infrasound sensors on balloons [@Kris18] or through airglow radiation perturbation [@Didi18].
Saturation of a wave occurs either through critical levels (when the speed of the background horizontal flow is equal to the wave phase speed) or wave breaking through convective instability. To quantify this probability, the saturation index $S$ of [@Hauc87] is used
$$S = \sqrt{\frac{F_0 N}{\rho k_x |\overline{u} - c|^3}}$$
where $F_0$ is the vertical momentum flux, $N$ is the Brunt-Väisälä frequency, $\rho$ the density of the atmosphere, $k_x$ the horizontal wave number $c$ the phase speed of the wave, and $\overline{\cdot}$ the average over our chosen mesoscale domain (thought to represent the large-scale component). When $S$ reaches values close to 1 the wave may be likely to break through critical level or saturation. For the case of Venus, the main wave is orographic and stationary with respect to the surface, with a phase speed $c$ equals to zero. The zonal wind speed is constantly increasing up to roughly 80 km of altitude, thus $|\overline{u} - c| \gg 0$ and superior to $F_0 N /k_x$, which causes $S$ to be much smaller than $1$, even for the largest values of stability $N$. The saturation index only reaches values up to 1 10 $^{-1}$, and most often values of 1 10 $^{-2}$, thus the probability of saturation of the mountain waves is low.
![Instantaneous vertical cross-section of the quantity $- \rho {u' w'}$ (Pa) at 1 of latitude, between the surface and approximately 55 km in the beginning of afternoon. Contours represent the potential temperature. Direction of the zonal is indicated with the black arrow.[]{data-label="fig:34"}](./Figure5.png){width="10cm"}
The propagation of the waves into the two mixed layers engender a dissipation of the waves: the part of the energy that is not transferred upward through tunneling effect is deposited in those mixed layers. This deposition of momentum can be quantified by calculating the vertical component of the Eliassen-Palm momentum flux [@Andr87] that is equal to $- \rho \overline{u' w'}$ where $\rho$ is the density and $u'$ and $w'$ the perturbations of the zonal wind $u$ and vertical wind $w$ (with respect to the mean defined as $\overline{\cdot}$). A postive value of $u'$ represents an eastward velocity perturbation and a positive value of $w'$ represents an upward velocity perturbation and therefore a positive value of $- \rho \overline{u' w'}$ represents an upward transport of westward momentum. The quantity displayed at Figure \[fig:34\] is an instantaneous snapshot of the quantity $- \rho {u' w'}$. This flux is around 10 times larger than the strongest measured on Earth in Antarctica [@Jewt15]. This is partly due to the fact that the density of the Venusian atmosphere is 65 times larger than the Earth atmosphere. Momentum is transported by the waves near the surface (where density is the highest), but also in the mixed layer around 30 km of altitude (where wave perturbations are strong). The trapped lee waves also transport momentum horizontally, while the vertical transport of momentum due to leakage is negligible. The momentum flux $-\overline{\rho u' w'}$ is the drag coefficient [@Smit79] calculated in sub-grid-scale parameterizations, such as the one used in the GCM runs by @Nava18, who employed the orographic-wave parameterization of @Lott97. In this parameterization, @Nava18 had to adopt values of 2 Pa for the threshold of the gravity wave mountain stress, and 35 km for the initial altitude of deposition of this mountain stress, to reproduce the bow-shape waves observed by Akatsuki. Our mesoscale modeling of the Venusian mountain waves, in which the propagation of gravity waves is resolved from the surface to 100 km, therefore validates the assumption of the @Nava18 study, both for the amplitude of the stress and the altitude of the stress deposition (a maximum of momentum is deposited by the waves at this altitude), and shows that they are physically based.
![Vertical cross-section of the vertical wind (m s$^{-1}$) on top panel and vertical cross-section of the Scorer parameter (km$^{-1}$) on middle panel and vertical cross-section of vertical momentum flux (Pa) on bottom panel for Aphrodite Terra on left column and Beta Regio on right column. Contour is potential temperature. Direction of the zonal is indicated with arrows.[]{data-label="fig:35"}](./Figure6.png){width="14cm"}
What we describe here for Alta Regio extends to the other regions considered in this study. Figure \[fig:35\] shows the vertical cross-section of the vertical wind, Scorer parameter and momentum flux for Aphrodite Terra (left) and Beta Regio (right). As for Atla Regio, the major terrain elevations generate gravity waves and the flow is in a non-linear regime. The two layers of low static stability are also present in those regions, which entails the generation of trapped lee waves. The amplitude of the vertical wind is smaller than in the Atla case due to lower slopes. Thus, the amplitude of the momentum flux is of the same order of magnitude, albeit slightly smaller. The vertical extent of the two mixed layers is very similar to Atla Region, in both the Aphrodite Terra and Beta Regio cases, but the horizontal wavelengths are slightly larger and resulting in a small decrease of the tunneling transmission $T$ but in the same order of magnitude, about 20 % for the first barrier (18-35 km altitudes) and 80 % for the second barrier (48-52 km altitudes).
Aphrodite Terra {#Sec:AT}
===============
Figure \[fig:411\] shows the selected domain for Aphrodite Terra with a resolution of 40 km, Ovda Terra is visible between 60 and 100 of longitude where the main bow-shape gravity waves have been observed by *Akatsuki*.
![Elevation map (km) of the selected domain for Aphrodite Terra with a resolution of 40 km.[]{data-label="fig:411"}](./Figure7.png){width="10cm"}
*Akatsuki* LIR measurements consist of a mean of the brightness temperature between 8 and 12 $\mu$m to which is applied a high-pass filter [@Fuku17; @Kouy17; @Tagu07]. To be able to compare the outputs of the model with the observations, we have to analyze in the model the deformation of the cloud top by the gravity waves. We consider that the characteristic timescale of the wave propagation is smaller than the radiative timescale, so that the deformation is adiabatic; it follows that potential temperature can be used as a tracer for the deformation of the cloud top by waves. By choosing one value for potential temperature to set a relevant material surface, a corresponding temperature map can be reconstructed, as well as a map of the corresponding altitude. To compare the waves resolved by the model to the ones observed by the *Akatsuki* spacecraft, anomaly temperature is calculated for potential temperature surface between 55 and 80 km, corresponding to a range of potential temperature from 800 to 1300 K, with a gaussian weighting function mimicking the LIR’s weight function. Then the temperature perturbations are vertically averaged. A high-pass gaussian filter is applied to filter out the global dynamics component [similarly to what is done in the maps derived from Akatsuki observations @Fuku17; @Kouy17], and a low-pass gaussian filter is also applied to remove the small-scale features smaller than a few tens of kilometer (i.e. approaching the effective horizontal resolution of our mesoscale simulations for Venus). Figure \[fig:422\] presents the associated residuals of the temperature anomaly after filtering. In the end, temperature maps such as Figure \[fig:421\] obtained from our mesoscale modeling by this method can be directly compared to the *Akatsuki* observations in @Fuku17 and @Kouy17.
Resulting bow-shape wave
------------------------
Figure \[fig:421\] shows the temperature anomaly modeled by our LMD Venus mesoscale model at the top of the cloud above Aphrodite Terra, close to the evening terminator. Cyan contours show the topography. The topography-induced perturbation at the top of the cloud expands from -40 to 40 of latitude with a bow-shaped morphology, resembling the measured signal by Akatsuki. Above Ovda Terra, the positive and negative anomaly measured by LIR [@Fuku17] is reproduced with a similar amplitude, 2 K. Outside of the bow-shape wave small-scale waves features are visible, similar structures are visible in the UV images [@Kita19]. This orographic gravity wave induces a deformation of about 600 m of the cloud top altitude. Temperature anomalies induced by mountain waves are visible in the middle and upper cloud layers whereas no temperature anomaly is discernible in the lower cloud layer due to the presence of the convective layer. This could explain the fact that VEx/VIRTIS measured on the nightside stationary waves in the upper cloud layer but not in the lower cloud layer [@Pera17].
![Left : Temperature anomaly (K) at the top of the cloud in late afternoon. Cyan contours show the topography every 1 km (Fig. \[fig:411\]). Direction of the zonal is indicated with the white arrow. Right : Temperature anomaly (K) at the top of the cloud observed by *Akatsuki*/LIR adapted from [@Fuku17]. White and cyan lines are topography.[]{data-label="fig:421"}](./Figure8.png){width="16cm"}
The divergence of the momentum flux indicates the acceleration (or equivalently the force per unit mass) caused by gravity waves on the mean flow when they break or encounter a critical level [@Frit03]. This acceleration can be calculated by the equation $$\frac{\partial \overline{u}}{\partial t} = \frac{1}{\rho}\frac{\partial}{\partial z} \rho \overline{u' w'}$$ Observations indicate that around cloud top, for altitudes above 67 km, there is a deceleration of the zonal wind.
[@Bert16] calculated that an deceleration of 13 m s$^{-1}$ per Venus day was necessary to explain longitudinal shift of zonal wind patterns. The *Akatsuki* spacecraft measured a longitudinal variability of the zonal wind around cloud top between 4 and 12 m s$^{-1}$ at the Equator [@Hori18] with no correlation with topography. The deceleration induced by the bow-shape waves resolved by our LMD Venus mesoscale model, integrated over a Venus day, reaches values around 3 m s$^{-1}$ consistent with the lower range of estimates based on *Akatsuki* measurements. This value is smaller than the computations in @Bert16 (and the higher range of Akatsuki estimates), which tends to indicate that the deceleration of zonal wind at cloud top cannot be explained solely by the impact of bow-shaped gravity waves.
![The scene displayed in this figure is similar to Figure \[fig:421\]. The low-pass filter signal (left panel) representing the global dynamics component has a maximum around 160 of longitude, where the local time is close to noon. The high-pass filter signal (right panel) shows small-scale wave features not visible on Akatsuki LIR images [@Kouy17] but discernible on UV images [@Kita19]. Cyan contours show the topography (Fig. \[fig:411\]).[]{data-label="fig:422"}](./Figure9.png){width="16cm"}
Variability with local time
---------------------------
The temperature anomaly of 2 K at the top of the cloud shown in Figure\[fig:421\] is the maximum signal obtained. This wave is visible in the model with an amplitude superior to 0.5 K for about 10 Earth days, against approximately 14 Earth days for the *Akatsuki* observation (between 14h et 17h in local time). Figure \[fig:431\] shows Temperature anomaly (K) at the top of the cloud 3.5 Earth day after Figure \[fig:421\], the bow-shape wave above Ovda Regio is still visible but with smaller amplitude meanwhile a bow-shape wave above Thetis Regio is discernible similarly to LIR observations [@Kouy17].
![Temperature anomaly (K) at the top of the cloud 3.5 Earth day after Figure \[fig:421\]. Cyan contours show the topography every 1 km (Fig. \[fig:411\]). Direction of the zonal wind is indicated with the white arrow.[]{data-label="fig:431"}](./Figure10.png){width="10cm"}
Additional simulations were performed at midnight and noon to study the variability of the waves at cloud top. At midnight, no significant wave (amplitude superior to 0.5 K) is observed. At noon, transient waves are visible with amplitude smaller than 1 K. The variation of the surface wind along the day, about 1 m s$^{-1}$ above Ovda Terra, is too small to explain this variability. This indicates that the near-surface conditions responsible for the emission of orographic gravity wave are not changing significantly along the day. The diurnal variability of conditions responsible for the propagation of orographic gravity waves towards high altitude is more likely to explain the observed variability with local time of the bow-shaped waves.
![Square of the Brunt-Väisälä frequency (s$^{-2}$) of the atmosphere from the surface to 6 km for midnight (left), noon (middle) and late afternoon (right) for Aphrodite Terra.[]{data-label="fig:432"}](./Figure11.png){width="16cm"}
Figure \[fig:432\] shows the square of the Brunt-Väisälä frequency (s$^{-2}$) above Aphrodite Terra between the surface and 6 km of altitude for the three local times : midnight (left), noon (center), late afternoon (right). Close to the surface the atmosphere is very stable, but this stability is decreasing with altitude by an order of magnitude. This decrease is different depending on the local time. At midnight, the static stability shows a strong gradient with altitude. At noon this gradient is smaller, and even smaller in late afternoon with constant value of 3 10$^{-6}$ s$^{-2}$ over more than one kilometer. This slow decrease has consequences over the Scorer parameter, it is superior to 2 km in the first 8 km in late afternoon, against only in the first 3 km at other local times. This strong value of the Scorer parameter over several kilometers favors the vertical propagation of the gravity waves. This is correlated to the vertical momentum flux: at night the amplitude is about 10$^{-2}$ Pa, while it is about 10$^{-1}$ Pa at noon. Meanwhile, the very-near surface (first hundred meters) stability is consistent with [@Nava18] study : the atmosphere is more stable at night, but with a smaller local-time variability. This behavior plays a role in the generation of the waves, while the first 4 to 5 km affect the vertical propagation.
Atla and Beta Regios
====================
Atla Regio
----------
The left panel of Figure \[fig:511\] shows the resulting temperature anomaly at cloud top for the Atla Terra (Fig. \[fig:31\]) simulation, using the same methodology as for Aphrodite Terra. Two waves are visible, one above the main mountain Maat Mons at -168 of longitude and an other over Ozza Mons at -163 of longitude. The amplitude of these waves is of about 1.5 K for the first one and about 1 K for the last one. The latitudinal expansion is about 20.
LIR observed one main wave above Ozza Mons with an amplitude of 2 K and an extension of 30 as well as a second wave of a few degrees of latitude associated with Maat Mons [@Kouy17]. The main resolved wave is similar to the observations, with however sharp morphology that may be due to the abrupt terrain elevation; the second wave is more extended in latitude than in the observations. These waves are obtained for the beginning of the afternoon when the amplitude is maximum, whereas LIR observed the maximum in the middle of the afternoon. The waves resolved by the model are symmetrical to the obstacle, meaning the latitudinal expansion towards the north and towards the south is the same, while LIR observed non-symmetrical waves attributed to the complex terrain.
![Left : Temperature anomaly (K) at the top of the cloud in the afternoon. Cyan contours show topography every 1.6 km (Fig. \[fig:31\]). Direction of the zonal is indicated with the white arrow. Right : Temperature anomaly (K) at the top of the cloud observed by *Akatsuki*/LIR adapted from [@Kouy17]. White and cyan lines are the topography. Blue lines on the right panel show the mesoscale domain.[]{data-label="fig:511"}](./Figure12.png){width="16cm"}
The generation and propagation of the waves are sensitive to the wind and stability conditions close to the surface and therefore to physical representations within the model, for example the radiative transfer. The Venus GCM may not represent fully accurate surface conditions, especially for winds, as well as their evolution with time. These differences could explain that, while the overall bow-shaped wave is reproduced in our Venus mesoscale simulation, some discrepancies with the observations do exist.
Beta Regio
----------
![Elevation map (km) of the selected domain for Beta Regio with a resolution of 30 km.[]{data-label="fig:521"}](./Figure13.png){width="10cm"}
Figure \[fig:521\] shows the selected domain Beta Regio with a resolution of 30 km, where the highest elevation point is Theia Mons at -80 of longitude and 30 of latitude. The temperature anomaly at the top of the cloud is shown in left panel of Figure \[fig:522\]. Two distinct stationary waves are visible, a main one above Theia Mons and a smaller one above Polik-mana Mons at 30 of latitude and -100 . These two waves are observed by *Akatsuki*/LIR with the same relative size visible in the right panel of Figure \[fig:522\]. The bow shape is visible with an amplitude of 1.5 K. A difference with previous anomaly is the asymmetry of the waves, which is more developed towards the North. This non-symmetrical shape is not clearly observed above Beta Regio in LIR images [@Kouy17] but visible in UV images [@Kita19]. The latitude of the region, 30, may have an impact on the morphology of the wave.
![Left : Temperature anomaly (K) at the top of the cloud in late afternoon. Cyan contours show topography every 850 m (Fig. \[fig:521\]). Direction of the zonal is indicated with the white arrow. Right : Temperature anomaly (K) at the top of the cloud observed by *Akatsuki*/LIR adapted from [@Kouy17]. White and cyan lines are the topography.[]{data-label="fig:522"}](./Figure14.png){width="16cm"}
Polar regions
=============
The highest topographical point is Maxwell Montes peaking at approximately 10 km at 65N latitude. This is a site of interest in orographic gravity waves on Venus. Unfortunately, the orbit of *Akatsuki* is equatorial and therefore the spacecraft cannot observe the atmospheric activity in polar regions; *Venus Express* had an elliptical polar orbit suited to study the south pole, but not the north pole. To study the generation of orographic waves above Maxwell Montes, and more generally in the polar regions, mesoscale simulations can be performed with polar stereographic projection over both poles for comparison. The chosen domains for North (right) and South pole (left) are shown in Figure \[fig:61\], with a resolution of 40 km for both domains. Recent improvements of the IPSL Venus GCM used to provide initial and boundary for our Venus mesoscale model, especially in the polar regions, ensure a correct large-scale forcing with the inclusion of the cold collar and mid-to-high-latitude jets [@Gara18].
![Elevation map (km) of the selected domains for the North pole (left) and the South pole (right) regions with a resolution of 40 km.[]{data-label="fig:61"}](./Figure15.png){width="16cm"}
Figure \[fig:62\] shows the temperature anomaly at the top of the cloud. Transient waves rotating around the planet are visible in both polar regions down to $\pm$ 60 with an amplitude larger than 5 K. The wave structure and amplitude are similar for both regions meaning that the underlying topography has hardly any impact on the temperature anomaly at cloud top. These waves resemble the planetary streak structure visible in the *Akatsuki*/IR2 images [@Lima18] and reproduced with GCM modelling [@Kash19].
![Temperature anomaly (K) at the top of the cloud for the North pole (left) and the South pole (right) regions. Black contours show topography every 1.6 km for the North pole and every 800 m for the South pole (Fig. \[fig:61\]).[]{data-label="fig:62"}](./Figure16.png){width="16cm"}
However, mountain waves are generated by Ishtar Terra. The vertical velocity above Maxwell Montes is shown in Figure \[fig:63\]. These waves propagate vertically and encounter the two low static stability regions but do not yield any stationary temperature anomaly at cloud-top. The two barriers are thicker than that for the other cases, 22 against 17 km for the first one and 7 against 4 km for the convective layer. The increase of the cloud convective layer thickness is expected from observations [@Tell09; @Haus14] and modelling [@Imam14; @Lefe18], yet few observations have been performed regarding the mixed layer in the deep atmosphere, especially regarding the variation of depth with latitude and local time. Only the temperature profiles of the Pioneer Venus probes [@Seiff80] would be available for such an analysis.
As is discussed in section \[Sec:gwg\], the two mixed layers in the lower atmosphere of Venus act as barriers for gravity waves. For the first barrier (altitudes 15-35 km) with $\omega$ equals to 2.3 10$^{-4}$ s$^{-1}$, $N$ to 1.6 10$^{-3}$ s$^{-1}$ and a horizontal wavelength of 200 km, the transmission is only of 13 %. For the second barrier (altitudes 45-55 km) with $\omega$ equals to 1.1 10$^{-3}$ s$^{-1}$, $N$ to 7.5 10$^{-3}$ s$^{-1}$ and the same horizontal wavelength, the transmission drops to 63 %. The orographic wave is therefore strongly affected by propagating through the two mixed layers (especially the lowermost one). As a result, the cloud top temperature anomaly induced by gravity waves in polar regions is smaller than the ones in the equatorial regions, and negligible against the cloud top transient waves. The large amplitude of the transient polar waves could prevent the visibility of the comparatively-low stationary gravity-wave perturbation and explain the non-observation of stationary waves in *Venus Express*/VIRTIS data at latitudes above 65S [@Pera17]. The saturation of the waves is here again unlikely, so contrary to mixed layers acting as efficient barriers, it does not seem a plausible mechanism to explain the lack of observed gravity-wave stationary wave in polar regions.
![Vertical cross-section of the vertical wind (m s$^{-1}$) between the surface and approximately 55 km. Please note that the cross-section is not at a constant latitude. Contours represent the potential temperature.[]{data-label="fig:63"}](./Figure17.png){width="10cm"}
Conclusion
==========
We present here the first mesocale model applied to Venus, in which the WRF dynamical core is coupled the set of physical parameterizations developed for the IPSL Venus GCM, notably the radiative transfer to compute the solar and IR rates during the simulations. Focusing on three areas of interest, the model resolves bow shape stationary waves with an amplitude around 1.5 K. The lifetime of the waves is about ten Earth days. The maximum amplitude of the waves are in the afternoon, close to the terminator for Aphrodite Terra, earlier for the two another cases. The characteristics and morphology of the waves observed by Akatsuki are overall well reproduced by the model.
Theses gravity waves are generated by the large-scale flow forced to go over Venus’ mountains. Propagating vertically, those orographic gravity waves encounter two layers of low static stability, the mixed layer (20-32 km) and the convective layer (47-55 km) where some energy is transmitted through the layers via tunneling phenomenon. The deep mixed layer is the most critical barrier for the wave, and comprehensive studies about the variability of this layer with latitude and local time would improve the understanding of the propagation of the mountain waves. The presence of these two layers generate trapped lee waves that propagate horizontally with some vertical leakage due to their large horizontal wavelength.
The variations of the wave characteristics with local time is imputed to the stability of the atmosphere close to the surface. In the afternoon, the first 4-5 km of the atmosphere are more stable that at night or at noon, which favors the vertical propagation of the waves.
The waves extract momentum from the surface with a flux around 2 Pa and deposit this momentum at around 30 km. Values of parameters used in the orographic parameterization of @Nava18 are therefore be confirmed by our mesoscale model which resolves the emission and propagation of gravity waves from the surface to 100-km altitude. At the cloud top, the waves induce a deceleration of several meters per second over a Venus day, consistent with *Akatsuki* measurements.
Temperature anomalies above Venus’ polar regions have also been explored with our mesoscale model. Transient waves rotating around the poles dominate the signal and the influence of the underlying topography is not noticeable. However, mountain waves are generated above Ishtar Terra – similarly to the waves generated by the other main mountains at lower latitudes. The increase of the deep atmosphere neutral barrier thickness strongly affects the amplitude of the orographic waves, which become negligible against the circumpolar transient waves.
Surface wind and the two mixed layers are key factors for the generation and propagation of mountain waves, a sensitivity study should be therefore considered in future work. Such study will imply complex changes in the IPSL Venus GCM dynamics and cloud model.
Additional simulations may be performed in the future for several other regions of interest, like Pheobe Regio at the Equator with a complex bow-shape morphology witnessed by Akatsuki, Gula Mons and Bell Regio, two sharp mountains at respectively 20N and 30N and Imdr Regio composed of two sharp mountains at 45S to investigate the influence of the impact of the morphology of the mountain, of the conditions near the surface and of the latitude on the temperature anomaly. Moreover the observations of stationary waves by VEx/VIRTIS in the southern hemisphere in the nightside with no apparent dependence with the local time [@Pera17] are features to be investigated.
The LMD Mesoscale model for Venus based on WRF [@Skam08] can be applied to the study of several other regional-scale phenomena, in particular slope winds, which were studied on Earth [@Brom01] and on Mars [@Spig11], and play a key role in the vertical extension of the planetary boundary layer convection [@Lebo18] and local meteorology in general.
Implementation of the photochemistry and microphysics schemes developed at IPSL in the Venus mesoscale model is planned to investigate the influence of mountains on the chemistry and the cloud formation. Explosive volcanism could explain SO$_2$ anomaly at the top of the cloud [@Espo88]: the study of the vertical transport and the impact of the waves on SO$_2$ [@Glaz99] and other volatile like water [@Aire15] could constrain this phenomenon, especially above Atla Regio where hot spots have been observed with *Venus Express* suggesting active volcanism [@Shal15]. Atmospheric variability on its own, without the need to invoke active volcanism, could also explain the observed SO$_2$ variability [@Marc13].
Acknowledgements {#acknowledgements .unnumbered}
================
This work was granted access to the High-Performance Computing (HPC) resources of Centre Informatique National de l’Enseignement Supérieur (CINES) under the allocations n°A0020101167 and A0040110391 made by Grand Équipement National de Calcul Intensif (GENCI). Simulation results used to obtain the figures in this paper are available in the open online repository https://figshare.com/s/0751293c8bbb6925a655. Full simulation results performed in this paper are available upon reasonable request (contact: [email protected]). The authors acknowledge Riwal Plougonven and Scot Rafkin for insightful discussions on gravity waves and mesoscale modeling. The authors thank the two anonymous reviewers that helped improve the quality of the paper. The authors thank Toru Kouyama for Akatusuki/LIR images. ML and SL acknowledge financial support from Programme National de Plan[é]{}tologie (PNP).
, M. W., [Mather]{}, T. A., [Pyle]{}, D. M., [Glaze]{}, L. S., [Ghail]{}, R. C., and [Wilson]{}, C. F. (2015). . , 113:33–48.
, D. G. (1987). . , 113:323–338.
, J.-L., [Khatuntsev]{}, I. V., [Hauchecorne]{}, A., [Markiewicz]{}, W. J., [Marcq]{}, E., [Lebonnois]{}, S., [Patsaeva]{}, M., [Turin]{}, A., and [Fedorova]{}, A. (2016). . , 121:1087–1101.
, J. E., [Young]{}, R. E., [Seiff]{}, A., [Ragent]{}, B., [Sagdeev]{}, R., [Linkin]{}, V. M., [Kerzhanovich]{}, V. V., [Ingersoll]{}, A. P., [Crisp]{}, D., [Elson]{}, L. S., [Preston]{}, R. A., [Golitsyn]{}, G. S., and [Ivanov]{}, V. N. (1986). . , 231:1422–1425.
, D. H., [Cassano]{}, J. J., [Klein]{}, T., [Heinemann]{}, G., [Hines]{}, K. M., [Steffen]{}, K., and [Box]{}, J. E. (2001). . , 129:2290–2309.
, P. R. A. (1983). . , 109:849–865.
Didion, A., Komjathy, A., Sutin, B., Nakazono, B., Karp, A., Wallace, M., Lantoine, G., Krishnamoorthy, S., Rud, M., Cutts, J., Lognonné, P., Kenda, B., Drilleau, M., Makela, J., Grawe, M., and Helbert, J. (2018). Remote sensing of venusian seismic activity with a small spacecraft, the vamos mission concept. In [*2018 IEEE Aerospace Conference*]{}, pages 1–14.
, D. (2003). . , pages 1161–1169.
, D. R., [Hills]{}, M. O. G., and [Blossey]{}, P. N. (2015). . , 72:1569–1584.
, C. (1961). . , 4:791–799.
, L. W., [Copley]{}, M., [Eckert]{}, R., [Gates]{}, L., [Stewart]{}, A. I. F., and [Worden]{}, H. (1988). . , 93:5267–5276.
, V., [Fournier]{}, R., [Dufresne]{}, J.-L., [Lebonnois]{}, S., [Hourdin]{}, F., and [Bullock]{}, M. A. (2009). . , 114:E11008.
, A., [Marcq]{}, E., [Luginin]{}, M., [Korablev]{}, O., [Bertaux]{}, J. L., and [Montmessin]{}, F. (2016). . , 275:143–162.
, P. G. and [Pettengill]{}, G. H. (1992). . , 97:13.
, D. C. and [Alexander]{}, M. J. (2003). . , 41:1112–1131.
, T., [Futaguchi]{}, M., [Hashimoto]{}, G. L., [Horinouchi]{}, T., [Imamura]{}, T., [Iwagaimi]{}, N., [Kouyama]{}, T., [Murakami]{}, S.-Y., [Nakamura]{}, M., [Ogohara]{}, K., [Sato]{}, M., [Sato]{}, T. M., [Suzuki]{}, M., [Taguchi]{}, M., [Takagi]{}, S., [Ueno]{}, M., [Watanabe]{}, S., [Yamada]{}, M., and [Yamazaki]{}, A. (2017). . , 10:85–88.
, I., [Garc[í]{}a Mu[ñ]{}oz]{}, A., [Hueso]{}, R., and [S[á]{}nchez-Lavega]{}, A. (2015). . , 245:16–31.
, I. and [Lebonnois]{}, S. (2018). . , 314:1–11.
, L. S. (1999). . , 104:18899–18906.
, A., [Chanin]{}, M. L., and [Wilson]{}, R. (1987). . , 14:933–936.
, R., [Kappel]{}, D., and [Arnold]{}, G. (2014). . , 232:232–248.
, R., [Kappel]{}, D., and [Arnold]{}, G. (2015). . , 117:262–294.
Horinouchi, T., Kouyama, T., Lee, Y. J., Murakami, S.-y., Ogohara, K., Takagi, M., Imamura, T., Nakajima, K., Peralta, J., Yamazaki, A., Yamada, M., and Watanabe, S. (2018). Mean winds at the cloud top of venus obtained from two-wavelength uv imaging by akatsuki. , 70(1):10.
, T., [Murakami]{}, S.-Y., [Satoh]{}, T., [Peralta]{}, J., [Ogohara]{}, K., [Kouyama]{}, T., [Imamura]{}, T., [Kashimura]{}, H., [Limaye]{}, S. S., [McGouldrick]{}, K., [Nakamura]{}, M., [Sato]{}, T. M., [Sugiyama]{}, K.-I., [Takagi]{}, M., [Watanabe]{}, S., [Yamada]{}, M., [Yamazaki]{}, A., and [Young]{}, E. F. (2017). . , 10:646–651.
, T., [Higuchi]{}, T., [Maejima]{}, Y., [Takagi]{}, M., [Sugimoto]{}, N., [Ikeda]{}, K., and [Ando]{}, H. (2014). . , 228:181–188.
, V., [Hertzog]{}, A., [Plougonven]{}, R., [C[á]{}mara]{}, A. d. l., and [Lott]{}, F. (2015). . , 72:3449–3468.
, H., [Sugimoto]{}, N., [Takagi]{}, M., [Matsuda]{}, Y., [Ohfuchi]{}, W., [Enomoto]{}, T., [Nakajima]{}, K., [Ishiwatari]{}, M., [Sato]{}, T. M., [Hashimoto]{}, G. L., [Satoh]{}, T., [Takahashi]{}, Y. O., and [Hayashi]{}, Y.-Y. (2019). . , 10:23.
Kitahara, T., Imamura, T., Sato, T. M., Yamazaki, A., Lee, Y. J., Yamada, M., Watanabe, S., Taguchi, M., Fukuhara, T., Kouyama, T., Murakami, S.-y., Hashimoto, G. L., Ogohara, K., Kashimura, H., Horinouchi, T., and Takagi, M. (2019). Stationary features at the cloud top of venus observed by ultraviolet imager onboard akatsuki. , 124.
, T., [Imamura]{}, T., [Taguchi]{}, M., [Fukuhara]{}, T., [Sato]{}, T. M., [Yamazaki]{}, A., [Futaguchi]{}, M., [Murakami]{}, S., [Hashimoto]{}, G. L., [Ueno]{}, M., [Iwagami]{}, N., [Takagi]{}, S., [Takagi]{}, M., [Ogohara]{}, K., [Kashimura]{}, H., [Horinouchi]{}, T., [Sato]{}, N., [Yamada]{}, M., [Yamamoto]{}, Y., [Ohtsuki]{}, S., [Sugiyama]{}, K., [Ando]{}, H., [Takamura]{}, M., [Yamada]{}, T., [Satoh]{}, T., and [Nakamura]{}, M. (2017). . , 44:12098–12105.
, S., [Komjathy]{}, A., [Pauken]{}, M. T., [Cutts]{}, J. A., [Garcia]{}, R. F., [Mimoun]{}, D., [Cadu]{}, A., [Sournac]{}, A., [Jackson]{}, J. M., [Lai]{}, V. H., and [Bowman]{}, D. C. (2018). . , 45:3393–3403.
, S., [Eymet]{}, V., [Lee]{}, C., and [Vatant d’Ollone]{}, J. (2015). . , 120:1186–1200.
, S., [Hourdin]{}, F., [Eymet]{}, V., [Crespin]{}, A., [Fournier]{}, R., and [Forget]{}, F. (2010). . , 115:E06006.
, S., [Schubert]{}, G., [Forget]{}, F., and [Spiga]{}, A. (2018). . , 314:149–158.
, M., [Lebonnois]{}, S., and [Spiga]{}, A. (2018). . , 123:2773–2789.
, M., [Spiga]{}, A., and [Lebonnois]{}, S. (2017). . , 122:134–149.
Limaye, S. S., Watanabe, S., Yamazaki, A., Yamada, M., Satoh, T., Sato, T. M., Nakamura, M., Taguchi, M., Fukuhara, T., Imamura, T., Kouyama, T., Lee, Y. J., Horinouchi, T., Peralta, J., Iwagami, N., Hashimoto, G. L., Takagi, S., Ohtsuki, S., Murakami, S.-y., Yamamoto, Y., Ogohara, K., Ando, H., Sugiyama, K.-i., Ishii, N., Abe, T., Hirose, C., Suzuki, M., Hirata, N., Young, E. F., and Ocampo, A. C. (2018). Venus looks different from day to night across wavelengths: morphology from akatsuki multispectral images. , 70(1):24.
, F. and [Miller]{}, M. J. (1997). . , 123:101–127.
, E., [Bertaux]{}, J.-L., [Montmessin]{}, F., and [Belyaev]{}, D. (2013). . , 6:25–28.
, L., [Brissaud]{}, Q., [Lai]{}, V. H., [Garcia]{}, R. F., [Martin]{}, R., [Krishnamoorthy]{}, S., [Komjathy]{}, A., [Cadu]{}, A., [Cutts]{}, J. A., [Jackson]{}, J. M., [Mimoun]{}, D., [Pauken]{}, M. T., and [Sournac]{}, A. (2018). . , 45:12.
, G. L. and [Yamada]{}, T. (1982). . , 20:851–875.
, T., [Schubert]{}, G., and [Lebonnois]{}, S. (2018). . , 11:487–491.
, J., [Hueso]{}, R., [S[á]{}nchez-Lavega]{}, A., [Piccioni]{}, G., [Lanciano]{}, O., and [Drossart]{}, P. (2008). . , 113:E00B18.
, J., [Hueso]{}, R., [S[á]{}nchez-Lavega]{}, A., [Lee]{}, Y. J., [Mu[ñ]{}oz]{}, A. G., [Kouyama]{}, T., [Sagawa]{}, H., [Sato]{}, T. M., [Piccioni]{}, G., [Tellmann]{}, S., [Imamura]{}, T., and [Satoh]{}, T. (2017). . , 1:0187.
, G. H., [Eliason]{}, E., [Ford]{}, P. G., [Loriot]{}, G. B., [Masursky]{}, H., and [McGill]{}, G. E. (1980). . , 85:8261–8270.
, F. M., [Neiman]{}, P. J., [Keller]{}, T. L., [Levinson]{}, D., and [Fedor]{}, L. (1997). , 54:1308–1333.
, T., [Sato]{}, T. M., [Nakamura]{}, M., [Kasaba]{}, Y., [Ueno]{}, M., [Suzuki]{}, M., [Hashimoto]{}, G. L., [Horinouchi]{}, T., [Imamura]{}, T., [Yamazaki]{}, A., [Enomoto]{}, T., [Sakurai]{}, Y., [Takami]{}, K., [Sawai]{}, K., [Nakakushi]{}, T., [Abe]{}, T., [Ishii]{}, N., [Hirose]{}, C., [Hirata]{}, N., [Yamada]{}, M., [Murakami]{}, S.-y., [Yamamoto]{}, Y., [Fukuhara]{}, T., [Ogohara]{}, K., [Ando]{}, H., [Sugiyama]{}, K.-i., [Kashimura]{}, H., and [Ohtsuki]{}, S. (2017). . , 69:154.
, G., [Covey]{}, C., [del Genio]{}, A., [Elson]{}, L. S., [Keating]{}, G., [Seiff]{}, A., [Young]{}, R. E., [Apt]{}, J., [Counselman]{}, C. C., [Kliore]{}, A. J., [Limaye]{}, S. S., [Revercomb]{}, H. E., [Sromovsky]{}, L. A., [Suomi]{}, V. E., [Taylor]{}, F., [Woo]{}, R., and [von Zahn]{}, U. (1980). . , 85:8007–8025.
, R. S. (1949). . , 75:41–56.
, A., [Kirk]{}, D. B., [Young]{}, R. E., [Blanchard]{}, R. C., [Findlay]{}, J. T., [Kelly]{}, G. M., and [Sommer]{}, S. C. (1980). . , 85:7903–7933.
, E. V., [Markiewicz]{}, W. J., [Basilevsky]{}, A. T., [Titov]{}, D. V., [Ignatiev]{}, N. I., and [Head]{}, J. W. (2015). . , 42:4762–4769.
, W. C. and [Klemp]{}, J. B. (2008). . , 227:3465–3485.
, R. B. (1979). . , 21:87–230.
, A. and [Forget]{}, F. (2009). . , 114:E02009.
, A., [Forget]{}, F., [Madeleine]{}, J.-B., [Montabone]{}, L., [Lewis]{}, S. R., and [Millour]{}, E. (2011). . , 212:504–519.
, A. and [Smith]{}, I. (2018). . , 308:197–208.
, B. R. and [Yewchuk]{}, K. (2004). . , 511:125–134.
, M., [Fukuhara]{}, T., [Imamura]{}, T., [Nakamura]{}, M., [Iwagami]{}, N., [Ueno]{}, M., [Suzuki]{}, M., [Hashimoto]{}, G. L., and [Mitsuyama]{}, K. (2007). . , 40:861–868.
, S., [Haeusler]{}, B., [Paetzold]{}, M., [Bird]{}, M. K., [Tyler]{}, G. L., [Andert]{}, T., and [Remus]{}, S. (2009). . , 114:E00B36.
, R. L., [Schubert]{}, G., and [Brinkman]{}, D. G. (2001). . , 106:31.
, R. E., [Walterscheid]{}, R. L., [Schubert]{}, G., [Pfister]{}, L., [Houben]{}, H., and [Bindschadler]{}, D. L. (1994). , 51:1857–1875.
, R. E., [Walterscheid]{}, R. L., [Schubert]{}, G., [Seiff]{}, A., [Linkin]{}, V. M., and [Lipatov]{}, A. N. (1987). . , 44:2628–2639.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Under $SL(2,R)$ electric-magnetic duality transformations the Bogomolnyi bound of dilaton-axion black holes is known to be invariant. In this paper we show that this invariance corresponds to the covariance of the $N=4$ supersymmetry transformation rules and their parameters. In particular this implies that Killing spinors transform covariantly into Killing spinors. As an example, we work out completely the case of the largest known family of axion-dilaton black holes which is $SL(2,R)$-invariant, finding the Killing spinors with the announced properties.'
---
QMW-PH-94-10\
[**hep-th/9404035**]{}\
April $6^{th}$, 1994
[ [^1]\
[*Department of Physics*]{}\
[*Queen Mary and Westfield College*]{}\
[*Mile End Road, London E1 4NS, U.K.*]{}]{}
Introduction {#introduction .unnumbered}
============
The SL(2,R)-duality invariance of the low-energy effective string theory equations of motion [@kn:STW; @kn:Sendual; @kn:Schdual] ($N=4$ supergravity [@kn:oldual]) is a fascinating symmetry that interchanges the strong and weak coupling regimes. Quantum effects (instantons [@kn:STW] or dyon charge quantization [@kn:Senquan; @kn:KO]) effectively break this symmetry to $SL(2,Z)$.
It was conjectured in [@kn:firstdual] and more recently in [@kn:Sendual] that $SL(2,Z)$ could be an exact, non-perturbative, symmetry of superstring theory. This idea is supported by the fact that the spectrum of dyon charges [@kn:Senquan] is $SL(2,Z)$-invariant. It is also extremely interesting the interplay between the $SL(2,Z)$-duality symmetry and other symmetries of compactified string theory: the $O(6,22;Z)$ symmetry, target-space duality and the string-fivebrane duality that has been explored in [@kn:SchSen] and [@kn:Binet].
In this letter we are going study the behavior of some supersymmetry properties under general $SL(2,R)$-duality transformations. In particular we will show in Sec. 1 the covariance of the $N=4$ supergravity transformation laws of the gravitino and dilatino fields under $SL(2,R)$-duality transformations. As a consequence, Killing spinors transform covariantly and the number of unbroken supersymmetries of $SL(2,R)$-related on-shell field configurations is invariant. Furthermore, the positivity bounds associated to the supersymmetry algebra are also invariant. Some of these results were hinted or proven in special cases in Ref. [@kn:TOM] and in a different context in [@kn:Senbogo].
In Sec. 2 we study the particular case of the $SL(2,R)$-invariant family of dilaton-axion black holes recently found in Ref. [@kn:KO]. We find the Killing spinors and check that their behavior under duality agrees with the results of Sec. 1. The invariance of the Bogomolnyi-Gibbons-Hull-type bound for axion-dilaton black holes is built in the solutions and it is, anyway, a straightforward generalization of the results of Ref. [@kn:TOM]. Therefore we will nos discuss it.
Our conclusions are in the last section.
Throughout all this paper we use the conventions and notations of Refs. [@kn:KLOPP; @kn:KO], but our definitions of the charges differ slightly. The Appendix contains, for the sake of completeness, a description of the axion-dilaton black holes and the definitions of charges we are using.
SL(2,R)-covariant Killing spinors
=================================
In the $SU(4)$ version of $N=4$ supergravity, the bosonic part of the supersymmetry rules of the gravitino and dilatino fields are respectively [@kn:oldual] $$\begin{aligned}
{\textstyle\frac{1}{2}}\delta_{\epsilon}\Psi_{\mu I} & = &
\nabla_{\mu}\epsilon_{I}
-{\textstyle\frac{i}{4}}e^{2\phi}\partial_{\mu}\epsilon_{I}
-{\textstyle\frac{1}{8}}\sigma^{\rho\sigma}
T_{\rho\sigma\, IJ}^{+}\gamma_{\rho}\epsilon^{J}\; ,
\label{eq:susytrans1}
\\
{\textstyle\frac{1}{2}}\delta_{\epsilon}\Lambda_{I} & = &
-{\textstyle\frac{i}{2}}(e^{2\phi}\not\!\partial\lambda)\epsilon_{I}
+{\textstyle\frac{1}{4}}\sigma^{\rho\sigma}T_{\rho\sigma\,
IJ}^{-}\epsilon^{J}\; ,
\label{eq:susytrans2}\end{aligned}$$ where $$\lambda = a+ie^{-2\phi}\, ,$$ is a complex scalar field built out of $\phi$, the dilaton field, and $a$, the axion field. Also $$T_{\rho\sigma\, IJ} =
{\textstyle2\sqrt{2}}e^{-\phi}[F_{\rho\sigma}\alpha_{IJ}
+iG_{\rho\sigma}\beta_{IJ}]\, ,$$ where $F_{\rho\sigma}=\partial_{\rho}A_{\sigma}-\partial_{\sigma}A_{\rho}$ and $G_{\rho\sigma}=\partial_{\rho}B_{\sigma}-\partial_{\sigma}B_{\rho}$ are the field strength tensors of two $U(1)$ vector fields $A_{\mu}$ and $B_{\mu}$. Finally, $\alpha_{IJ}$ and $\beta_{IJ}$ are two $SO(4)$ matrices with $SO(4)$ indices $I,J$.
Given any $SL(2,R)$ matrix $\left(\begin{array}{cc}\alpha & \beta \\
\gamma & \delta \\ \end{array} \right)$, $\alpha\delta-\beta\gamma=1$ and the complex scalar field $\lambda$ of the original field configuration we can define two functions $$\begin{aligned}
R & = & \alpha\lambda + \beta\, , \nonumber \\
S & = & \gamma\lambda + \delta\, .\end{aligned}$$ In terms of $R$ and $S$ and the old field configurations, the transformed fields can be conveniently written in this way: $$\begin{aligned}
\lambda^{\prime} & = & R/S\, , \nonumber \\
F^{\pm\prime}_{\rho\sigma} & = & S F^{\pm}_{\rho\sigma}\, ,\end{aligned}$$ and the analogous equation for $G$. In the Einstein frame (which we are using) $SL(2,R)$ does not act on the metric.
These transformations rotate continuously the “Maxwell law" into the Bianchi identity, hence the name electric-magnetic duality.
It is a well known fact (Refs. [@kn:oldual; @kn:Sendual]) that the equations of motion of this theory are invariant under the above transformations while the action is not.
In addition to this, it is easy to see that the equations (\[eq:susytrans1\]) and (\[eq:susytrans2\]) conserve their form (i.e. transform covariantly under $SL(2,R)$) if we assume the following transformation laws for the supersymmetry parameters and variations $$\begin{aligned}
\epsilon_{I}^{\prime} & = & e^{\frac{i}{2}Arg(S)}\epsilon_{I}\, ,
\label{eq:epsilon}
\\
(\delta_{\epsilon}\Psi_{\mu I})^{\prime} & = &
e^{\frac{i}{2}Arg(S)} \delta_{\epsilon}\Psi_{\mu I}\, ,
\\
(\delta_{\epsilon}\Lambda_{I})^{\prime} & = &
e^{\frac{-3i}{2}Arg(S)} \delta_{\epsilon}\Lambda_{I}\, .\end{aligned}$$ There is no need to use the equations of motion to prove this result and, thus, it holds for any on- or off-shell field configuration.
Although this result may not be too surprising we would like to stress that it is far from being trivial. $SL(2,R)$ is only an invariance of the classical equations of motion and so it transforms on-shell configurations into on-shell configurations. There is no apparent reason why arbitrary field configurations should behave nicely under $SL(2,R)$. Furthermore, the supersymmetry parameter $\epsilon$ is not even a field of the theory.
Observe that the transformation (\[eq:epsilon\]) is consistent with the invariance of the metric and (accordingly) of the Killing vector $\xi^{\mu}$ that can be built out of a (commuting) set of Killing spinors $\epsilon^{I}$: $$\xi^{\mu}=\overline{\epsilon}^{I}\gamma^{\mu}\epsilon_{I}\, .$$
Now let us study some consequences of this result.
Let us assume that for a particular field configuration the equations $\delta_{\epsilon}\Psi_{\mu I}= \delta_{\epsilon}\Lambda_{I}=0$ are satisfied by some set of spinors $\epsilon_{I}$, some (perhaps all) of them trivial (i.e. vanishing). We call this set of spinors “Killing spinors". Then the above result implies that in the duality-transformed field configurations these equations are also satisfied by another set of Killing spinors, the number of trivial (i.e. vanishing) and non-trivial Killing spinors being the same as in the original field configuration.
On-shell configurations admitting non-trivial asymptotically constant Killing spinors are said to have unbroken supersymmetries. Hence, the number of unbroken supersymmetries is the same for any two $SL(2,R)$-related classical solutions.
For off-shell configurations the situation is subtly different. They obey equations of the form $$\begin{aligned}
\frac{\delta S_{Class}}{\delta g^{\mu\nu}} & = & J_{\mu\nu}\, ,
\nonumber \\
\frac{\delta S_{Class}}{\delta A_{\mu}} & = & J_{A}{}^{\mu}\, ,
\nonumber \\
\frac{\delta S_{Class}}{\delta B_{\mu}} & = & J_{B}{}^{\mu}\, ,
\nonumber \\
\frac{\delta S_{Class}}{\delta \lambda} & = & J\, ,
\label{eq:sources}\end{aligned}$$ where $S_{Class}$ is the classical action $$S_{Class}=\int d^{4}x\sqrt{-g}\{-R
-2\frac{\partial_{\mu}\lambda
\partial^{\mu}\overline{\lambda}}{(\lambda-\overline{\lambda})^{2}} +
[i\lambda (F^{+2}+ G^{+2})+c.c.]\}$$ and $J_{\mu\nu},J_{A,B}{}^{\mu},J$ are non-vanishing functions (otherwise the configurations would be on-shell) that may describe additional matter sources or quantum corrections, for instance. The fact that a field configuration admits Killing spinors does not automatically imply that it obeys the classical equations of motion[^2] It does not imply, either, that the sources in the right-hand side of eqs. (\[eq:sources\]) are coupled in a way consistent with supersymmetry.
In Ref. [@kn:KSI] a set of consistency conditions for these sources to be consistent with supersymmetry was derived. In our case, these “Killing Spinor Identities" (KSI) take the form $$\begin{aligned}
\frac{e^{\phi}}{\sqrt{2}}
[J_{A}{}^{\mu}\alpha^{IJ}+i\gamma_{5}J_{B}{}^{\mu}\beta^{IJ}]
\overline{\epsilon}_{J}\gamma_{\mu}
+2iJ\ e^{-2 \phi}\overline{\epsilon}^{I} & = & 0 \, ,
\nonumber\\
2J^{\mu\nu}\overline{\epsilon}^{I}\gamma_{\nu}+
\frac{e^{\phi}}{\sqrt{2}}
[J_{A}{}^{\mu}\alpha^{IJ}+i\gamma_{5}J_{B}{}^{\mu}\beta^{IJ}]
\overline{\epsilon}_{J}
& = & 0 \, ,
\label{eq:ksi}\end{aligned}$$ where $\epsilon^{I}$ is a set of Killing spinors.
Under an $SL(2,R)$ rotation of the field configuration the sources transform as follows: $$\begin{aligned}
(J_{\mu\nu})^{\prime} & = & J_{\mu\nu}\, ,
\nonumber \\
(J_{A,B}{}^{\mu})^{\prime} & = & \alpha J_{A,B}{}^{\mu}\, ,
\nonumber \\
J^{\prime} & = & S^{2}J\, .\end{aligned}$$ It is obvious that the KSI will generally be violated after a general $SL(2,R)$ rotation, unless the sources vanish. This result is obviously related to the fact that a non-vanishing electric current $J$ produces a non-vanishing magnetic current after the rotation.
Thus, in general, only the unbroken supersymmetries of on-shell field configurations will be preserved by an electric-magnetic $SL(2,R)$ duality transformation.
This result can be read from a slightly different point of view, as in Ref. [@kn:trace]. If the sources come from quantum corrections and we select only supersymmetric configurations, then only those with no quantum corrections will survive. In our case ($N=4$) there is no trace anomaly and we don’t know what the quantum corrections would be like, but this is clearly a subject worth studying.
Finally, one is tempted to conjecture that in a supersymmetric theory with manifest $SL(2,R)$ invariant couplings the number of unbroken supersymmetries would always be invariant, but such a theory does not yet exist.
The Killing spinors of axion-dilaton black holes
================================================
Recently a quite general family of static black-hole solutions of the low-energy string theory equations of motion has been presented in [@kn:KO]. The main feature of this family, which contains many already known solutions [@kn:dilbh; @kn:STW; @kn:KLOPP], is that it constitutes a representation of the whole $SL(2,R)$-duality group in the sense that by applying any $SL(2,R)$ transformation to any solution in the family we get another member of the family[^3].
Our immediate goal is to find the non-trivial Killing spinors of this family of configurations and check explicitly Eq. (\[eq:epsilon\]). We already know that the non-extreme black-hole solutions in this class do not have non-trivial Killing spinors [@kn:KLOPP], so we will study only the solutions describing several extreme black holes in equilibrium, which can be found in the Appendix.
Note that only 6 of the $L$ $U(1)$ fields of the general solution fit into pure $N=4$ supergravity and we will take only 2 to be non-vanishing as in the previous section: $A^{(1)}_{\mu}\equiv A_{\mu},\,
A^{(2)}\equiv B_{\mu}$.
We are looking for time-independent ($\partial_{t}\epsilon_{I}=0$) solutions of the equations $$\begin{aligned}
\delta_{\epsilon}\Psi_{\mu I} & = 0\, ,
\label{eq:kill1} \\
\delta_{\epsilon}\Lambda_{I} & = & 0\, .
\label{eq:kill2}\end{aligned}$$ First we contract Eq.(\[eq:kill1\]) with $\gamma_{\mu}$ and obtain $$\not\!\nabla\epsilon_{I}-{\textstyle\frac{i}{4}e^{2\phi}\not\!\partial
a\epsilon_{I}}=0\, .$$ For time-independent Killing spinors of the above backgrounds this equation reduces to $$\gamma^{i}\partial_{\hat{\imath}}
[e^{-\frac{1}{2}(U+i\theta)}\epsilon_{I}] = 0\, ,
\label{eq:kill3}$$ where $${\cal H}_{2}/\overline{{\cal H}}_{2}=e^{2i\theta}\, ,$$ and we have used the property of these solutions $$\partial_{\hat{\imath}}\ln ({\cal H}_{2}/\overline{{\cal H}}_{2})
= ie^{2\phi}\partial_{\hat{\imath}}a\, .$$ If we apply the operator $\gamma^{i}\partial_{\hat{\imath}}$ to Eq. (\[eq:kill3\]) we get the integrability condition $$\partial_{\hat{\imath}}\partial_{\hat{\imath}}
[e^{-\frac{1}{2}(U+i\theta)}\epsilon_{I}] = 0\, ,$$ which means that the combination in brackets is a harmonic spinor. If we substitute now in Eq. (\[eq:kill3\]) a general harmonic spinor we see that only a constant one satisfies it, and thus $$\epsilon_{I}=e^{-\frac{1}{2}(U+i\theta)}\epsilon_{I(0)}\, ,
\label{eq:spinor}$$ where the $\epsilon_{I(0)}$s are constant spinors.
This result allows us to check partially the main result in the previous section. To do this, first observe that under $SL(2,R)$ the functions ${\cal H}_{1}$ and ${\cal H}_{2}$ a transform in this way: $$\begin{aligned}
{\cal H}_{1}^{\prime} & = & \frac{\alpha {\cal H}_{1} + \beta{\cal
H}_{2}}{\gamma \lambda_{0} + \delta}\, ,
\nonumber \\
{\cal H}_{2}^{\prime} & = & \frac{\gamma {\cal H}_{1} + \delta{\cal
H}_{2}}{\gamma \lambda_{0} + \delta}\, .\end{aligned}$$ Then, according to Eq. (\[eq:spinor\]) $$\begin{aligned}
\epsilon_{I}^{\prime} & = & e^{-\frac{1}{2}(U+i\theta^{\prime})}
\epsilon_{I(0)}^{\prime}
\nonumber \\
& = & e^{-\frac{1}{2}U} \left({\cal H}_{2}^{\prime}/\overline{{\cal
H}}_{2}^{\prime}\right)^{-\frac{1}{4}}
\left(\frac{\gamma \lambda_{0} + \delta}{\gamma \overline{\lambda}_{0} +
\delta}\right)^{\frac{1}{4}}\epsilon_{I(0)}^{\prime}
\nonumber \\
& = & e^{\frac{i}{2}Arg(S)}e^{-\frac{1}{2}(U+i\theta)}
e^{\frac{i}{2}Arg(S_{0})}\epsilon_{I(0)}^{\prime}\, ,\end{aligned}$$ and, looking at Eqs. (\[eq:epsilon\]) and (\[eq:spinor\]) we see that we only need to check that the asymptotic value of the Killing spinors transforms as follows: $$\epsilon_{I(0)}^{\prime}= e^{\frac{-i}{2}Arg(S_{0})}\epsilon_{I(0)}\, .
\label{eq:constantspinor}$$ where $S_{0}$ is the value of $S$ at infinity.
We only need to find the constant spinors. It is enough to consider the equations $\delta_{\epsilon}\Psi_{t I} = \delta_{\epsilon}\Lambda_{I} =
0$. A long but straightforward calculation leads to the following two equations for each black hole: $$\begin{aligned}
\frac{ie^{\phi_{0}}}{\sqrt{2}}
[\lambda_{0}M_{i}+\overline{\lambda}_{0}\Upsilon_{i}]
\epsilon_{I(0)}-\{\tilde{Q}_{i}\}_{IJ}\gamma^{0}\epsilon^{J}_{(0)}
& = & 0\, ,
\nonumber \\
\frac{ie^{\phi_{0}}}{\sqrt{2}}
[M_{i}+\Upsilon_{i}]\epsilon_{I(0)} -
\{P_{i}\}_{IJ}\gamma^{0}\epsilon^{J}_{(0)} & = & 0\, ,
\label{eq:constraints}\end{aligned}$$ where we have used abbreviated notation $$\{X\}_{IJ}=\{X^{A}\alpha_{IJ}+X^{B}\beta_{IJ}\}\, ,$$ for any quantity $X$ associated to the vector fields $A_{\mu}$ and $B_{\mu}$, the only ones we are considering here.
The algebraic constraints Eqs. (\[eq:constraints\]) tell us that the whole solution will have at most as many unbroken supersymmetries as the single black hole with less unbroken supersymmetries. Note that Eqs. (\[eq:Bogo\]), (\[eq:charges\]) and (\[eq:relation\]) guarantee that both Eqs. (\[eq:constraints\]) are always compatible for all black holes at once.
After using the explicit form of the $SO(4)$ matrices $\alpha_{IJ}$ and $\beta_{IJ}$ (see Ref. [@kn:KLOPP]) we can describe the final result as follows. First we define the central charges $$z_{\pm}=\sqrt{2}e^{-\phi_{0}}(\Gamma^{A}\pm i\Gamma^{B})\, .$$ There are three different cases
1. One unbroken supersymmetry in the $1,2$ sector with the following relations holding for each black hole and the following spinors: $$\begin{aligned}
M_{i} & = & |z_{-i}|\, ,
\nonumber \\
|\Upsilon_{i}| & = & |z_{+i}|\, ,
\nonumber \\
\mu_{-} & = &
-i\sqrt{2}e^{-\phi_{0}}\frac{\tilde{Q}^{A}
- i\tilde{Q}^{B}}{\lambda_{0} M + \overline{\lambda}_{0}\Upsilon} =
-i\sqrt{2}e^{-\phi_{0}}\frac{P^{A}-iP^{B}}{M + \Upsilon}\, ,
\nonumber \\
\epsilon_{3(0)} & = & \epsilon_{4(0)}=0\, ,
\nonumber \\
\epsilon_{1(0)} & = & \mu_{-i}\gamma^{0}\epsilon^{2}_{(0)}\, .\end{aligned}$$
2. One unbroken supersymmetry in the $3,4$ sector with the following relations holding for each black hole and the following spinors: $$\begin{aligned}
M_{i} & = & |z_{+i}|\, ,
\nonumber \\
|\Upsilon_{i}| & = & |z_{-i}|\, ,
\nonumber \\
\mu_{+} & = &
-i\sqrt{2}e^{-\phi_{0}}\frac{\tilde{Q}^{A}
+ i\tilde{Q}^{B}}{\lambda_{0} M + \overline{\lambda}_{0}\Upsilon} =
-i\sqrt{2}e^{-\phi_{0}}\frac{P^{A}+iP^{B}}{M + \Upsilon}\, ,
\nonumber \\
\epsilon_{1(0)} & = & \epsilon_{2(0)}=0\, ,
\nonumber \\
\epsilon_{3(0)} & = & \mu_{+i}\gamma^{0}\epsilon^{4}_{(0)}\, .\end{aligned}$$
3. Two unbroken supersymmetries, one in the $1,2$ sector and one in the $3,4$ sector with the following relations holding for each black hole and the following spinors: $$\begin{aligned}
M_{i} & = & |\Upsilon_{i}|=|z_{+i}|=|z_{-i}|\, ,
\nonumber \\
\epsilon_{1(0)} & = & \mu_{-i}\epsilon_{2(0)}\, ,
\nonumber \\
\epsilon_{3(0)} & = & \mu_{+i}\gamma^{0}\epsilon^{4}_{(0)}\, .\end{aligned}$$
For just one black hole, we always have supersymmetry. When there is more than just one, the situation is not so clear. To have supersymmetry the complex constants $\mu_{+i}$ or $\mu_{-i}$ must have the same value for the $N$ black holes. That this is indeed possible is clear: just take for one of them a set of charges $(M_{i},\Sigma_{i},\Delta_{i},Q^{A}_{i},Q^{B}_{i},P^{A}_{i},P^{B}_{i})$ that satisfies the Bogomolnyi-Gibbons-Hull bound Eq. (\[eq:Bogo\]) and take for the rest of the black holes sets of charges proportional to this one. This subclass of configurations has at least one unbroken supersymmetry. However, due to the large amount of charges and the involved relations between them we haven’t been able to prove that all of them are supersymmetric[^4], although this seems likely.
Back to our main problem, it’s easy to see that the constants $\mu_{\pm}$ transform exactly in the form required by the result of the previous section, or, equivalently, by Eq. (\[eq:constantspinor\]) $$\mu_{\pm}^{\prime} = e^{iArg(S_{0})}\mu_{\pm}\, .$$
This illustrates the main result of this paper.
Conclusions
===========
In this letter we have proven that if a field configuration admits Killing spinors, then all of its $SL(2,R$) images do, and that the Killing spinors of the images are related very simply to those of the original. This property is related to the $SL(2,R)$ invariance of the Bogomolnyi-Gibbons-Hull bound of axion-dilaton black holes. Since this symmetry does not act on the (Einstein-frame) metric, this result provides further examples of supersymmetry acting as a cosmic censor, although one needs to check some consistency conditions (the Killing Spinor Identities) to see whether supersymmetry is preserved when quantum corrections are taken into account.
There is a number of other duality symmetries in String Theory, and it would be interesting to find whether similar properties hold for them. Target-space duality, for instance, can be seen also as a symmetry of the low-energy String Theory equations of motion, which can be embedded in a theory with local supersymmetry. Although it looks very different from $SL(2,R)$ duality, there are some examples in which unbroken supersymmetries were preserved by it [@kn:BEK]. Note that $SL(2,R)$ does transform the stringy metric too. We hope to report on results on this problem soon.
Acknowledgements {#acknowledgements .unnumbered}
================
The author wish to thank R. Kallosh for many discussions and her support. This work has been partially supported by a Spanish Government MEC postdoctoral grant and by a postdoctoral European Communities Human Capital and Mobility program grant.
Axion-dilaton black-hole solutions
==================================
Here we briefly describe the extreme axion-dilaton multi-black-hole solutions found in Ref. [@kn:KO]. Our definitions of the charges differ slightly from those of that reference.
Apart from the metric and complex scalar field these solutions have $L$ $U(1)$ fields $F^{(n)}_{\mu\nu} = \partial_{\mu} A^{(n)}_{\nu}-
\partial_{\nu} A^{(n)}_{\mu}$.
We also use the auxiliary fields (the $SL(2,R)$-duals) $\tilde{F}^{(n)}_{\mu\nu} = e^{-2\phi} {}^{*}F^{(n)}_{\mu\nu}- ia
F^{(n)}_{\mu\nu}$. Owing to the “Maxwell law", which can be written $\nabla^{\mu} {}^{*}\tilde{F}_{\mu\nu} = 0$, there exist locally $L$ real vector fields $\tilde{A}^{(n)}_{\nu}$ such that $\tilde{F}_{\mu\nu}
= i(\partial_{\mu} \tilde{A}^{(n)}_{\nu}- \partial_{\nu}
\tilde{A}^{(n)}_{\mu}$). They simplify the description of the solutions, which are given by $$\begin{aligned}
ds^{2} & = &
e^{2U}dt^{2}-e^{-2U}d\vec{x}^{2}\, ,
\nonumber \\
e^{-2U}(\vec{x}) & = &
2 \,{\mbox{Im}}\,
({\cal H}_{1}(\vec{x})\, \overline{{\cal H}}_{2}(\vec{x}))\, ,
\nonumber \\
\lambda(\vec{x}) & = &
\frac{{\cal H}_{1}(\vec{x})}{{\cal H}_{2}(\vec{x})}\, ,
\nonumber \\
A_{t}^{(n)}(\vec{x}) & = &
e^{2U}(k^{(n)}{\cal H}_{2}(\vec{x})+c.c.)\, ,
\nonumber \\
\tilde{A}^{(n)}_{t}(\vec{x}) & = &
-e^{2U}(k^{(n)}{\cal H}_{1}(\vec{x})+c.c.)\, ,\end{aligned}$$ where ${\cal H}_{1}(\vec{x}),\, {\cal H}_{2}(\vec{x})$ are two complex harmonic functions with $N$ poles corresponding to $N$ labeled by $i=1,\ldots,N$ $$\begin{aligned}
{\cal H}_{1}(\vec{x}) & = &
\frac{e^{\phi_{0}}}{\sqrt{2}} \{\lambda_{0}+\sum_{i=1}^{N}
\frac{\lambda_{0}M_{i}+\overline{\lambda}_{0}
\Upsilon_{i}}{|\vec{x}-\vec{x}_{i}|}\}\, ,
\nonumber \\
{\cal H}_{2}(\vec{x}) & = &
\frac{e^{\phi_{0}}}{\sqrt{2}} \{1+\sum_{i=1}^{N}\frac{M_{i}+
\Upsilon_{i}}{|\vec{x}-\vec{x}_{i}|}\}\, ,
\nonumber \\
\partial_{\hat{\imath}}\partial_{\hat{\imath}}{\cal H}_{1} & = &
\partial_{\hat{\imath}}\partial_{\hat{\imath}}{\cal H}_{1} = 0\, .\end{aligned}$$ The different constants, which are defined later, have to satisfy several identities. First, they have to satisfy Bogomolnyi-Gibbons-Hull-type identities for each black hole and for the global solution: $$\begin{aligned}
M^{2}_{i}+|\Upsilon_{i}|^{2}+2i(\lambda_{0}-
\overline{\lambda}_{0})\sum_{n=1}^{L}|\Gamma^{(n)}_{i}|^{2} & = & 0\, ,
\nonumber \\
M^{2}+|\Upsilon|^{2}+2i(\lambda_{0}-
\overline{\lambda}_{0})\sum_{n=1}^{L}|\Gamma^{(n)}|^{2} & = & 0\, .
\label{eq:Bogo}\end{aligned}$$ Secondly, the consistency of the solution requires $$\begin{aligned}
k^{(n)} & = & -\sqrt{2}e^{-\phi_{0}}
\frac{\Gamma^{(n)}M+
\overline{\Gamma}^{(n)}\overline{\Upsilon}}{M^{2}-|\Upsilon|^{2}}
=k^{(n)}_{i}\, ,
\nonumber \\
Arg(\Upsilon) & = & Arg(\Upsilon_{i})\, ,
\label{eq:charges}\end{aligned}$$ for each black hole $i$.
Finally, $\Upsilon$ and $\Gamma$ are related by $$\Upsilon = i(\lambda_{0}-\overline{\lambda}_{0})
\frac{\sum_{n=1}^{L}\overline{\Gamma}^{(n)2}}{M}\, ,
\label{eq:relation}$$ for each black hole and for the global solutions.
The different constants are defined by the asymptotic expansions in the “upper sheet" (global charges) and in the $i^{th}$ black hole sheet: $$\begin{aligned}
g_{tt} & \sim & 1-\frac{2M}{r}\, .
\nonumber \\
F^{(n)+} & \sim & \frac{\Gamma^{(n)}}{r^{2}}\, ,
\nonumber \\
\tilde{F}^{(n)+} & \sim & \frac{\tilde{\Gamma}^{(n)}}{r^{2}}\, ,
\nonumber \\
\lambda & \sim & \lambda_{0}-
(\lambda_{0}-\overline{\lambda}_{0})\frac{\Upsilon}{r}\, .\end{aligned}$$ We can write these complex constants in terms of real constants: electric ($Q$) and magnetic ($P$) charges, dilaton ($\Sigma$) and axion ($\Delta$) charges and dilaton $\phi_{0}$ and axion $a_{0}$ asymptotic values: $$\begin{aligned}
\Gamma^{(n)} & = & \frac{1}{2}(Q^{(n)}+iP^{(n)})\, ,
\nonumber \\
\Upsilon & = & \Delta-i\Sigma\, ,
\nonumber \\
\lambda_{0} & = & a_{0}+ie^{-2\phi_{0}}\, .\end{aligned}$$ The charges in the upper sheet (the global charges) are the sum of the charges of the $N$ black holes.
[30]{}
A. Shapere, S. Trivedi and F. Wilczek, Mod. Phys. Lett. [ **A6**]{}, 2677 (1991). A. Sen, Nucl. Phys. [**B404**]{}, 109 (1993). J. Maharana and J.H. Schwarz, Nucl. Phys. [**B390**]{}, 3 (1993); J.H. Schwarz, preprint CALT-68-1815, hepth/9209125, presented at the International Workshop on String Theory, Quantum Gravity and the Unification of Fundamental Interactions, Rome, Italy, 21-26 Sep. 1992. S. Ferrara, J. Scherk and B. Zumino, Nucl. Phys. [**B121**]{}, 393 (1977);\
E. Cremmer, J. Scherk and S. Ferrara, Phys. Lett. [**74B**]{}, 61 (1978);\
B. de Wit, Nucl. Phys. [**B158**]{}, 189 (1979);\
E. Cremmer and B. Julia, Nucl. Phys. [**B159**]{}, 141 (1979);\
M.K. Gaillard and B. Zumino, Nucl. Phys. [**B193**]{}, 221 (1981). A. Sen, Phys. Lett. [**303B**]{}, 22 (1993). R. Kallosh and T. Ortín, Phys. Rev. [**D48**]{} 742-747 (1993). A. Font, L. Ibáñez, D. Lüst and F. Quevedo, Phys. Lett. [**249B**]{}, 35 (1990);\
S.J. Rey, Phys. Rev. [**D43**]{}, 526 (1991). J.H. Schwarz and A. Sen, Phys. Lett. [**312B**]{}, 105 (1993); Nucl. Phys. [**B411**]{}, 35 (1994). P. Binétruy, Phys. Lett. [**315B**]{}, 80 (1993). T. Ortín, Phys. Rev. [**D47**]{}, 3136 (1993). A. Sen, Mod. Phys. Lett. [**A8**]{}, 2023 (1993). R. Kallosh, A. Linde, T. Ortín, A. Peet, and A. Van Proeyen, Phys. Rev. [**D46**]{}, 5278 (1992). K.P. Tod, Phys. Lett. [**121B**]{}, 241 (1983). R. Kallosh and T. Ortín, [*Killing spinor identities*]{}, Stanford University Report SU-ITP-93-16, hepth/9306085. R. Kallosh and T. Ortín, [*Supersymmetry, trace anomaly and naked singularities*]{}, Isaac Newton Institute Report NI94002, Queen Mary and Westfield College Report QMW-PH-94-7, Stanford University Report SU-ITP-94-8, hepth/9404006. G. W. Gibbons, Nucl. Phys. [**B207**]{}, 337 (1982);\
G. W. Gibbons and K. Maeda, Nucl. Phys. [**B298**]{}, 741 (1988);\
D. Garfinkle, G. Horowitz and A. Strominger, Phys. Rev. [**D43**]{}, 3140 (1991). E. Bergshoeff, I. Entrop and R. Kallosh, [*Exact duality in string effective action*]{}, University of Groningen Report UG-8/93, Stanford University Report SU-ITP-93-37, and hepth/9401025.
[^1]: E-mail address: [[email protected]]{}
[^2]: It does not even imply that we have a “good" field configuration. A set of fields $\lambda,F_{\mu\nu},G_{\mu\nu},g_{\mu\nu}$ may admit Killing spinors and still it may not exist a vector fields $A_{\mu},B_{\mu}$ (which are the true dynamical fields) such that $F_{\mu\nu}=\partial_{\mu} A_{\nu} -
\partial_{\nu} A_{\mu}$ and $G_{\mu\nu}=\partial_{\mu} B_{\nu} -
\partial_{\nu} B_{\mu}$. See Ref. [@kn:Tod] for clear examples. In what follows we are going to exclude this possibility, so we will always have $\nabla^{\mu}{}^{*}F_{\mu\nu}=\nabla^{\mu}{}^{*}G_{\mu\nu}=0$.
[^3]: $SL(2,R)$ does not act irreducibly on this family. In particular, as we are going to see, this family contains backgrounds with different numbers of unbroken supersymmetries, and, as a consequence of the results of the previous section, the corresponding subfamilies provide smaller representations of the duality group. Further subdivisions are labeled by non-duality equivalent values of parameters like the asymptotic value of asymptotic of $\lambda$. We won’t pursue this issue farther in this letter.
[^4]: One might naively think that it would suffice to use the result of the first section, performing a duality transformation that would take us to purely electric or magnetic multi-black-hole solutions. However, these solutions were found through a duality rotation of multi-black-hole solutions with 2 electric and 2 magnetic charges but no axion. Thus, we could get rid of the axion charge at most. That simplification is not enough, and perhaps, a better expression for these solutions is need here.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the Semantic Robot Programming ([*SRP*]{}) paradigm as a convergence of robot programming by demonstration and semantic mapping. In [*SRP*]{}, a user can directly program a robot manipulator by demonstrating a snapshot of their intended goal scene in workspace. The robot then parses this goal as a scene graph comprised of object poses and inter-object relations, assuming known object geometries. Task and motion planning is then used to realize the user’s goal from an arbitrary initial scene configuration. Even when faced with different initial scene configurations, [*SRP*]{} enables the robot to seamlessly adapt to reach the user’s demonstrated goal. For scene perception, we propose the Discriminatively-Informed Generative Estimation of Scenes and Transforms ([*DIGEST*]{}) method to infer the initial and goal states of the world from RGBD images. The efficacy of [*SRP*]{} with [*DIGEST*]{} perception is demonstrated for the task of tray-setting with a Michigan Progress Fetch robot. Scene perception and task execution are evaluated with a public household occlusion dataset and our cluttered scene dataset.'
author:
- 'Zhen Zeng Zheming Zhou Zhiqiang Sui Odest Chadwicke Jenkins [^1]'
title: '**Semantic Robot Programming for Goal-Directed Manipulation in Cluttered Scenes** '
---
Introduction
============
Many service robot scenarios, such as setting up a dinner table or organizing a shelf, require a computational representation of a user’s desired world state. For example, how is the dinner table to be set, or how is the shelf to be organized. More specifically, what are the objects involved in the task, what are the desired poses of those objects, and what are the important spatial relationships between objects. Towards natural and intuitive modes of human-robot communication, we present the [**Semantic Robot Programming ([*SRP*]{})**]{} paradigm for declarative robot programming over user demonstrated scenes. In [*SRP*]{}, we assume a robot is capable of goal-directed manipulation [@Suietal2017ijrr] for realizing an arbitrary scene state in the world. A user can program such goal-directed robots by demonstrating their desired goal scene. assumes such scenes can be perceived from partial RGBD observations, which has proven a challenging problem in itself.
Goal-directed manipulation requires a true closing of the loop between perception and action, beyond the existing intellectual silos. Advances in object detection [@girshick2014rich; @ren2015faster] from appearance has improved greatly in filtering of background noise and focused attention to objects of interest. However, the applicability of such vision-based methods robot perception remains unclear, especially for the purposes of goal-directed manipulation. This circumstance has given rise to new approaches to semantic mapping [@kuipers2000spatial; @rusu2008towards; @herbst2014toward] to computationally model a robot’s environment into perceivable objects with robot-actionable affordances.
We posit semantic mapping offers a springboard to new forms of robot programming, such as Semantic Robot Programming, where semantic maps provide a generic abstraction layer for robot programming. In our approach to this problem, we must bridge the gap of interoperation between semantic mapping and existing methods for goal-directed task planning [@fikes1972strips; @laird1987soar], grasp planning [@ten2016localizing] and motion planning [@ompl]. We have previously proposed methods for scene estimation [@sui2015axiomatic; @Suietal2017ijrr] from robot RGBD sensing that used scene graphs expressed axiomatically as a semantic mapping abstraction. This abstraction allowed for ready use with modern task, grasp, and motion planning systems. The resulting of closing this loop with a semantic abstraction layer is envisioned to enable portable robot-executable expressions accessible across a variety of modalities, including: natural language, visual programming, and put-that-there gesturing [@cannon1993point; @kemp2008point]. However, the computational cost of inference over scenes is asymptotically intractable as the number of objects grows.
![A robot preparing a tray through goal-directed manipulations. Given the observation of the user desired goal state and the initial state of the tabletop workspace, the robot first perceives the axiomatic scene graph of the goal and initial state, and then plan and execute goal-directed actions to prepare the tray the way the user desires.[]{data-label="fig:illustration"}](illustration.png){width="\linewidth"}
In this paper, we propose the paradigm of Semantic Robot Programming for robot manipulators with a complementary method for more tractable scene perception. is a declarative approach to programming robots through demonstration, where users only need to demonstrate their desired state of the world. is general across methods of perception, given the perceived scene is represented axiomatically. For scene perception, we present the Discriminatively-Informed Generative Estimation of Scenes and Transforms ([*DIGEST*]{}) method to infer the initial and goal scene states for from RGBD images. [*DIGEST*]{} brings together discriminative object detection and generative pose estimation for inference of 6 DOF object poses in cluttered scenes, assuming the number of objects is known. Given perceived initial and goal scenes, the robot can plan and execute goal-directed manipulation to autonomously transit the world from the initial to the goal state.
We evaluate the paradigm in tray-setting task scenario with the Michigan Progress Fetch robot (Figure \[fig:illustration\]). We benchmark the performance of [*DIGEST*]{} on a household occlusion dataset [@aldoma2012point] and our cluttered scene dataset. We demonstrate that is effective in understanding the goal of a task given a demonstrated snapshot of the goal scene. And, the robot is able to plan and execute goal-directed manipulation actions to reach the goal from various initial states of the world. We additionally found [*DIGEST*]{} performs favorably in comparison with state-of-the-art methods for scene perception, such as D2P [@Narayanan-RSS-16], with fewer assumptions of prior knowledge.
Related Work
============
builds on much existing work in robot Programming by Demonstration (PbD) and scene perception for manipulation. Similar to robot PbD, aims to enable users to effectively communicate their objectives to robots for performing manipulation tasks. We posit advances in scene perception for manipulation offers new avenues for extending the ease and intuitiveness of robot PbD.
Programming by Demonstration
----------------------------
To improve communication of tasks from a user to a service robot, existing research has focused on learning low-level skills from users. Different approaches have been proposed in Programming by Demonstration (PbD) for low-level learning of skills, such as trajectories [@nakanishi2004learning] [@akgun2012keyframe] and control policy [@chernova2009interactive] [@grollman2010incremental] in robot *configuration space*. These methods are inherently limited to world states in *workspace* that are similar to the ones in the demonstrations. By representing the goal of a task in the *workspace* instead of in the *configuration space*, goal-directed manipulation can reason and plan its actions to reach the goal from arbitrary initial world states.
Other work has focused on the high-level aspects of a task. Veeraraghavan et al. [@veeraraghavan2008teaching] propose learning high level action plan for a repetitive ball collection task from demonstrations. Ekvall et al. [@ekvall2008robot] focus on learning task goals and use a task planner to reach the goal. Chao et al. [@chao2011towards] provide an interface for the user to teach task goals in a tabletop workspace. However, these methods wind up simplifying the scene perception problem by using planar objects, box-like objects or objects with distinguishing colors, that are far from real world scenarios. Recently, Yang et al. [@yang2015robot] have proposed learning action plans in real world scenario, similar to our robot programming paradigm that works with real world objects.
Scene Perception for Manipulation
---------------------------------
Being able to perceive objects in real world scenarios and act on them remains a challenge. Some works are able to extract grasping point [@ciocarlie2014towards; @lenz2015deep; @ten2015using] in point cloud data, however, their methods do not provide a structural understanding of the scene, failing to support goal-directed manipulation on objects.
Although not directly targeted at scene perception for manipulation, work on object pose estimation are highly related to our work. Feature-based object pose estimation methods suh as spin images [@johnson1999using], FPFH [@rusu2009fast], OUR-CVFH [@aldoma2012our] and VFH [@rusu2010fast], rely on feature matching between the object model and observation, however, the problem is that the performance of feature-based methods degrades as the environment becomes more cluttered and key features are occluded. Recently, Narayanan et al. proposed D2P [@Narayanan-RSS-16], which outperforms feature-based method OUR-CVFH on the household occlusion dataset [@aldoma2012point]. D2P renders multiple scene hypotheses, and use A\* to search for the hypothesis that best explains the observation. In our experiments, we demonstrate that our proposed scene estimation method [*DIGEST*]{} outperforms D2P on the household occlusion dataset.
To plan goal-directed manipulations, knowing the object poses is not sufficient, however. The robot must have a structural understanding of the scene, that is, the inter-object spatial relations. Given observations of the scene, our work estimates a scene graph that represent the scene structure. Liu et al. [@liu2015table] also estimate a scene graph given observations, however, their approach approximates objects as oriented bounding boxes. Sui et al. proposed a generative approach (AxMC) [@Suietal2017ijrr] for scene graph estimation and use Markov Chain Monte Carlo (MCMC) to search for the best scene graph hypothesis that explains the observations.
Both D2P and AxMC assume that the robot knows what objects are present in the scene, and objects are standing in their upright poses, thus both methods can only estimate 3 DOF poses of objects (i.e., $x, y, \theta$). However, these assumptions are too strong in real world scenarios. Instead, our scene estimation method [*DIGEST*]{} does not rely on any of these assumptions, and it can estimate 6 DOF poses of objects, as long as the number of objects in the scene is known.
{width="\linewidth"}
Problem Statement
=================
with assumes the number of objects $N_c$ present in the scene, 3D mesh models $\mathbf{M}=\{m_1,\cdots,m_l\}$ for a set of objects. The robot is assumed capable of performing a set of manipulation actions $\mathbf{A}=\{a_1,\cdots,a_n\}$ with known pre-conditions and post-conditions on these objects. We assume as given RGB-D observation of the goal scene $o_G$ specified by the user at time $t$, and the current scene $o_I$ at a later time $t+T$. The objective of is to plan a sequence of goal-directed manipulation actions $\{a_i,\cdots,a_j\}$ to rearrange objects in the world such that the inter-object relations in $s_G$ are satisfied; where infers the goal scene graph $s_G$ and the initial scene graph $s_I$, respectively.
We use a list of axiomatic assertions to describe a scene as a scene graph. The scene state at time $t$ is expressed as a scene graph $s_t=\{v^i(x_t)\}_{i=0}^K$, where $v^i \in \{exist, clear, on, in\}$ is an axiomatic assertion parameterized by $x_t=\{q^j_t\}_{j=0}^{N_c}$, with $q^j_t$ denoting the pose of $j$th object at time $t$, $N_c$ being the number of objects, and $K$ being the total number of axiomatic assertions. In our work, the assertions are limited to spatial relations that can be tested geometrically. The 6 DOF pose $q^j_t=[x^j_t,y^j_t,z^j_t, \phi^j_t, \psi^j_t, \theta^j_t]$ of each object is estimated, consisting 3D position ($x^j_t, y^j_t, z^j_t$) and orientation ($\phi^j_t, \psi^j_t, \theta^j_t$). The scene graph can be inferred from the estimated object poses, as explained later in Section \[sec:scene\_graph\_structure\].
Methods
=======
The paradigm consists of the perception of goal and initial scene states, and the planning and execution stages, as shown in Figure \[fig:pipeline\]. Given observations of a cluttered scene, the generative sampling inference process over object poses is informed by detections from a discriminative object detector. A scene graph encoding inter-object relations is geometrically inferred from an estimate of inferred object poses. The resulting scene graph is then expressed axiomatically for use in task planning and execution.
DIGEST Cluttered Scene Estimation
---------------------------------
Given observed RGB-D image pair of a cluttered scene at time $t$, the objective is to estimate the object poses $q_t^j, j=1,\cdots,N_c$. We utilize the discriminative power of a pre-trained object detector to first obtain a set of bounding boxes with object labels. These bounding boxes are used to guide the generative process of scene hypotheses sampling. An overview of the cluttered scene estimation is as illustrated in Figure \[fig:framework\].
### Object Detection and Scene Hypotheses Generation {#sec:object_detection_method}
Given an RGB image, $m$ bounding boxes are detected by the object detector. We use $B_i$ ($0 \leq i \leq m$) to denote the bounding box. In the output of the object detector, each $B_i$ is associated with a list of object detection confidence $v(L_j | B_i)$, where $L_j$ is the object class. For each $B_i$, we generate an object candidate $C_i$, $$\label{eq:candidate}
C_i = \{ \underset{L_j}{\arg\max} \, v( L_j | B_i ), \, B_i \}$$ which is a set including the object label with the highest confidence measure and the associated bounding box. For $m$ generated candidates, the number of scene hypotheses $h$ equals to $N_c$ chooses $m$, i.e., $$h =
\begin{cases}
{{}^{m}C_{N_c}}, & \text{if}\ N_c \leq m \\
1, & \text{otherwise}
\end{cases}$$ Thus, if the number of candidates is greater or equal to the number of objects in the scene, each scene hypothesis ${H_i}$ contains a combination of $N_c$ candidates selected from $m$ candidates. If the number of candidates is less than $N_c$, just one scene hypothesis with $m$ candidates will be generated.
{width="0.9\linewidth"}
### Bootstrap Filtering for Pose Estimation {#sec:bootstrap_filter_method}
Each scene hypothesis ${H_i}$ is modeled as a random state variable $x_t$, composed of a set of real-valued object poses. Object poses are assumed to be statistically independent. We model the inference of the state from robot observation as a Bayesian filter problem. Compared to traditional Bayesian filter problems, we have only one observation: a snapshot of the scene instead of a history of observations. Thus, we apply Iterated Likelihood Weighting [@mckenna2007tracking] to bootstrap the scene estimation process, where $z_1=z_2=\cdots=z_t$ and the state transition in the action model is replaced by a zero-mean Gaussian noise. We approximate the belief distribution by a collection of $N$ particles $\{x^{(j)}_t$ weighted by $w^{(j)}_t\}^{N}_{j=1}$, $$\label{eq:particle}
p( x_{t} | z_{1:{t}} )
\propto
p( z_{t} | x_{t} )\sum_{j} w^{(j)}_{t-1}p( x_{t} | x^{(j)}_{t-1}, u_{t-1})$$ $$x_t^{(j)} \sim \sum_{j} w^{(i)}_{t-1}p( x_{t} | x^{(i)}_{t-1}, u_{t-1})$$ as described by [@Dellaert_MCL]. To evaluate the weight $w_t^{(j)}$ for particle $x_t^{(j)}$, we render a depth image based on the object poses in $x_{t}^{(j)}$, and compare it against the observed depth image $\hat{z}_{t}^{(j)}$, $$w_t^{(j)} = e^{- \lambda_r \cdot \mathrm{d}(z, \hat{r}_{t}^{(j)})}$$ where $\lambda_r$ is a constant scaling factor. $\mathrm{d}(z,\hat{r}_{t}^{(j)})$ is the sum of the Euclidean distance between the 3D points projected back from depth images $z$, $\hat{r}_{t}^{(j)}$, using the intrinsic parameters of the camera. Pose estimation is performed over successive iterations that: 1) compute the weight of each particle, 2) normalize the weights to one, 3) draw $N$ particles by importance sampling, and 4) diffuse each sampled particle by a zero-mean Gaussian noise. After maximum number of iterations, the most likely particle as the scene estimate for scene hypothesis $H_i$:
$$x_t = \underset{{x}_t^{(j)}} {\arg\max} \, p(x_t^{(j)}|z_{1:t})$$
### Final Scene Ranking
After particle filtering for all scene hypotheses, we have a scene estimate $x_t$ for each scene hypothesis. We then rank them based on the likelihood of each $x_t$ as computed earlier. The most likely $x_t$ is taken as the scene estimate and is then used to derive the scene graph.
{width="0.9\linewidth"}
Scene Graph Structure {#sec:scene_graph_structure}
---------------------
The objects pose estimation of a cluttered scene can be turned into an axiomatic scene graph. We use following axiomatic assertions: $exist(q^j)$ for the assertion that object $j$ exists in the scene with pose $q^j$; $clear(q^i)$ for the assertion that the top of object $i$ is clear and no other objects are stacked on it; $on(q^i,q^j)$ for the assertion that object $i$ is stacked on object $j$; $in(q^i,q^j)$ for the assertion that object $i$ is in object $j$. An example of a scene graph is given in Figure \[fig:scene\_graph\_pddl\].
To assert the proximity relations between two objects $i, j$, we add a *virtual object* $q^\gamma$ with geometry $m^\gamma$ into the scene graph, with $m^\gamma$ being a shape that can be arbitrarily defined based on the application, and $q^\gamma$ being the identity pose in the frame of object $i$. Then, the proximity relation between objects $i,j$ can be encoded by $\{has(q^i,q^\gamma),\ in(q^\gamma,q^j)\}$, where $has(q^i,q^\gamma)$ asserts that object $i$ has a *virtual object* $q^\gamma$ attached to its frame. When the parent object $i$ is in a new location, the robot can adapt to the new scenario by placing the child object $j$ within the region of $m^\gamma$ attached to the frame of $i$.
To determine the stacking relations between the objects, we use simple heuristics. In the 3D mesh object models, the z-axis of each object is the gravitational axis when the object stands upright. The dimensions $\{h_x, h_y, h_z\}$ of the 3D box that encloses each object model are given as prior knowledge. In order to determine whether object $i$ is being supported by another object, two heuristics are tested: (1) if one of the object axes (e.g., x-axis) is aligned with the gravitational axis, then the height $h_i$ of the 3D volume occupied by the object equals to the corresponding dimension (e.g. $h_x$) of the provided 3D enclosing box. A simple rule $z^i - h^{table} > 0.5h^i$ is used to determine whether object $i$ is being supported by another object; (2) if none of the object axes are aligned with the gravitational axis, then object $i$ is being supported by another object.
The set of objects that is being supported by other objects is sorted with increasing $z$ values of the object pose, and is denoted as $O_{\text{s}}$, the remaining objects are denoted as $O_{\text{r}}$. For each object $i \in O_{\text{s}}$, a heuristic measure is used to determine which object $j \in O_{\text{r}}$ is supporting $i$, $$\operatorname{arg\,max}_j f(r_b(q^i),r_t(q^j))$$ where $f(r_1,r_2)$ measures the overlapping area of two regions $r_1, r_2$, and $r_t(q^i), r_b(q^i)$ represent the projected region on the table of the top and bottom surface of object $i$, respectively. Once the supporting object for $i \in O_{\text{s}}$ is identified, $i$ is moved from set $O_{\text{s}}$ to $O_{\text{r}}$. With the supporting relation between a pair of objects $i, j$ identified, the corresponding axiomatic assertion is expressed as either $on(q^i, q^j)$ or $in(q^i, q^j)$, depending on the geometry type of the supporting object $j$ being convex or concave.
Implementation
==============
RCNN object detector
--------------------
We employ R-CNN [@girshick14CVPR] as our discriminative object detector as described in section \[sec:object\_detection\_method\]. R-CNN first generates object bounding boxes given an image, then for each bounding box, it outputs the confidence measure through a deep convolutional neural network. For the sake of efficiency and performance, we replace the original selective search [@uijlings2013selective] with EdgeBox [@zitnick2014edge] for object proposal generation. We train an R-CNN object detector on our object dataset that includes 15 grocery objects. The dataset contains 8366 ground truth images (557 average ground truth images for one object) and 60563 background images. We fine tuned our object detector on a pre-trained model on ImageNet [@imagenet_cvpr09].
Particle Filtering and parallelization
--------------------------------------
During bootstrap filtering for pose estimation as described in section \[sec:bootstrap\_filter\_method\]. Each object in each particle $x_t^{(j)}$ is initialized by candidate $C_i$ in the scene hypothesis, the object label $l_i$ determines which 3D mesh object model to use, and the initial pose is uniformly sampled inside the bounding box $B_i$. A parallel graphics engine rapidly renders depth images given all particles. CUDA is used to compute the weights of all particles in parallel. Through our experiment, we fix particle filter iteration to 400 and use 625 particles.
In the particle filtering process, the pose of each object is estimated sequentially. For example, if there are four hypothesized objects and 400 particle filter iterations, the pose of the object with the maximum detection confidence is estimated in the first 100 iterations. Then the pose of the object with the 2nd largest detection confidence is estimated in the next 100 iterations, with the first object fixed at the most likely pose. We carry on the estimation process iteratively for the remaining objects.
Planning and Execution
----------------------
Given the observation of the goal state of the world, the robot estimates the goal scene graph, and stores the desired inter-object relations by PDDL [@mcdermott1998pddl]. Similarly, the robot estimates and stores the initial inter-object relations by PDDL. With sets of PDDL that describe the initial and goal state, the robot uses a task planner to plan a series of goal-directed actions to rearrange objects in the initial scene, such that the same inter-object relations in the goal scene graph are satisfied. We use breadth first search STRIPS[@fikes1971strips] as our task planner. Note that the robot does not need to rearrange the objects with the exact same poses as in the goal scene, as long as the same inter-object relations are achieved, similarly to how human would arrange a set of daily objects based on simple instructions.
The task planner gives a sequence of high-level pick-and-place actions. To pick an object, the robot is given a set of pre-computed grasp poses of the object using [@ten2015using], and uses Moveit! [@sucan2013moveit] to check which grasp pose it can generate a collision-free trajectory for, and use that for grasping. To place an object, the robot sample place poses in the empty space that satisfies the desired inter-object relations, and again use the place pose it can generate a collision-free trajectory for.
Experiments
===========
In our experiments, we first evaluate our scene estimation method on a public household occlusion dataset and our cluttered scene dataset, and then evaluate our overall semantic robot programming paradigm in tray setting tasks. [*DIGEST*]{} outperforms the state of the art method D2P on the household occlusion dataset, and outperforms FPFH on our cluttered scene dataset. We demonstrate the effectiveness of our system for programming a robot to complete various tray-setting tasks through goal-directed manipulations. We run all experiments on a computer with an Titan X Graphics card and CUDA 7.5.
DIGEST: Cluttered Scene Estimation
----------------------------------
To evaluate [*DIGEST*]{} on pose estimation, we benchmarked the performance of [*DIGEST*]{} on two different datasets: household occlusion dataset [@aldoma2012point], and our cluttered scene dataset. The household occlusion dataset contains objects standing up right, thus it only affords benchmarking on 3 DOF object pose estimation. In our cluttered scene dataset, objects can be in arbitrary pose, and we use it for benchmarking on 6 DOF object pose estimation. Object pose estimation accuracy is calculated as the percentage of correctly localized objects over the total number of objects in the dataset. An object is correctly localized if the pose error falls within certain position error threshold $\Delta t$ and rotation error threshold $\Delta \theta$. The position error is the Euclidean distance error in translation; the rotation error is the absolute angle error in orientation. For rotationally symmetric objects, the rotation error about the symmetric axis is ignored.
![Object pose estimation benchmark of on public household object dataset [@aldoma2012point], compared with three baseline methods: D2P, OUR-CVFH and BF-ICP for different correctness criteria $\Delta t$, $\Delta \theta$. [*DIGEST*]{} outperforms D2P for strict correctness criteria, and performs on par with D2P for relaxed correctness criteria.[]{data-label="fig:d2p_progress_exp_pose_err"}](benchmark_all_delta.png){width="\columnwidth"}
### Household Occlusion Dataset – 3 DOF Object Poses
The household occlusion dataset contains 22 test scenes with 80 objects in total. The test scenes include objects such as milk bottles, laundry items, mugs and etc; We compare against three baseline methods as described in [@Narayanan-RSS-16], that is, D2P, OUR-CVFH [@aldoma2012our], and Brute Force ICP (BF-ICP). D2P also uses an R-CNN object detector as part of their pose estimation process, but it is not clear what hyper parameters they choose during the training phase of the object detector. In order to avoid bias in the training of the object detector, we use their object detector on the household occlusion dataset.
![Object pose estimation benchmark of on our cluttered scene dataset, compared with baseline method FPFH under different correctness criteria $\Delta t$, $\Delta \theta$. [*DIGEST*]{} outperforms FPFH with large margin.[]{data-label="fig:digest_fpfh"}](benchmark_all_delta_6dof.png){width="\columnwidth"}
When only little error is allowed for an estimated pose to be counted as correct, as shown in the left upper plot in Figure \[fig:d2p\_progress\_exp\_pose\_err\], the accuracy of is nearly twice the accuracy of D2P. As we relax the tolerance on the pose estimation error, as shown in the other three plots in Figure \[fig:d2p\_progress\_exp\_pose\_err\], performs on par with D2P. Overall, outperforms D2P since (1) explores the state space a lot more than D2P, as we do not discretize the state space, and (2) does not use ICP for local search, which D2P employs for pose estimation. In terms of run time, takes around 30 seconds (varying with the number of objects and the size of object mesh), which is faster than 139.74 seconds reported in D2P.
### Cluttered Scene Dataset – 6 DOF Object Poses
We collect a cluttered scene dataset with 16 different sceness, and 72 objects in total. This dataset includes laundry items, kitchen items and toy with non-trivial geometry. The number of objects in each scene ranges from 3 to 7. This dataset is much more challenging than the household object dataset, as the objects can have random 6 DOF poses. We compare the performance of [*DIGEST*]{} with FPFH [@rusu2009fast], as shown in Figure \[fig:digest\_fpfh\].
{width="1.0\linewidth"}
Semantic Robot Programming: Tray Setting
----------------------------------------
We designed our experiments around a service robot scenario, as illustrated in Figure \[fig:illustration\]. The robot needs to prepare a tray as specified by the user int the goal scene. We tested our system on scenes of 4 to 6 objects including the tray, with different inter-object relations, such as stacking and proximity relations. The robot is able to perceive the initial and goal state, then plan and execute goal-directed actions to satisfy the inter-object relations in the goal scene graph.
An example of for goal-directed manipulation is shown in Figure \[fig:manipulation\_exp\]. Based on the scene graph inferred from the object pose estimates, the robot generates a sequence of goal-directed actions to achieve the goal state. Our tray setting experiments are shown in Figure \[fig:table\_digest\], and more detail in the accompanying video[^2]. The goal and start scenes are well estimated as a collection of 6DOF poses of objects. The robot successfully sets up a tray as the user desired in 10 out of 10 different tray setting experiments.
Conclusion
==========
We have presented Semantic Robot Programming as a paradigm for users to easily program robots in a declarative goal-directed manner. We demonstrate the effectiveness of using the proposed scene perception method on two datasets of objects in occlusion and clutter: both house occlusion dataset and our cluttered scene dataset. Through our approach to generative-discriminative perception, with is able to perceive, reason, and act to realize an arbitrary user-demonstrated goal in cluttered scenes.
There are many future directions to pursue, such as motion planning over sequences of general manipulation actions. Currently, grasp point localization [@ten2015using] is used to select good grasp poses for object picking. However, such selected grasp poses are not necessarily appropriate for a later placement actions. Visual inspection on selected grasps is done before robot execution. Ideally, appropriate grasp poses would be provided by a manipulation affordance mechanism, such as Affordance Templates [@hart2015affordance] associating robot action with an object. Such affordance mechanisms would allow for investigation of more flexible task and motion planning over sequences of actions. We further posit scene perception can be made to run in interactive-time through a thoughtful parallelized implementation, enabling potentially interactive planning and manipulation execution.
{width="\linewidth"}
[10]{}
B. Akgun, M. Cakmak, K. Jiang, and A. L. Thomaz. Keyframe-based learning from demonstration. , 4(4):343–355, 2012.
A. Aldoma, Z.-C. Marton, F. Tombari, W. Wohlkinger, C. Potthast, B. Zeisl, R. B. Rusu, S. Gedikli, and M. Vincze. Point cloud library. , 1070(9932/12), 2012.
A. Aldoma, F. Tombari, R. B. Rusu, and M. Vincze. Our-cvfh–oriented, unique and repeatable clustered viewpoint feature histogram for object recognition and 6dof pose estimation. In [*Joint DAGM (German Association for Pattern Recognition) and OAGM Symposium*]{}, pages 113–122. Springer, 2012.
D. J. Cannon. Point-and-direct telerobotics: Object level strategic supervisory control in unstructured interactive human-machine system environments. 1992.
C. Chao, M. Cakmak, and A. L. Thomaz. Towards grounding concepts for transfer in goal learning from demonstration. In [*2011 IEEE International Conference on Development and Learning (ICDL)*]{}, volume 2, pages 1–6. IEEE, 2011.
S. Chernova and M. Veloso. Interactive policy learning through confidence-based autonomy. , 34(1):1, 2009.
M. Ciocarlie, K. Hsiao, E. G. Jones, S. Chitta, R. B. Rusu, and I. A. [Ş]{}ucan. Towards reliable grasping and manipulation in household environments. In [*Experimental Robotics*]{}, pages 241–252. Springer, 2014.
F. Dellaert, D. Fox, W. Burgard, and S. Thrun. Monte carlo localization for mobile robots. In [*IEEE International Conference on Robotics and Automation (ICRA 1999)*]{}, May 1999.
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. . In [*CVPR09*]{}, 2009.
S. Ekvall and D. Kragic. Robot learning from demonstration: a task-level planning approach. , 5(3):33, 2008.
R. E. Fikes and N. J. Nilsson. Strips: A new approach to the application of theorem proving to problem solving. , 2(3-4):189–208, 1971.
R. E. Fikes and N. J. Nilsson. Strips: A new approach to the application of theorem proving to problem solving. , 2(3):189–208, 1972.
R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In [*Proceedings of the IEEE conference on computer vision and pattern recognition*]{}, pages 580–587, 2014.
R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In [*Computer Vision and Pattern Recognition*]{}, 2014.
D. H. Grollman and O. C. Jenkins. Incremental learning of subtasks from unsegmented demonstration. In [*Intelligent Robots and Systems (IROS), 2010 IEEE/RSJ International Conference on*]{}, pages 261–266. IEEE, 2010.
S. Hart, P. Dinh, and K. Hambuchen. The affordance template [ROS]{} package for robot task programming. In [*2015 IEEE International Conference on Robotics and Automation (ICRA)*]{}, pages 6227–6234. IEEE, 2015.
E. Herbst, P. Henry, and D. Fox. Toward online 3-d object segmentation and mapping. In [*Robotics and Automation (ICRA), 2014 IEEE International Conference on*]{}, pages 3193–3200. IEEE, 2014.
A. E. Johnson and M. Hebert. Using spin images for efficient object recognition in cluttered 3d scenes. , 21(5):433–449, 1999.
C. C. Kemp, C. D. Anderson, H. Nguyen, A. J. Trevor, and Z. Xu. A point-and-click interface for the real world: laser designation of objects for mobile manipulation. In [*Human-Robot Interaction (HRI), 2008 3rd ACM/IEEE International Conference on*]{}, pages 241–248. IEEE, 2008.
B. Kuipers. The spatial semantic hierarchy. , 119(1-2):191–233, 2000.
J. E. Laird, A. Newell, and P. S. Rosenbloom. Soar: An architecture for general intelligence. , 33, 1987.
I. Lenz, H. Lee, and A. Saxena. Deep learning for detecting robotic grasps. , 34(4-5):705–724, 2015.
Z. Liu, D. Chen, K. M. Wurm, and G. von Wichert. Table-top scene analysis using knowledge-supervised mcmc. , 33:110–123, 2015.
D. McDermott, M. Ghallab, A. Howe, C. Knoblock, A. Ram, M. Veloso, D. Weld, and D. Wilkins. -the planning domain definition language. 1998.
S. J. McKenna and H. Nait-Charif. Tracking human motion using auxiliary particle filters and iterated likelihood weighting. , 25(6):852–862, 2007.
J. Nakanishi, J. Morimoto, G. Endo, G. Cheng, S. Schaal, and M. Kawato. Learning from demonstration and adaptation of biped locomotion. , 47(2):79–91, 2004.
V. Narayanan and M. Likhachev. Discriminatively-guided deliberative perception for pose estimation of multiple 3d object instances. In [*Proceedings of Robotics: Science and Systems*]{}, AnnArbor, Michigan, June 2016.
S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In [*Advances in neural information processing systems*]{}, pages 91–99, 2015.
R. B. Rusu, N. Blodow, and M. Beetz. Fast point feature histograms (fpfh) for 3d registration. In [*Robotics and Automation, 2009. ICRA’09. IEEE International Conference on*]{}, pages 3212–3217. IEEE, 2009.
R. B. Rusu, G. Bradski, R. Thibaux, and J. Hsu. Fast 3d recognition and pose using the viewpoint feature histogram. In [*Intelligent Robots and Systems (IROS), 2010 IEEE/RSJ International Conference on*]{}, pages 2155–2162. IEEE, 2010.
R. B. Rusu, Z. C. Marton, N. Blodow, M. Dolha, and M. Beetz. Towards 3d point cloud based object maps for household environments. , 56(11):927–941, 2008.
I. A. Sucan and S. Chitta. Moveit! 2013. Available: [http://moveit. ros. org](http://moveit. ros. org).
I. A. [Ş]{}ucan, M. Moll, and L. E. Kavraki. The [O]{}pen [M]{}otion [P]{}lanning [L]{}ibrary. , 19(4):72–82, December 2012. <http://ompl.kavrakilab.org>.
Z. Sui, O. C. Jenkins, and K. Desingh. Axiomatic particle filtering for goal-directed robotic manipulation. In [*Intelligent Robots and Systems (IROS), 2015 IEEE/RSJ International Conference on*]{}, pages 4429–4436. IEEE, 2015.
Z. Sui, L. Xiang, O. C. Jenkins, and K. Desingh. Goal-directed robot manipulation through axiomatic scene estimation. , 36(1):86–104, 2017.
A. ten Pas and R. Platt. Using geometry to detect grasp poses in 3d point clouds. In [*Int’l Symp. on Robotics Research*]{}, 2015.
A. Ten Pas and R. Platt. Localizing handle-like grasp affordances in 3d point clouds. In [*Experimental Robotics*]{}, pages 623–638. Springer, 2016.
J. R. Uijlings, K. E. van de Sande, T. Gevers, and A. W. Smeulders. Selective search for object recognition. , 104(2):154–171, 2013.
H. Veeraraghavan and M. Veloso. Teaching sequential tasks with repetition through demonstration. In [*Proceedings of the 7th international joint conference on Autonomous agents and multiagent systems-Volume 3*]{}, pages 1357–1360. International Foundation for Autonomous Agents and Multiagent Systems, 2008.
Y. Yang, Y. Li, C. Ferm[ü]{}ller, and Y. Aloimonos. Robot learning manipulation action plans by “watching" unconstrained videos from the world wide web. In [*AAAI*]{}, pages 3686–3693, 2015.
C. L. Zitnick and P. Doll[á]{}r. Edge boxes: Locating object proposals from edges. In [*European Conference on Computer Vision*]{}, pages 391–405. Springer, 2014.
[^1]: Z. Zeng, Z. Zhou, Z. Sui and O.C. Jenkins are with the Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, USA, 48109-2121 [@umich.edu]{}
[^2]: <http://progress.eecs.umich.edu/projects/srp>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we consider surfaces in 4–dimensional pseudo–Riemannian space–forms with index 2. First, we obtain some of geometrical properties of such surfaces considering their relative null space. Then, we get classifications of quasi–minimal surfaces with positive relative nullity.'
address:
- 'Fatih Sultan Mehmet Vakif University, Faculty of Engineering, Department of Civil Engineering, Beyoğlu, Istanbul, Turkey, '
- 'Istanbul Technical University, Faculty of Science and Letters, Department of Mathematics, 34469 Maslak, Istanbul, Turkey'
author:
- Burcu Bektaş Demirci
- Nurettin Cenk Turgay
title: 'Quasi–minimal surfaces of pseudo–Riemannian space forms with positive relative nullity'
---
Introduction
============
A submanifolds of a pseudo–Riemannian manifold is said to be *quasi–minimal* if its mean curvature is light–like at every point. Since quasi–minimal submanifolds does not exist in Riemannian manifolds, they have taken attention of many geometers so far (See, for example, [@ChenE42Flat; @ChenE42PMCV; @Ganchev; @HaesenOrtega]). When the ambient space is a Lorentzian space–time, quasi–minimal submanifolds are also called *marginally trapped* in physics literature because they are closely related with the concept of trapped surfaces, introduced by Roger Penrose in [@Penrose65].
On the other hand, studying submanifolds by considering their relative null space was initiated by M. Dajczer and D. Gromoll in [@DajczerGro] where they obtained necessary and sufficent conditions for a spherical submanifold to have positive relative nullity. Recently, the complete classification of marginally trapped surfaces with positive relative nullity in Lorentzian space–forms was given by B.-Y.Chen and J. Van der Veken in [@ChenVeken1]. Further, they proved that there exists no quasi–minimal surface with positive relative nullity when the ambient space is a Robertson–Walker space–time with non–constant sectional curvatures, [@ChenVeken2].
In [@Chen], B.-Y.Chen mentioned some results concerning marginally trapped surfaces in Lorentzian space forms and in Lorentzian complex space forms and he also put forward some open problems about classification of such surfaces in a $4$–dimensional pseudo–Riemannian space forms with index $2$.
The main purpose of this paper is studying quasi–minimal surfaces of a $4$–dimensional pseudo–Riemannian space forms with index $2$ from in terms of their relative null spaces. In particular, we obtain the complete local classification of quasi–minimal surfaces in the pseudo–Euclidean space $\mathbb{E}^4_2$ and a pseudo–sphere $\mathbb{S}^4_2$, respectively. In Sect. 2, after we describe the notation that we will use, we give basic facts on quasi–minimal surfaces of a pseudo–Riemannian space forms. In Sect. 3, we present our main results.
Preliminaries
=============
Let $\mathbb E^n_s$ be the pseudo–Euclidean $n$–space defined by $\mathbb E^n_s=(\mathbb R^n, \hat{g})$, where $\hat{g}$ is the canonical metric tensor of index $s$ given by $$\hat{g}=\langle\cdot,\cdot\rangle=-\sum\limits_{i=1}^sdx_i\otimes dx_i+\sum\limits_{j=s+1}^ndx_j\otimes dx_j$$ for a Cartesian coordinate system $(x_1,x_2,\hdots,x_n)$ of $\mathbb R^n$. A non–zero vector $v$ is said to be space–like, light–like or time–like if $\langle v,v\rangle>0$, $\langle v,v\rangle=0$ or $\langle v,v\rangle<0$, respectively.
We put $$\begin{aligned}
\begin{split}
\mathbb S^n_s=&\{{\bf x}\in \mathbb E^{n+1}_s\;|\;
\langle{\bf x},{\bf x}\rangle=1\},\\
\mathbb H^n_s=&\{{\bf x}\in \mathbb E^{n+1}_{s+1}\;|\;
\langle{\bf x},{\bf x}\rangle=-1\}.\\
\end{split}\end{aligned}$$ Then, $\mathbb{S}^n_s$ and $\mathbb{H}^n_s$ are pseudo–Riemannian manifolds of constant sectional curvature $1$ and $-1$ known as a pseudo–sphere and a pseudo–hyperbolic space, respectively. For a non–zero real number $c>0$, we also denote a $n$–dimensional pseudo–Riemannian space form with index $s$ and constant sectional curvature $c$ by $R^n_s(c)$. It is known that $$R^n_s(c)=\left\{
\begin{array}{cl}
\mathbb E^n_s &\mbox{if $c=0$,} \\
\mathbb S^n_s &\mbox{if $c =1$,}\\
\mathbb H^n_s &\mbox{if $c=-1$.}
\end{array}
\right.$$ Let $\widetilde\nabla$ and $\tilde g$ stand for the Levi–Civita connection and the metric tensor of $R^n_s(c)$, respectively.
Pseudo–Riemannian Submanifolds of $R^n_s(c)$
--------------------------------------------
Consider an isometric immersion $f: (\Omega,\check{g})\hookrightarrow R^n_s(c)$ from an $m$–dimensional pseudo–Riemannian manifold $(\Omega,\check{g})$ and put $M=f(\Omega)$ with the metric $g=f^*(\check{g})$. If $\nabla$ denote the Levi–Civita connection of $M$, then for any vector fields $X,\ Y\in TM$ and $\xi\in N^f\Omega$, the Gauss and Weingarten formulas are given, respectively, by $$\begin{aligned}
\label{MEtomGauss}
\widetilde\nabla_X Y&=& \nabla_X Y + \alpha_f(X,Y),\\
\label{MEtomWeingarten}
\widetilde\nabla_X \xi&=& -A^f_\xi(X)+\nabla^\perp_X \xi,\end{aligned}$$ where $N^f\Omega$ stand for the normal bundle of $f$, $\alpha_f$ is the second fundamental form, $A^f_\xi$ is the shape operator along the normal direction $\xi$ and $\nabla^\perp$ is the normal connection of $f$. Also, $A^f_\xi$ and $\alpha_f$ are related by $$\begin{aligned}
\label{MinkAhhRelatedby}
g(A^f_\xi X,Y)=\tilde{g}(\alpha_f(X,Y),\xi).\end{aligned}$$ The mean curvature vector $H$ of $f$ is defined by $H=\frac{1}{m}{\mathrm{\,trace}\,}\;\alpha_f$. Note that $M$ is called quasi–minimal if the mean curvature vector $H$ is a light–like at each point of $M$.
On the other hand, the curvature tensor $R$ of $M$, the normal curvature tensor $R^\perp$ of $f$ and $\alpha_f$ satisfy
$$\begin{aligned}
\label{MinkGaussEquation} R(X,Y)Z&=&
c(X\wedge Y)Z+A^f_{ \alpha_f(Y,Z)}X-A^f_{\alpha_f(X,Z)}Y,\\
\label{MinkCodazzi} (\bar \nabla_X \alpha_f )(Y,Z)&=&(\bar \nabla_Y \alpha_f )(X,Z),\\
\label{MinkRicciEquation} R^{\perp}(X,Y)\xi&=&\alpha_f(X,A^f_\xi Y)-\alpha_f(A^f_\xi X,Y),\end{aligned}$$
which are called Gauss, Codazzi and Ricci equations, respectively, where $X\wedge Y$ and $\bar \nabla \alpha_f$ are defined respectively by $$\begin{aligned}
(X\wedge Y)Z&=&g(Y,Z)X-g(X,Z)Y,\\
(\bar \nabla_X \alpha_f)(Y,Z)&=&\nabla^\perp_X \alpha_f(Y,Z)-\alpha_f(\nabla_X Y,Z)-\alpha_f(Y,\nabla_X Z).\end{aligned}$$
The relative null space of $M$ at a point $p$ is defined by $$\mathcal{N}_p=\{X_p\in T_pM\;|\;\alpha_f(X_p,Y_p)=0
\mbox{ for all $Y_p\in T_pM$ }\}.$$ If the dimension of the relative null space $\mathcal{N}_p$ is non-zero for all $p\in M$, then $M$ is said to have positive relative nullity in $R^n_s(c)$, [@ChenVeken1].
Immersions into $\mathbb{S}^4_2$
--------------------------------
Let $f:(\Omega,\check{g})\hookrightarrow\mathbb{S}^4_2$ be an isometric immersion and $i:\mathbb S^4_2\subset\mathbb E^5_2$ be the inclusion. Call $\hat f=i\circ f$. We denote the Levi–Civita connection of $\mathbb E^5_2$ by $\hat{\nabla}$. Then, we have $$N^{\hat f}\Omega=i_*\left(N^f\Omega\right)\oplus \mathrm{span\,}\{\hat f\}$$ and shape operators of $f$ and $\hat f$ satisfy $
A^f_\eta=A^{\hat f}_{i_*\eta}\mbox{whenever $\eta\in N^f\Omega$}$ and $
A^{\hat f}_{\hat f}=-\mathrm{I},
$ where $\mathrm{I}$ is the identity. Moreover, the second fundamental forms of $f$ and $\hat f$ are related by $$\nonumber
\alpha_{\hat f}(X,Y)= i_*\left(\alpha_{f}(X,Y)\right)- g (X,Y)\hat f$$ whenever $X,Y\in TM$. Therefore, we have $$\label{nablaSrelatedby}
\hat\nabla_XY=i_*\left(\widetilde\nabla_XY\right)-g(X,Y)\hat f.$$
Surfaces with positive relative nullity in $R^4_2(c)$
=====================================================
In this section, we obtain complete local classification of quasi–minimal surface with positive relative nullity in a $4$–dimensional pseudo–Riemannian space forms. Throughout this section, a surface $M$ of $R^4_2(c)$ is defined by $M=f(\Omega)$ and we put $g=f^*(\check g)$ where $f:(\Omega,\check{g})\hookrightarrow(R^4_2(c),\tilde{g})$ is an isometric immersion and $c\in\{-1,0,1\}$.
\[lemma1\] Let $M$ be a quasi–minimal surface with positive relative nullity in a pseudo–Riemannian space form $R^4_2(c)$. Then, at each point $p\in M$, there exists an orthonormal basis $\{e_1,e_2\}$ for the tangent space of $M$ and a pseudo–orthonormal basis $\{e_3,e_4\}$ for the normal space of $M$ such that $$\begin{aligned}
\label{chosenbase}
\begin{split}
g(e_1,e_1) =-g(e_2,e_2)=\varepsilon, \quad &g(e_1,e_2)=0,\\
\tilde{g}(e_3,e_3)=\tilde{g}(e_4,e_4)=0,\quad &\tilde{g}(e_3,e_4)=-1,\\
\alpha_f(e_1,e_1)=\alpha_f(e_1,e_2)=0, \quad&\alpha_f(e_2,e_2)=e_3.
\end{split}\end{aligned}$$
Assume that $M$ is a quasi–minimal surface with positive relative nullity in $R^4_2(c)$. If $\dim\mathcal N_p=2$ for a $p\in M$, then $\alpha_f$ vanishes at $p$ which implies $H_p=0$. However, this is a contradiction because $M$ is quasi–minimal. Therefore, we have $$\dim\mathcal N_p=1 \quad \mbox{for all $p\in M$}.$$
On the other hand, if $\mathcal N_p$ is degenerated, i.e., $\mathcal N_p={\mathrm{span\,}}\{X_p\}$ for a light–like vector $X_p\in T_pM$, then we have $\alpha_f(X_p,Y_p)=0$ for any $Y_p\in T_pM$. If $Y_p$ is chosen to be a unique light–like tangent vector at $p$ such that $g_p(X_p,Y_p)=-1$, then we obtain $H_p=-\alpha_f(X_p,Y_p)=0$ which yields another contradiction. Consequently, there exists a tangent vector field $e_1\in \mathcal N_p$ with $g(e_1,e_1)=\varepsilon\in\{-1,1\}$. Let $e_2$ be a unit vector field orthogonal to $e_1$, which implies $g(e_2,e_2)=-\varepsilon,\ g(e_1,e_2)=0$. Since $e_1\in \mathcal N_p$, we have $\alpha_f(e_1,e_1)=\alpha_f(e_1,e_2)=0$ and $\dim\mathcal N_p=1$ implies $\alpha_f(e_2,e_2)\neq 0$. Now, we define a light–like vector field $e_3$ by $$e_3=-2\varepsilon H$$ and choose $e_4$ as the unique light–like vector field normal to $M$ such that $\tilde g(e_3,e_4)=-1$. Then, we obtain $\alpha_f(e_2,e_2)=e_3.$ Hence, we have obtained all of conditions appearing in .
Let us assume that $\{e_1, e_2\}$ is an orthonormal frame and $\{e_3, e_4\}$ is a pseudo–orthonormal frame on the quasi–minimal surface $M$ in $R^4_2(c)$ which satisfy the equation . With respect to chosen frame field $\{e_1,e_2,e_3,e_4\}$, we have $$\begin{aligned}
\label{levicon_1_R42}
\nabla_{e_i}e_1&=-\varepsilon\omega_{12}(e_i)e_2,\;\;\; \nabla_{e_i}e_2=-\varepsilon\omega_{12}(e_i)e_1\\
\label{levicon_2_R42}
\nabla^{\perp}_{e_i}e_3&=\phi(e_i)e_3,\;\;\;
\nabla^{\perp}_{e_i}e_4=-\phi(e_i)e_4.\end{aligned}$$ From now on, we denote $\phi(e_1)=\omega$ and $\phi(e_2)=\gamma$.
\[SurfR42cProp1\] Let $M$ be a quasi–minimal surface of a pseudo–Riemannian space forms $R^4_2(c)$ with positive relative nullity. Then, there exists a local coordinate system $(s,t)$ defined on a neighborhood of $p\in M$ such that the induced metric tensor $g$ of $M$ takes the form $$\label{SurfR42cProp1Eq1}
g=\varepsilon(ds\otimes ds-\phi^2dt\otimes dt), \qquad \varepsilon=\pm 1.$$ Moreover, the vector fields $e_1=\frac{\partial}{\partial s}$ and $e_2=\frac 1\phi\frac{\partial}{\partial t}$ satisfy
\[SurfR42cProp1Eq2ALL\] $$\begin{aligned}
\label{SurfR42cProp1Eq2a}\widetilde\nabla_{e_1}e_1=0,&\quad&\widetilde\nabla_{e_1}e_2=0\\
\label{SurfR42cProp1Eq2b}\widetilde\nabla_{e_2}e_1=-\omega e_2,&\quad&\widetilde\nabla_{e_2}e_2=-\omega e_1+e_3,\\
\label{SurfR42cProp1Eq2c}\widetilde\nabla_{e_1}e_3=\omega e_3,&\quad&\widetilde\nabla_{e_2}e_3=\gamma e_3, \end{aligned}$$
where the functions $\phi,\omega$ and $\gamma$ are defined by one of following forms:
- For $c=0$, $$\label{SurfR42cProp1Eq3ALL_1}
\left\{\begin{array}{l}
\phi(s,t)=A(t)(s+m(t))\\
\displaystyle\omega(s,t)=-\frac{1}{s+m(t)}\\
\displaystyle\gamma(s,t)=\frac{\gamma_0(t)}{s+m(t)}-
\frac{m'(t)}{(A(t)(s+m(t)))^2}
\end{array}
\right.$$
- For $\varepsilon c=1$, $$\begin{aligned}
\label{SurfR42cProp1Eq3ALL_2}
\left\{\begin{array}{l}
\phi(s,t)=A(t)\cos{(s+m(t))}\\
\omega(s,t)=\tan{(s+m(t))}\\
\displaystyle\gamma(s,t)=\sec{(s+m(t))}
\left(\gamma_0(t)+\tan{(s+m(t))}
\frac{m'(t)}{A(t)}\right)
\end{array}
\right.\end{aligned}$$
- For $\varepsilon c=-1$,
$$\begin{aligned}
\label{SurfR42cProp1Eq3ALL_3}
\left\{\begin{array}{l}
\phi(s,t)=A(t)\cosh{(s+m(t))}\\
\omega(s,t)=-\tanh{(s+m(t))}\\
\displaystyle\gamma(s,t)={\mathrm{sech\,}}{(s+m(t))}
\left(\gamma_0(t)-\tanh{(s+m(t))}
\frac{m'(t)}{A(t)}\right)
\end{array}
\right.\end{aligned}$$
for some smooth functions $m$ and $\gamma_0$, where $A$ is an arbitrarily chosen positive, smooth, non–vanishing function.
Suppose that $M$ is a quasi–minimal surface of a pseudo–Riemannian space forms $R^4_2(c)$ with positive relative nullity. From Lemma \[lemma1\], we choose a frame field $\{e_1, e_2, e_3, e_4\}$ which satisfy the conditions given by . Considering this, we obtain $A^f_{e_3}=0$. Calculating the Codazzi equation for $X=Z=e_1$, $Y=e_2$ and $X=Z=e_2$, $Y=e_1$, we get $\omega_{12}(e_1)=0\;\mbox{ and }\;
\varepsilon\omega_{12}(e_2)=\omega,$ respectively. Combining these equations with and , we get the equations in .
On the other hand, equations and gives $[e_1,e_2]=\omega e_2$ which implies $[e_1,\phi e_2]=0$ for a non–vanishing function $\phi$ satisfying $$\label{SurfR42cProp1ProofEq05}
e_1(\phi)=-\phi\omega.$$ Therefore, there exists a local coordinate system $(s,t)$ such that $e_1=\frac{\partial}{\partial s}$ and $\phi e_2=\frac{\partial}{\partial t}$ defined on a neighborhood of any point $p\in M$. Consequently, the induced metric tensor $g$ of $M$ takes the form given in and the equation turns into $$\label{SurfR42cProp1ProofEq06}
\phi_s=-\phi\omega.$$ From the Gauss equation and the Ricci equation , we obtain $$\label{SurfR42cProp1ProofEq07}
\omega_s=\omega^2+\varepsilon c\;\;\mbox{and}\;\;
\gamma_s=\omega\gamma+\frac{\omega_t}\phi.$$ By solving equations appearing in and for $c=0$, $\varepsilon c=1$ and $\varepsilon c=-1$, we get , and , respectively, for some smooth functions $\gamma_0,m$ and $A$. Note that if $\tilde A$ is another smooth, positive, non–vanishing function, then the local coordinate system $(\tilde s, \tilde t)$ defined by $\tilde s=s,\ \tilde t=\int_{c}^t\frac{A}{\tilde A}dt$ satisfies all conditions of the lemma for $A=\tilde A$. Hence, the function $A$ appearing in , and can be chosen arbitrarily.
Quasi–Minimal Surfaces in $\mathbb E^4_2$
-----------------------------------------
In this subsection, we get the following local classification theorem for a quasi–minimal surface with positive relative nullity in $\mathbb{E}^4_2$.
\[PRNE42ClassThm\] A quasi–minimal surface in $\mathbb E^4_2$ has positive relative nullity if and only if it is congruent to one of the followings:
1. A surface given by $$\begin{aligned}
\label{PRNE42Example1YV}
\begin{split}
f(s,t)=&\left(b(t)s+\int _{t_0}^tm(\xi)b'(\xi)d\xi,s\sinh t+\int _{t_0}^tm(\xi) \cosh\xi d\xi, \right.\\
&\left.s\cosh t+\int _{t_0}^tm(\xi) \sinh\xi d\xi,b(t)s+\int _{t_0}^tm(\xi)b'(\xi)d\xi\right),
\end{split}\end{aligned}$$
2. A surface given by $$\begin{aligned}
\label{PRNE42Example2YV}
\begin{split}
f(s,t)=&\left(b(t)s+\int _{t_0}^tm(\xi)b'(\xi)d\xi,s\cosh t+\int _{t_0}^tm(\xi) \sinh\xi d\xi, \right.\\
&\left.s\sinh t+\int _{t_0}^tm(\xi) \cosh\xi d\xi,b(t)s+\int _{t_0}^tm(\xi)b'(\xi)d\xi\right),
\end{split}\end{aligned}$$
where $m(t)$ is a smooth function and the function $b(t)$ is a solution of $b''(t)-b(t)=F(t)$ for a non–vanishing smooth function $F(t)$.
Let $M$ be a quasi–minimal surface of $\mathbb{E}^4_2$ with positive relative nullity and consider a local coordinate system $(s,t)$ satisfying the conditions given in Proposition \[SurfR42cProp1\]. Without loss of generality, we take $A(t)=1$.
By considering the first equation in , we obtain $$\label{PRNE42ClassThmProofEq01}
f(s,t)=sB(t)+B_1(t)$$ for some smooth $\mathbb R^4$–valued functions $B(t)$ and $B_1(t)$. Also, the equations and implies $$\label{PRNE42ClassThmProofEq03}
e_3=\frac{F(t)}{s+m(t)}C_0$$ for a non–zero constant vector $C_0$, where $F(t)$ is a smooth non–vanishing function defined by $F(t)=e^{\int_{t_0}^t \gamma_0(\xi)d\xi}$.
On the other hand, the second equation in gives $$\label{PRNE42ClassThmProofEq02}
B_1'(t)=m(t)B'(t).$$ Considering , and , we get $e_2=B'(t)$ which implies $\tilde g(B'(t),B'(t))=-\varepsilon$ and combining , and with the second equation in , we get $$\nonumber
B''(t)-B(t)=F(t)C_0.$$ The solution of this equation is $$\label{PRNE42ClassThmProofEq04b}
B(t)=\cosh t C_1+\sinh t C_2+b(t)C_0$$ for some constant vectors $C_1, C_2\in \mathbb{E}^4_2$, where the function $b(t)$ satisfies $b''(t)-b(t)=F(t)$. By combining , and with , we obtain $$\begin{aligned}
\label{PRNE42ClassThmProofEq01Re}
\begin{split}
f(s,t)=&\left(s\cosh t + \int _{t_0}^tm(\xi)\sinh\xi d\xi\right)C_1\\
&+\left(s\sinh t + \int _{t_0}^tm(\xi)\cosh\xi d\xi\right)C_2\\
&+\left(sb(t)+ \int _{t_0}^tm(\xi)b'(\xi)d\xi\right)C_0.
\end{split}\end{aligned}$$ Since $g(e_1,e_1)=\varepsilon$ and $\tilde{g}(e_1,e_3)=\tilde{g}(e_3,e_3)=0$, the equations and give $$\begin{aligned}
\nonumber
\begin{split}
\tilde{g}(C_0,C_1)=\tilde{g}(C_0,C_2)=0,\quad & \tilde{g}(C_0,C_0)=0,\\
\tilde g(C_1,C_1)=-\tilde g(C_2,C_2)=\varepsilon,\quad &\tilde g(C_1,C_2)=0.
\end{split}\end{aligned}$$ Therefore, up to a suitable isometry of $\mathbb E^4_2$, one can choose $C_0=(1,0,0,1)$, $C_1=(0,0,1,0)$, $C_2=(0,1,0,0)$ for the case $\varepsilon=1$. In this case, the equation turns into which gives the case (i) of the theorem. For $\varepsilon=-1$, up to a suitable isometry of $\mathbb E^4_2$, we choose $C_0=(1,0,0,1)$, $C_1=(0,1,0,0)$, $C_2=(0,0,1,0)$. Then, we obtain the case (ii) of the theorem. Hence, the necessary condition is proved.
Conversely, it can be shown that both of the isometric immersions given by and satisfy $\alpha_f(\partial_s,\partial_s)=\alpha_f(\partial_s,\partial_t)=0$ and $$\quad \alpha_f(\partial_t,\partial_t)=(s+m(t))F(t)(1,0,0,1)\neq 0.$$ Hence, the surfaces given in the theorem are quasi–minimal and they have positive relative nullity.
Quasi–Minimal Surfaces in $\mathbb S^4_2$
-----------------------------------------
In this subsection, we classify quasi–minimal surfaces of $\mathbb{S}^4_2$ with positive relative nullity, that is the case $c=1$. First, we want to present some examples of quasi–minimal surfaces in $\mathbb S^4_2$.
\[PRNS42Example1\] Consider the immersion $f:\Omega\hookrightarrow\mathbb S^4_2$ defined by $$\begin{aligned}
\label{PRNS42Example1YV}
(i\circ f)(s,t)
=\left(b(t)\cos s, \cos s \sinh t,\sin s, \cos s \cosh t,b(t)\cos s\right)\end{aligned}$$ for a smooth function $b(t)$ and put $M=f(\Omega)$, where $\Omega=(0,2\pi)\times \mathbb R$. Then, the mean curvature vector of $f$ is $$i_*H=\frac{b''(t)-b(t)}{\cos s}(1,0,0,0,1)$$ which implies that $M$ is a quasi–minimal surface of $\mathbb{S}^4_2$ if and only if $b''(t)-b(t)\neq 0$ for all $t\in M$. A direct computation yields that $M$ has positive relative nullity.
Similar as Example \[PRNS42Example1\], we give the following example.
\[PRNS42Example3\] Similar to Example \[PRNS42Example1\], the surface $M$ given by $$\begin{aligned}
\label{PRNS42Example3YV}
(i\circ f)(s,t)=\left(b(t)\cosh s, \sinh{s}, \cosh s\cos t,\cosh s\sin t, b(t)\cosh s\right)\end{aligned}$$ is quasi–minimal if and only if $b''(t)+b(t)\neq0$ for all $t\in M$. Moreover, $M$ has positive relative nullity.
\[PRNS42Example2\] Let $\alpha:(I,-dt^2)\hookrightarrow\mathbb{S}^2_1$ be a time–like curve parametrized by its arc–length with a non–vanishing curvature $\kappa$ and the unit normal vector field $N$. Then, the surface $M$ in $\mathbb{S}^4_2$, defined by the immersion $f:J\times I\hookrightarrow\mathbb S^4_2$ given by $$\begin{aligned}
\label{PRNS42Example2YV}
\begin{split}
\hat{f}(s,t)=&\cos{s}(b(t),\alpha_1(t),\alpha_2(t),\alpha_3(t), b(t))\\
&+\sin{s}\left(\int_{t_0}^t\kappa(\xi) b'(\xi)d\xi, N_{1}(t), N_{2}(t), N_{3}(t), \int_{t_0}^t\kappa(\xi) b'(\xi)d\xi\right)
\end{split}\end{aligned}$$ is a quasi–minimal surface with positive relative nullity if and only if $b$ is a smooth function satisfying the condition $$\label{PRNS42Example2Eq1}
b''(t)-\kappa(t)\int _{t_0}^t\kappa(\xi)b'(\xi)d\xi-b(t)\neq0$$ for all $t\in I$, where $j\circ\alpha=(\alpha_1,\alpha_2,\alpha_3)$, $j_*N=(N_1,N_2,N_3)$ and $j:\mathbb S^2_1\subset\mathbb E^3_1$ is the inclusion.
By a direct computation, we obtain $\widetilde\nabla_{e_1}e_1=\widetilde\nabla_{e_1}e_2=0$ which yields that $e_1{}_p\in\mathcal{N}_p$ for any $p\in M$, where $e_1=\partial_s$ and $e_2=\frac{1}{\kappa(t)\sin s+\cos s}\partial_t$. A further computation yields that $$i_*\left(h(e_2,e_2)\right)=-2i_*H=\frac{b''(t)-\kappa(t)\int _{t_0}^t\kappa(\xi)b'(\xi)d\xi-b(t)}{\kappa(t)\sin s+\cos s} (1,0,0,0,1).$$ Hence, $M$ is quasi–minimal if and only if the condition given in is satisfied.
\[PRNS42Example4\] Let $\alpha:(I,dt^2)\hookrightarrow\mathbb{S}^2_1$ be a space–like curve parametrized by its arc–length with a non–vanishing curvature $\kappa$ and the unit normal vector field $N$. Then, the surface $M$ in $\mathbb{S}^4_2$, defined by the immersion $f:J\times I\hookrightarrow\mathbb S^4_2$, given by $$\begin{aligned}
\label{PRNS42Exampl42YV}
\begin{split}
\hat{f}(s,t)=&\cosh s\left(b(t),\alpha_1(t),\alpha_2(t),\alpha_3(t),b(t)\right)\\
&+\varepsilon\sinh s\Big(\int_{t_0}^t\kappa(\xi) b'(\xi)d\xi,N_1(t),N_2(t),N_3(t),\int_{t_0}^t\kappa(\xi) b'(\xi)d\xi\Big)
\end{split}\end{aligned}$$ is a quasi–minimal surface with positive relative nullity if and only if $b$ is a smooth function satisfying the condition $$\label{PRNS42Example4Eq1}
b''(t)-\kappa(t)\int _{t_0}^t\kappa(\xi)b'(\xi)d\xi+b(t)\neq0.$$ for all $t\in I$, where $j\circ\alpha=(\alpha_1,\alpha_2,\alpha_3)$, $j_*N=(N_1,N_2,N_3)$ and $j:\mathbb S^2_1\subset\mathbb E^3_1$ is the inclusion.
Similar as the proof of Proposition , it can be seen that $M$ has positive relative nullity and $$i_*\left(h(e_2,e_2)\right)=2i_*H=\frac{b''(t)-\kappa(t)\int _{t_0}^t\kappa(\xi)b'(\xi)d\xi+b(t)}{\cosh s+\varepsilon\kappa(t)\sinh s} (1,0,0,0,1)$$ gives that $M$ is a quasi–minimal surface in $\mathbb{S}^4_2$ if and only if the condition given by the equation is valid.
Now, we are ready to prove the following local classification theorem.
\[PRNES42ClassThm\] A quasi–minimal surface in $\mathbb S^4_2$ has positive relative nullity if and only if it is congruent to one of the followings:
1. A surface given by .
2. A surface described in Proposition \[PRNS42Example2\].
3. A surface given by .
4. A surface described in Proposition \[PRNS42Example4\].
In order to prove necessary condition, assume that $M$ is a quasi–minimal surface with positive relative nullity in $\mathbb{S}^4_2$. We choose a local coordinate system $(s,t)$ which satisfies the equations and and define the tangent vector fields $e_1,e_2$ as in Proposition \[SurfR42cProp1\]. Using the equations and , we have
\[PRNES42ClassThmEq1\] $$\begin{aligned}
\label{PRNES42ClassThmEq1a}
\hat{\nabla}_{e_1}e_1=-\varepsilon \hat{f},
&\quad&\hat{\nabla}_{e_1}e_2=0\\
\label{PRNES42ClassThmEq1b}
\hat{\nabla}_{e_2}e_1=-\omega e_2,
&\quad&\hat{\nabla}_{e_2}e_2=-\omega e_1+e_3+\varepsilon \hat{f},\\
\label{PRNES42ClassThmEq1c}
\hat{\nabla}_{e_1}e_3=\omega e_3,
&\quad&\hat{\nabla}_{e_2}e_3=\gamma e_3.\end{aligned}$$
We are going to study the cases $\varepsilon=1$ and $\varepsilon=-1$ separately.
*Case (1)* $\varepsilon=1$. Considering the equations and , we get $$\label{PRNES42ClassThmEq3}
e_3=\frac{F(t)}{\cos{(s+m(t))}}C_0$$ where $C_0$ is a non–zero constant vector in $\mathbb{E}^5_2$ and $F(t)$ is a smooth function defined by $F(t)=e^{\int_{t_0}^t A(\xi)\gamma_0(\xi)d\xi}$. Also, $C_0$ is a light–like vector in $\mathbb{E}^5_2$ due to the fact that $\tilde{g}(e_3,e_3)=0$. Thus, up to isometries of $\mathbb{E}^5_2$, we can choose $C_0=(1,0,0,0,1)$.
On the other hand, the first equation in gives $$\label{PRNES42ClassThmEq2}
\hat{f}(s,t)=\cos{s}B(t)+\sin{s}B_1(t)$$ for some smooth $\mathbb{R}^5$–valued functions $B(t)$ and $B_1(t)$ which satisfy $$\label{PRNES42ClassThmEq4}
\sin{m(t)}B'(t)+\cos{m(t)}B_1'(t)=0$$ because of the second one of . By considering , we are going to consider subcases $\sin m(t)=0$, $\cos m(t)=0$ and $\sin m(t) \cos m(t)\neq0$ separately.
*Case(1a)*: $\sin{m(t)}=0$ on $M$. Thus, the equation implies $$\label{PRNES42ClassThmEq8c}
B_1(t)=B_{10}$$ for a non–zero constant vector in $B_{10}\in \mathbb{E}^5_2$. Without loss of generality, we put $A(t)=1$. Then, the equation $\hat{g}(\hat{f},\hat{f})=1$ and $\tilde{g}(e_1,e_3)=0$ imply $$\begin{aligned}
\begin{split}
\label{PRNES42ClassThmEq8b}
\tilde g(B(t),B(t))=\tilde g(B_{10},B_{10})=&1,\\
\tilde g(B(t),B_{10})=\tilde g(B(t),C_0)=\tilde g(B_{10},C_0)=&0.
\end{split} \end{aligned}$$
On the other hand, by combining the equations , and with the second equation in , we obtain $B''(t)-B(t)=F(t)C_0$ which implies $$\label{PRNES42ClassThmEq8}
B(t)=\cosh{t}C_1+\sinh{t}C_2+b(t)C_0$$ for some constant vectors $C_1, C_2\in\mathbb{E}^5_2$ and a smooth function $b$ satisfying $b''(t)-b(t)=F(t)$. Because of and , we have $$\begin{aligned}
\label{PRNE42ClassThmEq10}
\begin{split}
g(C_1,C_1)=-g(C_2,C_2)=&1, \\
g(C_1,C_2)=g(C_1,C_0)=g(C_2,C_0)=&0,\\
\end{split}\end{aligned}$$ Therefore, up to a suitable isometry of $\mathbb S^4_2$, we can choose $$\begin{aligned}
\nonumber
B_{10}=(0,0,1,0,0),\;
C_1=(0,0,0,1,0),\;C_2=(0,1,0,0,0). \end{aligned}$$ Consequently, the equation gives . Hence, we have the surface given in the case (i) of the theorem.
*Case(1b)* $\cos{m(t)}=0$ on $M$. Similar to Case (1a), we obtain that $M$ is congruent to the surface given by $f(\Omega)$, where $f:\Omega\hookrightarrow\mathbb S^4_2$ is defined by $$(i\circ f)(s,t)=(b(t)\sin s, \sin s \sinh t,\cos s, \sin s \cosh t,b(t)\sin s).$$ However, this surface is congruent to the surface given in the case (i) of the theorem.
*Case(1c)* $\sin{m(t)}\neq 0$ and $\cos{m(t)}\neq 0$ on an open subset $\mathcal O$ of $M$. By shrinking $\mathcal O$ if necessary, we assume $\mathcal O=I\times J$ for some open intervals $I$ and $J$. Without loss of generality, we choose $A(t)=\sec{m(t)}$. In this case, the equation implies $B'_1(t)=-\tan{m(t)}B'(t)$ and $B_1'$ does not vanish on $\mathcal O$. Then, we have $$\label{PRNES42ClassThmEq8d}
e_2=B'(t)$$ which implies $g(e_2,e_2)=g(B'(t),B'(t))=-1$. By combining the equations , and with the second equation in , we obtain $$\label{PRNES42ClassThmEq7_2}
B''(t)=B(t)- \tan m(t)B_1(t)+F(t) \sec m(t)C_0.$$
On the other hand, since $\hat g(\hat f,e_3)=0$, and imply $\tilde g(B_1,C_0)=\tilde g(B,C_0)=0$ which give
\[PRNES42ClassThmCaseacEq1ALL\] $$\begin{aligned}
\label{PRNES42ClassThmCaseacEq1a}B(t)&=&(b(t),\alpha_1(t),\alpha_2(t),\alpha_3(t), b(t)),\\
\label{PRNES42ClassThmCaseacEq1b}B_1(t)&=&(b_{10}(t), b_{11}(t), b_{12}(t), b_{13}(t), b_{10}(t)).\end{aligned}$$
for some smooth functions $b_0$, $\alpha_i$ and $b_{1j}$. Next, we define the smooth curves $\alpha,\gamma:J\to \mathbb E^3_1$ by $$\label{PRNES42ClassThmalpha-gamma}
\alpha=(\alpha_1,\alpha_2,\alpha_3),\quad \gamma=(b_{11}, b_{12}, b_{13})$$ Then, from we have $$\label{PRNES42ClassThmEq7_5} \alpha''(t)=\alpha(t)- \tan m(t)\alpha_1(t).$$
Next, by combining with , we obtain $g(B(t), B(t))=g(B_1(t),B_1(t))=1$ and $g(B'(t),B'(t))=-1$. We consider these equations and to obtain $\langle \alpha,\alpha\rangle=1$ and $\langle \alpha',\alpha'\rangle=1$ which yields that $\alpha$ is a time–like curve lying on $\mathbb S^2_1$ parametrized by its arc–length. Therefore, one can define (spherical) normal $N=(N_1,N_2,N_3)$ and (spherical) curvature $\kappa$ of $\alpha$ by $$\label{PRNES42ClassThmEq7_3}
\alpha''=\kappa N+\alpha,\;\; N'=\kappa \alpha'.$$ By combining and , we obtain $$\label{PRNES42ClassThmCase1cEq3a}
N(t)=\gamma(t),\quad \kappa(t)=-\tan m(t)$$ and by considering this equation and , from we obtain $$\label{PRNES42ClassThmCase1cEq3b}
b_{10}(t)=\int_{t_0}^t\kappa(\xi) b'(\xi)d\xi$$ for a constant $t_0$. Note that $\kappa$ does not vanish on $J$ because of the second equation in . By combining , and with , we obtain that $\mathcal O$ is congruent to the surface given by . Therefore, we have the case (ii) of the theorem.
*Case(2)* $\varepsilon=-1$. Similar to Case(1), considering the equations and , we get $$\label{PRNES42ClassThmEq13}
e_3=\frac{F(t)}{\cosh{(s+m(t))}}C_0$$ for a light-like constant vector $C_0\in \mathbb{E}^5_2$, where $F(t)$ is a non–vanishing smooth function defined by $\displaystyle F(t)=e^{\int_{t_0}^tA(\xi)\gamma_0(\xi)d\xi}$. Since $C_0$ is a light–like constant vector in $\mathbb{E}^5_2$, up to isometries of $\mathbb{E}^5_2$, we choose $C_0=(1,0,0,0,1)$.
On the other hand, the first equation in gives $$\label{PRNES42ClassThmEq11}
\hat{f}(s,t)=\cosh{s}B(t)+\sinh{s}B_1(t)$$ for some smooth $\mathbb{R}^5$–valued functions $B(t)$ and $B_1(t)$. Also, from the second equation in , we obtain $$\label{PRNES42ClassThmEq12}
\sinh{m(t)}B'(t)=\cosh{m(t)}B'_1(t).$$ We are going to study the subcases $\sinh m(t)=0$ and $\sinh m(t)\neq 0$ seperately.
*Case(2a)* $\sinh{m(t)}=0$ on $M$. Then, the equation implies $B_1(t)=B_{10}$ for a non–zero constant vector $B_{10}\in\mathbb{E}^5_2$. Without loss of generality, we put $A(t)=1$. Then, by considering $\hat{g}(\hat{f},\hat{f})=1$ and $\tilde{g}(e_1, e_3)=0$, from and we get $$\begin{aligned}
\label{PRNES42ClassThmCase2aEq1}
\begin{split}
\tilde g(B(t),B(t))=-\tilde g(B_{10},B_{10})=1, \quad &\tilde g(B(t), B_{10})=0,\\
\tilde{g}(B(t),C_0)=\tilde{g}(B_{10}, C_0)=0.&
\end{split}\end{aligned}$$
On the other hand, equations and give $$e_2=B'(t) \quad\mbox{ and }\quad e_3=\frac{F(t)}{\cosh{s}}C_0,$$ respectively. By combining these equations with the second equation in , we obtain $$\nonumber
B''( t)+B( t)=F( t)C_0$$ whose solution is $$\label{PRNES42ClassThmEq16}
B( t)=\cos{ t}C_1+\sin{ t}C_2+b( t)C_{0}$$ for some constant vectors in $C_1,C_2\in \mathbb{E}^5_2$, where $b$ is a smooth function satisfying $b''( t)+b( t)= F( t)$. By combining and , we get $$\begin{aligned}
\nonumber
\begin{split}
g(C_1,C_0)=g(C_2,C_0)=0 \quad &g(B_{10},C_1)=g(B_{10},C_2)=0\\
g(C_1,C_1)=g(C_2,C_2)=1, \quad &g(C_1, C_2)=0.
\end{split}\end{aligned}$$ Up to isometries of $\mathbb S^4_2$, we choose $$B_{10}=(0,1,0,0,0),\;\;
C_1=(0,0,1,0,0),\;\;
C_2=(0,0,0,1,0).$$ Then, the equation becomes . Hence $M$ is congruent to the surface given in the case (iii) of the theorem.
*Case(2b)* $\sinh {m(t)}\neq 0$ on an open subset $\mathcal O$ of $M$. Similar to Case (1c), we put $\mathcal O=I\times J$ and $A(t)={\mathrm{sech\,}}{m(t)}$, where $I$ and $J$ are some open intervals. In this case, because of , on $\mathcal O$ we have $$\label{PRNES42ClassThmEq20}
B'_1(t)=\tanh{m(t)}B'(t).$$ for vector–valued smooth functions $B(t)$ and $B_1(t)$. Consequently, we get $e_2=B'(t)$ and $g(e_2,e_2)=g(B'(t),B'(t))=1$. Using this and the equation in the second one of , we get $$\label{PRNES42ClassThmEq21}
B''(t)=-B(t)+\tanh{m(t)}B_1(t)+{\mathrm{sech\,}}{m(t)}F(t)C_0.$$
Similar to Case(1c), $B$ and $B_1$ satisfy and we define curves $\alpha,\gamma:J\to \mathbb E^3_1$ as in . Then, from we have $$\label{PRNES42ClassThmEq21b}
\alpha''(t)=-\alpha(t)+\tanh{m(t)}\gamma(t).$$ Note that since $\varepsilon=-1$, and implies $g(B(t),B(t))=g(B'(t),B'(t))=1$ from which we have $\langle \alpha,\alpha\rangle=\langle \alpha',\alpha'\rangle=1$. Therefore, $\alpha$ is a space–like curve lying on $\mathbb{S}^2_1$ and it is parametrized by its arc–length. Similar to Case (1c), we define $\kappa$ and $N$ by $$\begin{aligned}
\label{PRNES42ClassThmEq22}
\alpha''=\kappa N-\alpha,\quad N'=\kappa \alpha'.\end{aligned}$$ By combining and , we obtain $$\label{PRNES42ClassThmCase2bEq1}
N(t)=\gamma(t),\quad \kappa(t)=\tanh m(t).$$ Since $\sinh {m(t)}\neq 0$ on $\mathcal O$, $\kappa$ does not vanish on $J$. By considering and , we obtain $$\label{PRNES42ClassThmCase2bEq2}
b_{10}(t)=\int_{t_0}^t\kappa(\xi) b'(\xi)d\xi$$ for a constant $t_0$. By combining , and with , we obtain that $\mathcal O$ is congruent to the surface given by . Therefore, we have the case (iv) of the theorem.
We have completed the proof of the necessary condition. Conversely, as we describe in Example \[PRNS42Example1\], Example \[PRNS42Example3\], Proposition \[PRNS42Example2\] and Proposition \[PRNS42Example4\], all of the surfaces given in the theorem are quasi–minimal and they have positive relative nullity.
[99]{}
[ ]{}
[ ]{}
[ ]{}
[ ]{}
[ ]{}
[ ]{}
[ ]{}
,[24(22)]{}[2007]{}[5441–5452]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Magnetic resonance imaging (MRI) enables plant scientists to non-invasively study root system development and root-soil interaction. Challenging recording conditions, such as low resolution and a high level of noise hamper the performance of traditional root extraction algorithms, though. We propose to increase signal-to-noise ratio and resolution by segmenting the scanned volumes into root and soil in super-resolution using a 3D U-Net. Tests on real data show that the trained network is capable to detect most roots successfully and even finds roots that were missed by human annotators. Our experiments show that the segmentation performance can be further improved with modifications of the loss function.'
author:
- |
Yi Zhao$^1$, Nils Wandel$^1$, Magdalena Landl$^2$, Andrea Schnepf$^2$ and Sven Behnke$^1$[^1]\
1 – University of Bonn, Computer Science Institute VI, 53115 Bonn, Germany\
2 – FZ Jülich GmbH, Institute of Bio- and Geosciences 3, 52425 Jülich, Germany
bibliography:
- 'BibFile\_3D\_UNet.bib'
title: '3D U-Net for Segmentation of Plant Root MRI Images in Super-Resolution'
---
Introduction
============
In order to gain a better understanding of plant growth and its response to environmental influences such as droughts, sinking groundwater levels or climate change, plant scientists investigate root-soil interaction processes. Magnetic resonance imaging (MRI) non-invasively measures both roots and soil [@stingaciu2013situ]. Low resolution of MRI scans—compared to the diameter of thin roots—and high noise (Figure \[fig:mri\_and\_gt\]) due to ferromagnetic particles or low contrast between root and soil water signal hamper the application of traditional root extraction algorithms [@schulz2013plant]. For these reasons, root reconstruction is often done manually. To enable automatic reconstruction, the signal-to-noise ratio (SNR) and resolution of these MRI scans need to be enhanced.
In this work, we propose using a 3D U-Net [@cciccek20163d] to segment scans of pots containing plant roots into root and soil in super-resolution. As training data is scarce, we create synthetic MRIs by rendering roots in 3D and combine them with real MRI scans of pure soil. In this way, we improve segmentation performance in soil regions, compared to models that were trained on synthetic data only [@uzman2019learning; @horn2018superresolution]. Furthermore, we investigated the influence of different loss modifications. By doing this, the segmentation performance gets improved, especially for thin roots which are difficult to detect intactly.
![3D visualizations of noisy root images and original MRI image slices. (a)(c) each depicts the 3D visualization of a 3D MRI image, and (b)(d) each shows a horizontal slice of the 3D image on its left.[]{data-label="fig:mri_and_gt"}](images/mri-min.eps){width="315pt"}
Related Work
============
Recently, CNNs have achieved remarkable performance in 3D image segmentation. Havaei et al. [@havaei2017brain] use 2D CNNs to segment horizontal slices of 3D images and combine them into 3D outputs. However, 2D CNNs can only make limited use of spatial information along the vertical dimension. 3D CNNs can overcome this issue by learning 3D features from input images [@cciccek20163d]. Since their introduction in 2016, 3D U-Net [@cciccek20163d] has become a commonly used architecture for segmentation tasks of 3D data. The structure of 3D U-Net consists of a downsampling encoder and an upsampling decoder. The encoder extracts increasingly abstract features from an increasingly broader context, while the decoder restores the resolution from the downsampled feature maps. Moreover, to share the high-resolution features extracted in the encoder with the decoder, shortcut connections are established between them.
For image super-resolution, deep neural networks show state-of-the-art performance. To achieve a higher resolution in the network output, the input image needs to be upsampled. This can be done either by upsampling the input using interpolation [@dong2014learning], or within the network using methods like transposed convolution [@dong2016accelerating]. The latter results in better performance because the interpolation is learned directly from the data.
Root extraction algorithms were developed to automatically extract structural models of roots from 3D MRI images [@schulz2013plant]. However, their performances degrade significantly in the presence of low SNR or low resolution [@schulz2013plant]. Uzman et al. [@uzman2019learning] increased the SNR and resolution of MRI images using a 2D RefineNet. Due to the limited number of real images, synthetic MRI data was used during training. While the RefineNet is able to detect most roots, it also showed a non-negligible amount of false positives in some cases. Horn et al. [@horn2018superresolution] tried to complete the same task with a 3D CNN. However, because of memory constraints of the graphics processing unit (GPU), the network structure was relatively shallow. This resulted in fewer false positives, but at the same time significantly more false negatives for thin roots, leading to large gaps between detected root fragments.
Segmentation in Super-Resolution
================================
We segment the MRI scans in double resolution, i.e. the task is to map an input image $I\in \mathbb{R}^{x\times{y}\times{z}}$ to a binary segmentation output $S\in \mathbb{B}^{2x\times{2y}\times{2z}}$. To compute the mapping, a 3D U-Net architecture was chosen. It is relatively simple but still incorporates a large 3D context in its decision. During training, the network was fed randomly sampled 60$\times{}$60$\times{}$60 crops in order to stay within the limitations of GPU memory.
3D U-Net
--------
![3D Superresolution U-Net. The output has twice the input resolution.[]{data-label="fig:3d_unet"}](images/3D_unet_squeezed.eps){width="315pt"}
Like the original 3D U-Net, our network consists of an encoder and a decoder (Figure \[fig:3d\_unet\]). The encoder contains three convolutional modules, each with two convolutional layers and an increasing number of channels. To avoid introducing misleading information through padding, only valid convolutions are used. Adjacent convolutional modules are connected by a maxpooling layer for downsampling. The decoder part consists of three convolutional modules which are connected by 2$\times{}$ transposed convolutional layers for upsampling. Each upsampled tensor is concatenated with an intermediate output from the encoder part. In addition to entering the encoder-decoder path, the input is also directly upsampled with 2$\times{}$ transposed convolution and concatenated with the super-resolution intermediate output of the decoder. At the segmentation output, the channel number is reduced to one and a sigmoid function returns root probability estimates between 0 and 1.
Dataset
-------
The dataset used in this work builds on the synthetic dataset generated by Uzman et al. [@uzman2019learning]. The images in it are generated by combining augmentations of four reconstructed root structures with synthetic soil images which simulate real soil features.
#### Combining generated roots with real soil MRI
To increase the diversity of root structures, we generated 30 random roots of various shapes and root-diameters. Furthermore, we used MRI scans of pure soil to capture a larger range of 3D features of soil. For training, random crops of the generated roots were combined with crops of pure soil and augmented by varying parameters such as the contrast between root and soil. For validation, random crops of a different subset of the dataset were used. The final evaluation of the network’s performance was done on five real MRI images [@stingaciu2013situ] with human annotations.
Loss
----
The loss function for training is binary cross-entropy loss averaged among all image voxels. On top of that, we investigated two loss modifications. The first one is applying a higher weight on the root voxels than the soil voxels, forcing the network to focus more on the correct prediction of roots. The reason for trying this is the imbalance of the soil voxel quantity and the root voxel quantity. The second modification is using a don’t-care flag to label the root-soil border area, and ignore the voxels in it when calculating the loss. This is done to check the effect of making the learning task easier, because the this area is hard to segment precisely and also not so important.
Evaluation
----------
Since the MRI images contain far more soil voxels than root voxels, the F1-Score is used for evaluation because it is robust against class imbalance. The human annotations are slightly misaligned to the real MRI data, which is a common problem that also occurs in 2D root images [@smith2020segmentation]. Because of that, the Distance Tolerant F1-Score as introduced by Uzman et al. [@uzman2019learning] was used. This way, misalignments between the output and the ground truth will be tolerated if the distance between them is not larger than the value of the distance tolerance.
Results
=======
Figure \[fig:vis\_original\_vs\_improved\_dataset\]a shows the Distance Tolerant F1-Scores of the 3D U-Net model trained on the improved dataset, evaluated on the test set with five real MRI scans. For larger tolerances, the score reaches higher values. Figure \[fig:vis\_original\_vs\_improved\_dataset\]b shows the relative improvement of the score, compared to training with the original data set. The example segmentation results shown in Figure \[fig:vis\_original\_vs\_improved\_dataset\]c, \[fig:vis\_original\_vs\_improved\_dataset\]d demonstrate that the number of false positives decreases when using the improved dataset. The model trained with the improved dataset reaches an F1-Score of 0.964 on the validation set, and its segmentation results on the test data detect most root branches correctly.
![Comparison between the models trained with the original dataset and the improved dataset. (a) shows the Distance Tolerant F1-Scores of the model trained with the improved dataset. (b) shows the percentage of improvement of the Distance Tolerant F1-Scores when trained with the improved dataset compared to the original dataset. Error bar indicates the 95% confidence interval. (c)(d) are example segmentation results of the models trained with the original and improved dataset, respectively. Green, red, blue represent true positives (TPs), false positives (FPs), and false negatives (FNs), respectively.[]{data-label="fig:vis_original_vs_improved_dataset"}](images/original_vs_improved_dataset2-min.eps){width="315pt"}
Figure \[fig:compare\_with\_refinenet\] compares our 3D U-Net qualitatively to the segmentation results of RefineNet [@uzman2019learning]. There are fewer false positives in the results of our model, reducing the risk of those false positives being incorrectly extracted as roots by root extraction algorithms. However, there are also more false negatives in the 3D U-Net results, which appear as disconnections in the roots. Closing large disconnections can be difficult for the root extraction algorithms.
![Segmentation outputs compared between RefineNet and 3D U-Net on real MRI images. Green, red, blue represent TPs, FPs, and FNs, respectively. (a)(c) are results from RefineNet and (b)(d) are results from 3D U-Net.[]{data-label="fig:compare_with_refinenet"}](images/compare_with_refinenet.eps){width="315pt"}
Furthermore, we investigated two types of loss modifications to deal with the disconnection issue. Here, all compared models are trained with the improved dataset. The first modification is using a higher root weight of 10, which results in improved Distance-Tolerant F1-Scores (Figure \[fig:later\_time\_point\]a) as well as fewer disconnections in the segmented thin roots (Figure \[fig:later\_time\_point\]c). Some root-like false positive predictions are likely roots missed in the human annotations. The second loss modification is to ignore the root-soil border area using don’t-care flag. This also leads to improved F1-Scores (Figure \[fig:later\_time\_point\]a) but the roots in the segmentation result become visibly thicker which is slightly unrealistic (Figure \[fig:later\_time\_point\]d). The improvement of the F1-Scores is probably largely due to the higher recall because of the thicker root predictions.
![Effect of loss modifications. (a) shows the percentage of improvement of the Distance Tolerant F1-Scores of the 2 types of loss modification. Error bar indicates the 95% confidence interval. (b)(c)(d) shows the example segmentation results of the models trained with the original loss, root weight of 10, and don’t-care flag, respectively. Green, red, blue represent TPs, FPs, and FNs, respectively. []{data-label="fig:later_time_point"}](images/2lossModifications-min.eps){width="315pt"}
Conclusion
==========
We investigated semantic segmentation in super-resolution of plant root MRI images using a 3D U-Net model. When evaluated on the real MRI images, the model can detect most root branches correctly, despite some minor disconnections in thin roots. This disconnection problem can be mitigated by applying higher weights on root voxels or ignoring the voxels near the root-soil border when calculating the training loss.
[^1]: This research was supported by grants BE 2556/15 and SCHN 1361/3-1 of German Research Foundation (DFG).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that there is no classical regular Sturm-Liouville problem on a finite interval whose spectrum consists of infinitely many distinct primes numbers. In particular, this answers in the negative a question raised by Zettl in his book, [@zet]. We also show that there [*may*]{} exist such a problem if the parameter dependence is nonlinear.'
address: |
School of Mathematics and Statistics\
Carleton University, Ottawa, Ontario, Canada, K1S 5B6
author:
- 'Angelo B. Mingarelli'
title: 'A note on Sturm-Liouville problems whose spectrum is the set of prime numbers'
---
[^1] [^2]
\[section\] \[theorem\][Corollary]{} \[theorem\][Example]{}
Introduction
============
In this note a Sturm-Liouville equation on a finite interval $[a,b]$ is defined by a second order real differential expression of the form $$\label{1.1}
-(p(x)y^{\prime})^{\prime} + q(x)y = \lambda r(x) y$$ where $p, q, r : [a,b] \to \mathbb{R}$ and $1/p, q, r \in L[a,b]$, and $\lambda$ is a generally complex parameter. By a [*solution*]{} of we will mean, as is customary, a function $y$ defined and absolutely continuous on $[a,b] = I$ such that $(py^{\prime})(x)$ is also absolutely continuous and $y(x)$ satisfies the differential equation almost everywhere on $I$. In this setting one can allow $p(x)$ to have infinite values but only on a set of measure zero.
The classical regular Sturm-Liouville problem (SLP) associated with consists in finding those values of $\lambda \in \mathbb{C}$ such that has a non-trivial solution satisfying the [*separated homogeneous boundary conditions*]{} $$\begin{aligned}
\label{1.2}
y(a)\cos \alpha - (py^{\prime})(a)\sin \alpha =0,\\
\label{1.3}
y(b)\cos \beta - (py^{\prime})(b)\sin \beta =0,\end{aligned}$$ where $\alpha \in [0,\pi)$ and $\beta \in (0,\pi]$. Of course, this problem has a very long history dating back to Sturm’s original contributions in the 1830’s. It has been developed in many different directions but we shall not delve on this at the moment.
Still greater generality can be obtained by allowing $p(x)$ to be identically infinite on subintervals of $I$. In this case one needs to rewrite as a vector system in two dimensions, e.g., $$\begin{aligned}
\label{1.4}
u^{\prime} = - v/p, \\
\label{1.5}
v^{\prime} = (\lambda r - q) u.\end{aligned}$$
This defines a problem of [*Atkinson-type*]{}. The boundary conditions - now take the form $$\begin{aligned}
\label{1.6}
u(a)\cos \alpha + v(a)\sin \alpha =0,\\
\label{1.7}
u(b)\cos \beta + v(b)\sin \beta =0,\end{aligned}$$ where $\alpha \in [0,\pi)$ and $\beta \in (0,\pi]$.
The advantage of using the formulation - is that it can be used to study three-term recurrence relations as well, see \[[@bvp], Chapter 8\] and [@abm] for more details. Here, the lack of assumptions on the signs of the coefficients allows for the most generality. However, in this very case things can get pretty bad since there are coefficients and a corresponding interval $I$ with the property that the Dirichlet problem for on $I$ has discrete spectrum filling the whole complex plane (see [@atm], [@abm4]). For more information on Sturm-Liouville problems of Atkinson-type see [@bvp] and [@zet].
The aim of this note is to answer, in part, Problem IV raised by Zettl in \[[@zet], p. 299\]. We restate the problem here for ease of reference:
“IV: [*Find a SLP whose spectrum is the primes.*]{}"
It is then pointed out that given any [*finite*]{} set of distinct real numbers, a SLP of Atkinson-type can be found whose spectrum is precisely that set, \[[@zet], p. 299\].
On the basis of spectral asymptotics and the Prime Number Theorem we show that there does not exist a regular Sturm-Liouville problem whose spectrum consists of infinitely many prime numbers, thus answering said Problem IV in the negative. However, we show that there [*may*]{} exist such a problem if the parameter dependence in (or -) is [*nonlinear*]{}. Nevertheless, it is not clear how to choose the coefficients so that the primes are generated even in such nonlinear cases.
The main results and discussions
================================
In the sequel the set of rational primes refers to the usual set of prime numbers in the rational number field. For basic results about primes and their distribution we refer the reader to Hardy and Wright [@hw]. The following two results are, for the most part, independent of sign conditions on the coefficients in either or - (see [@atm] for specific details).
\[th1\] There does not exist a regular SLP with separated boundary conditions - whose spectrum is an infinite set of distinct rational primes.
Indeed, more is true. Theorem \[th1\] has a counterpart for SLP of Atkinson-type.
\[th2\] There does not exist a regular SLP of Atkinson-type - with separated boundary conditions - whose spectrum is an infinite set of distinct rational primes.
It follows that Zettl’s Problem IV as stated is unsolvable within this framework.
Now consider the Dirichlet problem associated with on $[0,1]$ with nonlinear dependence $$\label{1.8}
-y^{\prime\prime} + q(x)y = \left (\frac{\pi \lambda}{\log \lambda}\right )^2\, y.$$ The basic asymptotic estimate for the rational primes states that if $p_n$ denotes the $n$-th prime, then $$\label{2.3}p_n \sim n\, \log n, \quad n\to \infty,$$ the latter being a consequence of the Prime Number Theorem, cf., \[[@hw], p.10 and p.346\].
Choosing the simplest case (i.e., $q(x)=0$) we get that the eigenvalues $\lambda_ n$, for $n \geq 1$, of the Dirichlet problem on $[0,1]$ of must be the roots of the equation $$\frac{\lambda}{\log \lambda} = n,$$ from which we deduce the approximate values of $\lambda_n$, for large $n$, namely: $$\lambda_n = n\log n + n\log \log n + n \log\log\log n + \cdots$$ Even so, without any specific reference to prime numbers, the first two terms on the right of the preceding equation already agree with Cesàro’s approximate value for the $n$-th prime [@ec], i.e., $$p_n = n\log n + n\log \log n - n + n\frac{(\log\log n - 2)}{\log n} - \cdots$$
It is thus [*conceivable*]{}, but by no means obvious, that by modifying $q(x) \in L[0,1]$ on infinitely many subintervals on each of which $q(x)$ takes on constant values, we may come up with an asymptotic expansion for $p_n$ for large $n$. If so, we can hope to find a suitable $q$ such that the Dirichlet spectrum of agrees with the set of all rational primes. We emphasize that the existence of such a function $q(x)$ is an open question.
Proofs
======
(Theorem \[th1\]) We assume that $p(x)$ has a finite number of turning points and that $(r(x)/p(x))_+ $ is not a.e. equal to zero. Here, $(\cdot)_+$ denotes the positive part of the function in question. Then, by a direct application of the Atkinson-Mingarelli Theorem \[[@atm], Theorem 2.3\], we get that the positive eigenvalues, $\lambda_n^+$, of - satisfy $$\label{3.1}
\lambda_n^+ \sim \dfrac{n^2\,\pi^2}{\int_{a}^{b} \sqrt{\left ({r(x)}/{p(x)} \right)_+ \, dx}}.$$ It is now clear that and are incompatible if the primes (or an infinite subsequence of distinct such) were in fact the eigenvalues of such a problem.
(Theorem \[th2\]) Although the Atkinson-Mingarelli Theorem does not extend immediately to SLP of Atkinson-type one can resort to a simple order of magnitude argument. By assumption we take it that the spectrum is an infinite set at the outset (more on this below). We argue as in \[[@bvp], pp.206-207\] with minor changes. Assuming that for the moment $u(a), v(a)$ have fixed initial values, it is a known fact that the resulting unique solution, $u(x,\lambda)$, and indeed $v(x, \lambda)$, are analytic for all complex $\lambda$ and thus each is an entire function of $\lambda$ (see also [@abm2] and [@abm3] for extensions of these results).
Now, following \[[@bvp], pp.206-207\] a straightforward calculation shows that, even for indefinite coefficients, we still have $$\bigg | \frac{d}{dx} \log \{ |\lambda| |u|^2 + |v|^2\} \bigg | \leq \sqrt{|\lambda|}(|r| + |1/p|) + |q|/ \sqrt{|\lambda|}.$$ Integrating the latter over $[a,b]$, using the integrability assumptions on the coefficients, and exponentiating, we obtain the estimates (see \[[@bvp]; p. 206, eq. (8.2.5)\]) $$u(x,\lambda), v(x,\lambda) = \Large{O}\{exp(const. \sqrt{|\lambda|)}\},$$ thus showing that, as entire functions of $\lambda$, each must be of order not exceeding $1/2$. It follows from the theory of entire functions that the infinite sequence of complex zeros, $\lambda_n$, of (that define the eigenvalues) grow at such a rate that, for any $\varepsilon > 0$, $$\sum_{n=1}^\infty \frac{1}{|\lambda_n|^{\frac{1}{2} + \varepsilon}} < \infty.$$ However, for sufficiently small $\varepsilon > 0$, implies that the series of primes (or any infinite subsequence of distinct primes) must satisfy $$\sum_{n=1}^\infty \frac{1}{|p_n|^{\frac{1}{2} + \varepsilon}} = \infty.$$
This contradiction completes the proof.
[99]{}
F. V. Atkinson, [*Discrete and Continuous Boundary Problems*]{}, Academic Press, New York, (1964), xiv, 570 pp.
F. V. Atkinson and A. B. Mingarelli, [*Asymptotics of the number of zeros and of the eigenvalues of general weighted Sturm-Liouville problems*]{} J. für die Reine und Ang. Math., [**375/376**]{} (1987), 380-393.
E. Cesàro, [*Sur une formule empirique de M. Pervouchine*]{}, C. R. Acad. Sci. Paris, [**119** ]{}, (1894), 848-849.
G. H. Hardy and E. M. Wright, [*An Introduction to the Theory of Numbers*]{}, Fourth Edition, Oxford University Press, (1960), xvi, 421 pp.
A. B. Mingarelli, [*Volterra-Stieltjes Integral Equations and Generalized Ordinary Differential Expressions*]{}, Lecture Notes in Mathematics [**989**]{}, Springer-Verlag, New York, 1983.
A. B. Mingarelli, [*Some remarks on the order of an entire function associated with a second order differential equation*]{} in [*Ordinary Differential Equations and Operators*]{}, Lecture Notes in Mathematics [**1032**]{}, Springer-Verlag, New York, 1983: 384-389.
A. B. Mingarelli, [*Some remarks on the order of an entire function associated with a second order differential equation, II*]{} C.R. Math. Rep. Acad. Sci. Canada, [**6**]{} (1984), 79-83.
A. B. Mingarelli, [*Characterizing degenerate Sturm-Liouville problems* ]{}, Electronic J. Diff. Eqns., 2004 (130) (2004), 1-8.
A. Zettl, [*Sturm-Liouville Theory*]{}, Mathematical Surveys and Monographs [**121**]{}, American Mathematical Society, Rhode Island, 2005:
[^1]: Submitted August ??, 2011.
[^2]: The author is partially supported by a grant from NSERC Canada
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the paper, we propose a class of efficient momentum-based policy gradient methods for the model-free reinforcement learning, which use adaptive learning rates and do not require any large batches. Specifically, we propose a fast important-sampling momentum-based policy gradient (IS-MBPG) method based on a new momentum-based variance reduced technique and the importance sampling technique. We also propose a fast Hessian-aided momentum-based policy gradient (HA-MBPG) method based on the momentum-based variance reduced technique and the Hessian-aided technique. Moreover, we prove that both the IS-MBPG and HA-MBPG methods reach the best known sample complexity of $O(\epsilon^{-3})$ for finding an $\epsilon$-stationary point of the nonconcave performance function, which only require one trajectory at each iteration. In particular, we present a non-adaptive version of IS-MBPG method, i.e., IS-MBPG\*, which also reaches the best known sample complexity of $O(\epsilon^{-3})$ without any large batches. In the experiments, we apply four benchmark tasks to demonstrate the effectiveness of our algorithms.'
bibliography:
- 'MBPG.bib'
nocite: '[@langley00]'
---
Introduction
============
Reinforcement Learning (RL) has achieved great success in solving many sequential decision-making problems such as autonomous driving [@shalev2016safe], robot manipulation [@deisenroth2013survey], the game of Go [@silver2017mastering] and natural language processing [@wang2018deep]. In general, RL involves a Markov decision process (MDP), where an agent takes actions dictated by a policy in a stochastic environment over a sequence of time steps, and then maximizes the long-term cumulative rewards to obtain an optimal policy. Due to easy implementation and avoiding policy degradation, policy gradient method [@williams1992simple; @sutton2000policy] is widely used for finding the optimal policy in MDPs, especially for the high dimensional continuous state and action spaces. To obtain the optimal policy, policy gradient methods directly maximize the expected total reward (also called as performance function $J(\theta)$) via using the stochastic first-order gradient of cumulative rewards. Recently, policy gradient methods have achieved significant empirical successes in many challenging deep reinforcement learning applications [@li2017deep] such as playing game of Go and robot manipulation.
**Algorithm** **Reference** **Sample Complexity** **Batch Size** **Adaptive Learning Rate**
--------------- ----------------------- ----------------------- ------------------------------------------- ----------------------------
REINFORCE @williams1992simple $O(\epsilon^{-4})$ $O(\epsilon^{-2})$
SVRPG @papini2018stochastic $O(\epsilon^{-4})$ $O(\epsilon^{-2})$ & $O(\epsilon^{-2})$
SVRPG @xu2019improved $O(\epsilon^{-10/3})$ $O(\epsilon^{-4/3})$ & $O(\epsilon^{-2})$
HAPG @shen2019hessian $O(\epsilon^{-3})$ $O(\epsilon^{-1})$ & $O(\epsilon^{-2})$
SRVR-PG @xu2019sample $O(\epsilon^{-3})$ $O(\epsilon^{-1})$ & $O(\epsilon^{-2})$
IS-MBPG Ours $O(\epsilon^{-3})$ $O(1)$
HA-MBPG Ours $O(\epsilon^{-3})$ $O(1)$
IS-MBPG\* Ours $O(\epsilon^{-3})$ $O(1)$
Thus, policy gradient methods have regained much interest in reinforcement learning, and some corresponding algorithms and theory of policy gradient [@fellows2018fourier; @fujimoto2018addressing; @papini2018stochastic; @haarnoja2018soft; @xu2019improved; @shen2019hessian; @cheng2019predictor; @cheng2019trajectory; @wang2019neural] have been proposed and studied. Since the classic policy gradient methods (e.g., REINFORCE [@williams1992simple], PGT [@sutton2000policy], GPOMDP [@baxter2001infinite] and TRPO [@schulman2015trust]) approximate the gradient of the expected total reward based on a batch of sampled trajectories, they generally suffer from large variance in the estimated gradients, which results in a poor convergence. Following the standard stochastic gradient methods [@robbins1951stochastic; @ghadimi2013stochastic], these gradient-based algorithms require $O(\epsilon^{-4})$ samples for finding an $\epsilon$-stationary point of non-concave performance function $J(\theta)$ ( i.e., $\mathbb{E}\|\nabla J(\theta)\|\leq \epsilon$) [@ghadimi2013stochastic]. Thus, recently many works have begun to study to reduce variance in the policy gradient methods. For example, the early variance reduced policy methods [@greensmith2004variance; @peters2008reinforcement] mainly focused on using unbiased baseline functions to reduce the variance. @schulman2015high presented the generalized advantage estimation (GAE) to discover the balance between bias and variance of policy gradient. Then @gu2016q applied both the GAE and linear baseline function to reduce variance. Recently, @mao2018variance [@wu2018variance] proposed the input-dependent and action-dependent baselines to reduce the variance, respectively. More recently, @cheng2019predictor leveraged the predictive models to reduce the variance to accelerate policy learning.
Recently, the variance reduced gradient estimators such as SVRG [@johnson2013accelerating; @allen2016variance; @reddi2016stochastic], SAGA [@defazio2014saga], SARAH [@nguyen2017sarah], SPIDER [@fang2018spider] and SpiderBoost [@wang2019spiderboost] have been successful in the oblivious supervised learning. However, the RL optimization problems are *non-oblivious*, i.e., the distribution of the samples is non-stationarity and changes over time. Thus, @du2017stochastic [@xu2017stochastic; @wai2019variance] first transform the original non-oblivious policy evaluation problem into some oblivious subproblems, and then use the existing variance reduced gradient estimators (such as SVRG and SAGA) to solve these subproblems to reach the goal of reducing the large variance in the original RL problem. For example, @du2017stochastic first transforms the empirical policy evaluation problem into a quadratic convex-concave saddle-point problem via linear function approximation, and then applies the variants of SVRG and SAGA [@palaniappan2016stochastic] to solve this oblivious saddle-point problem.
More recently, @papini2018stochastic [@xu2019improved; @xu2019sample; @shen2019hessian] further have developed some variance reduced policy gradient estimators directly used in the non-oblivious model-free RL, based on the existing variance reduced techniques such as SVRG and SPIDER used in the oblivious supervised learning. Moreover, @xu2019improved [@xu2019sample; @shen2019hessian] have effectively improved the sample complexity by using these variance reduced policy gradients. For example, two efficient variance reduced policy gradient methods, i.e, SRVR-PG [@xu2019sample] and HAPG [@shen2019hessian] have been proposed based on the SARAH/SPIDER, and reach a sharp sample complexity of $O(\epsilon^{-3})$ for finding an $\epsilon$-stationary point, which improves the vanilla complexity of $O(\epsilon^{-4})$ [@williams1992simple; @ghadimi2013stochastic] by a factor of $O(\epsilon^{-1})$. Since a lower bound of complexity of $O(\epsilon^{-3})$ for recently proposed variance reduction techniques is established in [@arjevani2019lower], both the SRVR-PG and HAPG obtain a near-optimal sample complexity of $O(\epsilon^{-3})$. However, the practical performances of these variance reduced policy gradient methods are not consistent with their near-optimal sample complexity, because these methods require large batches and strict learning rates to achieve this optimal complexity.
In the paper, thus, we propose a class of efficient momentum-based policy gradient methods, which use adaptive learning rates and do not require any large batches. Specifically, our algorithms only need one trajectory at each iteration, and use adaptive learning rates based on the current and historical stochastic gradients. Note that @pirotta2013adaptive has studied the adaptive learning rates for policy gradient methods, which only focuses on Gaussian policy. Moreover, @pirotta2013adaptive did not consider sample complexity and can not improve it. While our algorithms not only provide the adaptive learning rates that are suitable for any policies, but also improve sample complexity.
Contributions {#contributions .unnumbered}
-------------
Our main contributions are summarized as follows:
- We propose a fast important-sampling momentum-based policy gradient (IS-MBPG) method with adaptive learning rate, which builds on a new momentum-based variance reduction technique of STORM/Hybrid-SGD [@cutkosky2019momentum; @tran2019hybrid] and the importance sampling technique.
- We propose a fast Hessian-aided momentum-based policy gradient (HA-MBPG) method with adaptive learning rate, based on the momentum-based variance reduction technique and the Hessian-aided technique.
- We study the sample complexity of our methods, and prove that both the IS-MBPG and HA-MBPG methods reach the best known sample complexity of $O(\epsilon^{-3})$ without any large batches (see Table \[tab:1\]).
- We propose a non-adaptive version of IS-MBPG method, i.e., IS-MBPG\*, which has a simple monotonically decreasing learning rate. We prove that it also reaches the best known sample complexity of $O(\epsilon^{-3})$ without any large batches.
After our paper is accepted, we find that three related papers [@xiong2020non; @pham2020hybrid; @yuan2020stochastic] more recently are released on arXiv. @xiong2020non has studied the adaptive Adam-type policy gradient (PG-AMSGrad) method, which still suffers from a high sample complexity of $O(\epsilon^{-4})$. Subsequently, @pham2020hybrid [@yuan2020stochastic] have proposed the policy gradient methods, i.e., ProxHSPGA and STORM-PG, respectively, which also build on the momentum-based variance reduced technique of STORM/Hybrid-SGD. Although both the ProxHSPGA and STORM-PG reach the best known sample complexity of $O(\epsilon^{-3})$, these methods still rely on large batch sizes to obtain this sample complexity and do not provide an effective adaptive learning rate as our methods.
Notations {#notations .unnumbered}
---------
Let $\|\cdot\|$ denote the vector $\ell_2$ norm and the matrix spectral norm, respectively. We denote $a_n=O(b_n)$ if $a_n\leq cb_n$ for some constant $c>0$. $\mathbb{E}[X]$ and $\mathbb{V}[X]$ denote the expectation and variance of a random variable $X$, respectively. $\mathbb{E}_{\tau_t} [\cdot]= \mathbb{E}_{\tau_t} [\cdot|\tau_1,\cdots,\tau_{t-1}]$ for any $t\geq 2$.
Background
==========
In the section, we will review some preliminaries of standard reinforcement learning and policy gradient.
Reinforcement Learning
----------------------
Reinforcement learning is generally modeled as a discrete time Markov Decision Process (MDP): $\mathcal{M}=\{\mathcal{S},\mathcal{A},\mathcal{P},\mathcal{R},\gamma,\rho_0\}$. Here $\mathcal{S}$ is the state space, $\mathcal{A}$ is the action space, and $\rho_0$ denotes the initial state distribution. $\mathcal{P}(s'|s,a)$ denotes the probability that the agent transits from the state $s$ to $s'$ under taking the action $a\in \mathcal{A}$. $\mathcal{R}(s,a): \mathcal{S} \times \mathcal{A}\mapsto [-R,R] \ (R>0)$ is the bounded reward function, *i.e.,* the agent obtain the reward $\mathcal{R}(s,a)$ after it takes the action $a$ at the state $s$, and $\gamma \in (0,1)$ is the discount factor. The policy $\pi (a|s)$ at the state $s$ is represented by a conditional probability distribution $\pi_{\theta}(a|s)$ associated to the parameter $\theta \in \mathbb{R}^{d}$.
Given a time horizon $H$, the agent can collect a trajectory $\tau=\{s_0,a_0, \cdots, s_{H-1}, a_{H-1}\}$ under any stationary policy. Following the trajectory $\tau$, a cumulative discounted reward can be given as follows: $$\begin{aligned}
\mathcal{R}(\tau) = \sum^{H-1}_{h=0}\gamma^h\mathcal{R}(s_h,a_h),\end{aligned}$$ where $\gamma$ is the discount factor. Assume that the policy $\pi_{\theta}$ is parameterized by an unknown parameter $\theta\in \mathbb{R}^d$. Given the initial distribution $\rho_0=\rho(s_0)$, the probability distribution over trajectory $\tau$ can be obtain $$\begin{aligned}
\label{eq:0}
p(\tau|\theta) = \rho(s_0)\prod_{h=0}^{H-1}\mathcal{P}(s_{h+1}|s_h,a_h)\pi_{\theta}(a_h|s_h).\end{aligned}$$
Policy Gradient
---------------
The goal of RL is to find an optimal policy $\pi_{\theta}$ that is equivalent to maximize the expected discounted trajectory reward: $$\begin{aligned}
\label{eq:1}
\max_{\theta \in \mathbb{R}^d} J(\theta):= \mathbb{E}_{\tau\sim p(\tau|\theta)} [\mathcal{R}(\tau)]=\int\mathcal{R}(\tau) p(\tau|\theta)d\tau.\end{aligned}$$ Since the underlying distribution $p$ depends on the variable $\theta$ and varies through the whole optimization procedure, the problem is a *non-oblivious* learning problem, which is unlike the traditional supervised learning problems that the underlying distribution $p$ is stationary. To deal with this problem, the policy gradient method [@williams1992simple; @sutton2000policy] is a good choice. Specifically, we first compute the gradient of $J(\theta)$ with respect to $\theta$, and obtain $$\begin{aligned}
\label{eq:2}
\nabla J(\theta) & \!=\! \int\! \mathcal{R}(\tau) \nabla p(\tau|\theta)d\tau \!=\!\int\! \mathcal{R}(\tau) \frac{\nabla p(\tau|\theta)}{p(\tau|\theta)}p(\tau|\theta)d\tau \nonumber \\
& \! = \mathbb{E}_{\tau\sim p(\tau|\theta)} \big[\nabla\log p(\tau|\theta)\mathcal{R}(\tau)\big].\end{aligned}$$ Since the distribution $p(\tau | \theta)$ is unknown, we can not compute the exact full gradient of . Similar for stochastic gradient descent (SGD), the policy gradient method samples a batch of trajectories $\mathcal{B}=\{\tau_i\}_{i=1}^{|\mathcal{B}|}$ from the distribution $p(\tau | \theta)$ to obtain the stochastic gradient as follows: $$\begin{aligned}
\hat{\nabla} J(\theta) = \frac{1}{|\mathcal{B}|} \sum_{i\in \mathcal{B}} \nabla\log p(\tau_i|\theta)\mathcal{R}(\tau_i). \nonumber\end{aligned}$$ At the $t$-th iteration, the parameter $\theta$ can be updated: $$\begin{aligned}
\theta_{t+1} = \theta_t + \eta_t\hat{\nabla}_{\theta} J(\theta),\end{aligned}$$ where $\eta_t>0$ is a learning rate. In addition, since the term $\nabla\log p(\tau_i|\theta)$ is independent of the transition probability $\mathcal{P}$, we rewrite the stochastic gradient $ \hat{\nabla} J(\theta)$ as follows: $$\begin{aligned}
\label{eq:3}
&\hat{\nabla} J(\theta) = \frac{1}{|\mathcal{B}|} \sum_{i\in \mathcal{B}} g(\tau_i,\theta) \\
&= \frac{1}{|\mathcal{B}|} \sum_{i\in \mathcal{B}}\big(\sum_{h=0}^{H-1} \nabla_{\theta}\log \pi_{\theta}(a^i_h,s^i_h)\big) \big(\sum_{h=0}^{H-1}\gamma^h\mathcal{R}(s^i_h,a^i_h) \big), \nonumber\end{aligned}$$ where $g(\tau_i,\theta)$ is an unbiased stochastic gradient based on the trajectory $\tau_i$, *i.e.,* $\mathbb{E}[g(\tau_i,\theta)] = \nabla J(\theta)$. Based on the above gradient estimator in , we can obtain the existing well-known gradient estimators of policy gradient such as the REINFORCE, the PGT and the GPOMDP. Due to $\mathbb{E}[\nabla_{\theta} \log \pi_{\theta}(a,s)]=0$, the REINFORCE adds a constant baseline $b$ and obtains a gradient estimator as follows: $$\begin{aligned}
g(\tau_i,\theta) \!=\! \big(\sum_{h=0}^{H-1} \nabla_{\theta}\log \pi_{\theta}(a^i_h,s^i_h)\big) \! \big(\sum_{h=0}^{H-1}\gamma^h\mathcal{R}(s^i_h,a^i_h)\!-\!b \big). \nonumber\end{aligned}$$ Further, considering the fact that the current actions do not rely on the previous rewards, the PGT refines the REINFORCE and obtains the following gradient estimator: $$\begin{aligned}
g(\tau_i,\theta) \!=\! \sum_{h=0}^{H-1} \!\sum_{j=h}^{H-1}\big(\gamma^j\mathcal{R}(s^i_j,a^i_j)\!-\!b_j \big)\nabla_{\theta}\log \pi_{\theta}(a^i_h,s^i_h). \nonumber\end{aligned}$$ Meanwhile, the PGT estimator is equivalent to the popular GPOMDP estimator defined as follows: $$\begin{aligned}
g(\tau_i,\theta) \!=\! \sum_{h=0}^{H-1}\!\sum_{j=0}^{h}\nabla_{\theta}\log \pi_{\theta}(a^i_j,s^i_j)(\gamma^h\mathcal{R}(s^i_h,a^i_h)\!-\!b_h). \nonumber\end{aligned}$$
Momentum-Based Policy Gradients
===============================
In the section, we propose a class of fast momentum-based policy gradient methods based on a new momentum-based variance reduction method, i.e., STORM [@cutkosky2019momentum]. Although the STORM shows its effectiveness in the *oblivious* learning problems, it is not well suitable for the *non-oblivious* learning problem \[eq:1\], where the underlying distribution $p(\cdot)$ depends on the variable $\theta$ and varies through the whole optimization procedure. To deal with this challenge, we will apply two effective techniques, i.e., *importance sampling* [@metelli2018policy; @papini2018stochastic] and *Hessian-aided* [@shen2019hessian], and propose the corresponded policy gradient methods, respectively.
Important-Sampling Momentum-Based Policy Gradient
---------------------------------------------------
In the subsection, we propose a fast important-sampling momentum-based policy gradient (IS-MBPG) method based on the importance sampling technique. Algorithm \[alg:1\] describes the algorithmic framework of IS-MBPG method.
Total iteration $T$, parameters $\{k,m,c\}$ and initial input $\theta_1$;\
Sample a trajectory $\tau_1$ from $p(\tau |\theta_1)$, and compute $u_1 = g(\tau_1|\theta_1)$;\
Sample a trajectory $\tau_t$ from $p(\tau |\theta_t)$, and compute $u_{t} = \beta_t g(\tau_t|\theta_t) + (1-\beta_t)\big[u_{t-1} + g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\big]$, where the importance sampling weight $w(\tau_t|\theta_{t-1},\theta_t)$ can be computed by using ;\
Compute $G_t = \|g(\tau|\theta_t)\|$; Compute [$\eta_t = \frac{k}{(m+\sum_{i=1}^tG^2_i)^{1/3}}$]{}; Update $\theta_{t+1} = \theta_t + \eta_t u_t$; Update [$\beta_{t+1} = c\eta_t^2$]{}; $\theta_{\zeta}$ chosen uniformly random from $\{\theta_t\}_{t=1}^{T}$.
Since the problem is *non-oblivious* or *non-stationarity* that the underlying distribution $p(\tau|\theta)$ depends on the variable $\theta$ and varies through the whole optimization procedure, we have $\mathbb{E}_{\tau\sim p(\tau|\theta)}[g(\tau | \theta)-g(\tau|\theta')] \neq \nabla J(\theta) - \nabla J(\theta')$. Given $\tau$ sampled from $p(\tau|\theta)$, we define an importance sampling weight $$\begin{aligned}
\label{eq:5}
w(\tau|\theta',\theta)=\frac{p(\tau|\theta')}{p(\tau|\theta)}= \prod_{h=0}^{H-1}\frac{\pi_{\theta'}(a_h|s_h)}{\pi_{\theta}(a_h|s_h)}\end{aligned}$$ to obtain $\mathbb{E}_{\tau\sim p(\tau|\theta)}\big[g(\tau | \theta)-w(\tau|\theta',\theta)g(\tau|\theta')\big] = \nabla J(\theta) - \nabla J(\theta')$. In Algorithm \[alg:1\], we use the following momentum-based variance reduced stochastic gradient $$\begin{aligned}
u_{t} =& (1-\beta_t)\big[ \underbrace{ u_{t-1} + g(\tau_t | \theta_{t}) - w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})}_{\mbox{SARAH}} \big]\nonumber \\
& + \beta_t \underbrace{g(\tau_t|\theta_t)}_{\mbox{SGD}}, \nonumber\end{aligned}$$ where $\beta_t\in (0, 1]$. When $\beta_t=1$, $u_t$ will reduce to a vanilla stochastic policy gradient used in the REINFORCE. When $\beta_t=0$, it will reduce to the SARAH-based stochastic policy gradient used in the SRVR-PG.
Let $e_t = u_t-\nabla J(\theta_t)$. It is easily verified that $$\begin{aligned}
&\mathbb{E}[e_{t}] \!=\! \mathbb{E}\big[(1-\beta_t)e_{t-1} \!+\! \beta_t( \underbrace{g(\tau_t|\theta_t)\!-\!\nabla J(\theta_t)}_{=T_1}) +(1-\beta_t) \nonumber \\
& \cdot \!\big(\! \underbrace{g(\tau_t | \theta_{t})\!-\! w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\!-\! \nabla J(\theta_t) \!+\! \nabla J(\theta_{t-1})}_{=T_2} \! \big)\!\big] \nonumber \\
& = (1-\beta_t)\mathbb{E}[e_{t-1}],\end{aligned}$$ where the last equality holds by $\mathbb{E}_{\tau_t\sim p(\tau|\theta_t)}[T_1]=0$ and $\mathbb{E}_{\tau_t\sim p(\tau|\theta_t)}[T_2]=0$. By Cauchy-Schwarz inequality, we can obtain $$\begin{aligned}
\mathbb{E}\|e_t\|^2 \leq &(1-\beta_t)^2\mathbb{E}\|e_{t-1}\|^2 +2\beta_t^2\mathbb{E}\|T_1\|^2 \nonumber \\
&+ 2(1-\beta_t)^2\mathbb{E}\|T_2\|^2.\end{aligned}$$ Since $O(\|T_2\|^2)=O(\|\theta_{t}-\theta_{t-1}\|^2)=O(\eta_t^2\|u_t\|^2)$, we can choose appropriate $\eta_t$ and $\beta_t$ to reduce the variance of stochastic gradient $u_t$. From the following theoretical results, our IS-MBPG algorithm can generate the adaptive and monotonically decreasing learning rate $\eta_t \in (0,\frac{1}{2L}]$, and the monotonically decreasing parameter $\beta_t \in (0,1]$.
Hessian-Aided Momentum-Based Policy Gradient
----------------------------------------------
In the subsection, we propose a fast Hessian-aided momentum-based policy gradient (HA-MBPG) method based on the Hessian-aided technique. Algorithm \[alg:2\] describes the algorithmic framework of HA-MBPG method.
Total iteration $T$, parameters $\{k,m,c\}$ and initial input $\theta_1$;\
Sample a trajectory $\tau_1$ from $p(\tau|\theta_1)$, and compute $u_1 = g(\tau_1|\theta_1)$;\
Choose $\alpha$ uniformly at random from $[0,1]$, and compute $\theta_t(\alpha) = \alpha \theta_t + (1-\alpha)\theta_{t-1}$; Sample a trajectory $\tau_t$ from $p(\tau|\theta_t(\alpha))$, and compute $u_{t} = \beta_t w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t) + (1-\beta_t)\big(u_{t-1} + \Delta_t\big)$, where $w(\tau|\theta_{t},\theta_t(\alpha))$ and $\Delta_t$ can be computed by using and , respectively;\
Compute $G_t = \|g(\tau|\theta_t)\|$; Compute [$\eta_t = \frac{k}{(m+\sum_{i=1}^tG^2_i)^{1/3}}$]{}; Update $\theta_{t+1} = \theta_t + \eta_t u_t$; Update [$\beta_{t+1} = c\eta_t^2$]{}; $\theta_{\zeta}$ chosen uniformly random from $\{\theta_t\}_{t=1}^{T}$.
In Algorithm \[alg:2\], at the $7$-th step, we use an unbiased term $\Delta^t$ (*i.e.,* $\mathbb{E}_{\tau_t\sim p(\tau|\theta_t(\alpha))}[\Delta^t]
= \nabla J(\theta_t) - \nabla J(\theta_{t-1})$) instead of the biased term $g(\tau | \theta_{t})-g(\tau|\theta_{t-1})$. To construct the term $\Delta^t$, we first assume that the function $J(\theta)$ is twice differentiable as in [@furmston2016approximate; @shen2019hessian]. By the Taylor’s expansion (or Newton-Leibniz formula), the gradient difference $\nabla J(\theta_t) - \nabla J(\theta_{t-1})$ can be written as $$\begin{aligned}
\label{eq:13}
\nabla J(\theta_t) - \nabla J(\theta_{t-1}) = \big[\int^1_0\nabla^2 J(\theta_t(\alpha))d\alpha\big]v_t,\end{aligned}$$ where $v_t = \theta_t-\theta_{t-1}$ and $\theta_t(\alpha) = \alpha\theta_t + (1-\alpha)\theta_{t-1}$ for some $\alpha\in [0,1]$. Following [@furmston2016approximate; @shen2019hessian], we obtain the policy Hessian $\nabla^2 J(\theta)$ as follows: $$\begin{aligned}
&\nabla^2 J(\theta) = \mathbb{E}_{\tau\sim p(\tau|\theta)}\big[\big(\nabla \log p(\tau|\theta)\nabla
\log p(\tau|\theta)^T\nonumber \\
& \quad +\nabla^2 \log p(\tau|\theta)\big)\mathcal{R}(\tau)\big] \nonumber \\
& = \mathbb{E}_{\tau\sim p(\tau|\theta)}\big[\nabla \Phi(\tau|\theta)\nabla \log p(\tau|\theta)^T +\nabla^2 \Phi(\tau|\theta)\big], \nonumber\end{aligned}$$ where $\Phi(\tau|\theta) = \sum_{h=0}^{H-1}\sum_{j=h}^{H-1}\gamma^jr(s_j,a_j)\log\pi_{\theta}(a_h,s_h)$. Given the random tuple $(\alpha,\tau)$, where $\alpha$ samples uniformly from $[0,1]$ and $\tau$ samples from the distribution $p(\tau|\theta_t(\alpha))$, we can construct $\Delta_t$ as follows: $$\begin{aligned}
\label{eq:6}
\Delta_t: = \hat{\nabla}^2(\theta_t(\alpha),\tau)v_t,\end{aligned}$$ where $\mathbb{E}_{\tau\sim p(\tau|\theta_t(\alpha))}[\hat{\nabla}^2(\theta_t(\alpha),\tau)] = \nabla^2J(\theta_t(\alpha))$ and $$\begin{aligned}
\hat{\nabla}^2(\theta_t,\tau) = & \nabla \Phi(\tau|\theta_t(\alpha))\nabla \log p(\tau|\theta_t(\alpha))^T \nonumber \\
& + \nabla^2 \Phi(\tau|\theta_t(\alpha)). \nonumber\end{aligned}$$ Note that $\mathbb{E}_{\alpha \sim U[0,1]} [ \nabla^2J(\theta_t(\alpha))]= \int^1_0\nabla^2 J(\theta_t(\alpha))d\alpha$ implies the unbiased estimator $\nabla^2 J(\theta(\bar{\alpha}))$ with $\bar{\alpha}$ uniformly sampled from $[0,1]$. Given $\bar{\alpha}$, we have $\mathbb{E}_{\tau\sim p(\tau|\theta_t(\bar{\alpha}))}[\hat{\nabla}^2(\theta_t(\bar{\alpha}),\tau)] = \nabla^2J(\theta_t(\bar{\alpha}))$. According to the equation , thus we have $\mathbb{E}_{\alpha \sim U[0,1], \ \tau\sim p(\tau|\theta_t(\alpha)) }[\Delta_t]=\nabla J(\theta_t) - \nabla J(\theta_{t-1})$, where $U[0,1]$ denotes the uniform distribution over $[0,1]$.
Next, we rewrite as follows: $$\begin{aligned}
\label{eq:7}
\Delta_t &= \big(\nabla \log p(\tau|\theta_t(\alpha))^Tv_t\big)\nabla \Phi(\tau|\theta_t(\alpha)) \nonumber \\
&\quad + \nabla^2 \Phi(\tau|\theta_t(\alpha))v_t.\end{aligned}$$ Considering the second term in is a time-consuming Hessian-vector product, in practice, we use can the finite difference method to estimate $ \nabla^2 \Phi(\tau|\theta_t(\alpha))v_t$ as follows: $$\begin{aligned}
& \nabla^2 \Phi(\tau|\theta_t(\alpha))v_t \nonumber \\
& \approx \frac{\nabla \Phi(\tau|\theta_t(\alpha) + \delta v_t) - \nabla \Phi(\tau|\theta_t(\alpha) - \delta v_t)}{2\delta}v_t \nonumber \\
& = \nabla^2 \Phi(\tau|\tilde{\theta}_t(\alpha))v_t,\end{aligned}$$ where $\delta>0$ is very small and $\tilde{\theta}_t(\alpha)\in \big[\theta_t(\alpha)-\delta v_t,\theta_t(\alpha)+\delta v_t]$ is obtained by the mean-value theorem. Suppose $\Phi(\tau|\theta)$ is $L_2$-second-order smooth, we can upper bound the approximated error: $$\begin{aligned}
\|\nabla^2 \Phi(\tau|\theta_t(\alpha))v_t-\nabla^2 \Phi(\tau|\tilde{\theta}_t(\alpha))v_t\| \leq L_2\|v_t\|\delta.\end{aligned}$$ Thus, we take a sufficiency small $\delta$ to obtain arbitrarily small approximated error.
In Algorithm \[alg:2\], we use the following momentum-based variance reduced stochastic gradient $$\begin{aligned}
u_{t} = \beta_t w(\tau|\theta_{t},\theta_t(\alpha))g(\tau|\theta_t) + (1-\beta_t)\big(u_{t-1} + \Delta_t\big), \nonumber\end{aligned}$$ where $\beta_t\in (0, 1]$. When $\beta_t=1$, $u_t$ will reduce to a vanilla stochastic policy gradient used in the REINFORCE. When $\beta_t=0$, it will reduce to the Hessian-aided stochastic policy gradient used in the HAPG.
Let $e_t = u_t-\nabla J(\theta_t)$. It is also easily verified that $$\begin{aligned}
\mathbb{E}[e_{t}] & = \mathbb{E}\big[(1\!-\!\beta_t)e_{t-1} \!+\! \beta_t( \underbrace{w(\tau|\theta_{t},\theta_t(\alpha))g(\tau|\theta_t)\!-\!J(\theta_t)}_{=T_3}) \nonumber \\
& \quad +(1-\beta_t)\big( \underbrace{\Delta_t- J(\theta_t) + J(\theta_{t-1})}_{=T_4} \big)\big] \nonumber \\
& = (1-\beta_t)\mathbb{E}[e_{t-1}],\end{aligned}$$ where the last equality holds by $\mathbb{E}_{\tau\sim p(\tau|\theta_t(\alpha))}[T_3]=0$ and $\mathbb{E}_{\tau\sim p(\tau|\theta_t(\alpha))}[T_4]=0$. Similarly, by Cauchy-Schwarz inequality, we can obtain $$\begin{aligned}
\mathbb{E}\|e_t\|^2 \leq &(1-\beta_t)^2\mathbb{E}\|e_{t-1}\|^2 +2\beta_t^2\mathbb{E}\|T_3\|^2 \nonumber \\
&+ 2(1-\beta_t)^2\mathbb{E}\|T_4\|^2.\end{aligned}$$ Since $O(\|T_4\|^2)=O(\|\theta_{t}-\theta_{t-1}\|^2)=O(\eta_t^2\|u_t\|^2)$, we can choose appropriate $\eta_t$ and $\beta_t$ to reduce the variance of stochastic gradient $u_t$. From the following theoretical results, our HA-MBPG algorithm can also generate the adaptive and monotonically decreasing learning rate $\eta_t \in (0,\frac{1}{2L}]$, and the monotonically decreasing parameter $\beta_t \in (0,1]$.
Non-Adaptive IS-MBPG\*
----------------------
In this subsection, we propose a non-adaptive version of IS-MBPG algorithm, i.e., IS-MBPG\*. The IS-MBPG\* algorithm is given in Algorithm \[alg:3\]. Specifically, Algorithm \[alg:3\] applies a simple monotonically decreasing learning rate $\eta_t$, which only depends on the number of iteration $t$.
Total iteration $T$, parameters $\{k,m,c\}$ and initial input $\theta_1$;\
Sample a trajectory $\tau_1$ from $p(\tau |\theta_1)$, and compute $u_1 = g(\tau_1|\theta_1)$;\
Sample a trajectory $\tau_t$ from $p(\tau |\theta_t)$, and compute $u_{t} = \beta_t g(\tau_t|\theta_t) + (1-\beta_t)\big[u_{t-1} + g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\big]$;\
Compute [$\eta_t = \frac{k}{(m+t)^{1/3}}$]{}; Update $\theta_{t+1} = \theta_t + \eta_t u_t$; Update [$\beta_{t+1} = c\eta_t^2$]{}; $\theta_{\zeta}$ chosen uniformly random from $\{\theta_t\}_{t=1}^{T}$.
Convergence Analysis
====================
In this section, we will study the convergence properties of our algorithms, i.e., IS-MBPG, HA-MBPG and IS-MBPG\*. *All related proofs are provided in supplementary document.* We first give some assumptions as follows:
Gradient and Hessian matrix of function $\log\pi_{\theta}(a|s)$ are bounded, *i.e.,* there exist constants $M_g, M_h >0$ such that $$\begin{aligned}
\|\nabla_{\theta}\log\pi_{\theta}(a|s)\| \leq M_g, \ \|\nabla^2_{\theta}\log\pi_{\theta}(a|s)\| \leq M_h.\end{aligned}$$
Variance of stochastic gradient $g(\tau|\theta)$ is bounded, *i.e.,* there exists a constant $\sigma >0$, for all $\pi_{\theta}$ such that $\mathbb{V}(g(\tau|\theta)) = \mathbb{E}\|g(\tau|\theta)-\nabla J(\theta)\|^2 \leq \sigma^2$.
Variance of importance sampling weight $w(\tau|\theta_1,\theta_2)=p(\tau|\theta_1)/p(\tau|\theta_2)$ is bounded, *i.e.,* there exists a constant $W >0$, it follows $\mathbb{V}(w(\tau|\theta_1,\theta_2)) \leq W$ for any $\theta_1, \theta_2 \in \mathbb{R}^d$ and $\tau\sim p(\tau|\theta_2)$.
Assumptions 1 and 2 have been commonly used in the convergence analysis of policy gradient algorithms [@papini2018stochastic; @xu2019improved; @xu2019sample; @shen2019hessian]. Assumption 3 has been used in the study of variance reduced policy gradient algorithms [@papini2018stochastic; @xu2019improved; @xu2019sample]. Note that the bounded importance sampling weight in Assumption 3 might be violated in practice. For example, when using neural networks (NNs) as the policy, small perturbations in $\theta$ might raise a large gap in the point probability due to some activation functions in NNs, which results in very large importance sampling weights. Thus, we generally clip the importance sampling weights to make our algorithms more effective. Based on Assumption 1, we give some useful properties of stochastic gradient $g(\tau|\theta)$ and $\hat{\nabla}^2(\theta_t,\tau)$, respectively.
\[pro:1\] (Proposition 4.2 in [@xu2019sample]) Suppose $g(\tau|\theta)$ is the PGT estimator. By Assumption 1, we have
- $g(\tau|\theta)$ is $\hat{L}$-Lipschitz differential, i.e., $\|g(\tau|\theta)-g(\tau|\theta')\|\leq L\|\theta-\theta'\|$ with $\hat{L}=M_hR/(1-\gamma)^2$;
- $J(\theta)$ is $\hat{L}$-smooth, i.e., $\|\nabla^2 J(\theta)\|\leq \hat{L}$;
- $g(\tau|\theta)$ is bounded, i.e., $\|g(\tau|\theta)\|\leq G$ for all $\theta\in \mathbb{R}^d$ with $G=M_gR/(1-\gamma)^2$.
Since $\|\nabla J(\theta)\|=\|\mathbb{E}[g(\tau|\theta)]\|\leq \mathbb{E}\|g(\tau|\theta)\|\leq G$, Proposition \[pro:1\] implies that $J(\theta)$ is $G$-Lipschitz. Without loss of generality, we use the PGT estimator to generate the gradient $g(\tau|\theta)$ in our algorithms, so $G_t = \|g(\tau|\theta)\| \leq G$.
\[pro:2\] (Lemma 4.1 in [@shen2019hessian]) Under Assumption 1, we have for all $\theta$ $$\begin{aligned}
\|\hat{\nabla}^2(\theta,\tau)\|^2 \leq \frac{H^2M_g^4R^2+M_h^2R^2}{(1-\gamma)^4}=\tilde{L}^2.\end{aligned}$$
Since $\|\nabla^2J(\theta)\|=\|\mathbb{E}[\hat{\nabla}^2(\theta,\tau)]\|\leq \mathbb{E}\|\hat{\nabla}^2(\theta,\tau)\|\leq \tilde{L}$, Proposition \[pro:2\] implies that $J(\theta)$ is $\tilde{L}$-smooth. Let $L=\max(\hat{L},\tilde{L})$, so $J(\theta)$ is $L$-smooth.
Convergence Analysis of IS-MBPG Algorithm
-----------------------------------------
In the subsection, we analyze the convergence properties of the IS-MBPG algorithm. The detailed proof is provided in Appendix \[Appendix:A1\]. For notational simplicity, let $B^2 = L^2 + 2G^2C^2_w$ with $C_w = \sqrt{H(2HM_g^2+M_h)(W+1)}$.
\[th:1\] Assume that the sequence $\{\theta_t\}_{t=1}^T$ be generated from Algorithm \[alg:1\]. Set $k=O(\frac{G^{2/3}}{L})$, $c=\frac{G^2}{3k^3L}+104B^2$, $m = \max\{2G^2,(2Lk)^3,(\frac{ck}{2L})^3\}$ and $\eta_0 = \frac{k}{m^{1/3}}$, we have $$\begin{aligned}
\mathbb{E}\|\nabla J(\theta_\zeta)\| \leq \frac{\sqrt{2\Omega}m^{1/6} + 2\Omega^{3/4}}{\sqrt{T}} + \frac{2\sqrt{\Omega}\sigma^{1/3}}{T^{1/3}},\end{aligned}$$ where $\Omega=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}}{4B^2}\ln(T+2)\big)$ with $J^*=\sup_{\theta}J(\theta) <+\infty$.
Since $\Omega=O(\ln(T))$, Theorem \[th:1\] shows that the IS-MBPG algorithm has $O(\sqrt{\ln(T)}/T^{\frac{1}{3}})$ convergence rate. The IS-MBPG algorithm needs $1$ trajectory to estimate the stochastic policy gradient $u_t$ at each iteration, and needs $T$ iterations. Without loss of generality, we omit a relative small term $\sqrt{\ln(T)}$. By $T^{-\frac{1}{3}} \leq \epsilon$, we choose $T=\epsilon^{-3}$. Thus, the IS-MBPG has the sample complexity of $1\cdot T = O(\epsilon^{-3})$ for finding an $\epsilon$-stationary point.
Convergence Analysis of HA-MBPG Algorithm
-----------------------------------------
In the subsection, we study the convergence properties of the HA-MBPG algorithm. The detailed proof is provided in Appendix \[Appendix:A2\].
\[th:2\] Assume that the sequence $\{\theta_t\}_{t=1}^T$ be generated from Algorithm \[alg:2\], and let $k=O(\frac{G^{2/3}}{L})$, $c=\frac{G^2}{3k^3L}+52L^2$, $m = \max\{2G^2,(2Lk)^3,(\frac{ck}{2L})^3\}$ and $\eta_0 = \frac{k}{m^{1/3}}$, we have $$\begin{aligned}
\mathbb{E}\|\nabla J(\theta_\zeta)\| \leq \frac{\sqrt{2\Lambda}m^{1/6} + 2\Lambda^{3/4}}{\sqrt{T}} + \frac{2\sqrt{\Lambda}\sigma^{1/3}}{T^{1/3}}, \nonumber\end{aligned}$$ where $\Lambda=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{4L^2k}\sigma^2 + \frac{(W+1)c^2k^{3}}{2L^2}\ln(T+2)\big)$ with $J^*=\sup_{\theta}J(\theta) <+\infty$.
Since $\Lambda=O(\ln(T))$, Theorem \[th:2\] shows that the HA-MBPG algorithm has $O(\sqrt{\ln(T)}/T^{\frac{1}{3}})$ convergence rate. The HA-MBPG algorithm needs $1$ trajectory to estimate the stochastic policy gradient $u_t$ at each iteration, and needs $T$ iterations. Without loss of generality, we omit a relative small term $\sqrt{\ln(T)}$. By $T^{-\frac{1}{3}} \leq \epsilon$, we choose $T=\epsilon^{-3}$. Thus, the HA-MBPG has the sample complexity of $1\cdot T = O(\epsilon^{-3})$ for finding an $\epsilon$-stationary point.
Convergence Analysis of IS-MBPG\* Algorithm
-------------------------------------------
In the subsection, we give the convergence properties of the IS-MBPG\* algorithm. The detailed proof is provided in Appendix \[Appendix:A3\].
\[th:3\] Assume that the sequence $\{\theta_t\}_{t=1}^T$ be generated from Algorithm \[alg:3\], and let $B^2 = L^2 + 2G^2C^2_w$, $k> 0$ $c=\frac{1}{3k^3L}+104B^2$, $m = \max\{2,(2Lk)^3,(\frac{ck}{2L})^3\}$ and $\eta_0 = \frac{k}{m^{1/3}}$, we have $$\begin{aligned}
\mathbb{E}\|\nabla J(\theta_\zeta)\|=\frac{1}{T}\sum_{t=1}^T\mathbb{E}\|\nabla J(\theta_t)\| \leq \frac{\sqrt{\Gamma}m^{1/6}}{\sqrt{T}} + \frac{\sqrt{\Gamma}}{T^{1/3}}, \nonumber\end{aligned}$$ where $\Gamma=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}\sigma^2}{4B^2}\ln(T+2)\big)$ with $J^*=\sup_{\theta}J(\theta) <+\infty$.
Since $\Gamma=O(\ln(T))$, Theorem \[th:3\] shows that the IS-MBPG\* algorithm has $O(\sqrt{\ln(T)}/T^{\frac{1}{3}})$ convergence rate. The IS-MBPG\* algorithm needs $1$ trajectory to estimate the stochastic policy gradient $u_t$ at each iteration, and needs $T$ iterations. Without loss of generality, we omit a relative small term $\sqrt{\ln(T)}$. By $T^{-\frac{1}{3}} \leq \epsilon$, we choose $T=\epsilon^{-3}$. Thus, the IS-MBPG\* also has the sample complexity of $1\cdot T = O(\epsilon^{-3})$ for finding an $\epsilon$-stationary point.
\
\
Experiments
=============
In this section, we demonstrate the performance of our algorithms on four standard reinforcement learning tasks, which are CartPole, Walker, HalfCheetah and Hopper. The first one is a discrete task from classic control, and the later three tasks are continuous RL task, which are popular MuJoCo environments [@todorov2012mujoco]. Detailed description of these environments is shown in Fig. \[fig:1\]. Our code is publicly available on https://github.com/gaosh/MBPG.
Experimental Setup
--------------------
In the experiment, we use Categorical Policy for CartPole, and Gaussian Policy for all the other environments. All Policies are parameterized by the fully connected neural network. The detail of network architecture and activation function used are shown in the Appendix A. The network settings are similar to the HAPG [@shen2019hessian] algorithm. We implement our algorithms by using garage [@garage] and pytorch [@paszke2019pytorch]. Note that Previous works mostly use environments implemented by old versions of garage, while latest version of garage directly use environments from gym [@gym]. As a result, there might be an inconsistency of the reward calculation between this paper and previous works due to the difference of environment implementation.
In the experiments, we compare our algorithm with the existing two best algorithms: Hessian Aided Policy Gradient (HAPG) [@shen2019hessian], Stochastic Recursive Variance Reduced Policy Gradient (SRVR-PG) [@xu2019sample] and a baseline algorithm: REINFORCE [@sutton2000policy]. For a fair comparison, the policies of all methods use the same initialization, which ensures that they have similar start point. Moreover, to ease the impact of randomness, we run each method 10 times, and plot mean as well as variance interval for each of them.
In addition, for the purpose of fair comparison, we use the same batch size $\abs{\mathcal{B}}$ for all algorithms, though our algorithms do not have a requirement on it. HAPG and SRVR-PG have sub-iterations (or inner loop), and requires additional hyper-parameters. The inner batch size for HAPG and SRVR-PG is also set to be the same value. For all the other hyper-parameters, we try to make them be analogous to the settings in their original paper. One may argue that our algorithms need three hyper-parameters $k$, $m$ and $c$ to control the evolution of learning rate while for other algorithms one hyper parameter is enough to control the learning rate. However, it should be noticed that our algorithms do not involve any sub-iterations unlike HAPG and SRVR-PG. Introducing sub-iterations itself naturally bring more hyper-parameters such as the number of sub-iteration and the inner batch size. From this perspective, the hyper-parameter complexity of our algorithms resembles HAPG and SRVR-PG. The more details of hyper-parameter selection are shown in Appendix \[Appendix:A\]. Similar to the HAPG algorithm, we use the **system probes** (i.e., the number of state transitions) as the measurement of sample complexity instead of number of trajectories. The reason of doing so is because each trajectory may have different length of states due to a failure flag returned from the environment (often happens at the beginning of training). Besides this reason, if using the number of trajectories as complexity measurement and the environment can return a failure flag, a faster algorithm may have a lot more system probes given the same number of trajectories. We also use average episode return as used in HAPG [@shen2019hessian].
Experimental Results
----------------------
The results of experiments are presented in Fig. \[fig:2\]. In the CartPole environment, our IS-MBPG and HA-MBPG algorithms have better performances than the other methods. In the Walker environment, our algorithms start to have more advantages. Specifically, the average return of IS-MBPG and HA-MBPG grows rapidly at the beginning of training. Moreover, our IS-MBPG algorithm achieves the best final performance with a obvious margin. HA-MBPG performs similar compared to SRVR-PG and HAPG, though it has an advantage at the beginning. In Hopper environment, our IS-MBPG and HA-MBPG algorithms are significantly faster compared to all other methods, while the final average reward are similar for different algorithms. In HalfCheetah environment, IS-MBPG, HA-MBPG and SRVR-PG performs similarly at the beginning. In the end of training, IS-MBPG can achieve the best performance. We note that HAPG performs poorly on this task, which is probably because of the normalized gradient and fixed learning rate in their algorithm. For all tasks, HA-MBPG are always inferior to the IS-MBPG. One possible reason for this observation is that we use the estimated Hessian vector product instead of the exact Hessian vector product in HA-MBPG algorithm, which brings additional estimation error to the algorithm.
In Fig. \[fig:3\], we plot the average reward when changing batch size in CartPole environment. From Fig. \[fig:3\], we find that when $20\%$ of the original batch size, our HA-MBPG and IS-MBPG algorithms still outperform the HAPG and SRVR-PG algorithms, respectively. When the batch size is 1, our HA-MBPG and IS-MBPG algorithms still reach a good performance. These results demonstrate that our HA-MBPG and IS-MBPG algorithms are not sensitive to the selection of batch size. Fig. \[fig:4\] shows that the non-adaptive IS-MBPG\* algorithm also has similar performances as the adaptive IS-MBPG algorithm.
Conclusion
==========
In the paper, we proposed a class of efficient momentum-based policy gradient methods (i.e., IS-MBPG and HA-MBPG), which use adaptive learning rates and do not require any large batches. Moreover, we proved that both IS-MBPG and HA-MBPG methods reach the best known sample complexity of $O(\epsilon^{-3})$, which only require one trajectory at each iteration. In particular, we also presented a non-adaptive version of IS-MBPG method (i.e., IS-MBPG\*), which has a simple monotonically decreasing learning rate. We proved that the IS-MBPG\* also reaches the best known sample complexity of $O(\epsilon^{-3})$ only required one trajectory at each iteration.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the anonymous reviewers for their helpful comments. We also thank the IT Help Desk at University of Pittsburgh. This work was partially supported by U.S. NSF IIS 1836945, IIS 1836938, IIS 1845666, IIS 1852606, IIS 1838627, IIS 1837956.
Environments CartPole Walker Hopper HalfCheetah
------------------------------------------ --------------- --------------- --------------- ---------------
Horizon 100 500 1000 500
Baseline None Linear Linear Linear
Neural Network (NN) sizes $8\times8$ $64\times64$ $64\times64 $ $64\times64$
NN activation function Tanh Tanh Tanh Tanh
Number of timesteps $5\times10^5$ $1\times10^7$ $1\times10^7$ $1\times10^7$
Batch size $\abs{\mathcal{B}}$ 50 100 50 100
HAPG $\abs{\mathcal{B}_{\text{sub}}}$ 10 10 10 10
SRVR-PG $\abs{\mathcal{B}_{\text{sub}}}$ 10 10 10 10
HAPG $n_{\text{sub}}$ 5 10 10 10
SRVR-PG $n_{\text{sub}}$ 3 2 2 2
IS-MBPG/HA-MBPG $k$ 0.75 0.75 0.75 0.75
IS-MBPG/HA-MBPG $c$ 2 2 1 1
IS-MBPG/HA-MBPG $m$ 2 12 3 3
IS-MBPG\* $k$ 0.9 0.9 - -
IS-MBPG\* $c$ 2 2 - -
IS-MBPG\* $m$ 2 12 - -
REINFORCE learning rate $\eta$ 0.01 0.01 0.01 0.01
HAPG learning rate $\eta$ 0.01 0.01 0.01 0.01
SRVR-PG learning rate $\eta$ 0.1 0.1 0.1 0.1
Supplementary Materials for “Momentum-Based Policy Gradient Methods" {#Appendix:A}
======================================================================
In this section, we first provide the details of hyper-parameter selection for the algorithms in Table \[tab:2\]. Table \[tab:2\] also shows that the detail of network architecture and activation function used in the experiments. Next, we study the convergence properties of our algorithms. We begin with giving some useful lemmas.
\[lem:A1\] (Lemma 1 in [@cortes2010learning]) Let $w(x)=P(x)/Q(x)$ be the importance weight for distributions $P$ and $Q$. The following identities hold for the expectation, second moment, and variance of $w(x)$ $$\begin{aligned}
\mathbb{E}[w(x)] = 1, \ \mathbb{E}[w^2(x)]= d_2(P||Q), \ \mathbb{V}[w(x)] = d_2(P||Q)-1,
\end{aligned}$$ where $d_2(P||Q)=2^{D(P||Q)}$, and $D(P||Q)$ is the $R\acute{e}nyi$ divergence between distributions $P$ and $Q$.
\[lem:A2\] Under Assumptions 1 and 3, let $w(\tau|\theta_{t-1},\theta_t)= g(\tau|\theta_{t-1})/g(\tau | \theta_{t})$, we have $$\begin{aligned}
\mathbb{V}[w(\tau|\theta_{t-1},\theta_t)] \leq C^2_w\|\theta_t-\theta_{t-1}\|^2,
\end{aligned}$$ where $C_w = \sqrt{H(2HM_g^2+M_h)(W+1)}$.
This proof can easy follow the proof of Lemma 6.1 in [@xu2019improved].
\[lem:A3\] Under Assumption 1, let $e_t = \nabla J(\theta_t)-u_t$. Given $0\leq \eta_t \leq \frac{1}{2L}$ for all $t\geq 1$, we have $$\begin{aligned}
\mathbb{E} [J(\theta_{t+1})] \geq \mathbb{E} [J(\theta_{t}) - \frac{3\eta_t}{4}\|e_t\|^2 + \frac{\eta_t}{8}\|\nabla J(\theta_t)\|^2 ].\end{aligned}$$
Let $e_t = \nabla J(\theta_t)-u_t$. By using $J(\theta)$ is $L$-smooth, we have $$\begin{aligned}
\mathbb{E} [J(\theta_{t+1})] & \geq \mathbb{E} [J(\theta_{t}) + \nabla J(\theta_{t})^T(\theta_{t+1}-\theta_{t}) - \frac{L}{2}\|\theta_{t+1}-\theta_{t}\|^2] = \mathbb{E} [J(\theta_{t}) + \eta_t\nabla J(\theta_{t})^Tu_t - \frac{L\eta_t^2}{2}\|u_{t}\|^2] \nonumber \\
& = \mathbb{E} [J(\theta_{t}) + \eta_t(\nabla J(\theta_{t})-u_t)^Tu_t + \eta_t\|u_t\|^2 - \frac{L\eta_t^2}{2}\|u_{t}\|^2] \nonumber \\
& \geq \mathbb{E} [J(\theta_{t}) - \frac{\eta_t}{2}\|\nabla J(\theta_{t})-u_t\|^2 + \frac{\eta_t}{2}(1-L\eta_t)\|u_t\|^2] \nonumber \\
& \geq \mathbb{E} [J(\theta_{t}) - \frac{\eta_t}{2}\|\nabla J(\theta_{t})-u_t\|^2 + \frac{\eta_t}{4}\|u_t\|^2] \nonumber \\
& \geq \mathbb{E} [J(\theta_{t}) - \frac{3\eta_t}{4}\|e_t\|^2 + \frac{\eta_t}{8}\|\nabla J(\theta_t)\|^2 ],\end{aligned}$$ where the second inequality holds by Young’s inequality, and the third inequality holds by $0< \eta_t \leq \frac{1}{2L}$, and the last inequality follows by $\|\nabla J(\theta_t)\|^2 \leq 2\|u_t\|^2 + 2\|\nabla J(\theta_t)-u_t\|^2 = 2\|u_t\|^2 + 2\|e_t\|^2$.
Convergence Analysis of IS-MBPG Algorithm {#Appendix:A1}
-------------------------------------------
In this subsection, we analyze the convergence properties of IS-MBPG algorithm. For notational simplicity, let $\nabla J(\theta)$ denote $\nabla_{\theta} J(\theta)$.
\[lem:A4\] Assume that the stochastic policy gradient $u_t$ be generated from Algorithm \[alg:1\], and let $e_t = u_t-\nabla J(\theta_t)$, we have $$\begin{aligned}
\mathbb{E}\big[ \eta^{-1}_{t-1}\|e_t\|^2 \big]
\leq 2\beta^2_t\eta^{-1}_{t-1}G_t^2 + \eta^{-1}_{t-1}(1-\beta_t)^2\big(1+8\eta^2_{t-1}B^2\big)\mathbb{E} \|e_{t-1}\|^2 + 8(1-\beta_t)^2B^2\eta_{t-1}\|\nabla J(\theta_{t-1})\|^2, \nonumber
\end{aligned}$$ where $B^2 = L^2 + 2G^2C^2_w$ with $C_w = \sqrt{H(2HM_g^2+M_h)(W+1)}$.
By the definition of $u_t$ in Algorithm \[alg:1\], we have $$\begin{aligned}
u_t - u_{t-1} = -\beta_tu_{t-1}+ \beta_t g(\tau_t|\theta_t) + (1-\beta_t)\big( g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\big).
\end{aligned}$$ Then we have $$\begin{aligned}
\label{eq:A1}
\mathbb{E}\big[ \eta^{-1}_{t-1} \|e_t\|^2\big]& = \mathbb{E}\big[\eta^{-1}_{t-1} \|\nabla J(\theta_{t-1}) - u_{t-1} + \nabla J(\theta_t)-\nabla J(\theta_{t-1}) -(u_t-u_{t-1})\|^2\big] \\
& = \mathbb{E}\big[ \eta^{-1}_{t-1} \|\nabla J(\theta_{t-1}) - u_{t-1} + \nabla J(\theta_t)-\nabla J(\theta_{t-1}) + \beta_tu_{t-1}- \beta_t g(\tau_t|\theta_t)\nonumber \\
& \quad - (1-\beta_t)\big( g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\big)\|^2\big] \nonumber \\
& = \mathbb{E}\big[ \eta^{-1}_{t-1} \|(1-\beta_t)(\nabla J(\theta_{t-1}) - u_{t-1}) + \beta_t(\nabla J(\theta_t)- g(\tau_t|\theta_t))\nonumber \\
& \quad - (1-\beta_t)\big( g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})-(\nabla J(\theta_t)-\nabla J(\theta_{t-1}))\big)\|^2\big] \nonumber \\
& = \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|\nabla J(\theta_{t-1}) - u_{t-1}\|^2 + \eta^{-1}_{t-1}\mathbb{E}\big[\|\beta_t(\nabla J(\theta_t)- g(\tau_t|\theta_t))\nonumber \\
& \quad - (1-\beta_t)\big( g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})-(\nabla J(\theta_t)-\nabla J(\theta_{t-1}))\big)\|^2\big] \nonumber \\
& \leq \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|\nabla J(\theta_{t-1}) - u_{t-1}\|^2 + 2\beta^2_t\eta^{-1}_{t-1}\mathbb{E}\|\nabla J(\theta_t)- g(\tau_t|\theta_t)\|^2\nonumber \\
& \quad +2(1-\beta_t)^2 \mathbb{E}\| g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})-(\nabla J(\theta_t)-\nabla J(\theta_{t-1}))\|^2 \nonumber \\
& \leq \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 2\beta^2_t\eta^{-1}_{t-1}\mathbb{E}\|g(\tau_t|\theta_t)\|^2
+ 2(1-\beta_t)^2\eta^{-1}_{t-1} \mathbb{E}\| g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\|^2 \nonumber \\
& = \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 2\beta^2_t\eta^{-1}_{t-1}G_t^2 + 2(1-\beta_t)^2\eta^{-1}_{t-1}
\underbrace{ \mathbb{E}\| g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\|^2 }_{=T_1}, \nonumber
\end{aligned}$$ where the forth equality holds by $\mathbb{E}_{\tau_t\sim p(\tau|\theta_t)}[g(\tau_t|\theta_t)]=\nabla J(\theta_t)$ and $\mathbb{E}_{\tau_t\sim p(\tau|\theta_t)}[g(\tau_t|\theta_t)- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})]=\nabla J(\theta_t)-\nabla J(\theta_{t-1})$; the first inequality follows by Young’s inequality; and the last inequality holds by $\mathbb{E}\|\zeta-\mathbb{E}[\zeta]\|^2=\mathbb{E}\|\zeta\|^2-(\mathbb{E}[\zeta])^2 \leq \mathbb{E}\|\zeta\|^2$.
Next, we give an upper bound of the term $T_1$ as follows: $$\begin{aligned}
\label{eq:A2}
T_1 & = \mathbb{E}\| g(\tau_t | \theta_{t})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\|^2 \nonumber \\
& = \mathbb{E}\| g(\tau_t | \theta_{t})-g(\tau_t|\theta_{t-1}) + g(\tau_t|\theta_{t-1})- w(\tau_t|\theta_{t-1},\theta_t) g(\tau_t|\theta_{t-1})\|^2 \nonumber \\
& \leq 2\mathbb{E}\| g(\tau_t | \theta_{t})-g(\tau_t|\theta_{t-1})\|^2 + 2\mathbb{E}\|(1- w(\tau_t|\theta_{t-1},\theta_t)) g(\tau_t|\theta_{t-1})\|^2 \nonumber \\
& \leq 2L^2\|\theta_t - \theta_{t-1}\|^2 + 2G^2\mathbb{E}\|1- w(\tau_t|\theta_{t-1},\theta_t)\|^2 \nonumber \\
& = 2L^2\|\theta_t - \theta_{t-1}\|^2 + 2G^2\mathbb{V}\big(w(\tau_t|\theta_{t-1},\theta_t)\big) \nonumber \\
& \leq 2(L^2 + 2G^2C^2_w)\|\theta_t-\theta_{t-1}\|^2,
\end{aligned}$$ where the second inequality holds by Proposition \[pro:1\], and the third equality holds by Lemma \[lem:A1\], and the last inequality follows by Lemma \[lem:A2\].
Combining the inequalities with , let $B^2 = L^2 + 2G^2C^2_w$, we have $$\begin{aligned}
\mathbb{E}\big[ \eta^{-1}_{t-1} \|e_t\|^2\big]& \leq \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 2\beta^2_t\eta^{-1}_{t-1}G_t^2+ 4(1-\beta_t)^2\eta^{-1}_{t-1}B^2\|\theta_t-\theta_{t-1}\|^2 \nonumber \\
& = \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 2\beta^2_t\eta^{-1}_{t-1}G_t^2 + 4(1-\beta_t)^2B^2\eta_{t-1}\|u_{t-1}\|^2 \nonumber \\
& = \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 2\beta^2_t\eta^{-1}_{t-1}G_t^2 + 4(1-\beta_t)^2B^2\eta_{t-1}\|e_{t-1} + \nabla J(\theta_{t-1})\|^2 \nonumber \\
& \leq 2\beta^2_t\eta^{-1}_{t-1}G_t^2 + \eta^{-1}_{t-1}(1-\beta_t)^2\big(1+8\eta^2_{t-1}B^2\big)\mathbb{E} \|e_{t-1}\|^2 + 8(1-\beta_t)^2B^2\eta_{t-1}\|\nabla J(\theta_{t-1})\|^2.
\end{aligned}$$
\[th:A1\] Assume that the sequence $\{\theta_t\}_{t=1}^T$ be generated from Algorithm \[alg:1\], and let $B^2 = L^2 + 2G^2C^2_w$, $k=O(\frac{G^{2/3}}{L})$ $c=\frac{G^2}{3k^3L}+104B^2$, $m = \max\{2G^2,(2Lk)^3,(\frac{ck}{2L})^3\}$ and $\eta_0 = \frac{k}{m^{1/3}}$, we have $$\begin{aligned}
\mathbb{E}\|\nabla J(\theta_\zeta)\|=\frac{1}{T}\sum_{t=1}^T\mathbb{E}\|\nabla J(\theta_t)\| \leq \frac{\sqrt{2\Omega}m^{1/6} + 2\Omega^{3/4}}{\sqrt{T}} + \frac{2\sqrt{\Omega}\sigma^{1/3}}{T^{1/3}}, \nonumber\end{aligned}$$ where $\Omega=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}}{4B^2}\ln(T+2)\big)$ with $J^*=\sup_{\theta}J(\theta) <+\infty$.
Due to $m\geq (2Lk)^3$, we have $\eta_t\leq \frac{k}{m^{1/3}}\leq \frac{1}{2L}$. Since $\eta_t\leq \frac{1}{2L}$ and $m\geq (\frac{ck}{2L})^3$, we have $\beta_{t+1}=c\eta^2_t\leq \frac{c\eta_t}{2L}\leq \frac{ck}{2Lm^{1/3}}\leq 1$. By Lemma \[lem:A4\], we have $$\begin{aligned}
\label{eq:A4}
\mathbb{E}[\eta^{-1}_{t}\|e_{t+1}\|^2-\eta^{-1}_{t-1}\|e_t\|^2] & \leq \mathbb{E}\big[ 2\beta^2_{t+1}\eta^{-1}_{t}G_{t+1}^2 + \eta^{-1}_{t}(1-\beta_{t+1})^2\big(1+8\eta^2_{t}B^2\big) \|e_{t}\|^2 \nonumber \\
& \quad + 8(1-\beta_{t+1})^2B^2\eta_{t}\|\nabla J(\theta_{t})\|^2 -\eta^{-1}_{t-1}\|e_t\|^2 \big] \nonumber \\
& \leq \mathbb{E}\big[ 2c^2\eta^{3}_{t}G_{t+1}^2 + \underbrace{ \big( \eta^{-1}_{t}(1-\beta_{t+1})(1+8\eta^2_{t}B^2) -\eta^{-1}_{t-1} \big) \|e_{t}\|^2 }_{=T_2}
+ 8B^2\eta_{t}\|\nabla J(\theta_{t})\|^2 \big],
\end{aligned}$$ where the last inequality holds by $0< \beta_{t+1} \leq 1$. Since the function $x^{1/3}$ is cancave, we have $(x+y)^{1/3}\leq x^{1/3} + yx^{-2/3}/3$. Then we have $$\begin{aligned}
\eta^{-1}_{t} - \eta^{-1}_{t-1} &= \frac{1}{k}\bigg( \big(m+\sum_{i=1}^tG^2_i\big)^{1/3} - \big(m+\sum_{i=1}^{t-1}G^2_i\big)^{1/3} \bigg) \leq \frac{G^2_t}{3k(m+\sum_{i=1}^{t-1}G^2_i)^{2/3}} \nonumber \\
& \leq \frac{G^2_t}{3k(m - G^2+\sum_{i=1}^{t}G^2_i)^{2/3}} \leq \frac{G^2_t}{3k(m/2 +\sum_{i=1}^{t}G^2_i)^{2/3}} \leq \frac{2^{2/3}G^2_t}{3k(m +\sum_{i=1}^{t}G^2_i)^{2/3}} \nonumber \\
& \leq \frac{2^{2/3}G^2}{3k^3}\eta_t^2 \leq \frac{2^{2/3}G^2}{6k^3L}\eta_t\leq \frac{G^2}{3k^3L}\eta_t,
\end{aligned}$$ where the third inequality holds by $m\geq 2G^2$, and the sixth inequality holds by $0<\eta\leq \frac{1}{2L}$.
Next, considering the upper bound of the term $T_2$, we have $$\begin{aligned}
\label{eq:A5}
T_2 &= \big( \eta^{-1}_{t}(1-\beta_{t+1})(1+8\eta^2_{t}B^2) -\eta^{-1}_{t-1} \big) \|e_{t}\|^2 \nonumber \\
& = \big( \eta^{-1}_{t} - \eta^{-1}_{t-1} + 8B^2\eta_t - \beta_{t+1}\eta^{-1}_{t} - 8\eta_t\beta_{t+1}B^2 \big) \|e_{t}\|^2 \nonumber \\
& \leq \big( \eta^{-1}_{t} - \eta^{-1}_{t-1} + 8B^2\eta_t - \beta_{t+1}\eta^{-1}_{t} \big) \|e_{t}\|^2 \nonumber \\
& \leq \big( \frac{G^2}{3k^3L}\eta_t + 8B^2\eta_t - c\eta_{t} \big) \|e_{t}\|^2= -96B^2\eta_t\|e_{t}\|^2,
\end{aligned}$$ where the last equality holds by $c=\frac{G^2}{3k^3L}+104B^2$. Combining the inequalities with , we have $$\begin{aligned}
\label{eq:A6}
\mathbb{E}[\eta^{-1}_{t}\|e_{t+1}\|^2-\eta^{-1}_{t-1}\|e_t\|^2] \leq \mathbb{E}\big[ 2c^2\eta^{3}_{t}G_{t+1}^2 -96B^2\eta_t\|e_{t}\|^2
+ 8B^2\eta_{t}\|\nabla J(\theta_{t})\|^2 \big].
\end{aligned}$$
We define a *Lyapunov* function $\Phi_t = J(\theta_t) - \frac{1}{128B^2\eta_{t-1}}\|e_t\|^2$ for any $t\geq 1$. Then we have $$\begin{aligned}
\label{eq:A7}
\mathbb{E}[\Phi_{t+1} - \Phi_t] & = \mathbb{E}\big[ J(\theta_{t+1}) - J(\theta_t) - \frac{1}{128B^2\eta_{t}} \|e_{t+1}\|^2 + \frac{1}{128B^2\eta_{t-1}}\|e_t\|^2 \big] \nonumber \\
& \geq \mathbb{E}\big[- \frac{3\eta_t}{4}\|e_t\|^2 + \frac{\eta_t}{8}\|\nabla J(\theta_t)\|^2 - \frac{1}{128B^2}(\eta^{-1}_{t}\|e_{t+1}\|^2 - \eta^{-1}_{t-1}\|e_t\|^2)\big] \nonumber \\
& \geq -\frac{c^2\eta^{3}_{t}G_{t+1}^2 }{64B^2}+ \frac{\eta_t}{16}\mathbb{E}\|\nabla J(\theta_t)\|^2,
\end{aligned}$$ where the first inequality holds by the Lemma \[lem:A3\], and the second inequality follows by the above inequality . Summing the above inequality over $t$ from $1$ to $T$, we obtain $$\begin{aligned}
\label{eq:A8}
\sum_{t=1}^T\mathbb{E}[\eta_t\|\nabla J(\theta_t)\|^2] & \leq \mathbb{E}[16(\Phi_{T+1} - \Phi_1)] + \sum_{t=1}^T\frac{c^2\eta^{3}_{t}G_{t+1}^2 }{4B^2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{8B^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{c^2k^{3}}{4B^2}\sum_{t=1}^T\frac{G_{t+1}^2 }{m+\sum_{i=1}^tG_i^2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{8B^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{c^2k^{3}}{4B^2}\sum_{t=1}^T\frac{G_{t+1}^2 }{G^2+\sum_{i=1}^{t+1}G_i^2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{8B^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{c^2k^{3}}{4B^2}\sum_{t=1}^T\big(\ln(G^2+ \sum_{i=1}^{t+1}G_i^2) - \ln(G^2+ \sum_{i=1}^{t}G_i^2)\big) \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{8B^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{c^2k^{3}}{4B^2}(\ln(G^2+ \sum_{i=1}^{T+1}G_i^2) - \ln(G^2)) \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}}{4B^2}\ln(1+ \sum_{i=1}^{T+1}\frac{G_i^2}{G^2}) \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}}{4B^2}\ln(T+2),\end{aligned}$$ where $J^*=\sup_{\theta}J(\theta) <+\infty$, and the fourth inequality holds by the concavity of the function $\ln(x)$, and the sixth inequality holds by the definition of $e_1$ and $\eta_0$.
By Cauchy-Schwarz inequality, we have $\mathbb{E}[XY]^2 \leq\mathbb{E}[X^2]\mathbb{E}[Y^2]$. Let $X=\sqrt{\eta_T\sum_{t=1}^{T-1}\|\nabla J(\theta_t)\|^2}$ and $Y=\sqrt{1/\eta_T}$, we have $$\begin{aligned}
\mathbb{E}\big[\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big] \leq \mathbb{E}[1/\eta_T]\mathbb{E}\big[ \eta_T\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big].\end{aligned}$$ Since $\eta_t$ is decreasing, we have $$\begin{aligned}
\label{eq:A9}
\mathbb{E}\big[\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big] \leq \mathbb{E}[1/\eta_T]\mathbb{E}\big[\sum_{t=1}^T\eta_T\|\nabla J(\theta_t)\|^2\big] \leq \mathbb{E}[1/\eta_T]\mathbb{E}\big[\sum_{t=1}^T \eta_t\|\nabla J(\theta_t)\|^2\big].\end{aligned}$$ Combining the inequalities and , we obtain $$\begin{aligned}
\label{eq:A10}
\mathbb{E}\big[\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big] &\leq \mathbb{E}\big[ \frac{16(J^* - J(\theta_1)) + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}}{4B^2}\ln(T+2)}{\eta_T} \big] \nonumber \\
& = \mathbb{E}\big[ \Omega\big(m+\sum_{t=1}^TG^2_t\big)^{1/3} \big]\end{aligned}$$ where $\Omega=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}}{4B^2}\ln(T+2)\big)$.
By Assumption 2, we have $G^2_t = \|g(\tau|\theta_t)-\nabla J(\theta_t)+\nabla J(\theta_t)\|^2 \leq 2\|g(\tau|\theta_t)-\nabla J(\theta_t)\|^2 + 2\|\nabla J(\theta_t)\|^2 \leq 2\sigma^2+2\|\nabla J(\theta_t)\|^2$. Then using the inequality $(a+b)^{1/3}\leq a^{1/3} + b^{1/3}$ for all $a,b>0$ to the inequality , we obtain $$\begin{aligned}
\bigg(\mathbb{E}\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg)^2 & \leq \mathbb{E}\bigg[\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg]^2 = \mathbb{E}\big[\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big] \nonumber \\
& \leq \mathbb{E}\bigg[\Omega(m+2T\sigma^2)^{1/3} + 2^{1/3}\Omega\big(\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big)^{1/3}\bigg] \nonumber \\
& =\Omega(m+2T\sigma^2)^{1/3} + 2^{1/3}\Omega\mathbb{E}\bigg[\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg]^{2/3} \nonumber \\
& \leq \Omega(m+2T\sigma^2)^{1/3} + 2^{1/3}\Omega\bigg[\mathbb{E}\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg]^{2/3},\end{aligned}$$ where the first inequality holds by the convexity of the function $x^2$, and the last inequality holds by the concavity of the function $x^{2/3}$. For simplicity, let $Z=\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}$, we have $$\begin{aligned}
\label{eq:A11}
\big(\mathbb{E}[Z]\big)^2 \leq \Omega(m+2T\sigma^2)^{1/3} + 2^{1/3}\Omega\big(\mathbb{E}[Z]\big)^{2/3}.\end{aligned}$$ The inequality implies that $ \big(\mathbb{E}[Z]\big)^2 \leq 2\Omega(m+2T\sigma^2)^{1/3}$ or $ \big(\mathbb{E}[Z]\big)^2 \leq 2\cdot2^{1/3}\Omega\big(\mathbb{E}[Z]\big)^{2/3}$. Thus, we have $$\begin{aligned}
\mathbb{E}[Z] \leq \sqrt{2\Omega}(m+2T\sigma^2)^{1/6} + 2\Omega^{3/4}.\end{aligned}$$ By Cauchy-Schwarz inequality, then we have $$\begin{aligned}
\frac{1}{T}\sum_{t=1}^T\mathbb{E}\|\nabla J(\theta_t)\| &= \mathbb{E} \big[\frac{1}{T}\sum_{t=1}^T\|\nabla J(\theta_t)\|\big] \leq \mathbb{E}\bigg[\sqrt{\frac{1}{T}}\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg] \nonumber \\
& \leq \frac{\sqrt{2\Omega}(m+2T\sigma^2)^{1/6} + 2\Omega^{3/4}}{\sqrt{T}} \nonumber \\
& \leq \frac{\sqrt{2\Omega}m^{1/6} + 2\Omega^{3/4}}{\sqrt{T}} + \frac{2\sqrt{\Omega}\sigma^{1/3}}{T^{1/3}},\end{aligned}$$ where the last inequality follows by the inequality $(a+b)^{1/6}\leq a^{1/6} + b^{1/6}$ for all $a,b>0$.
Convergence Analysis of HA-MBPG Algorithm {#Appendix:A2}
-------------------------------------------
In this subsection, we analyze the convergence properties of HA-MBPG algorithm.
\[lem:B1\] Assume that the stochastic policy gradient $u_t$ be generated from Algorithm \[alg:2\]. Let $e_t = u_t-\nabla J(\theta_t)$, we have $$\begin{aligned}
\mathbb{E}\big[ \eta^{-1}_{t-1}\|e_t\|^2 \big]
\leq 4(W+1)\beta^2_t\eta^{-1}_{t-1}G_t^2 + \eta^{-1}_{t-1}(1-\beta_t)^2\big(1+4\eta^2_{t-1}L^2\big)\mathbb{E} \|e_{t-1}\|^2 + 4(1-\beta_t)^2L^2\eta_{t-1}\|\nabla J(\theta_{t-1})\|^2. \nonumber
\end{aligned}$$
By the definition of $u_t$ in Algorithm \[alg:2\], we have $$\begin{aligned}
u_t - u_{t-1} = -\beta_tu_{t-1}+ \beta_t w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t) + (1-\beta_t) \Delta_t.
\end{aligned}$$ Then we have $$\begin{aligned}
\label{eq:B1}
\mathbb{E}\big[ \eta^{-1}_{t-1} \|e_t\|^2\big]& = \mathbb{E}\big[\eta^{-1}_{t-1} \|\nabla J(\theta_{t-1}) - u_{t-1} + \nabla J(\theta_t)-\nabla J(\theta_{t-1}) -(u_t-u_{t-1})\|^2\big] \\
& = \mathbb{E}\big[ \eta^{-1}_{t-1} \|\nabla J(\theta_{t-1}) - u_{t-1} + \nabla J(\theta_t)-\nabla J(\theta_{t-1}) + \beta_tu_{t-1} - \beta_t w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t)
- (1-\beta_t) \Delta_t\big] \nonumber \\
& = \mathbb{E}\big[ \eta^{-1}_{t-1} \|(1-\beta_t)(\nabla J(\theta_{t-1}) - u_{t-1}) + \beta_t(\nabla J(\theta_t)- w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t))\nonumber \\
& \quad - (1-\beta_t)\big( \Delta_t -(\nabla J(\theta_t)-\nabla J(\theta_{t-1}))\big)\|^2\big] \nonumber \\
& = \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|\nabla J(\theta_{t-1}) - u_{t-1}\|^2 + \eta^{-1}_{t-1}\mathbb{E}\big[\|\beta_t(\nabla J(\theta_t)- w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t))\nonumber \\
& \quad - (1-\beta_t)\big( \Delta_t-(\nabla J(\theta_t)-\nabla J(\theta_{t-1}))\big)\|^2\big] \nonumber \\
& \leq \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|\nabla J(\theta_{t-1}) - u_{t-1}\|^2 + 2\beta^2_t\eta^{-1}_{t-1}\mathbb{E}\|\nabla J(\theta_t)- w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t)\|^2\nonumber \\
& \quad +2(1-\beta_t)^2 \mathbb{E}\| \Delta_t -(\nabla J(\theta_t)-\nabla J(\theta_{t-1}))\|^2 \nonumber \\
& \leq \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 2\beta^2_t\eta^{-1}_{t-1}\underbrace{\mathbb{E}\|w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t)\|^2}_{=T_3}
+ 2(1-\beta_t)^2\eta^{-1}_{t-1} \mathbb{E}\| \Delta_t\|^2, \nonumber
\end{aligned}$$ where the forth equality holds by $\mathbb{E}_{\tau_t\sim p(\tau|\theta_t(\alpha))}[w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t)]=\nabla J(\theta_t)$ and $\mathbb{E}_{\tau_t\sim p(\tau|\theta_t(\alpha))}[\Delta_t]=\nabla J(\theta_t)-\nabla J(\theta_{t-1})$; the first inequality follows by Young’s inequality; and the last inequality holds by $\mathbb{E}\|\zeta-\mathbb{E}[\zeta]\|^2=\mathbb{E}\|\zeta\|^2-(\mathbb{E}[\zeta])^2 \leq \mathbb{E}\|\zeta\|^2$.
Next, we give an upper bound of the term $T_3$ as follows: $$\begin{aligned}
\label{eq:B2}
T_3 & = \mathbb{E}\|w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t)\|^2 \nonumber \\
& = \mathbb{E}\|w(\tau_t|\theta_{t},\theta_t(\alpha))g(\tau_t|\theta_t) -g(\tau_t|\theta_t) + g(\tau_t|\theta_t)\|^2\nonumber \\
& \leq 2\mathbb{E}\|\big(w(\tau_t|\theta_{t},\theta_t(\alpha))-1\big)g(\tau_t|\theta_t)\|^2 + 2\mathbb{E}\|g(\tau_t|\theta_t)\|^2 \nonumber \\
& \leq 2\mathbb{E}\|w(\tau_t|\theta_{t},\theta_t(\alpha))-1\|^2 \mathbb{E}\|g(\tau_t|\theta_t)\|^2 + 2\mathbb{E}\|g(\tau_t|\theta_t)\|^2 \nonumber \\
& \leq 2(W+1)G_t^2,
\end{aligned}$$ where the last inequality holds by Proposition \[pro:1\] and Assumption 3.
Finally, combining the inequalities with , we have $$\begin{aligned}
\mathbb{E}\big[ \eta^{-1}_{t-1} \|e_t\|^2\big]& \leq \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 4\beta^2_t\eta^{-1}_{t-1}(W+1)G_t^2
+ 2(1-\beta_t)^2\eta^{-1}_{t-1} \mathbb{E}\| \hat{\nabla}^2(\theta_t,\tau)v\|^2 \nonumber \\
& \leq \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 4\beta^2_t\eta^{-1}_{t-1}(W+1)G_t^2 + 2(1-\beta_t)^2\eta^{-1}_{t-1} L^2 \mathbb{E}\| \theta_t - \theta_{t-1}\|^2 \nonumber \\
& = \eta^{-1}_{t-1}(1-\beta_t)^2\mathbb{E} \|e_{t-1}\|^2 + 4\beta^2_t\eta^{-1}_{t-1}(W+1)G_t^2 + 2(1-\beta_t)^2L^2\eta_{t-1}\|e_{t-1} + \nabla J(\theta_{t-1})\|^2 \nonumber \\
& \leq 4(W+1)\beta^2_t\eta^{-1}_{t-1}G_t^2 + \eta^{-1}_{t-1}(1-\beta_t)^2\big(1+4\eta^2_{t-1}L^2\big)\mathbb{E} \|e_{t-1}\|^2 + 4(1-\beta_t)^2L^2\eta_{t-1}\|\nabla J(\theta_{t-1})\|^2,\nonumber
\end{aligned}$$ where the second inequality holds by the Proposition 2.
\[th:C1\] Assume that the sequence $\{\theta_t\}_{t=1}^T$ be generated from Algorithm \[alg:2\], and let $k=O(\frac{G^{2/3}}{L})$ $c=\frac{G^2}{3k^3L}+52L^2$, $m = \max\{2G^2,(2Lk)^3,(\frac{ck}{2L})^3\}$ and $\eta_0 = \frac{k}{m^{1/3}}$, we have $$\begin{aligned}
\mathbb{E}\|\nabla J(\theta_\zeta)\|=\frac{1}{T}\sum_{t=1}^T\mathbb{E}\|\nabla J(\theta_t)\| \leq \frac{\sqrt{2\Lambda}m^{1/6} + 2\Lambda^{3/4}}{\sqrt{T}} + \frac{2\sqrt{\Lambda}\sigma^{1/3}}{T^{1/3}}, \nonumber\end{aligned}$$ where $\Lambda=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{4L^2k}\sigma^2 + \frac{(W+1)c^2k^{3}}{2L^2}\ln(T+2)\big)$ with $J^*=\sup_{\theta}J(\theta) <+\infty$.
This proof mainly follows the proof of the above Theorem \[th:A1\]. Due to $m\geq (2Lk)^3$, we have $\eta_t\leq \frac{k}{m^{1/3}}\leq \frac{1}{2L}$. Since $\eta_t\leq \frac{1}{2L}$ and $m\geq (\frac{ck}{2L})^3$, we have $\beta_{t+1}=c\eta^2_t\leq \frac{c\eta_t}{2L}\leq \frac{ck}{2Lm^{1/3}}\leq 1$. By Lemma \[lem:B1\], we have $$\begin{aligned}
\label{eq:C2}
\mathbb{E}[\eta^{-1}_{t}\|e_{t+1}\|^2-\eta^{-1}_{t-1}\|e_t\|^2] & \leq \mathbb{E}\big[ 4(W+1)\beta^2_{t+1}\eta^{-1}_{t}G_{t+1}^2 + \eta^{-1}_{t}(1-\beta_{t+1})^2\big(1+4\eta^2_{t}L^2\big) \|e_{t}\|^2 \nonumber \\
& \quad + 4(1-\beta_{t+1})^2L^2\eta_{t}\|\nabla J(\theta_{t})\|^2 -\eta^{-1}_{t-1}\|e_t\|^2 \big] \nonumber \\
& \leq \mathbb{E}\big[ 4(W+1)c^2\eta^{3}_{t}G_{t+1}^2 + \underbrace{ \big( \eta^{-1}_{t}(1-\beta_{t+1})(1+4\eta^2_{t}L^2) - \eta^{-1}_{t-1} \big) \|e_{t}\|^2 }_{=T_4}
+ 4L^2\eta_{t}\|\nabla J(\theta_{t})\|^2 \big],
\end{aligned}$$ where the last inequality holds by $0< \beta_{t+1} \leq 1$. Since the function $x^{1/3}$ is cancave, we have $(x+y)^{1/3}\leq x^{1/3} + yx^{-2/3}/3$. Then we have $$\begin{aligned}
\eta^{-1}_{t} - \eta^{-1}_{t-1} &= \frac{1}{k}\bigg( \big(m+\sum_{i=1}^tG^2_i\big)^{1/3} - \big(m+\sum_{i=1}^{t-1}G^2_i\big)^{1/3} \bigg) \leq \frac{G^2_t}{3k(m+\sum_{i=1}^{t-1}G^2_i)^{2/3}} \nonumber \\
& \leq \frac{G^2_t}{3k(m - G^2+\sum_{i=1}^{t}G^2_i)^{2/3}} \leq \frac{G^2_t}{3k(m/2 +\sum_{i=1}^{t}G^2_i)^{2/3}} \leq \frac{2^{2/3}G^2_t}{3k(m +\sum_{i=1}^{t}G^2_i)^{2/3}} \nonumber \\
& \leq \frac{2^{2/3}G^2}{3k^3}\eta_t^2 \leq \frac{2^{2/3}G^2}{6k^3L}\eta_t\leq \frac{G^2}{3k^3L}\eta_t,
\end{aligned}$$ where the third inequality holds by $m\geq 2G^2$, and the sixth inequality holds by $0<\eta\leq \frac{1}{2L}$.
Next, considering the upper bound of the term $T_4$, we have $$\begin{aligned}
\label{eq:C3}
T_4 &= \big( \eta^{-1}_{t}(1-\beta_{t+1})(1+4\eta^2_{t}L^2) -\eta^{-1}_{t-1} \big) \|e_{t}\|^2 \nonumber \\
& = \big( \eta^{-1}_{t} - \eta^{-1}_{t-1} + 4L^2\eta_t - \beta_{t+1}\eta^{-1}_{t} - 4\eta_t\beta_{t+1}L^2 \big) \|e_{t}\|^2 \nonumber \\
& \leq \big( \eta^{-1}_{t} - \eta^{-1}_{t-1} + 4L^2\eta_t - \beta_{t+1}\eta^{-1}_{t} \big) \|e_{t}\|^2 \nonumber \\
& \leq \big( \frac{G^2}{3k^3L}\eta_t + 4L^2\eta_t - c\eta_{t} \big) \|e_{t}\|^2= -48L^2\eta_t\|e_{t}\|^2,
\end{aligned}$$ where the last equality holds by $c=\frac{G^2}{3k^3L}+52L^2$. Combining the inequalities with , we have $$\begin{aligned}
\label{eq:C4}
\mathbb{E}[\eta^{-1}_{t}\|e_{t+1}\|^2-\eta^{-1}_{t-1}\|e_t\|^2] \leq \mathbb{E}\big[ 4(W+1)c^2\eta^{3}_{t}G_{t+1}^2 -48L^2\eta_t\|e_{t}\|^2
+ 4L^2\eta_{t}\|\nabla J(\theta_{t})\|^2 \big].
\end{aligned}$$
We define a *Lyapunov* function $\Psi_t = J(\theta_t) - \frac{1}{64L^2\eta_{t-1}}\|e_t\|^2$ for any $t\geq 1$. Then we have $$\begin{aligned}
\label{eq:C5}
\mathbb{E}[\Psi_{t+1} - \Psi_t] & = \mathbb{E}\big[ J(\theta_{t+1}) - J(\theta_t) - \frac{1}{64L^2\eta_{t}} \|e_{t+1}\|^2 + \frac{1}{64L^2\eta_{t-1}}\|e_t\|^2 \big] \nonumber \\
& \geq \mathbb{E}\big[- \frac{3\eta_t}{4}\|e_t\|^2 + \frac{\eta_t}{8}\|\nabla J(\theta_t)\|^2 - \frac{1}{64L^2}(\eta^{-1}_{t}\|e_{t+1}\|^2 - \eta^{-1}_{t-1}\|e_t\|^2)\big] \nonumber \\
& \geq -\frac{(W+1)c^2\eta^{3}_{t}G_{t+1}^2 }{32L^2}+ \frac{\eta_t}{16}\mathbb{E}\|\nabla J(\theta_t)\|^2,
\end{aligned}$$ where the first inequality holds by the Lemma \[lem:A3\], and the second inequality follows by the above inequality . Summing the above inequality over $t$ from $1$ to $T$, we obtain $$\begin{aligned}
\label{eq:C6}
\sum_{t=1}^T\mathbb{E}[\eta_t\|\nabla J(\theta_t)\|^2] & \leq \mathbb{E}[16(\Psi_{T+1} - \Psi_1)] + \sum_{t=1}^T\frac{(W+1)c^2\eta^{3}_{t}G_{t+1}^2 }{2L^2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{4L^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{(W+1)c^2k^{3}}{2L^2}\sum_{t=1}^T\frac{G_{t+1}^2 }{m+\sum_{i=1}^tG_i^2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{4L^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{(W+1)c^2k^{3}}{2L^2}\sum_{t=1}^T\frac{G_{t+1}^2 }{G^2+\sum_{i=1}^{t+1}G_i^2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{4L^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{(W+1)c^2k^{3}}{2L^2}\sum_{t=1}^T\big(\ln(G^2+ \sum_{i=1}^{t+1}G_i^2) - \ln(G^2+ \sum_{i=1}^{t}G_i^2)\big) \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{4L^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{(W+1)c^2k^{3}}{2L^2}(\ln(G^2+ \sum_{i=1}^{T+1}G_i^2) - \ln(G^2)) \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{m^{1/3}}{4L^2k}\sigma^2 + \frac{(W+1)c^2k^{3}}{2L^2}\ln(1+ \sum_{i=1}^{T+1}\frac{G_i^2}{G^2}) \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{m^{1/3}}{4L^2k}\sigma^2 + \frac{(W+1)c^2k^{3}}{2L^2}\ln(T+2),\end{aligned}$$ where $J^*=\sup_{\theta}J(\theta) <+\infty$, and the fourth inequality holds by the concavity of the function $\ln(x)$, and the sixth inequality holds by the definition of $e_1$ and $\eta_0$.
By Cauchy-Schwarz inequality, we have $\mathbb{E}[XY]^2 \leq\mathbb{E}[X^2]\mathbb{E}[Y^2]$. Let $X=\sqrt{\eta_T\sum_{t=1}^{T-1}\|\nabla J(\theta_t)\|^2}$ and $Y=\sqrt{1/\eta_T}$, we have $$\begin{aligned}
\mathbb{E}\big[\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big] \leq \mathbb{E}[1/\eta_T]\mathbb{E}\big[ \eta_T\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big].\end{aligned}$$ Since $\eta_t$ is decreasing, we have $$\begin{aligned}
\label{eq:C7}
\mathbb{E}\big[\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big] \leq \mathbb{E}[1/\eta_T]\mathbb{E}\big[\sum_{t=1}^T\eta_T\|\nabla J(\theta_t)\|^2\big] \leq \mathbb{E}[1/\eta_T]\mathbb{E}\big[\sum_{t=1}^T\eta_t\|\nabla J(\theta_t)\|^2\big].\end{aligned}$$ Combining the inequalities and , we obtain $$\begin{aligned}
\label{eq:C8}
\mathbb{E}\big[\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big] &\leq \mathbb{E}\big[ \frac{16(J^* - J(\theta_1)) + \frac{m^{1/3}}{4L^2k}\sigma^2 + \frac{(W+1)c^2k^{3}}{2L^2}\ln(T+2)}{\eta_T} \big] \nonumber \\
& = \mathbb{E}\big[ \Lambda\big(m+\sum_{t=1}^TG^2_t\big)^{1/3} \big]\end{aligned}$$ where $\Lambda=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{4L^2k}\sigma^2 + \frac{(W+1)c^2k^{3}}{2L^2}\ln(T+2)\big)$.
By Assumption 2, we have $G^2_t = \|g(\tau|\theta_t)-\nabla J(\theta_t)+\nabla J(\theta_t)\|^2 \leq 2\|g(\tau|\theta_t)-\nabla J(\theta_t)\|^2 + 2\|\nabla J(\theta_t)\|^2 \leq 2\sigma^2+2\|\nabla J(\theta_t)\|^2$. Then using the inequality $(a+b)^{1/3}\leq a^{1/3} + b^{1/3}$ for all $a,b>0$ to the inequality , we obtain $$\begin{aligned}
\bigg(\mathbb{E}\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg)^2 & \leq \mathbb{E}\bigg[\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg]^2 = \mathbb{E}\big[\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big] \nonumber \\
& \leq \mathbb{E}\bigg[\Lambda(m+2T\sigma^2)^{1/3} + 2^{1/3}\Lambda\big(\sum_{t=1}^T\|\nabla J(\theta_t)\|^2\big)^{1/3}\bigg] \nonumber \\
& =\Lambda(m+2T\sigma^2)^{1/3} + 2^{1/3}\Lambda\mathbb{E}\bigg[\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg]^{2/3} \nonumber \\
& \leq \Lambda(m+2T\sigma^2)^{1/3} + 2^{1/3}\Lambda\bigg[\mathbb{E}\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg]^{2/3},\end{aligned}$$ where the first inequality holds by the convexity of the function $x^2$, and the last inequality holds by the concavity of the function $x^{2/3}$. For simplicity, let $Z=\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}$, we have $$\begin{aligned}
\label{eq:C9}
\big(\mathbb{E}[Z]\big)^2 \leq \Lambda(m+2T\sigma^2)^{1/3} + 2^{1/3}\Lambda\big(\mathbb{E}[Z]\big)^{2/3}.\end{aligned}$$ The inequality implies that $ \big(\mathbb{E}[Z]\big)^2 \leq 2\Lambda(m+2T\sigma^2)^{1/3}$ or $ \big(\mathbb{E}[Z]\big)^2 \leq 2\cdot2^{1/3}\Lambda\big(\mathbb{E}[Z]\big)^{2/3}$. Thus, we have $$\begin{aligned}
\mathbb{E}[Z] \leq \sqrt{2\Lambda}(m+2T\sigma^2)^{1/6} + 2\Lambda^{3/4}.\end{aligned}$$ By Cauchy-Schwarz inequality, then we have $$\begin{aligned}
\frac{1}{T}\sum_{t=1}^T\mathbb{E}\|\nabla J(\theta_t)\| &= \mathbb{E} \big[\frac{1}{T}\sum_{t=1}^T\|\nabla J(\theta_t)\|\big] \leq \mathbb{E}\bigg[\sqrt{\frac{1}{T}}
\sqrt{\sum_{t=1}^T\|\nabla J(\theta_t)\|^2}\bigg] \nonumber \\
& \leq \frac{\sqrt{2\Lambda}(m+2T\sigma^2)^{1/6} + 2\Lambda^{3/4}}{\sqrt{T}} \nonumber \\
& \leq \frac{\sqrt{2\Lambda}m^{1/6} + 2\Lambda^{3/4}}{\sqrt{T}} + \frac{2\sqrt{\Lambda}\sigma^{1/3}}{T^{1/3}},\end{aligned}$$ where the last inequality follows by the inequality $(a+b)^{1/6}\leq a^{1/6} + b^{1/6}$ for all $a,b>0$.
Convergence Analysis of IS-MBPG\* Algorithm {#Appendix:A3}
--------------------------------------------
In this subsection, we detailedly provide the convergence properties of our IS-MBPG\* algorithm.
\[lem:D1\] Assume that the stochastic policy gradient $u_t$ be generated from Algorithm \[alg:3\], and let $e_t = u_t-\nabla J(\theta_t)$, we have $$\begin{aligned}
\mathbb{E}\big[ \eta^{-1}_{t-1}\|e_t\|^2 \big]
\leq 2\beta^2_t\eta^{-1}_{t-1}\sigma^2 + \eta^{-1}_{t-1}(1-\beta_t)^2\big(1+8\eta^2_{t-1}B^2\big)\mathbb{E} \|e_{t-1}\|^2 + 8(1-\beta_t)^2B^2\eta_{t-1}\|\nabla J(\theta_{t-1})\|^2, \nonumber
\end{aligned}$$ where $B^2 = L^2 + 2G^2C^2_w$ with $C_w = \sqrt{H(2HM_g^2+M_h)(W+1)}$.
The proof is the similar to that of Lemma \[lem:A4\]. The only difference is that instead of using $2\beta^2_t\eta^{-1}_{t-1}\mathbb{E}\|\nabla J(\theta_t) - g(\tau_t|\theta_t)\|^2 \leq 2\beta^2_t\eta^{-1}_{t-1}\sigma^2$ instead of $2\beta^2_t\eta^{-1}_{t-1}\mathbb{E}\|\nabla J(\theta_t) -g(\tau_t|\theta_t)\|^2 \leq 2\beta^2_t\eta^{-1}_{t-1}\mathbb{E}\|g(\tau_t|\theta_t)\|^2 = 2\beta^2_t\eta^{-1}_{t-1}G_t^2$.
\[th:D1\] Assume that the sequence $\{\theta_t\}_{t=1}^T$ be generated from Algorithm \[alg:3\], and let $B^2 = L^2 + 2G^2C^2_w$, $k> 0$ $c=\frac{1}{3k^3L}+104B^2$, $m = \max\{2,(2Lk)^3,(\frac{ck}{2L})^3\}$ and $\eta_0 = \frac{k}{m^{1/3}}$, we have $$\begin{aligned}
\mathbb{E}\|\nabla J(\theta_\zeta)\|=\frac{1}{T}\sum_{t=1}^T\mathbb{E}\|\nabla J(\theta_t)\| \leq \frac{\sqrt{\Gamma}m^{1/6}}{\sqrt{T}} + \frac{\sqrt{\Gamma}}{T^{1/3}}, \nonumber\end{aligned}$$ where $\Gamma=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}\sigma^2}{4B^2}\ln(T+2)\big)$ with $J^*=\sup_{\theta}J(\theta) <+\infty$.
This proof mainly follows the proof of the above Theorem \[th:A1\]. Due to $m\geq (2Lk)^3$, we have $\eta_t\leq \frac{k}{m^{1/3}}\leq \frac{1}{2L}$. Since $\eta_t\leq \frac{1}{2L}$ and $m\geq (\frac{ck}{2L})^3$, we have $\beta_{t+1}=c\eta^2_t\leq \frac{c\eta_t}{2L}\leq \frac{ck}{2Lm^{1/3}}\leq 1$. By Lemma \[lem:D1\], we have $$\begin{aligned}
\label{eq:D1}
\mathbb{E}[\eta^{-1}_{t}\|e_{t+1}\|^2-\eta^{-1}_{t-1}\|e_t\|^2] & \leq \mathbb{E}\big[ 2\beta^2_{t+1}\eta^{-1}_{t}\sigma^2 + \eta^{-1}_{t}(1-\beta_{t+1})^2\big(1+8\eta^2_{t}B^2\big) \|e_{t}\|^2 \nonumber \\
& \quad + 8(1-\beta_{t+1})^2B^2\eta_{t}\|\nabla J(\theta_{t})\|^2 -\eta^{-1}_{t-1}\|e_t\|^2 \big] \nonumber \\
& \leq \mathbb{E}\big[ 2c^2\eta^{3}_{t}\sigma^2 + \underbrace{ \big( \eta^{-1}_{t}(1-\beta_{t+1})(1+8\eta^2_{t}B^2) -\eta^{-1}_{t-1} \big) \|e_{t}\|^2 }_{=T_5}
+ 8B^2\eta_{t}\|\nabla J(\theta_{t})\|^2 \big],
\end{aligned}$$ where the last inequality holds by $0< \beta_{t+1} \leq 1$. Since the function $x^{1/3}$ is cancave, we have $(x+y)^{1/3}\leq x^{1/3} + yx^{-2/3}/3$. Then we have $$\begin{aligned}
\eta^{-1}_{t} - \eta^{-1}_{t-1} &= \frac{1}{k}\bigg( \big(m+t\big)^{1/3} - \big(m+t-1\big)^{1/3} \bigg) \leq \frac{1}{3k(m+t-1)^{2/3}} \nonumber \\
& \leq \frac{1}{3k(m/2 +t)^{2/3}} \leq \frac{2^{2/3}}{3k(m +t)^{2/3}} \nonumber \\
& \leq \frac{2^{2/3}}{3k^3}\eta_t^2 \leq \frac{2^{2/3}}{6k^3L}\eta_t\leq \frac{1}{3k^3L}\eta_t,
\end{aligned}$$ where the second inequality holds by $m\geq 2$, and the fifth inequality holds by $0<\eta\leq \frac{1}{2L}$.
Next, considering the upper bound of the term $T_5$, we have $$\begin{aligned}
\label{eq:D2}
T_5 &= \big( \eta^{-1}_{t}(1-\beta_{t+1})(1+8\eta^2_{t}B^2) -\eta^{-1}_{t-1} \big) \|e_{t}\|^2 \nonumber \\
& = \big( \eta^{-1}_{t} - \eta^{-1}_{t-1} + 8B^2\eta_t - \beta_{t+1}\eta^{-1}_{t} - 8\eta_t\beta_{t+1}B^2 \big) \|e_{t}\|^2 \nonumber \\
& \leq \big( \eta^{-1}_{t} - \eta^{-1}_{t-1} + 8B^2\eta_t - \beta_{t+1}\eta^{-1}_{t} \big) \|e_{t}\|^2 \nonumber \\
& \leq \big( \frac{1}{3k^3L}\eta_t + 8B^2\eta_t - c\eta_{t} \big) \|e_{t}\|^2= -96B^2\eta_t\|e_{t}\|^2,
\end{aligned}$$ where the last equality holds by $c=\frac{1}{3k^3L}+104B^2$. Combining the inequalities with , we have $$\begin{aligned}
\label{eq:D3}
\mathbb{E}[\eta^{-1}_{t}\|e_{t+1}\|^2-\eta^{-1}_{t-1}\|e_t\|^2] \leq \mathbb{E}\big[ 2c^2\eta^{3}_{t}\sigma^2 -96B^2\eta_t\|e_{t}\|^2
+ 8B^2\eta_{t}\|\nabla J(\theta_{t})\|^2 \big].
\end{aligned}$$
We define a *Lyapunov* function $\Phi_t = J(\theta_t) - \frac{1}{128B^2\eta_{t-1}}\|e_t\|^2$ for any $t\geq 1$. Then we have $$\begin{aligned}
\label{eq:D4}
\mathbb{E}[\Phi_{t+1} - \Phi_t] & = \mathbb{E}\big[ J(\theta_{t+1}) - J(\theta_t) - \frac{1}{128B^2\eta_{t}} \|e_{t+1}\|^2 + \frac{1}{128B^2\eta_{t-1}}\|e_t\|^2 \big] \nonumber \\
& \geq \mathbb{E}\big[- \frac{3\eta_t}{4}\|e_t\|^2 + \frac{\eta_t}{8}\|\nabla J(\theta_t)\|^2 - \frac{1}{128B^2}(\eta^{-1}_{t}\|e_{t+1}\|^2 - \eta^{-1}_{t-1}\|e_t\|^2)\big] \nonumber \\
& \geq -\frac{c^2\eta^{3}_{t}\sigma^2 }{64B^2}+ \frac{\eta_t}{16}\mathbb{E}\|\nabla J(\theta_t)\|^2,
\end{aligned}$$ where the first inequality holds by the Lemma \[lem:A3\], and the second inequality follows by the above inequality . Summing the above inequality over $t$ from $1$ to $T$, we obtain $$\begin{aligned}
\label{eq:D5}
\sum_{t=1}^T\mathbb{E}[\eta_t\|\nabla J(\theta_t)\|^2] & \leq \mathbb{E}[16(\Phi_{T+1} - \Phi_1)] + \sum_{t=1}^T\frac{c^2\eta^{3}_{t}\sigma^2 }{4B^2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{8B^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{c^2k^{3}\sigma^2}{4B^2}\sum_{t=1}^T\frac{1}{m+t} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{8B^2\eta_0}\mathbb{E}\|e_1\|^2 + \frac{c^2k^{3}\sigma^2}{4B^2}\sum_{t=1}^T\frac{1}{t+2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{1}{8B^2\eta_0}\sigma^2 + \frac{c^2k^{3}\sigma^2}{4B^2}\sum_{t=1}^T\frac{1}{t+2} \nonumber \\
& \leq \mathbb{E}[16(J^* - J(\theta_1))] + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}\sigma^2}{4B^2}\ln(T+2),\end{aligned}$$ where $J^*=\sup_{\theta}J(\theta) <+\infty$, and the third inequality is due to $m\geq 2$, and the last inequality holds by $\sum_{t=1}^T\frac{1}{t+2} \leq \int^T_1\frac{1}{t+2}dt \leq \ln(T+2)$.
Since $\eta_t=\frac{k}{(m+t)^{1/3}}$ is decreasing, we have $$\begin{aligned}
\label{eq:D6}
\sum_{t=1}^T\mathbb{E}\|\nabla J(\theta_t)\|^2 &\leq 1/\eta_T\sum_{t=1}^T\mathbb{E}\big[ \eta_t\|\nabla J(\theta_t)\|^2\big] \nonumber \\
&\leq \frac{16(J^* - J(\theta_1)) + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}}{4B^2}\ln(T+2)}{\eta_T}\nonumber \\
& = \Omega\big(m+T\big)^{1/3}\end{aligned}$$ where $\Gamma=\frac{1}{k}\big(16(J^* - J(\theta_1)) + \frac{m^{1/3}}{8B^2k}\sigma^2 + \frac{c^2k^{3}\sigma^2}{4B^2}\ln(T+2)\big)$.
According to Jensen’s inequality, we have $$\begin{aligned}
\label{eq:D6}
\frac{1}{T}\sum_{t=1}^T \mathbb{E}\|\nabla J(\theta_t)\| & \leq \big( \frac{1}{T}\sum_{t=1}^T \mathbb{E}\|\nabla J(\theta_t)\|^2\big)^{1/2}
\leq\frac{\sqrt{\Gamma}\big(m+T\big)^{1/6}}{\sqrt{T}}
\leq \frac{\sqrt{\Gamma}m^{1/6}}{\sqrt{T}} + \frac{\sqrt{\Gamma}}{T^{1/3}},\end{aligned}$$ where the last inequality follows by the inequality $(a+b)^{1/6}\leq a^{1/6} + b^{1/6}$ for all $a,b>0$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Financial markets for Liquified Natural Gas (LNG) are an important and rapidly-growing segment of commodities markets. Like other commodities markets, there is an inherent spatial structure to LNG markets, with different price dynamics for different points of delivery hubs. Certain hubs support highly liquid markets, allowing efficient and robust price discovery, while others are highly illiquid, limiting the effectiveness of standard risk management techniques. We propose a joint modeling strategy, which uses high-frequency information from thickly-traded hubs to improve volatility estimation and risk management at thinly traded hubs. The resulting model has superior in- and out-of-sample predictive performance, particularly for several commonly used risk management metrics, demonstrating that joint modeling is indeed possible and useful. To improve estimation, a Bayesian estimation strategy is employed and data-driven weakly informative priors are suggested. Our model is robust to sparse data and can be effectively used in any market with similar irregular patterns of data availability.'
author:
- 'Michael Weylandt[^1]'
- Yu Han
- 'Katherine B. Ensor'
bibliography:
- 'references.bib'
date: 'Last Updated: '
title: Multivariate Modeling of Natural Gas Spot Trading Hubs Incorporating Futures Market Realized Volatility
---
Introduction
============
Financial markets for Liquefied Natural Gas (LNG) are among the most important commodities markets in the United States (U.S.), with their importance rising as natural gas makes up an increasingly large share of U.S. energy consumption. Unlike other major commodities, the pricing of natural gas is largely driven by transportation and storage costs, which are reflected in the correlated price dynamics of different spot prices. In this paper, we develop a joint model for natural gas spot prices which takes advantage of the observable market structure and pools information across different time scales to more accurately forecast volatility for thinly traded spot prices. Our results indicate that joint modeling of disparate natural gas spot prices is able to predict future volatility more accurately than standard univariate models.
Natural gas is a naturally-occurring mixture of hydrocarbons, principally methane ($\text{CH}_4$), which is widely used for both large-scale commercial power generation and domestic use. In 2017, natural gas was used to generate approximately 1.273 petawatthours (PWh) or 1,273 billion killowatthours (kWh) of electricity in the U.S., comprising approximately 31.7% of all electricity generated in the U.S. [@EIA-FAQ]. In 2015, natural gas surpassed coal as the principal source of electricity generation in the U.S. and is forecast to continue to make up a larger proportion of the U.S.’s electricity generation mixture in coming years [@AP:2015; @EIA-STEO-2018-01]. This increase in usage has been primarily driven by the development of hydraulic fracturing (“fracking”) technology, which has driven down the cost of domestic production significantly by allowing natural gas to be extracted from shale efficiently and at relatively low cost [@BLS-2013].
Natural gas is principally stored and transported in a cooled liquid form and, as such, so-called Liquefied Natural Gas markets are the main venue for large-scale commercial trade in natural gas. This trade is centered around a network of many standardized LNG transit and storage centers throughout the United States, colloquially known as “hubs.” (The use of the term “hub” may be somewhat confusing here since it does not imply that the storage center is centrally located or otherwise important, but we will use it throughout this paper as it is a standard terminology in these markets.) These hubs are connected by a nation-wide network of pipelines which connect the hubs to each other, as well as to important population centers and power generation plants, as shown in Figure \[fig:lng\_pipeline\_map\] below. Like any commodity, LNG prices fluctuate unpredictably in response to market forces, sometimes rather significantly. Both upstream producers and downstream consumers, as well as investors in LNG markets from outside the supply chain, have a material interest in measuring and managing their exposure to these fluctuations.
Over-the-counter “spot” markets exist for almost every hub in the U.S., but the liquidity and transparency of these markets varies widely from hub to hub. For a few thickly-traded hubs, the spot market is highly liquid and supports a market of standardized futures contracts and other derivatives, not unlike equity (stock) or foreign exchange (FX) markets. The quality of these markets make them an attractive venue for speculators and market makers, who absorb risk from producers and consumers, improving overall market efficiency. For many other hubs, however, the spot market is thinly traded, with only a few large trades occurring each day, inhibiting price discovery, limiting the attractiveness of spot markets to third parties, and making risk management more difficult.
It has been observed by many authors that the use of high(er)-frequency “realized” volatility measures, based on intra-day price movements, significantly improves volatility estimation and allows for more efficient risk measurement, particularly in settings where a model based solely on daily data would otherwise be slow to react [@Anderson:2005; @Hansen:2012; @Hansen:2016]. It is therefore natural to ask whether similar techniques can be applied to LNG markets, particularly at thinly traded hubs where intra-day data may not be available. This paper answers that question in the affirmative, demonstrating that it is possible to use intra-day data from LNG futures markets to improve volatility estimation at thinly-traded hubs.
Taking the “Realized Beta GARCH” model of @Hansen:2014 as our starting point, we develop a predictive single-factor multivariate volatility model for use in LNG markets. Our model is fully predictive and fully multivariate in nature, estimating volatility for multiple LNG spot and futures markets simultaneously. We adopt a Bayesian approach to parameter estimation in order to address two well-known difficulties of GARCH models: i) by incorporating information from a large pre-existing financial literature into our prior specifications, we are able to regularize our parameter estimation; and ii) by working in a fully Bayesian framework, we are able to coherently propagate uncertainty in our estimated parameters into our risk-management calculations.
The remainder of this paper is organized as follows: Section \[sec:mkts\] provides additional background on LNG markets, after which Section \[sec:lit\] reviews related work. Section \[sec:data\] describes the data set used in the following analysis and highlights several stylized facts that motivate our proposed model. Section \[sec:model\] specifies our proposed model in detail, with Section \[sec:priors\] providing additional detail on suggested priors. Section \[sec:application\] demonstrates the usefulness of our model with an application to U.S. LNG markets, highlighting the improved accuracy on standard risk-management benchmarks. Finally, Section \[sec:disc\] closes with a discussion and a consideration of some directions for future work. Supplemental materials give additional background, data description, and results.
U.S. Natural Gas Markets {#sec:mkts}
------------------------
Conceptually, the simplest mechanism for trading LNG is through so-called “spot” markets. Trades in this market lead to (essentially) immediate exchange of cash for delivery of LNG to the purchaser’s account at a pre-determined hub. Historically, the most important of these hubs is the so-called “Henry Hub,” located immediately outside the town of Erath in southern Louisiana. While LNG markets have expanded nationwide, Henry Hub, and the associated distribution network, remains a key transit nexus and price reference. The transit and storage facilities at Henry Hub are connected to several major interstate LNG pipelines, facilitating easy transit throughout the U.S., as shown in Figure \[fig:lng\_pipeline\_map\]. As such, spot prices at Henry Hub are generally understood to serve as a proxy for U.S. LNG market prices more broadly.
As with many commodities, markets for LNG are heavily financialized, with a wide array of futures, options, and other derivative securities being heavily traded. Because of the central role Henry Hub plays in spot markets, the vast majority of these derivatives are directly or indirectly based on Henry Hub spot prices. The New York Mercantile Exchange introduced futures contracts based on Henry Hub prices in 1990; these contracts are traded at a variety of maturities up to 18 months at an average daily volume of approximately \$14 billion USD and comprise the third largest commodity market in the U.S., exceeded only by crude oil and gold [@CME-2018-Q1]. In recent years, the Intercontinental Exchange has also become a leading venue for LNG derivatives, currently serving as the venue for approximately 40% of total open interest in Henry Hub-referenced futures [@ICE-LNG]. Non-U.S. hubs are of increasing interest, recognizing the evolution and growth of international LNG markets [@EIA-AsiaHubs:2017], but we restrict our focus to hubs in the continental United States.
![U.S. LNG Pipelines (as of 2008). There are approximately 3 million miles of LNG pipelines in the United States, connecting LNG storage facilities with customers. The majority of these pipelines concentrated along the Gulf of Mexico and in western Pennsylvania / northern West Virginia. Note the large number of interstate pipelines originating at or near Henry Hub in southern Louisiana. Figure originally published by the @EIA-MAP.[]{data-label="fig:lng_pipeline_map"}](lng_map_eia_cut.pdf){width="4.5in" height="2.6in"}
Unlike equity or bond markets, the mechanisms by which LNG is stored and transported have a significant effect on price behaviors. @Mohammadi:2011 gives a detailed survey of the pricing structure from well-heads (extraction) to end-consumers. The relationship between LNG spot and futures prices was investigated by @Ghoddusi:2016 who found, *inter alia* that short-maturity futures are Granger-causal for physical prices, a fact which is consistent with our findings on the usefulness of futures realized volatility in spot price volatility modeling. To the best of our knowledge, the subject of this paper – the relationship between the instantaneous volatilities of different spot prices – has not been previously examined in the literature.
Previous Work {#sec:lit}
-------------
Financial time series, including LNG prices, exhibit complex and well-studied patterns, notably significantly non-constant patterns of volatility (heteroscedasticity). This motivates the use of instantaneous volatility, typically understood as the instantaneous standard deviation of returns, as an explicit and important quantity in any realistic model. Direct measurement of volatility, however, is a difficult task, as volatility can change as rapidly or more rapidly than prices can be observed. To address this limitation, two parallel lines of research have been developed in the econometrics literature. The first, latent volatility process models, posit dynamics for the unobserved volatility process and attempt to estimate the parameters of those dynamics directly from returns data; these models are typically further sub-divided into stochastic volatility [@Kim:1998; @Vankov:2019] and (generalized) autoregressive conditional heteroscedasticity (GARCH), depending on the form of the assumed dynamics [@Engle:1982; @Bollerslev:1986]. We focus on GARCH models here as they form the basis of our proposed approach. These models typically assume a daily volatility $\sigma_t^2$ which is never observed directly, but can be estimated over long periods of time using many realizations of the squared daily return $r_t^2 =(\log P_t / P_{t-1})^2$, essentially using only the estimator $\hat{\sigma_t}^2 \approx (r_t - \overline{r})^2$ and using a model to pool estimates through time. Given the fact that these models allow the instantaneous volatility to change rapidly and have so little information per time point, they are often more effective for recovering the parameters of the dynamics than for estimating the instantaneous volatility at a given time point. Despite this, these models have proven immensely popular in the econometric literature, spawning an enormous number of variants and refinements; see, *e.g.*, the thirty-page glossary of @Bollerslev:2010 or the review of @Bauwens:2006.
The second line of research, so-called “realized volatility” models, attempts to enlarge the information set used to estimate the integrated (average) volatility over a fixed period of time. Early work in this direction incorporated additional commonly-cited summary statistics, such as the open (first) price, the intra-day trading range, or combinations thereof [@Rogers:1991]. @Yang:2000 derived an optimal (minimum-variance unbiased) estimator allowing non-zero drift and overnight jumps, which can be interpreted as a weighted sum of the overnight, intra-day, and Rogers-Sachell [-@Rogers:1991] volatility estimators. When appropriately tuned, the Yang-Zhang estimator is up to fourteen times more efficient than the standard close-to-close volatility estimate [@Yang:2000], though its assumption of continuous price dynamics leads it to slightly underestimate volatility in the presence of jumps. Moving beyond “OHLC” (Open, High, Low, Close) data, more recent work uses even higher-frequency data, correcting for the idiosyncrasies of market micro-structure, but these adjustments are typically developed for equities and are quite sensitive to market microstructure, so we do not pursue them here.
Given the successes of both of these lines of research, it is natural to combine them into a single model, in which instantaneous volatility is still assumed to evolve according to some form of GARCH dynamics, but for which we have more accurate estimators than the squared daily return. The Realized GARCH framework introduced by P.R. Hansen and co-authors [@Hansen:2012; @Hansen:2016] is a recent entry in this vein, which allows for slightly more general specifications than the previously proposed Multiplicative Error Model (MEM) or High-Frequency Based Volatility (HEAVY) frameworks [@Engle:2006; @Shephard:2010]. In particular, we build upon the “realized beta GARCH” model of @Hansen:2014 which proposes a model with two series, termed the “market” and the “asset,” where the market has a realized volatility measure available but the asset does not, and uses information from the market to improve estimation for the asset. We adapt this approach for predictive risk management in the LNG market, describing a coherent Bayesian estimation strategy complete with data-driven priors and an out-of-sample VaR forecasting scheme. @Contino:2017 previously considered Bayesian approaches to estimating realized GARCH models, though they do not consider the multivariate realized beta GARCH model that we use here.
The Network GARCH model of @Zhou:2018 is similar to the model we present in Section \[sec:model\], but they assume that the observations are conditionally independent and have the squared return of one asset directly influence the latent volatility of another asset. As we show below, daily returns of LNG spot prices are highly correlated, so the Network GARCH is clearly misspecified for our application. Interpreted as a network, our model encodes the marginal volatilities as a star graph, with Henry Hub at its center and other nodes connected only to it. We leave the problem of allowing and estimating a more general graph structure to future work.
Empirical Characteristics of LNG Prices {#sec:data}
=======================================
Commercial data vendors recognize over one-hundred and fifty tradeable spot LNG prices in the continental U.S. and Canada (see, *e.g.*, the `BGAS` screen on a Bloomberg terminal). Not surprisingly, a full data history is not available for each of these price series. We curate a data set of forty spot prices for which we were able to obtain a full daily price history over the ten-year period ranging from January 2006 to December 2015 (2209 trading days). This period roughly aligns with the third subsample period identified by @Li:2019, who describes it as a large, active, and growing market with lower levels of volatility than observed during the initial emergence of LNG markets. Additionally, we use OHLC data from Henry Hub futures at one, two, three, six, nine, and twelve month maturities to calculate the market return and realized volatility measures. We highlight several key properties of our data set below and give additional details in Section \[app:descrip\] of the Supplementary Materials. Section \[app:series\] of the Supplementary Materials describes our data curation and gives replication instructions.
LNG markets exhibit many of the same well-known characteristics as equity markets, including asymmetric heavy-tailed returns, volatility clustering, and long-lasting autocorrelation of squared returns. Figure \[fig:hhub\_facts\] shows the daily return of the Henry Hub spot, as well as its absolute value and autocorrelation function. From this, it is clear that Henry Hub exhibits volatility clustering and log-lasting autocorrelation in its second moment. This series exhibits strong evidence of heavy tails (sample excess kurtosis of 11.3), but only limited evidence of skewness (sample skewness 0.72). Other spot prices exhibit similar behaviors.
![Henry Hub spot returns. Henry Hub clearly exhibits many of the same properties as equity returns: volatility clustering, heavy-tails, and autocorrelation in the second moment.[]{data-label="fig:hhub_facts"}](hh_facts.pdf){width="6.5in" height="2in"}
Spot prices at different trading hubs tend to move together, with shocks at one hub being quickly reflected in connected hubs and eventually throughout the entire LNG market. As can be seen in Figure \[fig:spots\_together\], the daily price returns at different spots are clearly highly correlated and exhibit similar volatility clustering patterns. For the 40 spot prices considered in Section \[sec:application\], the first principal component explains approximately 74% of the total Spearman covariance, suggesting that a single market factor drives most of the observed variability.
![Return series of several LNG spots. (Details of each spot series are given in Section \[app:series\] of the Supplementary Materials.) Returns are clearly highly correlated across series, with the major volatility events (late 2009 and early 2014) being observed at all hubs. This high degree of correlation suggests that a single-factor model will suffice.[]{data-label="fig:spots_together"}](multi_spots.pdf){width="6.5in" height="2.2in"}
LNG spot prices, particularly those for Henry Hub, are also closely connected to the associated futures market. As can be seen in the top panel of Figure \[fig:futures\], Henry Hub spot and futures are highly correlated, with the correlation highest for the shortest maturity futures. Similarly, spot and futures volatility estimates are also highly correlated, as shown in the bottom panel of Figure \[fig:futures\]. Because the futures market is typically more liquid than the spot market, we can use the additional information provided by this market to improve estimation of spot volatilities.
Putting these results together, we see that the spot LNG market exhibits equity-like volatility patterns with strong correlations among different spots and between the spot and futures markets. Spot prices exhibit quite heavy tails, but relatively little skewness, suggesting the use of a skewed observation distribution rather than an asymmetric GARCH formulation. Finally, the dominant leading principal component, indicative of market-wide volatility shifts, suggests that a single factor model will perform well. We will incorporate each of these observations as we develop our proposed model in the following section.
![Comparison of Henry Hub Spot and Futures. As shown in the top panel, spot and futures returns are moderately correlated (25-29%), particularly at shorter maturities (29.2%). As shown in the bottom panel, spot and futures volatitlities are highly correlated (74-99%) for all of the realized volatility measures considered, including the close-to-close, Garman-Klass [-@Garman:1980], Rogers-Satchell [-@Rogers:1991], and Yang-Zhang [-@Yang:2000] estimators. In both cases, we observe strong correlation between the low frequency spot data and the high frequency futures data.[]{data-label="fig:futures"}](futures.pdf){width="6.5in" height="3.2in"}
Model Specification {#sec:model}
===================
In this section, we introduce our model piecewise, giving motivation for our modeling choices as we proceed. The complete specification is given in Equation at the end of this section. Conceptually, our model has three major components: a standard GARCH specification for the volatility at Henry Hub ($\sigma_{M, t}$), multiple realized volatility measures ($\varsigma_{j, t}$), and an augmented GARCH specification for the non-Henry Hub spots which includes an additional term ($\sigma_{M, t}^2$) to capture market-wide changes in the volatility. Because our eventual aim is risk management, our model is fully predictive: in particular, volatilities at time $t$ ($\sigma_{M, t}, \sigma_{i, t}$) depend only on past values, allowing for out-of-sample forecasting, rather than having $\sigma_{M, t}$ influence $\sigma_{i, t}$ as in the specification of @Hansen:2014.
We adopt a slight variant of a standard linear GARCH(1, 1) specification for the instantaneous volatility at Henry Hub: $$\sigma_{M, t}^2 = \omega_M + \gamma_M \sigma_{M, t-1}^2 + \sum_{j=1}^J\zeta_j \varsigma_{j, t - 1}^2 + \tau_{M, 1} |r_{M, t - 1}| + \tau_{M, 2}(r_{M, t-1} - \mu_M)^2$$ where $r_{M, t}$ is the daily Henry Hub spot price return at $t$, $\mu_M$ is the long-run average return, and $\varsigma_{j, t - 1}$ is the value of the $j$^th^ realized volatility measure at time $t - 1$. @Chan:2016 note that, contrary to crude oil markets, the presence of a leverage term has minimal predictive value in LNG markets. As such, we follow @Hansen:2012 and use a second-order Hermite polynomial in $|r_{\cdot, t}|$ to allow for flexible modeling of leverage, rather than enforcing leverage directly in our specification. Note that, since we are using a linear rather than log-linear formulation, we include an absolute value term to disallow negative volatilities. The realized volatility term ($\zeta_j^2$) allows our model to indirectly incorporate aspects of volatility captured by the realized volatility measurements which are not otherwise captured by the close-to-close return.
The individual asset volatilities follow a similar specification, though we introduce an additional “coupling” term to capture market-wide changes in volatility where Henry Hub is used as a proxy for the market as a whole: $$\sigma_{i, t}^2 = \omega_i + \gamma_{i} \sigma_{i, t-1}^2 + \tau_{i, 1} |r_{i, t-1}| + \tau_{M, 2}(r_{i, t-1} - \mu_i)^2 + \beta_{i} \sigma^2_{M, t-1} \quad \forall i=1, \dots, I$$ where $r_{i, t}$ is the return of (non-Henry Hub) spot price $i$ at time $t$. The coupling parameter $\beta_i$ measures the influence of Henry Hub volatility on the other hubs. If $\beta_i = 0$, then Henry Hub volatility is minimally informative of the volatility at hub $i$ (conditionally independent given the past observables), while a large value of $\beta_i$ indicates that the secondary hub experiences significant “spill-over” volatility from Henry Hub.
Conditional on these volatilities, we assume that the daily returns follow a multivariate skew-normal distribution, though any skewed elliptical distribution, *e.g.,* the multivariate skewed $t$ distribution, could be used as well. In other words, $$\begin{aligned}
\vec{r}_t &= \begin{pmatrix} r_{M, t} \\ \{r_i, t\}_{i=1}^I \end{pmatrix}\sim \text{Multi-Skew-Normal}\left(\alpha, \mu, \Sigma_t \right) \\
\Sigma_t &= \text{diag}\left(\sigma_{M, t}, \{\sigma_{i, t}\}_{i=1}^k\right)\; \Omega\; \text{diag}\left(\sigma_{M, t}, \{\sigma_{i, t}\}_{i=1}^k\right)\end{aligned}$$ where $\alpha$ and $\mu$ are fixed (non-time varying) skewness and mean parameters and $\Omega$ is the fixed (non-time varying) correlation of returns. Because our data exhibits relatively low degrees of skewness, we do not impose strong skewness through our model specification, instead allowing for the possibility of skewness by combining the linear GARCH specification with a skew-normal return distribution [@Azzalini:1999]. Empirically, we have found this scheme to perform better than the log-linear specification described by @Hansen:2012, but, as they note, the success of the realized GARCH framework does not depend critically on the specification used.
Finally, we include an explicit model for the realized volatility measures: $$\varsigma_{j, t} \sim \mathcal{N}(\xi_j + \phi_j \sigma_{M, t} + \delta_{j, 1} |r_{M, t}| + \delta_{j, 2} r_{M, t}^2, \nu_j^2) \quad \forall j=1, \dots, J.$$ As before, we incorporate a leverage effect using a second order Hermite polynomial. We emphasize that the realized volatility measures need not be unbiased estimates of the true market volatility at time $t$. The bias ($\xi_j$) and scaling ($\phi_j$) factors allow for mis-scaled or mis-aligned volatility measures. In other words, an intra-day volatility measure does not need to be re-scaled to daily volatility. This is especially important in our application below, where realized volatility measures from futures markets are used for spot markets. (See Figure \[fig:futures\] for evidence of different scalings.) The bias and scaling terms allow these estimates to inform our estimation, even though our model does not directly account for the particular dynamics of the futures markets (*i.e.*, cost of carry, interest rate dynamics, *etc.*; see the discussion by @Li:2019). The use of drift-adaptive volatility measure of @Yang:2000 also helps to ameliorate these effects.). As we will see, our model is able to adapt to the mismatch between the futures and spot markets, demonstrating that, despite this discrepancy, the futures market realized volatility does indeed constitute a useful addition to spot models.
Putting these pieces together, we obtain the complete specification: $$\begin{aligned}
\vec{r}_t &= \begin{pmatrix} r_{M, t} \\ \{r_i, t\}_{i=1}^I \end{pmatrix}\sim \text{Multi-Skew-Normal}\left(\alpha, \mu, \Sigma_t \right) \nonumber \\
\Sigma_t &= \text{diag}\left(\sigma_{M, t}, \{\sigma_{i, t}\}_{i=1}^k\right)\; \Omega\; \text{diag}\left(\sigma_{M, t}, \{\sigma_{i, t}\}_{i=1}^k\right) \nonumber \\
\sigma_{M, t}^2 &= \omega_M + \gamma_{M} \sigma_{M, t-1}^2 + \sum_{j=1}^J \zeta_j \varsigma_{j, t-1}^2 + \tau_{M, 1} |r_{M, t-1}| + \tau_{M, 2}(r_{M, t-1} - \mu_M)^2 \label{eqn:model}\\
\sigma_{i, t}^2 &= \omega_i + \gamma_{i} \sigma_{i, t-1}^2 + \tau_{i, 1} |r_{i, t-1}| + \tau_{i, 2}(r_{i, t-1} - \mu_i)^2 + \beta_{i} \sigma^2_{M, t-1} \quad \forall i=1, \dots, I \nonumber \\
\varsigma_{j, t} &\sim \mathcal{N}(\xi_j + \phi_j \sigma_{M, t} + \delta_{j, 1} |r_{M, t}| + \delta_{j, 2} r_{M, t}^2, \nu_j^2) \quad \forall j=1, \dots, J \nonumber\end{aligned}$$ where
$r_{M, t}$
: is the market return at time $t$ (observed);
$r_{i, t}$
: is the return of (non-Henry Hub) spot price $i$ at time $t$ (observed);
$\alpha, \mu$
: are fixed (non-time-varying) return parameters (unobserved);
$\Sigma_{t}$
: is the conditional variance at time $t$ (unobserved);
$\sigma_{M, t}$
: is the instantaneous market volatility at time $t$ (unobserved);
$\sigma_{i, t}$
: is the instantaneous volatility of spot price $i$ at time $t$ (unobserved);
$\beta_i$
: measures the effect of market volatility on the volatility of spot price $i$ (unobserved);
$\zeta_i$
: measures the influence of realized volatility measure $j$ on market volatility (unobserved);
$\omega, \gamma, \tau$
: are fixed (non-time-varying) GARCH parameters (unobserved);
$\varsigma_{j, t}$
: is a realized measure of market volatility at time $t$ (observed); and
$\xi, \phi, \delta, \nu$
: are fixed (non-time-varying) parameters of the realized volatility measurement (unobserved).
Bayesian Estimation and Prior Selection {#sec:priors}
---------------------------------------
To estimate the parameters of our model, we adopt a Bayesian approach. As we will see below, this poses several advantages, perhaps the most obvious of which is that it allows us to incorporate information from the large GARCH literature in the form of a prior. A less obvious, but equally important, advantage is that the Bayesian framework allows for coherent propagation of parameter uncertainty into subsequent analyses. In the financial context, this typically produces more extreme, but ultimately more accurate, risk measures [@Ardia:2017].
The choice of priors is fundamental to any Bayesian model. Rather than developing priors based on theoretical considerations, we derive priors from a related but independent data set where possible. In particular, we fit a (univariate) realized GARCH model to the returns of the S&P 500, an index of major U.S. stocks, and use the maximum likelihood estimates to center our priors for LNG. This process is repeated over 250 of randomly-chosen year-long periods during our sample period (2006-2015) and the median estimate is used as the mean of the prior. The prior standard deviation is matched to ten times the median absolute deviation of the MLEs. This yields the prior specifications shown in Table \[tab:priors\]. For priors which can be well estimated from the data, *e.g.*, the mean return $\mu$, or the fixed correlation matrix, $\Omega$, we use a standard weakly informative prior. Throughout, we also constrain our priors, and hence our posterior, to ensure stationarity of the underlying process.
To ensure the reasonableness of our priors, we generate simulated trajectories with GARCH parameters drawn from our priors as a “prior predictive check.” On visual inspection of simulated price series, we observe that our priors imply realistic GARCH dynamics, though the volatility is higher than what we actually observe, consistent with our use of weakly informative (wide) priors on the GARCH parameters, which allow for higher-than-realistic volatilities. Table \[tab:prior\_pc\] compares the 90% prior predictive intervals of several relevant statistics against the S&P 500 data used to calibrate our priors and against Henry Hub spot returns. Our priors imply a kurtosis below the observed kurtosis of the S&P 500 and Henry Hub. This divergence appears to be driven by the model not accounting for large exogenous jumps rather than an incorrect GARCH specification. Overall, we see that our priors are realistic, even though they often imply a higher “volatility of volatility” than actually observed.
Computational Concerns
----------------------
Full Bayesian inference for Model requires use of a Markov Chain Monte Carlo (MCMC) sampler. Because the latent volatilities, $\sigma_{M, T}$ and $\sigma_{i, T}$, are typically highly correlated across time, Hamiltonian Monte Carlo samplers are particularly well suited for this problem. We make use of the `Stan` probabilistic programming language and its variant of the No-U-Turn Sampler [@Carpenter:2017; @Hoffman:2014]. While `Stan` is able to sample the posterior relatively efficiently for this problem, simultaneous inference for a large number of assets, *e.g.*, the 40 spot prices considered in the next section, is still computationally quite demanding. To avoid this, we take advantage of the factorizable structure of Model and compute partial posteriors for each asset separately and combine our partial posteriors to obtain an approximate posterior. The parameters of the joint return distribution can then be estimated efficiently after by treating the returns as independent samples after standardization by the estimated instantaneous volatilities. Additional details of our Markov Chain Monte Carlo procedure and convergence diagnostics are included in Section \[app:computation\] of the Supplementary Materials.
Application to Financial Risk Management {#sec:application}
========================================
We demonstrate the effectiveness of our model with an application to LNG spot prices at non-Henry hubs. We fit our model, under the priors described in Section \[sec:priors\], on a 10 year historical period, refitting every 50 days using a 250-day look-back window. This periodic refitting strategy implicitly allows for time-varying coefficients, unlike the approach of @Hansen:2014, who only allow for time-varying correlations ($\Omega$) with a parametric specification for the temporal dynamics. Unless otherwise stated, all volatility measures are posterior expectations taken over 4000 posterior samples. Throughout, we compare the results of Model , which we denote as [`RBGARCH`]{}for realized beta GARCH, with a pair of standard skew-normal GARCH(1, 1) models, which we denote as [`GARCH`]{}. To ensure a fair comparison, the same model specification and priors are used for both [`GARCH`]{}and [`RBGARCH`]{}. While the [`RBGARCH`]{}model is able to convincingly outperform the [`GARCH`]{}model in-sample, the superior out-of-sample risk-management performance is particularly compelling evidence as to the usefulness of the realized beta GARCH model in LNG markets. Additional comparisons are given in Section \[app:additional\_results\] of the Supplementary Materials.
In-Sample Model Fit
-------------------
We first compare the in-sample model fit of the [`GARCH`]{}and [`RBGARCH`]{}models. To assess improvement in in-sample performance, we use @Watanabe:2010’s [-@Watanabe:2010] *Widely Applicable Information Criterion* (WAIC) . Essentially a Bayesian analogue of AIC, WAIC is an approximate measure of the leave-one-observation-out predictive performance of a model. The likelihood of the [`RBGARCH`]{}model contains terms for both the daily returns and the realized volatility measure. To ensure a fair comparison with the [`GARCH`]{}model, we calculate the WAIC of the [`RBGARCH`]{}model on a *partial* likelihood, using only the likelihood of the daily returns. The partial likelihood is analogous to the complete likelihood of the [`GARCH`]{}model, which does not include a term for the realized volatility measure, and allows for an apples-to-apples WAIC comparison of the two models.
In Figure \[fig:waic\], we compare the in-sample performance of the [`GARCH`]{}and [`RBGARCH`]{}models using a signed-rank transform to account for the highly non-normal sampling distribution of the WAIC differences. The left portion of Figure \[fig:waic\] shows a small number of sub-periods where the WAIC comparison appears to strongly favor the [`GARCH`]{}model. A closer investigation of these sub-periods reveals that these irregularities are caused by short-lived outliers in the underlying data which corrupt the WAIC estimates. @Vehtari:2017 suggest that if the posterior variance of posterior log-predictive density exceeds 0.4 for any single observation, WAIC does not provide a robust estimate to out-of-sample predictive accuracy. Applying this heuristic to our data, we see that almost all of the sub-periods where [`GARCH`]{}appears to out-perform [`RBGARCH`]{}have WAIC instabilities and that, ignoring those outliers, the [`RBGARCH`]{}model consistently has superior in-sample performance, even after adjusting for the added complexity of the [`RBGARCH`]{}model.
A paired one-sided Wilcoxon test comparing the WAICs of the [`GARCH`]{}and [`RBGARCH`]{}models was performed, with the alternative hypothesis that the [`RBGARCH`]{}model has a higher median WAIC across our 73 time periods (not omitting the outlier-corrupted periods). We see consistent evidence that [`RBGARCH`]{}outperforms, with 37 out of 40 $p$-values less than 0.10 and 34 out of 40 less than 0.05. (Note that the non-parametric Wilcoxon test was used because the sampling distribution of WAIC is highly non-normal in small samples and extremely sensitive to the heteroscedasticity of the underlying data.) Overall, we see consistent evidence that, by jointly modeling the volatility of spot and futures returns, the [`RBGARCH`]{}model is able to consistently capture price dynamics more accurately than the [`GARCH`]{}model.
![Signed-rank-transformed differences of 40 spots over 73 periods. Ignoring periods where our WAIC estimates are corrupted by outliers (purple), the [`RBGARCH`]{}model performs better the [`GARCH`]{}model in 2333 out of 2444 (95.4%) sub-samples.[]{data-label="fig:waic"}](waic_with_stability_main.pdf){width="\textwidth"}
In-Sample Risk Management
-------------------------
We consider a multi-asset portfolio containing a mixture of spot LNG and one-month Henry Hub futures. In-sample VaR estimates are computed by taking quantiles of the posterior predictive distribution of portfolio returns using the posterior distribution of the mean return parameters $\mu, \alpha$ and the estimated conditional variance $\Sigma_t$. To assess the accuracy of our VaR estimates, we use the (unconditional) test of @Kupiec:1995, which compares the number of times that the actual losses are larger than the estimated quantile (VaR) with the theoretical rate using a binomial proportion test. The Kupiec test is a misspecification test, where smaller values indicate stronger evidence that the VaR estimates are not correctly calibrated. Results of this test are shown in Figure \[fig:in\_sample\_var\]. As this figure shows, the [`RBGARCH`]{}model consistently provides more accurate in-sample VaR estimates than the [`GARCH`]{}model. We note that, because it models the volatilities jointly and does not assume independence, the [`RBGARCH`]{}model does particularly well for balanced portfolios where the correlation among assets is particularly important.
![In-Sample VaR Accuracy of the [`GARCH`]{}and [`RBGARCH`]{}models, as measured by @Kupiec:1995’s [-@Kupiec:1995] unconditional exceedances test. Larger $p$-values indicate better performance. The [`RBGARCH`]{}model performs well for all portfolios considered, while the [`GARCH`]{}model shows clear signs of inaccurate VaR calculation for all portfolios except the 20% Spot/80% Futures portfolio. Displayed $p$-values are for in-sample VaR estimates for the entire sample period.[]{data-label="fig:in_sample_var"}](in_sample_kupiec_500_small.pdf){width="\textwidth" height="2.5in"}
Out-of-Sample Risk Management
-----------------------------
The ultimate test of a financial model is its out-of-sample performance, so we compare the out-of-sample performance of the [`GARCH`]{}and [`RBGARCH`]{}models, again focusing on VaR forecasting. Forward-looking (out-of-sample) volatility predictions were obtained by (forward) filtering observed returns for each posterior sample and then taking the quantile of the posterior predictive distribution for each day. In other words, out-of-sample observations were used to update the conditional forward-looking volatility estimates, but were not used for parameter re-estimation. While we only consider one-day ahead prediction here, it is straightforward to extend these results to longer forecast horizons using standard techniques, *e.g.*, the data-driven aggregation method proposed by @Hamidieh:2010.
Results of this test are shown in Figure \[fig:out\_sample\_var\] for the same two-asset portfolios as considered in the previous section. The [`RBGARCH`]{}model continues to consistently outperform the [`GARCH`]{}model, particularly for the portfolios in which the spot asset is heavily weighted. In particular, while the performance of the [`GARCH`]{}model is significantly worse out-of-sample, the [`RBGARCH`]{}model performs almost as well in-sample as it does out-of-sample performance. A closer examination reveals that, while the [`GARCH`]{}model significantly underestimates volatility, the [`RBGARCH`]{}model produces slightly over-conservative out-of-sample VaR estimates, a feature which we consider acceptable, and possibly even desirable, in applications.
![Rolling Out-Of-Sample VaR Accuracy of the [`GARCH`]{}and [`RBGARCH`]{}models, as measured by @Kupiec:1995’s [-@Kupiec:1995] unconditional exceedances test. Larger $p$-values indicate better performance. The [`RBGARCH`]{}model outperforms the [`GARCH`]{}model for the seven different portfolio weights considered here. Displayed $p$-values are for out-of-sample filtered VaR estimates for the entire sample period.[]{data-label="fig:out_sample_var"}](out_sample_kupiec_500_small.pdf){width="\textwidth" height="2.5in"}
Conclusions {#sec:disc}
===========
We present a model for joint volatility estimation at several LNG hubs and demonstrate that, despite the apparent mismatch, information from Henry Hub futures can be used to improve volatility estimation for non-Henry Hub spot prices. We provide suggested priors for use in Bayesian estimation and demonstrate the effectiveness of the Bayesian approach for accurate risk prediction. Our model produces improved volatility estimates, which exhibit more accurate estimates of unconditional volatility levels, increased responsiveness to external market shocks, and improved VaR estimation. Notably, our joint model performs as well out-of-sample as it does in-sample, making it particularly promising for risk management applications. This impressive out-of-sample performance can be traced back to our use of realized volatility information from the futures market: by taking advantage of an additional highly-accurate realized volatility measure, the Realized Beta GARCH is able to adapt to changing market conditions much more rapidly than standard approaches. Joint modeling of different aspects of LNG markets is an effective and flexible strategy for incorporating inconsistently available data and provides useful insights into the degree and dynamics of LNG spot price volatility and inter-asset correlation. The joint modeling framework may be easily extended to include additional realized volatility measures or to allow more complex market dynamics.
Acknowledgements {#acknowledgements .unnumbered}
----------------
We thank the El Paso Corporation Finance Center at the Jones Graduate School of Business at Rice University and the Rice Center for Computational Finance and Economic Systems (CoFES) for assistance acquiring historical spot and futures price data. We also thank CoFES for providing computational resources. MW acknowledges support from NSF Graduate Research Fellowship Program under grant number 1842494. The authors have no conflicts of interest to declare.
Supplemental Materials {#supplemental-materials .unnumbered}
----------------------
Supplemental Materials to be published alongside the online version of this paper include additional descriptive analysis of our sample data, more details of our in- and out-of-sample performance, code necessary to replicate our data set, and a more detailed discussion of computational aspects of our approach, including the `Stan` code implementing Model .
[**Supplemental Materials**]{}
Additional Data Description {#app:descrip}
===========================
In this section, we provide some additional characterization of the data used for the empirical study in Section \[sec:application\] of our paper. As discussed in Sections \[sec:mkts\] and \[sec:data\], Henry Hub spot prices play an important role in both the spot and futures markets. Figure \[fig:hhub\_spot\] shows the spot price of Henry Hub futures for the sample period, as well as (annualized) rolling return and volatility estimates, sample skewness, and sample (non-excess) kurtosis. As @Li:2019 notes our sample period is characterized by relatively low volatility, with changes in volatility, skewness, and kurtosis being driven by periods of elevated volatility in late 2009 and in early 2014. Given this low level of volatility, we do not include a jump component in our model, contrary to the findings of @Mason:2014 who find the presence of a jump component significantly improves model fit in an earlier, more volatile, period.
LNG futures prices are strongly correlated across different maturities, with the correlation remaining roughly constant across time as shown in Figure \[fig:hhub\_futures\]. Futures prices exhibit less pronounced jump behavior than the underlying spot price, particularly in the latter portion of our sample, leading to reduced correlation between future and sport return at a daily frequency. At lower frequencies, the effects of these jumps are reduced and the correlation between spot and futures returns is higher, as shown in Figure \[fig:hhub\_futures\_cor\].
Given that the effects of storage and production shocks are felt throughout LNG markets, we would expect the *volatilities* of spot prices are correlated as well. As can be seen in Figure \[fig:spot\_vol\], this is indeed the case – spot price volatilities are typically highly correlated, though not to quite the same degree as spot price returns. As before, there is clear evidence of a subgroup of spot prices that are more closely correlated among each other than to the rest of the market.
While the spot and futures markets generally move together, they are still somewhat disjoint, with shocks to the futures market not necessarily being reflected in the spot market and *vice versa*. Despite this disjoint behavior, there is still a strong correlation between realized volatility in the futures markets and close-to-close volatility in the spot markets. Figure \[fig:yz\_vol\] illustrates this correlation, using several different realized volatility measures and futures maturities. Henry Hub clearly displays a higher level of correlation with futures volatility than the other spot prices in our sample, consistent with the fact that LNG futures are based on Henry Hub prices. Regardless, futures volatility is indeed a useful predictor of volatility at non-Henry spots, as demonstrated in Section \[sec:application\].
The observed correlation does not appear to be sensitive to the choice of realized volatility measure, suggesting our results are robust to the exact form of external information being supplied. Shorter maturity futures, especially one-month futures (NG1), consistently show the highest correlation with spot prices, so we use the Yang-Zhang estimator applied to the generic fixed-maturity NG1 in our application. The high-degree of correlation suggests that single-factor volatility models, including our proposed model, will be able to capture most of the second-order dynamics. This is consistent with the findings of @Kyj:2009 who, like us, combine a single-factor model with realized volatility measures and find significant improvements in equity portfolio construction.
Data Set Replication Information {#app:series}
================================
Table \[tab:series\] lists the Bloomberg identifiers used to obtain the data for our analysis. For spot prices, only the last traded price of the day (Bloomberg key `PX_LAST`) was used. These spot prices were selected from Bloomberg’s full listing of LNG spot prices (available through the `BGAS` screen) as having a relatively complete trading history (more than 245 trading days per year) for the entire sample period (2006 - 2015). Days when only certain spots traded are omitted from our analysis to ensure an aligned data set. After screening, the `NGCDAECO` spot price (Alberta, Canada) was manually removed as it is the same spot as `NGCAAECO`, but reported in Canadian dollars instead of U.S. dollars. For futures prices, Open, High, Low, and Close prices were obtained using Bloomberg keys `PX_OPEN`, `PX_HIGH`, `PX_LOW`, and `PX_CLOSE` respectively.
As described in the main text, the reported Open and Close prices are occasionally outside of the daily trading range (the High-Low range). This is not a data error, but is instead an artifact of the Open and Close prices being set by an auction process which is not included in the High and Low calculations, rather than the standard market mechanisms. Despite this, the Yang-Zhang volatility estimator is derived assuming a continuous time model for the price series and makes no distinction between traded and non-traded prices. As such, it is not well-defined when the Open or Close prices fall outside of the intraday trading range. This occurs somewhat regularly in our data set, as detailed in Table \[app:tab:inconsistencies\], particularly at longer maturities. To address this, we restrict the Open and Close prices to the intraday trading range. This potentially results in a small loss of estimation efficiency, but empirically appears to have minimal effect, particularly for our application (Section \[sec:application\]), where only the one month futures are used.
![Henry Hub spot prices and rolling estimates of the first four moments. Spot prices peaked in mid-2008, declined from late 2008 to early 2009 and have remained at a relatively constant level since. Spot returns exhibit moderate heteroscedasticity, skewness and kurtosis, with large changes in the sample moments being primarily driven by high-volatility periods in Fall 2009 and Spring 2014.[]{data-label="fig:hhub_spot"}](hh_summary.pdf){width="\textwidth"}
![Pearson correlation among daily spot price returns for 40 hubs in the sample period. Spot prices are typically highly correlated with each other and with Henry Hub (not shown here). The Cheyenne, Northeast Transco Zone 6, Texas Eastern Transmission, Tennessee Gas Pipeline, Northeast, and Iroquois Transmission System spot prices, however, are more highly correlated among themselves than to other spots during the sample period.[]{data-label="fig:corrgram"}](corrplot.pdf){width="\textwidth"}
![Henry Hub futures move in tandem with spot prices, but generally exhibit fewer large jumps. In particular, the futures markets do not exhibit the large jumps in early 2014 that are a significant driver of the elevated kurtosis observed in spot prices during the same period.[]{data-label="fig:hhub_futures"}](hh_futures.pdf){width="\textwidth"}
![Daily returns of Henry Hub futures are highly correlated, but only exhibit moderate correlation with Henry Hub spot returns. Correlations among spot and futures returns are higher at lower frequencies (weekly, monthly, *etc.*).[]{data-label="fig:hhub_futures_cor"}](corrplot_futures.pdf){width="\textwidth"}
![Pearson correlation among the estimated volatilities of 40 spot prices during the sample period. As we see, there is generally a strong positive correlation among spot volatilities, though it is not as pronounced as the correlation among spot returns. There is some evidence that the structure of volatility is changing over time, as can be seen by the emergence of subgroups in the 2012-2015 period.[]{data-label="fig:spot_vol"}](corrplot_spotvol.pdf){width="\textwidth"}
![Pearson correlation among the Henry Hub spot volatility, non-Henry Hub spot volatility, and futures realized volatility during the sample period. The reported correlations with non-Henry spots are the average correlation over the 40 non-Henry spots in our sample. The correlation between Henry Hub volatility and short-maturity (NG1) realized volatility is consistently the highest. The degree of correlation appears robust to the choice of realized volatility measure.[]{data-label="fig:yz_vol"}](corrplot_futuresvol.pdf){width="\textwidth"}
[**Data Type**]{} [**Identifier**]{} [**Full Name**]{}
------------------- -------------------- ------------------------------------------------------
NAGAALLI Alliance Pipeline Delivered
NAGAANRL Mid-Continent / ANR Lavrne (Custer, OK)
NAGAMICG Michigan Consolidated Gas (Detroit, MI)
NAGANGMC Mid-Continent Natural Gas Spot Price
NAGANGPL Mid-Continent / Chicago Citygate (Chicago, IL)
NAGANGTO Texas-Oklahoma East (Montgomery County, TX)
NAGANOND N. Natural Mainline (Clay County, KS)
NAGANORB N. Border Natural Gas Spot Price (Ventura, IA)
NGCAAECO AECO C Hub (Alberta, Canada)
NGCGNYNY TETCO M3 (New York, NY)
NGGCANRS Gulf Coast / ANR Southeast
NGGCCGLE Columbia Transmission TCO Pool (Leach, KY)
NGGCCOLG Columbia Gulf Onshore Louisiana Pool
NGGCT800 Tennessee 800 Leg (SE Louisiana)
NGGCTR30 Transco Station 30 (Wharton County, TX)
NGGCTRNZ Transco Station 65 (Beauregard Parish, LA)
NGGCTXEW TETCO West Louisiana
NGGCTXEZ TETCO East Louisiana
NGNECNGO Dominion South Point (Lebanon, OH)
NGNEIROQ Iroquois Gas Transmission System (Waddington, NY)
NGNEIZN2 Iroquois Zone 2 (Wright, NY)
NGNETNZ6 Tennessee Gas Pipeline Zone 6
NGNETRNZ Northeast Transco Zone 6 (Linden, NJ)
NGRMCHEY Cheyenne Hub (Cheyenne, WY)
NGRMDENV Colorado Interstate Gas Mainline
NGRMELPS Non-Bondad San Juan Basin (El Paso, TX / Blanco, NM)
NGRMEPBD Bondad San Juan Basin (El Paso, TX)
NGRMKERN Rocky Mountains / Kern River (Opal, WY)
NGRMNWES Northwest Pipeline (Stanfield, OR)
NGTXEPP2 Permian Basin (West Texas)
NGTXOASI Waha Hub (Waha, TX)
NGTXPERY Columbia Mainline (Perryville, LA)
NGUSHHUB Henry Hub (Erath, LA)
NGWCEPEB El Paso South Mainline (El Paso, TX)
NGWCPGNE Pacific Gas and Electric Citygate (N. California)
NGWCPGSP Northwest Transmission (Malin, OR)
NGWCPGTP Pacific Gas and Electric Topock (Topock, AZ)
NGWCSCAL Southern California Border
NTGSTXKA Katy Texas Area
SNNWPIPA Eastern Oklahoma Panhandle Field Zone (Haven, KS)
NG1 Generic 1st NG Future
NG2 Generic 2nd NG Future
NG3 Generic 3rd NG Future
NG6 Generic 6th NG Future
NG9 Generic 9th NG Future
NG12 Generic 12th NG Future
: Bloomberg identifiers of the price series used in our analysis. LNG Futures were used to compute realized volatility measures and served as the market proxy while LNG Spot prices were used for the sparse-data asset.[]{data-label="tab:series"}
Additional Results {#app:additional_results}
==================
In this section, we provide additional description of the in- and out-of-sample performance of our [`RBGARCH`]{}model, extending the results presented in Section \[sec:application\].
In-Sample Model Fit
-------------------
We begin by considering the in-sample fit, as measured by the Widely Applicable Information Criterion (WAIC) proposed by @Watanabe:2010. As discussed in more detail by @Gelman:2014, WAIC is a fully Bayesian analogue of AIC [@Akaike:1974], which attempts to estimate out-of-sample expected log-posterior density. Unlike the Deviance Information Criterion (DIC) of @Spiegelhalter:2002 [@Spiegelhalter:2014], WAIC is calculated using the entire posterior distribution, rather than just the posterior mean, taking fuller account of posterior uncertainty. Note that the Bayesian WAIC should not be confused with the so-called *Bayesian* Information Criterion (BIC) of @Schwarz:1978 or its actually Bayesian analogue WBIC [@Watanabe:2013], both of which attempt to estimate the marginal likelihood of a proposed model [@Neath:2012]. Because we do not assume that either the [`GARCH`]{}or [`RBGARCH`]{}models are true, we focus more on predictive accuracy than model recovery.
![Signed-rank-transformed differences of 40 spots over 73 periods. Positive values indicate that the [`RBGARCH`]{}model outperformed the [`GARCH`]{}model for that sample period. $p$-values from a paired one-sample Wilcoxon test applied to WAIC differences for each sample are displayed on the left axis. The [`RBGARCH`]{}model outperformed the [`GARCH`]{}model across all spots, but was somewhat more sensitive to outliers in the data.[]{data-label="fig:waic_app"}](waic_with_stability_app.pdf){width="\textwidth"}
Figure \[fig:waic\_app\] shows in-sample WAIC comparisons of the [`GARCH`]{}and [`RBGARCH`]{}models for each of our 40 spot prices and 73 sub-sample periods. As discussed above, for a small fraction of sub-samples 5-15%, depending on the spot, the WAIC estimates are corrupted by the presence of extreme outliers (see, *e.g.*, Figure \[app:fig:outlier\]). We identify these corrupted sub-periods using the 0.4 posterior variance heuristic of @Vehtari:2017 and highlight them in purple. This heuristic identifies several periods in which the WAIC estimates of the [`RBGARCH`]{}model are unstable, but those of the [`GARCH`]{}model are not. We conjecture that this is due to the additional structure of the [`RBGARCH`]{}model: because the [`RBGARCH`]{}model expects volatilities to move together, an outlier in the spot price which does not have a corresponding jump in the futures market is even less likely under the [`RBGARCH`]{}model than a simultaneous jump in both markets. Regardless, for our specific goal of VaR forecasting, the [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model with no outlier filtering necessary, as shown in the following two sections.
![Price history for Custer, OK LNG trading hub. The large price jump on February 5^th^, 2014 causes WAIC instabilities for several sample sub-periods, as highlighted in Figure \[fig:waic\_app\]. Because we use a 50-day rolling window with a 250-day history, a single outlier of this form can impact up to 5 sub-periods.[]{data-label="app:fig:outlier"}](outlier_spot.pdf){height="2in"}
In Figure \[app:fig:waic\], we show the fraction of subsample periods for which the [`RBGARCH`]{}model outperforms the [`GARCH`]{}model by $K$ standard errors, broken down by year. (Because WAIC calculations include a standard error, it is possible to obtain a standard error on the difference in WAIC values. For details, see Section 5 of @Vehtari:2017.) The [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model, typically by one or more standard errors, for all years in the sample period.
![In-Sample WAIC Comparison of [`GARCH`]{}and [`RBGARCH`]{}models, aggregated over all 40 spot prices in our sample. The reported probabilities are the fraction of 250 day sample periods in which the the WAIC of the [`RBGARCH`]{}exceeded that of the [`GARCH`]{}model by at least $K$ times the estimated standard error of the difference for $K = 0, 1, 2$. For all years in the sample period, the [`RBGARCH`]{}model has a consistently higher WAIC than the [`GARCH`]{}model, typically by several standard errors.[]{data-label="app:fig:waic"}](waic_fig.pdf){height="2.6in"}
In-Sample Risk Management
-------------------------
In addition to the unconditional test of @Kupiec:1995, we can also assess VaR accuracy using the conditional test of @Christoffersen:1998. While the Kupiec test is a marginal goodness of fit test, evaluating whether the samples appear to be draws from a Bernoulli distribution with nominal coverage, the Christoffersen test considers dependence between samples and tests whether the probability of an exceedance on day $T$ is independent of the probability of an exceedance on day $T + 1$. Figures \[app:fig:in\_500\] and \[app:fig:in\_200\] display $p$-values for each of the 40 spots, at 99.8% and 99.5% levels respectively. Figure \[app:fig:in\_var\_paired\] gives a simplified version of this information, from which we can clearly see the improved performance of the [`RBGARCH`]{}model, as indicated by the upward sloping lines comparing the $p$-values for each model.
From these, we see that the [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model, particularly for portfolios with approximately 50% allocation in the spot and futures, taking advantage of its multivariate structure. While it is clear from these plots that the [`RBGARCH`]{}model performs well, it is interesting to examine the limitations of the [`RBGARCH`]{}model. To this end, Figure \[app:fig:in\_sample\_var\] shows the expected and observed number of exceedances, corresponding to different confidence levels in the VaR calculation, for the [`GARCH`]{}and [`RBGARCH`]{}models. As can be seen from these figures, neither the [`GARCH`]{}nor the [`RBGARCH`]{}are perfectly calibrated (signified by the red line), though the two models fail differently. The [`GARCH`]{}model typically has more observed exceedances than expected, indicating a systematic underestimation of VaR, while the [`RBGARCH`]{}has generally fewer observed than expected, indicating overly conservative estimates. This is consistent with what we would expect for a well-fitting Bayesian forecast, as discussed in @Ardia:2017. (Note that the VaR estimates provided by both [`GARCH`]{}and [`RBGARCH`]{}approaches are more conservative than would be obtained from their non-Bayesian counterparts, but the [`RBGARCH`]{}model is especially conservative as it accounts for parameter uncertainty in many more parameters than the [`GARCH`]{}model.)
![In-sample VaR accuracy measures of the [`GARCH`]{}and [`RBGARCH`]{}models at the 99.8% (1 in 500 days) confidence level. The @Kupiec:1995 and @Christoffersen:1998 tests both indicate that the [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model for all 40 spots in our sample, with the advantage being particularly pronounced for approximately equally weighted portfolios. In order to accurately evaluate VaR predictions at this extreme quantile, the reported $p$-values were obtained by taking the concatenated in-sample predictions over the entire 10-year sample period.[]{data-label="app:fig:in_500"}](in_sample_combined_500){width="\textwidth"}
![In-sample VaR accuracy measures of the [`GARCH`]{}and [`RBGARCH`]{}models at the 99.5% (1 in 200 days) confidence level. Compared to Figure \[app:fig:in\_500\], the in-sample advantage of the [`RBGARCH`]{}model is still present, albeit less pronounced, as both models are able to capture this portion of the distribution well.[]{data-label="app:fig:in_200"}](in_sample_combined_200){width="\textwidth"}
![In-sample VaR accuracy measures of the [`GARCH`]{}and [`RBGARCH`]{}models at the 99.8% (1 in 500 days) confidence level. Consistent with Figure \[app:fig:in\_500\], the [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model, as can be seen from the upward sloping lines connecting corresponding $p$-values from the two models.[]{data-label="app:fig:in_var_paired"}](in_sample_var_500){width="\textwidth"}
![Assessment of the In-Sample VaR estimates of the [`GARCH`]{}and [`RBGARCH`]{}models for a range of sample portfolios. The [`GARCH`]{}model typically underestimates the true VaR (lies above the red line) while the [`RBGARCH`]{}model is more accurate, but with a conservative bias (lies below the red line).[]{data-label="app:fig:in_sample_var"}](in_sample_exceed.pdf){width="\textwidth"}
Out-of-Sample Risk Management
-----------------------------
Figures \[app:fig:out\_500\], \[app:fig:out\_200\], \[app:fig:out\_var\_paired\], \[app:fig:out\_sample\_var\] repeat the analysis of the previous subsection on the out-of-sample VaR forecasts. As described in Section \[sec:application\], these forecasts are obtained by filtering the volatility for each posterior sample and then calculating an overall VaR estimate by marginalizing over the posterior samples. Not surprisingly, our out-of-sample forecasts are generally less accurate than our in-sample estimates. Despite this, we see essentially the same results as before: the [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model, doing particularly well on portfolios with large amounts of both spot and futures. When the [`RBGARCH`]{}is mis-calibrated, it tends to be slightly conservative, but overall it is quite accurate.
The systematic underestimation of standard GARCH models is well-known to practitioners and model based estimates of VaR are typically multiplied by a “fudge factor” to account for this. In 1996, the Basel Committee on Banking Supervision recommended the use of a fudge factor of 3 [@BIS:1996 Paragraph B.4(j)], cementing the widespread use of this sort of *ad hoc* adjustment. While this style of adjustment is sufficient for conservative risk management, if applied uncritically it can greatly increase the financial burden imposed on market participants due to the increased capital requirements associated with higher VaR levels. The superior performance of the [`RBGARCH`]{}model suggests that a fudge factor may not be necessary if additional information is used to improve out-of-sample predictive performance.
![Out-of-sample VaR accuracy measures of the [`GARCH`]{}and [`RBGARCH`]{}models at the 99.8% (1 in 500 days) confidence level. The @Kupiec:1995 and @Christoffersen:1998 tests both indicate that the [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model for all 40 spots in our sample, with the advantage being particularly pronounced for approximately equally weighted portfolios. Somewhat surprisingly, the [`RBGARCH`]{}model appears to perform almost as well out-of-sample as it does in-sample, while the [`GARCH`]{}model is noticeably less accurate.[]{data-label="app:fig:out_500"}](out_sample_combined_500){width="\textwidth"}
![Out-of-sample VaR accuracy measures of the [`GARCH`]{}and [`RBGARCH`]{}models at the 99.5% (1 in 200 days) confidence level. The @Kupiec:1995 and @Christoffersen:1998 tests both indicate that the [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model for all 40 spots in our sample, with the advantage being particularly pronounced for approximately equally weighted portfolios. Unlike the in-sample case (Figure \[app:fig:in\_200\]), the [`RBGARCH`]{}model here clearly outperforms the [`GARCH`]{}model.[]{data-label="app:fig:out_200"}](out_sample_combined_200){width="\textwidth"}
![Out-of-sample VaR accuracy measures of the [`GARCH`]{}and [`RBGARCH`]{}models at the 99.8% (1 in 500 days) confidence level. Consistent with Figure \[app:fig:out\_500\], the [`RBGARCH`]{}model consistently outperforms the [`GARCH`]{}model, as can be seen from the upward sloping lines connecting corresponding $p$-values from the two models.[]{data-label="app:fig:out_var_paired"}](out_sample_var_500){width="\textwidth"}
![Assessment of the Out-of-Sample VaR estimates of the [`GARCH`]{}and [`RBGARCH`]{}models for a range of sample portfolios. The [`GARCH`]{}model typically underestimates the true VaR (lies above the red line) while the [`RBGARCH`]{}model is more accurate, but with a conservative bias (lies below the red line).[]{data-label="app:fig:out_sample_var"}](out_sample_exceed.pdf)
Computational Details {#app:computation}
=====================
In this section, we give additional details of the computational approach used to estimate Model . To perform estimation, we use the No-U-Turn Sampler variant of Hamiltonian Monte Carlo [@Hoffman:2014; @Neal:2011; @Betancourt:2017a] as implemented in `Stan` [@Carpenter:2017]. Hamiltonian Monte Carlo is particularly well-suited for models such as ours, as it can take advantage of gradient information to more efficiently explore high-dimensional and highly-correlated posterior distributions. The relevant `Stan` code is given in Section \[app:code\]. Section \[app:mcmc\] describes *ex post* diagnostics to assess the efficiency of Hamiltonian Monte Carlo for our analysis. We confirm that `Stan` can sample effectively from the posterior using the Simulation-Based Calibration (SBC) approach of @Talts:2018, as described in Section \[app:sbc\]. Finally, we examine the finite sample estimation performance of our model in Section \[app:power\].
`Stan` Code {#app:code}
-----------
The probabilistic programming language `Stan` [@Carpenter:2017] provides a modeling language for high-performance Markov Chain Monte Carlo (‘MCMC’), using the No-U-Turn Sampler variant of Hamiltonian Monte Carlo [@Hoffman:2014]. Unless otherwise stated, `Stan` was used for all posterior inference reported in this paper. The `Stan` manual [@Stan:2017] provides an exhaustive introduction to the use of `Stan`. The `Stan` code presented below was used to fit Model . The `Stan` code for the univariate model used for comparisons in Section \[sec:application\] can be obtained by removing certain sections of this code.
`Stan` does not provide a built-in *multivariate* skew-normal density, so we implement it as a user-defined function. For computational reasons, we model the marginal variances and the correlation matrix separately, so we use a four-parameter formulation of the multivariate skew-normal density. Our specification is essentially that of @Azzalini:1999, except we use the Cholesky factor of the correlation matrix, which we denote $\Omega$, instead of the full correlation matrix (which they denote $\Omega_z$) to more efficiently evaluate the multivariate normal density. Additionally, we denote the marginal variances by $\sigma$, as opposed to $\omega$.
``` {.stan mathescape="" fontsize="\footnotesize"}
functions{
real multi_skew_normal_lpdf(vector y, vector mu, vector sigma, vector alpha, matrix omega){
real retval = 0;
int K = rows(y);
retval += multi_normal_cholesky_lpdf(y | mu, diag_pre_multiply(sigma, omega));
for (i in 1:K){
retval += normal_lcdf(dot_product(alpha, (y - mu) ./ sigma) | 0, 1);
}
return retval;
}
}
```
We combine the market (Henry Hub futures) and asset (non-Henry spot) returns into a $T$-length array of 2-vectors.
``` {.stan mathescape="" fontsize="\footnotesize"}
data {
int<lower=1> T;
vector[T] return_market;
vector[T] return_asset;
vector<lower=0>[T] realized_vol_market;
}
transformed data{
vector[2] returns[T];
for(t in 1:T){
returns[t][1] = return_market[t];
returns[t][2] = return_asset[t];
}
}
```
As discussed above, we use the Cholesky factor of the correlation matrix for computational efficiency. This allows for more efficient evaluation of the multivariate normal density (by avoiding an expensive determinant calculation) and allows `Stan` to use a positive (semi)-definiteness-enforcing transformation automatically [@Pinheiro:1996].
``` {.stan mathescape="" fontsize="\footnotesize"}
parameters {
vector[2] mu;
vector[2] alpha;
cholesky_factor_corr[2] L;
```
We enforce stationarity by constraining the parameters to sections of the parameter space which yield a stationary process. Note that we do not use variance targeting.
``` {.stan mathescape="" fontsize="\footnotesize"}
real<lower=0> omega_market;
real<lower=0,upper=1> gamma_market;
real<lower=0,upper=(1-gamma_market)> tau_1_market;
real<lower=0,upper=(1-gamma_market-tau_1_market)> tau_2_market;
```
We treat the initial volatility as an unknown parameter to be inferred. In practice, this is not particularly important since we use a long (250 day) window to fit our model.
``` {.stan mathescape="" fontsize="\footnotesize"}
real<lower=0> sigma_market1;
real<lower=0> omega_asset;
real<lower=0,upper=1> beta_asset;
real<lower=0,upper=(1-beta_asset)> gamma_asset;
real<lower=0,upper=(1-beta_asset-gamma_asset)> tau_1_asset;
real<lower=0,upper=(1-beta_asset-gamma_asset-tau_1_asset)> tau_2_asset;
real<lower=0> sigma_asset1;
real xi;
real phi;
real<lower=0> delta_1_rv;
real<lower=0> delta_2_rv;
real<lower=0> rv_sd;
}
```
Because the instantaneous volatilities in a GARCH model are deterministic, conditional on the returns, model parameters, and initial condition, we calculate them in the `transformed parameters` block. Note the use of local variables `sigma_market` and `sigma_asset` to reduce memory pressure.
``` {.stan mathescape="" fontsize="\footnotesize"}
transformed parameters {
vector<lower=0>[2] sigma[T];
vector[T] rv_market_mean;
{
vector[T] sigma_market;
vector[T] sigma_asset;
sigma_market[1] = sigma_market1;
sigma_asset[1] = sigma_asset1;
for(t in 2:T){
sigma_market[t] = sqrt(omega_market +
gamma_market * square(sigma_market[t - 1]) +
tau_1_market * fabs(return_market[t-1]) +
tau_2_market * square(return_market[t-1]));
sigma_asset[t] = sqrt(omega_asset +
gamma_asset * square(sigma_asset[t - 1]) +
beta_asset * square(sigma_market[t - 1]) +
tau_1_asset * fabs(return_asset[t-1]) +
tau_2_asset * square(return_asset[t-1]));
}
for(t in 1:T){
sigma[t][1] = sigma_market[t];
sigma[t][2] = sigma_asset[t];
}
rv_market_mean = xi + phi * sigma_market +
delta_1_rv * fabs(return_market) +
delta_2_rv * square(return_market);
}
}
```
The priors from Table \[tab:priors\] are used. Note that `Stan` uses a mean-standard deviation parameterization of the normal distribution as opposed to the more common mean-variance or mean-precision parameterizations.
``` {.stan mathescape="" fontsize="\footnotesize"}
model {
// Priors
mu ~ normal(0, 1);
alpha ~ normal(0, 1);
L ~ lkj_corr_cholesky(1);
beta_asset ~ normal(0, 1);
// Market (Henry Hub) GARCH Dynamics
// $\sigma_{t}^2 = \omega + \gamma \sigma_{t-1}^2 + \tau_1 |r_{t-1}| + \tau_2 r_{t-1}^2$
omega_market ~ normal(0.002, 0.025);
gamma_market ~ normal(0.8, 0.6);
tau_1_market ~ normal(0, 0.1);
tau_2_market ~ normal(0.1, 0.7);
// Asset (non-Henry Hub Spot) GARCH Dynamics
// $\sigma_{t}^2 = \omega + \gamma \sigma_{t-1}^2 + \tau_1 |r_{t-1}| + \tau_2 r_{t-1}^2$
omega_asset ~ normal(0.002, 0.025);
gamma_asset ~ normal(0.8, 0.6);
tau_1_asset ~ normal(0, 0.1);
tau_2_asset ~ normal(0.1, 0.7);
// Realized volatility dynamics
// $\varsigma_t = \xi + \phi \sigma_t + \delta_1 |r_t| + \delta_2 r_t^2 + \mathcal{N}(0, \nu^2)$
xi ~ normal(0.02, 0.6);
phi ~ normal(15, 60);
delta_1_rv ~ normal(1.15, 8);
delta_2_rv ~ normal(1.15, 14);
rv_sd ~ normal(0.05, 0.25);
```
A weak data-dependent prior is on the initial conditions of the GARCH volatility.
``` {.stan mathescape="" fontsize="\footnotesize"}
// Initialize initial vol with weak prior
sigma_market1 ~ normal(sd(return_asset), 5 * sd(return_asset));
sigma_asset1 ~ normal(sd(return_asset), 5 * sd(return_asset));
// Likelihood
realized_vol_market ~ normal(rv_market_mean, rv_sd);
for(t in 1:T){
returns[t] ~ multi_skew_normal(mu, sigma[t], alpha, L);
}
}
```
$\Sigma_t = \text{diag}(\sigma_t) L^TL \text{diag}(\sigma_t)$ is the covariance matrix of the returns model.
``` {.stan mathescape="" fontsize="\footnotesize"}
generated quantities{
cov_matrix[2] Sigma[T];
for(t in 1:T){
Sigma[t] = tcrossprod(diag_pre_multiply(sigma[t], L));
}
}
```
MCMC Diagnostics {#app:mcmc}
----------------
We propose the use of the Hamiltonian variant of Markov Chain Monte Carlo to estimate Model . While Hamiltonian Monte Caro (HMC), like all commonly-used MCMC methods, is guaranteed to *asymptotically* recover the true posterior, one should always carefully assess its performance in a given simulation before performing inference based on its output. @Gelman:2011 give a succinct review of standard MCMC diagnostics. We supplement these with various recently-developed HMC-specific diagnostics, such as the Expected Bayesian Fraction of Missing Information (E-BFMI) [@Betancourt:2016] and divergent transitions [@Betancourt:2017a Section 6.2]. While the use of Hamiltonian Monte Carlo for Realized GARCH models has not been previously examined, our results suggest that it is an efficient and robust sampling scheme for this type of model.
For each run of HMC, we run four separate chains of 2000 iterations.[^2] While 2000 iterations may seem insufficient, particularly as compared to the tens of thousands typically used with other samplers, HMC typically mixes far more efficiently than other methods, reducing the total number of iterations required. As we will see below, this is sufficient for our problem. To increase the robustness of our results, we used a slightly higher target adaptation rate ($\texttt{adapt\_delta} = 0.99$) and maximum tree-depth ($\texttt{max\_treedepth} = 12$) than `Stan`’s default settings ($0.8$ and $10$ respectively). On the vast majority of our fits, no divergent transitions were encountered, the maximum treedepth was never hit, and the average E-BMFI across chains was above 0.9 (values below 0.2 are typically considered indicative of a sampling pathology). Taken together, these results suggest that the sampler was able to efficiently explore the posterior distribution.
Having confirmed that the sampler was able to efficiently explore the posterior distribution, we are now in a position to assess whether the sampler output provides enough precision for subsequent analyses. We compute both the effective sample size and the split-$\hat{R}$ diagnostic [@Gelman:2013 Section 11.4-11.5] for each of the parameters of our model as well as the estimated volatilities $\sigma_t$. Our results are presented in Table \[tab:mcmc\]. The diagnostics suggest that we are able to explore the posterior efficiently and that our results are sufficiently precise for downstream analyses.
Simulation-Based Calibration {#app:sbc}
----------------------------
@Talts:2018 propose a general scheme for validating Bayesian inference, based on the idea of the *data-averaged posterior*. They note that, if the parameters are indeed drawn from the prior, the average of the *exact* posterior distributions is exactly the prior. As such, any deviation between the data-averaged posterior produced by a sampling scheme and the prior indicates biases in the sampling scheme. Applying this technique to Model yields the results in Figure \[fig:sims\_sbc\], which show no systematic deviations from uniformity, suggesting that our inference is unbiased and that our sampling scheme is well-suited for the model.
![Simulation-Based Calibration Analysis of Hamiltonian Monte Carlo applied to Model with $T = 200$. The blue bars indicate the number of SBC samples in each rank bin while the green bars give a 99% marginal confidence band for the number of samples in each bin. While there is significant variability in the histograms, commensurate with a complex model, we do not observe any *systematic* deviations.[]{data-label="fig:sims_sbc"}](sbc_t_200.pdf){width="\textwidth"}
Additional Simulation Studies {#app:power}
-----------------------------
In this section, we provide additional evidence that Model is able to effectively capture the underlying GARCH dynamics, even with relatively small data sizes. Since we are concerned with performance across a wide range of possible parameter values, we use wider priors for these simulation studies than we do in the main body of this paper. In particular, we use a weakly informative $\mathcal{N}(0, 1)$-prior on all GARCH parameters for each study in this section.
The natural first question to ask of any Bayesian model is whether the posterior credible intervals are indeed accurately computed. If the parameters are drawn from the prior distribution, the posterior credible intervals should be perfectly calibrated and, *ceteris paribus*, should decrease in length as the sample size increases. Figures \[fig:sims\_coverage\] and \[fig:sims\_width\] confirm that our posterior credible intervals are correctly calculated and well calibrated. @Hansen:2012 demonstrated that the univariate Realized GARCH framework satisfies the conditions for standard ($1/\sqrt{T}$) asymptotic convergence of the Quasi-Maximum Likelihood Estimate. Figure \[fig:sims\_width\] suggests that Model is similarly well-behaved and that Bernstein-von Mises-type convergence rates can be obtained.
A close examination of Figure \[fig:sims\_width\] will reveal that parameters associated with the mean model – namely the $\mu$ and $\alpha$ parameters – have wider posterior credible intervals than the parameters of the GARCH and realized volatility processes ($\omega$, $\gamma$, $\tau$, $\beta$, *etc.*) While this may seem counter-intuitive at first, it appears to be a consequence of the relatively high-levels of volatility considered, which make precise inference on the mean difficult. Despite this difficulty, our simulations suggest that Model \[eqn:model\] is able to recover the underlying dynamics effectively even with relatively small samples.
[**Parameter**]{} [**Effective Sample Size ($\widehat{n_{\text{eff}}}$)**]{} [**Potential Scale Reduction Factor ($\widehat{R}$)**]{}
--------------------------------------- ------------------------------------------------------------ ----------------------------------------------------------
$L_{21} = \textsf{chol}(\Omega)_{21}$ 2704.386 0.999
$L_{22} = \textsf{chol}(\Omega)_{22}$ 2705.967 0.999
$\omega_M$ 1491.062 1.001
$\gamma_M$ 2345.653 1.000
$\tau_{1, M}$ 2223.725 1.000
$\tau_{2, M}$ 2705.410 1.001
$\sigma_M$ 2051.442 1.000
$\omega_i$ 1994.433 1.000
$\beta_i$ 2604.040 1.000
$\gamma_i$ 2625.324 1.000
$\tau_{1, i}$ 2331.136 1.000
$\tau_{2, i}$ 2372.701 1.000
$\sigma_i$ 2084.207 1.000
$\xi$ 1715.700 1.001
$\phi$ 1662.953 1.001
$\delta_1$ 2395.295 1.000
$\delta_2$ 2297.632 1.000
$\nu$ 4000.000 1.000
: Markov Chain Monte Carlo Diagnostics for Hamiltonian Monte Carlo estimation of Model \[eqn:model\]. These diagnostics are taken from a representative run of four chains of 2000 iterations each. Reported values for $\sigma_M$ and $\sigma_i$ are averages over the entire fitting period.[]{data-label="tab:mcmc"}
![Empirical coverage probabilities of posterior credible intervals associated with Model under a weakly informative prior distribution. Because we drew the parameters from the prior, these intervals are correctly calibrated.[]{data-label="fig:sims_coverage"}](posterior_interval_coverage.pdf){width="\textwidth"}
![Mean width of symmetric posterior intervals associated with Model under a weakly informative prior distribution. We observe standard $n^{-1/2}$-type convergence for all parameters as the length of the sample period ($T$) increases. Somewhat surprisingly, we note that parameters associated with the mean model ($\mu, \alpha$) are typically more uncertain than the parameters of the GARCH dynamics.[]{data-label="fig:sims_width"}](posterior_interval_width.pdf){width="\textwidth"}
Additional Background {#app:additional_background}
=====================
In this section, we give some additional background for the reader who may be interested in learning more about volatility models or the structure of LNG markets.
Volatility Models
-----------------
The first latent volatility process model was the Autoregressive Conditional Heteroscdasticity (ARCH) model of @Engle:1982, which modeled the instantaneous volatility $\sigma_t^2$ as a weighted sum of the $q$-previous (standardized) returns. @Bollerslev:1986 proposed the *Generalized* ARCH (GARCH) model which instead uses an autoregressive moving average (ARMA) model for the instantaneous volatility, combining previous volatility levels and (standarized) returns. As demonstrated by @Bollerslev:1986, the use of previous volatility levels significantly reduces the amount of history required, giving a more accurate and more parsimonious model. For this reason, GARCH-type models have generally supplanted ARCH models in applied work.
GARCH-type models have had an enormous impact on financial econometrics and many variants have been proposed, including the integrated GARCH (I-GARCH) of @Engle:1986; the exponential GARCH (E-GARCH) of @Nelson:1991; the GJR-GARCH of @Glosten:1993, which allows for asymmetric positive and negative effects; and the asymmetric power GARCH of @Ding:1993, which introduces a Box-Cox transform [@Box:1964] into the model specification and unifies several previous proposals. The family GARCH model of @Hentschel:1995 gives a very general specification comprising a wide range of GARCH variants. More recent variants attempt to decompose latent volatilities into long- and short-term components, yielding the additive and multiplicative component GARCH models of @Engle:1999 and @Engle:2012, respectively. The glossary of @Bollerslev:2010 provides a useful and comprehensive list of variants.
Multivariate extensions of GARCH models are equally numerous and we refer the reader to the survey of @Bauwens:2006 for a more detailed review. The simplest multivariate GARCH model is perhaps that of @Bollerslev:1990, who assumes a constant conditional correlation (CCC) matrix among the various assets. @Engle:2002 extended the CCC model was extended to allow for slowly-varying conditional correlations, yielding the popular dynamic conditional correlation (DCC) specification. Our proposed model uses the CCC approach, though the possibility of misspecification is mitigated by our rolling refitting strategy which yields an approximate DCC specification.
A popular alternative to GARCH models is the class of “stochastic volatility” models first introduced by @Kim:1998. These models allow the volatility to evolve according to a (typically independent) stochastic process, giving their name. This is in contrast to GARCH-type models where the next day’s volatility is deterministic, conditional on the current return and volatility history. We do not review these models in detail here, instead referring the reader to the paper of @Asai:2006 and the discussion in @Vankov:2019
A commonly noted shortcoming of both SV- and GARCH-type models is their slow responsiveness to rapid changes in the volatility level [@Engle:2002b; @Anderson:2003; @Anderson:2005]. This slowness is a consequence of the “smoothing” nature of both SV and GARCH models, which balance the information provided by a single day’s return against a long history. By using several volatility measures, it becomes possible to weight the current time period more heavily and to develop more responsive models, as shown by, *e.g.*, @Visser:2011 and @Anderson:2011 for GARCH models and @Takahashi:2009 for SV models.
Driven by market conventions, the most common additional volatility measures are those based on OHLC (Open, High, Low, Close) data, such as the high-low range proposed by @Parkinson:1980, the open-close difference proposed by @Garman:1980, or high, low, and close data as suggested by @Rogers:1991. The estimator of @Yang:2000 is optimal (minimum variance unbiased) among estimators based solely on OHLC data. If higher frequency data is available, even more accurate estimators have been proposed, though their efficacy is very sensitive to the structure of the market being considered [@Barndorff-Neilsen:2002; @Barndorff-Neilsen:2003; @Zhang:2005; @Barndorff-Nielsen:2011].
While realized volatility measures can be used to improve GARCH estimation by replacing the standardized squared return with an improved estimate, this naive approach does not capture the structural relationships among different volatility measures. Several classes of “complete” models which jointly model prices and realized volatilty measures have been proposed, including the Multiplicative Error Model (MEM) framework of @Engle:2006 and the High-Frequency Based Volatility (HEAVY) framework of @Shephard:2010. In this paper, we consider the Realized GARCH framework of @Hansen:2012 [@Hansen:2014; @Hansen:2016], which is among the most flexible specification proposed to date. In addition to standard volatility estimation, the realized GARCH framework has proven useful for risk management, with @Watanabe:2012 demonstrating its usefulness in conditional quantile estimation and @Contino:2017 [@Wang:2018] showing its usefulness for estimating VaR and Expected Shortfall [@Artzner:1999; @Acerbi:2002; @Jorion:2006].
Natural Gas Markets
-------------------
The past 20 years have seen significant developments in the structure and size of US LNG markets. Much of this development has been driven by the development of shale gas extraction technologies, in particular hydraulic fracturing (“fracking”), which have lowered production costs and increased demand for LNG and LNG derivatives [@Caporin:2017]. At the same time, increasing awareness of the environmental impacts of various energy sources has prompted additional investment into LNG: while a fossil fuel and not renewable, LNG is widely considered to be a cleaner energy source than coal and a possible bridge to a fully renewable energy sector. Over the same period, investor interest in LNG and other commodities has increased, spurring the growth of liquid futures and derivatives markets [@Tang:2012].
@Li:2019, following @Narayan:2015, identifies September 2008 as the “coming of age” of modern LNG markets. Modern LNG markets are characterized by active, but relatively stable, spot trading at a large number of hubs and by informative and liquid futures markets. For historical reasons, Henry Hub plays a particularly prominent role in LNG markets, *inter alia* serving as the reference price for much of the LNG derivatives market, it is far from the only hub of economic interest. Practitioners recognize over one hundred trading hubs in the continental United States and Canada, with the Chicago Citygate, Algonquin Citygate (serving the Boston area), Opal (Lincoln County, Wyoming), Southern California, and NOVA (Alberta, Canada) hubs being particularly closely monitored by market participants. (Strictly speaking, some of these are commonly reported regional average indices, not physical hubs, but we will continue to refer to them as hubs.)
@Mohammadi:2011 gives a detailed survey of the pricing structure from well-heads (extraction) to end-consumers, while @Hou:2018 and @Hailemariam:2019 attempt to quantify the effect of supply and demand shocks on observed prices. Unlike equities, LNG spot and futures prices do not move in perfect synchrony due to storage and transpot costs. The relationship between LNG spot and futures prices was investigated by @Ghoddusi:2016 who found, *inter alia* that short-maturity futures are Granger-causal for physical prices, a fact which is consistent with our findings on the usefulness of futures realized volatility in spot price volatility modeling. measurable impact on observed price dynamics.
The relationship between the prices of LNG and other commodities has been particularly well-studied, and has been found to be surprisingly complex: @Hartley:2008 found important dependencies on foreign exchange rates, spot inventories, and unpredictable (weather) shocks; @Brigida:2014 and @Batten:2017 found that technological advancements weakened, but did not eliminate the underlying relationship; and @Caporin:2017 investigated the effect of the shale revolution on the relationship.
[^1]: To whom correspondence should be addressed: <[email protected]>.
[^2]: Following `Stan`’s default settings, the first 1000 iterations were used as an adaptation period and discarded, while the second 1000 iterations were stored. No thinning was performed.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The result of Boyce and Huneke gives rise to a 1-dimensional continuum, which is the intersection of a descending family of disks, that admits two commuting homeomorphisms without a common fixed point.'
address: 'National Supercomputing Centre IT4Innovations, Division of the University of Ostrava, Institute for Research and Applications of Fuzzy Modeling, 30. dubna 22, 701 03 Ostrava, Czech Republic – and – Faculty of Applied Mathematics, AGH University of Science and Technology, al. Mickiewicza 30, 30-059 Kraków, Poland'
author:
- 'J. P. Boroński'
title: A note on fixed points of abelian actions in dimension one
---
Two commuting $C^1$-diffeomorphisms of a disk must have a common fixed point, and if they preserve a nonseparating plane continuum $Z$ (i.e. a compact and connected set such that $\mathbb{R}^2\setminus Z$ is connected), then there is a common fixed point in $Z$ [@Ribon]. An arc-like continuum is the one that is the inverse limit of arcs. Bing showed that every arc-like continuum embeds in the plane as the intersection of a descending family of disks [@Bing]. Hamilton showed that every arc-like continuum has the fixed point property [@Ha]. In particular, any $\mathbb{Z}$-action on such a continuum fixes a point. For more general abelian actions on arc-like continua no such result has been known, although it holds for example for dendrites and uniquely arcwise connected continua [@Shi]. For self-maps of the arc the examples of Boyce [@Boyce] and Huneke [@Huneke] provide two commuting surjective maps without a common fixed point[^1]. Noteworthy their result gives the following.
**Example.** There exists an arc-like continuum $X$ and a pair of homeomorphisms $F,G:X\to X$ such that $F\circ G=G\circ F$ and $\operatorname{Fix}(F)\cap \operatorname{Fix}(G)=\emptyset$.
The argument for the existence of the above example is quite short. However, we did not find such a result in the literature, and none of the experts we consulted had been aware of such a result before. After personal discussions during the 52nd Spring Topology and Dynamics Conference, held at Auburn University in March of 2018, we were encouraged to publish this note.
Let $f,g:[0,1]\to[0,1]$ be two commuting surjections without a common fixed point. Let $h=f\circ g$ and consider the inverse limit space
$$X=\lim_{\leftarrow}\{[0,1],h\}=\{(x_1,x_2,...)\in[0,1]^{\mathbb{N}}:h(x_{i+1})=x_i,\textrm{ for all }i\in\mathbb{N}\}.$$ Since $f$ and $g$ commute, all three maps induce maps on $X$, given by
$$H(x_1,x_2,...)=(h(x_1),h(x_2),....),$$
$$F(x_1,x_2,...)=(f(x_1),f(x_2),....),$$
and
$$G(x_1,x_2,...)=(g(x_1),g(x_2),....).$$
To see that $F$ and $G$ are homeomorphisms one can use the following, particularly elegant, argument from [@Mouron]. Since $H=F\circ G=G\circ F$, and $H$ is just the shift homeomorphism, the maps $F$ and $G$ are also homeomorphisms. Since $f$ and $g$ have no common fixed point, the same is true for $F$ and $G$.
We note that by Barge-Martin embedding theorem [@BM] both $F$ and $G$ extend (up to conjugacy) to homeomorphisms of the plane. Therefore the following question is of interest.
Can $F$ and $G$ be extended to two commuting planar homeomorphisms?
An affirmative answer to the above problem would prove the necessity of the $C^1$ setting in the generalization of the Cartwright-Littlewood fixed point theorem in [@Ribon]. We also note, that the example in [@Huneke] is constructed as a limit of piecewise linear functions, each of which has a well-defined derivative of constant absolute value greater than $3$ everywhere, outside of a finite set of critical points. This gives a good starting point to a potential modification, that would result in the continuum $X$ presented here, being homeomorphic to the pseudo-arc; see [@MT].
Can $f$ and $g$ be modified so that the example holds on the pseudo-arc?
The interested reader might want to consult [@Mouron] for related results.
**Acknowledgments.** The author is supported by University of Ostrava grant lRP201824 “Complex topological structures” and the NPU II project LQ1602 IT4Innovations excellence in science. \[lastpage\]
[12]{} , **Proc. Amer. Math. Soc.**, 110 (1990), 523–525. **Duke Math. J.** 18, (1951), 653–663. , **Trans. Amer. Math. Soc.** 137 (1969), 77–92. **Proc. Amer. Math. Soc.** 2, (1951), 173–174. **Trans. Amer. Math. Soc.** 139 (1969) 371–381. P. Minc and W. R. R. Transue, *A transitive map on $[0,1]$ whose inverse limit is the pseudoarc*, **Proc. Amer. Math. Soc.**, 111 (1991), 1165–1170. **Colloq. Math.** 121 (2010), 63–77. **Ergodic Theory Dynam. Systems** 36 (2016), 173–197. **Proc. Amer. Math. Soc.** 137 (2009), 771–775. arXiv:1709.00852.
[^1]: I am grateful to Benjamin Vejnar for bringing this result to my attention. An interested reader might want to consult [@Vejnar] for some related recent results.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In what follows we improve an inequality related to matrix theory.\
T. Laffey proved (2013) a weaker form of this inequality [@2].
---
[On a Certain Function with Negative Coefficients ]{}
Dov Aharonov
Given $\rho = \sum\limits_{k=1}^{n}\mu_k \leq 1$ , $\mu_k > 0$ , $k=1,2,\ldots,n$,
$c_1 >c_2>\cdots >c_n>0$
Then for $n \geq 2$ , or $n=1$ $\mu_1<1$:
$$\sum\limits_{j=1}^{\infty} \pi (1-c_jt)^{\mu_j} = 1-\sum\limits_{j=1}^{\infty}D_j t^j
\ \mbox{we have} \
D_j >0 \ , \ j=1,2,\ldots$$
In order to prove the theorem it is enough to prove it for $\rho =1$.
[**Proof of Lemma 1**]{}
Assume that $\sum\limits_{j=1}^{n}\mu_j=1$ and that for this case the theorem is proved. Then for $0 < \rho <1$ we have
$$(1-\sum D_jt^j)^{\rho}=
1-\rho\sum D_jt^j + \rho(\rho-1)\left(\sum D_jt^j\right)^{2}-
\frac{1}{6}\rho(\rho-1)(\rho-2)\left(\sum D_jt^j\right)^{3}+ \cdots$$ $$=1-\sum\limits_{j=1}^{\infty}\tilde{D}_j t^j$$ Hence $\tilde{D}_j > 0$
$\blacksquare$
From now on we may assume, without loss of generality, $$\sum\limits_{j=1}^{n}\mu_j =1$$
Assuming $(1)$ and $(2)$ we have $$\sum\limits_{j=1}^{n}c_j \mu_j= D_1 \ , \ D_1 > c_n$$
[**Proof of Lemma 2**]{}
$$\prod\limits_{j=1}^{n}(1-c_jt)^{\mu_j} =
\prod\limits_{j=1}^{n}(1-\mu_jc_jt+O(t^2)) =
1-\left(\sum\limits_{j=1}^{n}\mu_jc_j\right)t+ O(t^2)=
1-D_1t+O(t^2)$$ Hence, the first part is confirmed.
For the second part recall the monotonicity of $c_j$.
We have
$$D_1 = \sum\limits_{j=1}^{n}c_j\mu_j > c_n
\sum\limits_{j=1}^{n} \mu_j = c_n$$
$\blacksquare$
The proof of our theorem will be by induction. For $n=1,2$ the proof follows very simply. Details are omitted.
Now assume that the theorem is proved for $n-1$ ($n \geq 3$). We want to prove it for $n$. Suppose that this is not the case.
It will be convenient to use the notation $$L_k = \prod\limits_{j=1}^{k}(1-c_jt)^{\mu_j}$$
Obviously, $L_k = L_{k-1}(1-c_kt)^{\mu_k}$. In particular, $$L_n = L_{n-1}(1-c_nt)^{\mu_n}=1-\sum\limits_{j=1}^{\infty}D_jt^j$$ By our assumption, there exists among the coefficients at least one which is not negative. Take the smallest index, say $m$, for which this is true. Hence $$-D_j < 0 \ , \ 1 \leq j \leq m-1 \ , \ -D_m \geq 0$$ Time has come to use a simple idea, but useful.
From $(4)$ $$L_n = L_{n-2}(1-c_{n-1}t)^{\mu_{n-1}}(1-c_nt)^{\mu_n}$$ Then for $c_n=0$ $L_n=L_{n-1}$ and we are back to the case of $n-1$ factors, i.e. the theorem is assumed to be correct by the induction assumption. The same follows if we assume $c_{n-1}=c_n$.
Indeed $(1-c_{n-1}t)^{\mu_{n-1}}(1-c_nt)^{\mu_n}=
(1-c_{n-1}t)^{\mu_{n-1}+\mu_n}$, and again we have $n-1$ factors. (Note that if we want to prove a weaker result for $\mu_1=\mu_2=\cdots=\mu_n = \frac{1}{n}$ the proof does not work!!)
[**Figure 1**]{}
(0,0)–(0,8); (0,4)–(10,4); (8,0)–(8,4.25); (4,3.75)–(4,4.25); (0,2)–(8,2); at (4,4.25) [$c_n$]{}; at (8,4.25) [$c_{n-1}$]{}; at (0,2) [$-D_j$]{}; at (0,-1) [$j=1,2,\ldots, n-1$]{};
By the induction assumption.
[**Figure 2**]{}
(0,0)–(0,8); (0,4)–(10,4); at (4,3.8) [$\bar{c}_n$]{}; (4,3.75)–(4,4.25); at (6,4.25) [$c_{n-1}$]{}; (6,3.75)–(6,4.25); (3,4) arc (180:0:.5); (2,4) arc (-180:0:.5); (1,4) arc (180:0:.5); (4,4) arc (-180:-90:.5); (4.5,3.5)–(6,3); (0,3)–(.5,3.5); (.5,3.5)–(1,4);
$$\begin{array}{l}
-D_m(0) <0\\
-D_m(c_{n-1}) <0
\end{array}$$
Consider now $c_n$ as a variable and $c_1, c_2, \ldots, c_{n-1}$ as fixed. $$0 \leq c_n \leq c_{n-1}$$ Since $-D_m \geq 0$ at some point, say $\bar{c}_n$, for this interval and also $-D_m(0) < 0$, $-D_m(c_{n-1}) < 0$ it follows by the mean value theorem that for some point in this interval, say $c^*_n$: $-D_m(c^*_n)=0$.
But there are only finite number of zeros of $-D_m$. This is true due to the fact that it is a polynomial function of the variable $c_n$. Thus, without loss of generality, $c^*_n$ may be taken as the largest zero in the interval $(0,c_{n-1})$.
[**Figure 3**]{}
(0,0)–(0,8); (0,4)–(10,4); (7,6)–(7,2); at (0,7) [$-D_m$]{}; (0,2)– (1,4);
(7,3) arc (270:180:1); (6,4) circle (0.07); (1,4)–(6,4); at (6,5) [$c^*_n$]{}; at (7,2) [$c_{n-1}$]{};
[**Figure 4**]{}
(0,0)–(0,8); (0,0)–(10,0); (8,-4)–(8,4); (8,-3) arc (270:180:3); at (0,7) [$-D_m$]{}; at (8,-4) [$c_{n-1}$]{}; at (4,-4) [$c_n^*$]{};
Thus (due to the fact that there are finite number of zeros as explained above) we have that the interval $(c^*_n,c_{n-1})$ is free of zeros.
In what follows we show a contradiction. this will be done by considering a sufficiently small interval $(c^*_n,c_{n}^*+ \varepsilon)$ for which $-D_m$ is changed from zero at $c_n^*$ to a positive value.
We now present the calculations leading to this assertion.
Indeed $$L_n(c_1,c_2,\ldots,c_{n-1},c^*_{n}+\varepsilon)=
L_{n-1}(c_1,c_2,\ldots,c_{n-1})(1-(c^*_{n}+\varepsilon)t)^{\mu_n}=$$ $$= L_{n-1}(1-c^*_nt-\varepsilon t)^{\mu_n}$$ Also:
$$(1-c_n^*t-\varepsilon t)^{\mu_n}=
(1-c_n^*t)^{\mu_n}\left(1-\frac{\varepsilon t}{1-c^*_nt}\right)^{\mu_n}=
(1-c_n^*t)^{\mu_n}\left(1-\frac{\varepsilon t}{1-c^*_nt} \mu_n +O(\varepsilon^2)\right)$$ Thus $$L_n(c_1,c_2, \ldots ,c_{n-1},c^*_{n}+\varepsilon)=
[L_{n-1}(c_1,c_2, \ldots ,c_{n-1})(1-c_n^*t)^{\mu_n}]
\left(1-\frac{\varepsilon t}{1-c^*_nt} \mu_n +O(\varepsilon^2)\right)$$ $$=(1-D_1 t - D_2t^2-\cdots -D_{m-1}t^{m-1}+0-D_{m+1}t^{m+1}+
\cdots)
\left(1-\frac{\varepsilon t\mu_n}{1-c^*_nt}+O(\varepsilon^2)\right)
=R \ \mbox{(notation)}$$ To continue it will be convenient to use the notation $\{F\}_K$ for the $k$-th coefficient of $F$.
Also denote $\mu_n\varepsilon$ by $\varepsilon^*$. $$\{R\}_m = \left\{\left(1-\sum\limits_{j=1}^{m-1}D_j t^j\right)\left(1-\frac{\varepsilon^* t}{1-c^*_nt}\right)\right\}_m +O(\varepsilon^2)$$ Positive terms do not destroy positivity if ignored, hence $$\{R\}_m = \left\{-\frac{\varepsilon^*t}{1-c^*_nt}\right\}_m+
\left\{(D_1t)\left(\frac{\varepsilon^*t}{1-c^*_nt}\right)\right\}_m+ \ \mbox{positive terms} \ +O(\varepsilon^2)$$ $$\begin{array}{rl}
\{R\}_m \geq& -\varepsilon(c_n^*)^{m-1}+
(\varepsilon^*)\left\{D_1t^2\left(1+\sum\limits_{k=1}^{\infty}
(c_n^*t)^k\right)\right\}_m+O(\varepsilon^2)
\\
\{R\}_m \geq& -\varepsilon(c_n^*)^{m-1}+
\varepsilon^*D_1(c_n^*)^{m-2}+O(\varepsilon^*)
\\
\vspace{-10pt}\\
\{R\}_m \geq& \varepsilon^*(c_n^*)^{m-2}[-c_n^*+D_1]+O(\varepsilon^*)
\end{array}$$ For $\varepsilon$ small enough this is positive by $(3)$ applied for the special value $c_n^*$ of $c_n$.
Hence, positivity of the $n$-th coefficient is established provided $\varepsilon$ is small enough. Thus we arrived at a contradiction, which ends the proof. $\blacksquare$
[**Figure 5**]{}
(0,0)–(0,8); (0,4)–(10,4); (4,5)–(4,3); at (4,3) [$c^*_n$]{}; (4,4)–(8,3); (8,5)–(8,3); at (8,3) [$c_{n-1}$]{}; (5.76,4.3) circle (.07); at (7,5) [$c^*_n+\varepsilon$]{};
Contradiction
Compare with Figure 4
\
I want to thank Prof Raphi Loewy for bringing to my attention the papers [@1] and [@2].
[99999]{}
F. Holland: [*A contribution to the inverse eigenvalue problem for non- negative matrices*]{}, Mathematical Proceedings of the Royal Irish Academy 110A (2013) 13-19;\
doi:10.3318/PRIA.2010.110.1.13.
T. J. Laffey: [*F*]{}, Mathematical Proceedings of the Royal Irish Academy 113A (2013) (2), 97-106;\
doi:10.3318/PRIA.2013.113.10.
Department of mathematics-I.I.T,Haifa 32000.Israel.\
E-mail address: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Alexander Adam, Sam Kitchen and Toby Wiseman\
Blackett Laboratory, Imperial College London, London SW7 2AZ, UK\
bibliography:
- 'stationary.bib'
date: June 2011
title: A numerical approach to finding general stationary vacuum black holes
---
Introduction
============
Stationary vacuum black holes in spacetime dimension $D > 4$ have received much attention over the past decade, and many remarkable results have emerged [@Kol:2004ww; @Harmark:2007md; @Emparan:2008eg]. In asymptotically flat $D=4$ spacetime Kerr is unique. In $D>4$ beyond the Myers-Perry solution [@MyersPerry] which generalizes Kerr, the space of solutions thought to exist is very exotic [@Emparan:2007wm] and there may yet be further solutions discovered. String theory famously requires more than 4 dimensions for consistency, and hence provides a powerful physical motivation to think about these exotic solutions. Perhaps the strongest motivation to consider higher dimensions and also exotic boundary conditions is the $AdS$-$CFT$ correspondence, which relates many of these interesting solutions and other gravitational phenomena to physical phenomena found in certain strongly coupled conformal field theories and gauge theories where direct field theory calculations are extremely hard. The majority of the solutions conjectured to exist are unlikely to be found analytically using current techniques. The problem of solving the Einstein equations to find them reduces to solving partial differential equations for the components of the metric. Since these exotic solutions are of cohomogeneity greater than one it is not clear this can be done analytically in general. Thus we are forced to consider numerical methods to construct such solutions. The aim of this paper is to find numerical methods that are geometrically elegant, apply for arbitrary cohomogeneity and topology, and that function well in practice.
From the classic 4$D$ stationary uniqueness theorems (see for example [@Carter]) we understand that the stationary problem of finding vacuum black holes should be thought of as an elliptic PDE system which is treated as a boundary value problem, where asymptotic conditions are imposed together with regularity at the horizon. Rigidity implies the existence of axisymmetry, and then the problem reduces to one depending non-trivially on two coordinates (‘cohomogeneity two’). Taking the Weyl-Papapetrou form for the metric then manifests the PDE problem as an elliptic one. Black hole uniqueness theorems have been extended to $D>4$ in the stationary case of $D-2$ commuting Killing directions [@Morisawa:2004tc; @Hollands:2007aj] where by reducing on these Killing directions the problem can be viewed as a 2$D$ elliptic one on the orbit space of these isometries. For non-trivial dependence on only two dimensions the stationary numerical problem has been extensively studied in the context of relativistic stars for decades [@W72; @BS74; @BI76; @Stergioulas], using formulations based on the Weyl-Papapetrou form of the metric where a sufficient subset of the Einstein equations are elliptic. These methods have been applied to exotic static and stationary 4$D$ black holes [@Kleihaus:1997ic; @Kleihaus:2000kg]. In [@star; @Wiseman:2002zc] it was shown how to apply cohomogeneity two methods to higher dimensions, without having $(D-2)$ commuting Killing vectors. A form for the metric analogous to the Weyl ansatz is taken and the Einstein equations reduce to a set of elliptic ones for the metric components, together with constraints which may be satisfied provided appropriate boundary conditions are imposed.
The main challenge to formulating the problem for general cohomogeneity is to present the Einstein equations as an elliptic system in an elegant and covariant manner. In this paper we build on the previous algorithms proposed in [@KitchenHeadrickTW] to find static vacuum black holes. In this method a static solution is analytically continued to a Riemannian manifold. If we are considering a black hole, then provided it has a single non-extremal horizon or multiple horizons sharing the same (non-zero) surface gravity, then the periodicity of Euclidean time can be chosen such that this Euclidean signature manifold has no boundary at the location of the horizon and is perfectly smooth there. The only boundaries in the problem arise from the region far from the horizon where one may consider various boundary or asymptotic conditions [@PFLuciettiTW]. An elegant feature of this approach is that static vacuum Lorentzian geometries can then be treated simply as a particular case of the problem of finding Riemannian Einstein metrics which is more general and has other interesting applications [@CY; @dP3]. Instead of solving the Einstein equation directly, one solves the Harmonic Einstein equation with DeTurck’s choice of gauge fixing term.[^1] For Euclidean signature this is elliptic and can be treated as a boundary value problem. The Ricci-DeTurck flow or Newton method provide explicit algorithms to find solutions [@KitchenHeadrickTW]. Numerical Ricci flows have been considered in [@Garfinkle:2003an; @HeadrickTW; @Holzegel:2007zz; @dP3; @Headrick:2007fk; @KitchenHeadrickTW].
This approach is closely related to the use of generalized harmonic coordinates in the hyperbolic context [@Friedrich; @Garfinkle:2001ni]. One solves the Harmonic Einstein equation to yield a solution to the Einstein equation in generalized harmonic gauge. There is an important and subtle difference between the hyperbolic and elliptic contexts. For dynamics the vanishing of this gauge fixing term is imposed on the initial data, and the contracted Bianchi identity implies it remains zero at later times. Whilst one solves the Harmonic Einstein equation, one only considers this equation in the neighbourhood of a solution to the Einstein equation. In contrast in the Riemannian elliptic case one does not have this luxury. One must impose physical boundary conditions and then solve the Harmonic Einstein equation subject to these. Since one does not know the solutions which correspond to solutions of the Einstein equation, the whole problem being to find these, one cannot consider the Harmonic Einstein equation only in the neighbourhood of such Einstein solutions, but must consider it generally subject to the boundary data. In principle there may exist solutions to the Harmonic Einstein equation, termed Ricci solitons, where the solution is not Einstein, and the DeTurck gauge fixing term does not vanish. In practice one can check whether the solution found is a soliton or not. However, the existence of solitons is rather constrained. Bourguignon [@Bourguignon] has proven on a compact manifold without boundary such solitons cannot exist, and in [@PFLuciettiTW] it was argued using a simple maximum principle that in this static case with various boundary and asymptotic conditions no solitons exist.
Our aim here is to develop general methods to find stationary black holes. However this Riemannian approach for static solutions cannot be applied to the stationary case since there is no analytic continuation of the stationary vacuum Einstein equations. The key observation is that the Harmonic Einstein equation of the Riemannian signature black hole gives precisely the same equations for the metric components if one continues back to Lorentzian signature provided one uses coordinates adapted to the static symmetry. Whilst in general the Lorentzian Harmonic Einstein equation has hyperbolic signature, this shows that when one considers the Harmonic Einstein equation restricted to static Lorenztian metrics it is elliptic and can be solved as a boundary value problem. The only difference is that now the Lorentzian horizon represents a physical boundary - in the Riemannian continuation the static adapted coordinate chart has a boundary at the horizon, in analogy to the origin of polar coordinates. One may determine the boundary conditions in these static coordinates by considering smoothness in a chart that does cover the horizon (analogous to Cartesian coordinates). The key result of this paper is to show that the Harmonic Einstein equation restricted to a class of Lorentzian stationary metrics motivated by the Rigidity theorem [@Hollands:2006rj; @Moncrief:2008mr], general enough to consider wide classes of stationary black holes which include the asymptotically flat case, is elliptic. The horizon must be thought of as a boundary, and there are well posed boundary conditions that can be given there. The resulting elliptic system can be solved by Ricci flow on this space of Lorentzian stationary metrics, or by Newton’s method.
For simplicity we restrict our discussion here to vacuum solutions with no cosmological term and only non-extremal Killing horizons. The structure of the paper is then as follows. We begin in section §\[sec:method\] by reviewing the approach to finding static solutions using the Euclidean continuation and Harmonic Einstein equation. Then in section §\[sec:static\] we show how to reinterpret this from an entirely Lorentzian perspective, demonstrating the equation is elliptic and that regularity conditions can be given at the horizon, which in Lorentzian signature appears as a boundary. In §\[sec:stationary\] we consider the character of the Lorentzian stationary Harmonic Einstein equation. We show that for globally timelike stationary Killing vector (i.e. stationary spacetimes without horizon or ergo-regions) this is again elliptic. From this analysis it is clear that the threat to ellipticity in the stationary case is due to ergo-regions. In §\[sec:blackholes\] we make an ansatz for stationary black hole spacetimes motivated by the Rigidity theorem, general enough to include the asymptotically flat case. This is the ansatz that Harmark [@Harmark:2009dh] has used to consider classification of higher dimensional black holes. Following from Rigidity, Killing symmetry in the directions of angular or linear motion of the horizon is assumed, and this is crucial for maintaining ellipticity in the presence of ergo-regions. Based on the analytic treatment of the stationary problem for the uniqueness theorems, we assume the orbit space is a smooth Riemannian manifold, and the horizon and any axes of symmetries of the angular motion are boundaries of this space. We are then able to conclude that the Harmonic Einstein equation is elliptic when restricted to such stationary spacetimes. We determine the boundary conditions at the horizon and rotational symmetry axes in an analogous manner to those in the Lorentzian static case, and show they are compatible with finding non-soliton solutions to the Harmonic Einstein equations ie. solutions of the stationary Einstein equations. We conclude the paper in section \[sec:example\] by giving a simple numerical example of these techniques. We numerically find 4$D$ rotating black holes in a cavity. We use Anderson’s boundary condition, taking the boundary to be conformal to the product of time and a round sphere with fixed trace of its extrinsic curvature [@Anderson1]. We demonstrate the use of both Newton’s method and Ricci flow as algorithms to find the solutions.
Review of the Harmonic Einstein equation, Ricci flow and Newton’s method for static geometries {#sec:method}
==============================================================================================
We now review in more detail the framework for finding static vacuum black holes that we wish to extend to the stationary case. Since we will consider vacuum solutions with no cosmological term we are interested in finding Ricci flat static metrics. We analytically continue time to give a Euclidean metric. We take all the Killing horizons to have the same surface gravity so that we may choose Euclidean time to have the appropriate period such that we have a smooth Riemannian manifold, and no boundaries associated to the horizons. Thus the problem is now to find Ricci flat Riemannian metrics with a $U(1)$ isometry which is generated by a hypersurface orthogonal Killing vector, and which has fixed points of the isometry at the locations which continue to Killing horizons.
Following our previous work, instead of solving the Ricci flatness condition $R_{\mu\nu} = 0$, we instead introduce the tensor, $$\begin{aligned}
{R}^H_{\mu\nu} \equiv R_{\mu\nu} - \nabla_{(\mu} \xi_{\nu)} \end{aligned}$$ which we will refer to as the Ricci-DeTurck tensor, where the vector field $\xi^\mu$ is constructed as, $$\begin{aligned}
\xi^\mu = g^{\alpha\beta} \left( \Gamma_{\alpha\beta}^{~~\mu} - \bar{\Gamma}_{\alpha\beta}^{~~\mu} \right)\end{aligned}$$ from a fixed smooth reference connection $\bar{\Gamma}_{\alpha\beta}^{~~\mu}$ on the manifold. The term $\nabla_{(\mu} \xi_{\nu)}$ is the ‘DeTurck term’ introduced by DeTurck in the context of Riemannian geometry and later used by him to show Ricci flow is parabolic [@DeTurck]. We note that the vector field $\xi$ is globally defined, being the difference of our Levi-Civita and reference connection. This term is a global analog of the local term considered in generalized harmonic gauge fixing [@Friedrich; @Garfinkle:2001ni] which is now extensively used in numerical simulation following Garfinkle’s work.
We focus on solving the ‘Harmonic Einstein equation’ $R^H_{\mu\nu} = 0$. The Ricci-DeTurck tensor has the property that it exhibits a simple structure in its second derivative terms, so that, $$\begin{aligned}
{R}^H_{\mu\nu} \equiv R_{\mu\nu} - \nabla_{(\mu} \xi_{\nu)} \sim -\frac{1}{2} g^{\alpha\beta} \partial_\alpha \partial_\beta {g}_{\mu\nu} + \ldots\end{aligned}$$ where $\ldots$ refer to terms of lower order in derivatives. Then for a Riemannian signature manifold the Harmonic Einstein equation is clearly elliptic, and can be solved as an elliptic boundary value problem in the metric components $g_{\mu\nu}$. Whilst the Ricci tensor $R_{\mu\nu}$ shares the static symmetry of the metric $g_{\mu\nu}$, $R^H_{\mu\nu}$ does not unless we choose the reference connection appropriately. We do so by taking the reference connection to be the Levi-Civita connection of a smooth reference metric, $\bar{g}_{\mu\nu}$, so that, $$\begin{aligned}
\xi_\alpha = g^{\mu\nu} \left( \bar{\nabla}_\mu g_{\nu \alpha} - \frac{1}{2} \bar{\nabla}_\alpha g_{\mu\nu} \right)\end{aligned}$$ and we take this reference metric to share the static isometry. Here, $\bar{\nabla}$ is the covariant derivative with respect to the metric $\bar{g}_{\mu\nu}$. This ensures $R^H_{\mu\nu}$ is symmetric with respect to the static isometry. Both metric and reference metric continue to Lorentzian spacetimes with the same static Killing horizons with equal surface gravities with respect to Killing time.
The aim is to solve the Harmonic Einstein equation to yield a Ricci flat solution with $\xi^\mu = 0$ which specifies that the metric is presented in the generalized harmonic gauge defined by our reference metric. However in general there will exist solutions which are not Ricci flat, but solve $R_{\mu\nu} = \nabla_{(\mu} \xi_{\nu)}$ so that $\xi^\mu$ does not vanish. Such solutions are called Ricci solitons. Fortunately the existence of Ricci solitons is quite constrained as we shall now discuss. The contracted Bianchi identity for the Harmonic Einstein equation implies $\xi$ obeys the elliptic linear equation, $$\begin{aligned}
\label{eq:bianchi}
\mathcal{O}_\mu^{~\nu} \xi_\nu \equiv \nabla^2 \xi_\mu + R_{\mu}^{~\nu} \xi_\nu = 0 \, .\end{aligned}$$ The boundary conditions for the metric may be taken as giving boundary conditions for the vector field $\xi$. Let us consider a vector field $v^\mu$ with the same boundary behaviour as that of $\xi^\mu$. We expect that our boundary conditions are such that the problem, $$\begin{aligned}
\label{eq:Ov}
\mathcal{O}_\mu^{~\nu} v_{\nu} = 0\end{aligned}$$ is a well posed elliptic linear PDE in the vector components $v^\mu$, and that $v^\mu = 0$ is a solution. For example, taking asymptotically flat boundary conditions then $\xi^\mu \to 0$ in the asymptotic region [@PFLuciettiTW]. Alternatively taking Anderson’s boundary conditions where the conformal class of the metric is fixed on a boundary, together with the trace of the extrinsic curvature [@Anderson1; @PFLuciettiTW] then one imposes $\xi^\mu = 0$. For the boundary conditions relevant for a vacuum Randall-Sundrum brane [@RSI; @RSII], the normal component of $\xi^\mu$ to the brane boundary vanishes, and the tangential components of $\xi^\mu$ have Neumann boundary conditions. In all these three examples the associated problem has elliptic boundary conditions for the vector $v^\mu$. In the first two these ensure $v^\mu$ vanishes, whereas in the latter one, they are consistent with $v^\mu$ vanishing but do not impose it.
A necessary condition for a Ricci soliton to exist is that the linear elliptic operator $\mathcal{O}$ must have a non-trivial kernel. For given boundary conditions on the metric we may view this as an obstruction to the existence of Ricci solitons. A well known result is that for compact manifolds without boundary there are no Ricci soliton solutions for any choice of vector field. Using a simple maximum principle based on the contracted Bianchi identity one can show that with a variety of boundary conditions, including asymptotically flat, Kaluza-Klein and Anderson’s boundary conditions, then again no Ricci solitons can exist [@PFLuciettiTW].
However, even assuming Ricci solitons do exist, then the well-posedness of the Harmonic Einstein equation implies that generically[^2] solutions should be locally unique. Hence a Ricci flat solution cannot be perturbatively close to a soliton solution. Therefore even if solitons exist, it is straightforward in principle to identify whether a solution that has been found is a soliton or not, the most obvious way being to compute $\xi$ and determine if it vanishes.
Here we will now briefly review Anderson’s boundary conditions since we will use them later in our numerical example. Anderson has shown the surprising result that one cannot impose the most obvious Dirichlet boundary condition, namely fixing the induced metric on a boundary, since it does not give rise to a regular elliptic system. Instead he has argued that one can fix the conformal class of the induced metric on a boundary together with the trace of the extrinsic curvature [@Anderson1]. The metric in the vicinity of such a boundary can be written as, $$\begin{aligned}
ds^2 = N dr^2 + N_a dr d {x}^a + {g}_{ab} d{x}^a d{x}^b\end{aligned}$$ where the boundary is the hypersurface $r = 0$ and the remaining tangential coordinates are ${x}^a$ for $a = 1, \ldots, D-1$. For a static non-extremal black hole continued to Euclidean signature, the boundary will be a product of the Euclidean time circle with some spatial geometry. Fixing the conformal class of the induced metric and trace of the extrinsic curvature locally gives $D(D-1)/2$ conditions on the metric functions. However since the metric has $D(D+1)/2$ independent components we require another $D$ conditions. Recall our previous discussion, that on any boundaries or asymptotic regions we must ensure that $\xi$ has a behaviour such that the associated linear problem in equation defined above is well posed and has trivial solution $v^\mu = 0$. In order to achieve this the remaining $D$ conditions are fixed by imposing $\xi^\mu = 0$ on the boundary.
For Riemannian metrics we may solve the Harmonic Einstein equation using two distinct methods. Firstly we may use the DeTurck flow, $$\begin{aligned}
\frac{d}{d\lambda} g_{\mu\nu} = -2 R^H_{\mu\nu}\end{aligned}$$ beginning with an initial guess metric, and then flowing in the auxiliary time $\lambda$, hoping to end at a fixed point of the flow, $R^H_{\mu\nu} = 0$. The flow is strongly parabolic provided $R^H_{\mu\nu}$ is elliptic. Furthermore for our choice of reference metric it preserves the static isometry of $g$. A very elegant feature of this flow is that it is diffeomorphic to the Ricci flow $\frac{d}{d\lambda} g_{\mu\nu} = -2 R_{\mu\nu}$, since the DeTurck term $\nabla_{(\mu} \xi_{\nu)}$ just introduces an infinitessimal diffeomorphism at each point along the flow. Hence while the flow in the space of metrics will depend explicitly on the choice of reference connection $\bar{\Gamma}$, geometrically the flows do not depend on this choice. An important consequence is that the basin of attraction of a fixed point of the flow is geometrically invariant, and doesn’t depend on the choice of $\bar{\Gamma}$. One important drawback of this method is that, as discussed in [@HeadrickTW] vacuum black holes of interest are often unstable fixed points of Ricci flow, due to the existence of Euclidean negative modes of the Lichnerowicz operator [@GPY]. However, the Ricci-DeTurck flow may still be used as an algorithm. If the fixed point of interest has $n$ negative modes, then as described in [@HeadrickTW], an $n$ parameter family of initial guesses must be taken, and the $n$ parameters tuned in the sense of a shooting problem in order to reach the fixed point of interest.
The second method is to consider solving the equations $R^H_{\mu\nu}$ in a non-local way, using the canonical method to solve non-linear equations, the Newton’s method. Writing the linearisation of $R^H$ as, $$\begin{aligned}
R^H_{\mu\nu}[ g + \epsilon \delta g](x) = R^H_{\mu\nu}[g](x) + \epsilon {\Delta}[g](x)_{\mu\nu}^{~~~\alpha\beta} \delta g_{\alpha\beta}(x) + O(\epsilon^2)\end{aligned}$$ we may think of this as an (infinite dimensional) vector equation, with $O(g)$ a matrix, with indices constructed from the spacetime index pair $\mu\nu$ and the point $x$ on the manifold. In practice one will introduce a truncation of the spacetime, for example using finite difference or spectral methods. Let us denote the finite dimensional collective index which represents the spacetime and position/mode using $A,B$, so that we may write $g_{\mu\nu}(x)$ as the vector $g_A$. Then we write the above equation as, $$\begin{aligned}
R^H_A( g + \epsilon \delta g) = R^H_A (g) + \epsilon {\Delta}(g)_{A}^{~B} \delta g_{B} + O(\epsilon^2)\end{aligned}$$ The Newton method iteratively improves a guess metric $g^{(i)}_A$ as, $$\begin{aligned}
g^{(i+1)}_A - g^{(i)}_A = - ({\Delta}(g^{(i)})^{-1})_A^{~B} R^H_{B}(g^{(i)})\end{aligned}$$ As for the Ricci flow method, the Newton method preserves the static isometry of the metric. This method has the important advantage over the Ricci-DeTurck flow method that it is not sensitive to negative modes of the Lichnerowicz operator. However it does assume that the linear problem ${\Delta} \cdot v = R^H$ can be solved for $v$. In practice robust methods exist to solve such (finite dimensional) linear systems, such as biconjugate gradient, which are insensitive to the spectrum of ${\Delta}$, provided there are no zero modes. Thus a single initial guess will suffice, rather than having to tune a family of initial guesses. An important disadvantage of the Newton method over the Ricci-DeTurck flow is that it is not geometric in the sense that the path taken by the algorithm in the space of geometries will depend explicitly on the choice of reference connection $\bar{\Gamma}$. This implies that the basin of attraction of a solution will also depend on this choice.
Static spacetimes from a Lorentzian perspective {#sec:static}
===============================================
Instead of immediately now considering stationary spacetimes, it is instructive to first consider static spacetimes from a Lorentzian perspective. The Harmonic Einstein equation is not elliptic for a general Lorentzian manifold, and without ellipticity one would not expect to be able to impose the various boundary conditions that we require physically in a well posed manner. However, consider a chart away from any horizon which manifests the static symmetry, $$\begin{aligned}
\label{eq:staticLor}
ds^2 = g_{\mu\nu} dx^\mu dx^\nu = - N(x) dt^2 + h_{ij}(x) dx^i dx^j\end{aligned}$$ so that $N > 0$. We may regard Euclidean time as being fibered over a base, the Riemannian manifold which we shall denote $\mathcal{M}$ with Euclidean metric $ds^2_{\mathcal{M}} = h_{ij}(x) dx^i dx^j$. Furthermore with the choice that our reference metric is also static with respect to $\partial / \partial t$, so that, $$\begin{aligned}
\bar{ds}^2 = \bar{g}_{\mu\nu} dx^\mu dx^\nu= - \bar{N}(x) dt^2 + \bar{h}_{ij}(x) dx^i dx^j\end{aligned}$$ again with $\bar{N} > 0$ and $\bar{h}_{ij}$ a smooth Euclidean metric, then $R^H_{\mu\nu}$ shares the static symmetry. Due to this static symmetry the Harmonic Einstein equation $R^H_{\mu\nu} = 0$ thought of as PDEs for the metric components of $g$ is invariant under an analytic continuation $t \rightarrow \tau = i t$. Hence we immediately see that the Harmonic Einstein equation restricted to Lorentzian static metrics and reference metrics is elliptic. Furthermore, Ricci-DeTurck flow yields precisely the same flow equations for the metric functions $N$ and $h_{ij}$ above in either Lorentzian signature, or under continued Euclidean signature.
Explicitly the static Ricci-DeTurck tensor has components, $$\begin{aligned}
R^H_{tt} &=& - \frac{1}{2} \hat{\nabla}^{i}( \partial_{i} N) + \frac{1}{2 N} ( \partial^{i} N)( \partial_{i} N) - \frac{1}{2} \hat{\xi}^{k} \partial_{k} N - \frac{1}{4 N} \bar{h}_{(-1)}^{mi} ( \partial_{m} \bar{N})( \partial_{i} N) \nonumber \\
R^H_{ti}& = & 0 \nonumber \\
R^H_{ij} &=& \hat{R}_{ij} - \hat{ \nabla}_{(i} \hat{\xi}_{j)} - \frac{1}{4 N^2} ( \partial_{i} N)( \partial_{j} N) - \frac{1}{2} h_{k(i} \hat{ \nabla}_{j)} \left( \frac{1}{N} \bar{h}_{(-1)}^{km} \partial_{m} \bar{N} \right)
\end{aligned}$$ where indices are contracted and covariant derivatives $\hat{\nabla}$ are with respect to the base metric $h_{ij}$. To avoid confusion we use the notation $\bar{h}_{(-1)}^{ij}$ for the inverse metric to $\bar{h}_{ij}$, so that $\bar{h}_{ik} \, \bar{h}_{(-1)}^{kj} = \bar{h}^{(-1)ij} \bar{h}_{jk} = \delta^{i}_{k}$. The vector field $\hat{\xi}^i$ is the DeTurck vector of the base metric, namely, $$\begin{aligned}
\label{eq:baseDeTurck}
\hat{\xi}^i = h^{jk} \left( \hat{\Gamma}^i_{~ jk} - \bar{\hat{\Gamma}}^i_{~ jk} \right)\end{aligned}$$ where $\hat{\Gamma}^i_{~ jk} $ is the connection for $h_{ij}$ and $\bar{\hat{\Gamma}}^i_{~ jk} $ is the connection for $\bar{h}_{ij}$. We see the terms that contain two derivatives acting on the metric components of $g$ are of elliptic form provided $h_{ij}$ is a Riemannian metric. We also explicitly see that $R^H_{\mu\nu}$ is symmetric under the static isometry.
In the exterior of any horizons we have $N, \bar{N} > 0$. However now consider a non-extremal Killing horizon, where $\partial / \partial t$ has fixed action. Now $N$ and $\bar{N}$ vanish at the horizon. The base $h_{ij}$ must remain a smooth Riemannian geometry where $N = 0$ for the spatial horizon to have a well defined geometry, and we choose the same for $\bar{h}_{ij}$. In the Riemannian case we know that if we have have a single non-extremal horizon, or multiple horizons with the same surface gravity, then we may make Euclidean time periodic as, $\tau \sim \tau + 2 \pi / \kappa$ for some appropriate choice of constant $\kappa$, such that there is no boundary at the horizon(s) and the geometry is smooth there. Furthermore we know that $R^H_{\mu\nu}$ is a smooth tensor on this geometry, and hence the Ricci-DeTurck flow and Newton methods preserve the smoothness and lack of boundary at the Riemannian horizon. However, in the chart above with $t \rightarrow \tau = i t$, the metric becomes, $$\begin{aligned}
\label{eq:staticEuc}
ds^2 = + N(x) d\tau^2 + h_{ij}(x) dx^i dx^j\end{aligned}$$ and the chart does not cover the horizon where $N = 0$. Such coordinates adapted to the static symmetry are analogous to polar coordinates, and fail at the polar origin, the horizon. To manifest the smoothness of the Riemannian manifold one must go to ‘Cartesian’ coordinates. Taking coordinates in the base adapted to the horizon such that $x^i = (r , x^a)$ where $r =0 $ is the horizon, we write, $$\begin{aligned}
\label{eq:staticpolarEuc}
ds^2 = + r^2 V d\tau^2 + U \left( dr + r \, U_a dx^a \right)^2 + h_{ab} dx^a dx^b\end{aligned}$$ where the metric functions are functions of $r$ and $x^a$. Changing to coordinates, $$\begin{aligned}
a = r \sin{\kappa \tau} \, , \quad b = r \cos{\kappa \tau} \end{aligned}$$ provides a good chart covering the horizon, such that the metric components are smooth functions, provided that $V, U, U_a, h_{ab}$ are smooth ($C^\infty$) functions of $r^2$ and $x^a$, and, $$\begin{aligned}
V = \kappa^2 U\end{aligned}$$ at the horizon $r =0$ [@KitchenHeadrickTW; @PFLuciettiTW]. Precisely the same conditions will apply to the reference metric which is also required to be smooth with no boundary at $r = 0$. We now see that instead of using a good chart which does not manifest the static isometry, we might just as well use the original ‘polar’ chart and simply treat the horizon as a boundary, and determine the boundary behaviour using the regular chart, namely that $V, U, U_a, h_{ab}$ are smooth in $r^2$ and $x^a$, and that $V = \kappa ^2 U$ at $r = 0$ where the constant $\kappa$ determines the angular period of Euclidean time, and hence the temperature of the solution. In practice, adapting coordinates to the static symmetry, and indeed any other isometries, is important numerically to simplify the problem, and this is precisely the approach taken in previous work [@HeadrickTW; @KitchenHeadrickTW]. As noted above taking smooth coordinates we know $R^H_{\mu\nu}$ preserves smoothness and lack of boundary at the horizon. Whilst in adapted coordinates we have a boundary at the horizon the same must be true, namely that the smoothness and regularity conditions above apply equally well to the tensor $R^H_{\mu\nu}$ which must be a smooth tensor on $g$ and we may explicitly check that, $$\begin{aligned}
R^H = + r^2 f d\tau^2 + g \left( dr + r \, g_a dx^a \right)^2 + r_{ab} dx^a dx^b\end{aligned}$$ where the functions $f, g, g_a$ and $r_{ab}$ are smooth in $r^2, x^a$, and in addition $f = \kappa^2 g$ at $r = 0$. However, the simpler way to see that this must be the case is to remember that in the smooth Cartesian coordinates $(a,b, x^a)$ then $g_{\mu\nu}$ is smooth and hence $R^H_{\mu\nu}$ will be too, and since $R^H_{\mu\nu}$ with our reference metric preserves the static isometry, then it follows that $R^H_{\mu\nu}$ must have the behaviour stated above.
In Lorentzian signature, if we are to study only the exterior of the horizon, then the horizon should be regarded as a physical boundary. There is no analog chart to the Riemannian case where the boundary can be smoothly removed without introducing the black hole interior. However, since the Harmonic Einstein equation and the solutions for the metric components are invariant under continuation $t \rightarrow \tau = i t$, then precisely the same boundary conditions for regularity apply in the Lorentzian case as in the Euclidean case in the static adapted chart . Hence we may work directly in Lorentzian signature, where the equations are the same and so are elliptic, and then provide the same boundary conditions there for the metric components at the horizon, taking, $$\begin{aligned}
\label{eq:staticpolarLor}
ds^2 = - r^2 V dt^2 + U \left( dr + r \, U_a dx^a \right)^2 + h_{ab} dx^a dx^b\end{aligned}$$ where $r = 0$ at the horizon, $V > 0 $ outside the horizon and vanishes on it. Then at $r =0$ we again require $V, U, U_a, h_{ab}$ are smooth in $r^2$ and $x^a$, and that $V = \kappa^2 U$. Now the constant $\kappa$ precisely gives the surface gravity of the Killing horizon with respect to $\partial / \partial t$. We may manifest the regularity of this horizon in a similar manner to the Euclidean case by performing a change of coordinates, $$\begin{aligned}
a = r \sinh{\kappa t} \, , \quad b = r \cosh{\kappa t} \end{aligned}$$ giving a chart with coordinates $a,b,x^a$ that now covers the $t = 0$ slice of the Killing horizon, and whose metric components are smooth functions. The essential difference with the Euclidean case is that if we are interested in the exterior of the horizon, then the horizon remains a boundary in this good chart.
Due to the invariance of the components of $R^H_{\mu\nu}$ under $t \rightarrow \tau = i t$, the tensor $R^H_{\mu\nu}$ shares the same regularity properties as the metric in the Lorentzian context too so that, $$\begin{aligned}
R^H = - r^2 f dt^2 + g \left( dr + r \, g_a dx^a \right)^2 + r_{ab} dx^a dx^b\end{aligned}$$ near the chart boundary at $r = 0$ where again $f, g, g_a$ and $r_{ab}$ are smooth in $r^2, x^a$, and in addition $f = \kappa^2 g$. In the Riemannian picture Ricci flow and the Newton method preserve smoothness and lack of boundary at the horizon. We see that equivalently in the Lorentzian picture we have the very nice property that Ricci flow and the Newton method will preserve the surface gravity of the solution.
We conclude this discussion with some comments. One attractive feature of the Riemannian approach to static black holes where one removes the horizon boundary by taking periodic time is that there is no boundary associated to the horizon, and the boundary conditions imposed far from the horizon may be taken to fix the size of the time circle. Since the size of the time circle is interpreted as inverse temperature, we see that viewing the system as a boundary value problem we are naturally led to fix physical data. A very nice consequence of our static Lorentzian discussion is that although we must now view the horizon as a boundary, we are again naturally lead to impose physical data there, namely the surface gravity with respect to $\partial / \partial t$. Asymptotically or on a boundary away from the horizon we impose conditions to fix the value of the function $N$ in , and this then determines the normalization of $\partial / \partial t$. Thus together these fix the physical data specifying the black hole, and moreover this data is preserved by the Ricci flow and Newton method.
Since in practice (for example [@KitchenHeadrickTW]) we choose to use adapted coordinates that manifest the static symmetry even when thinking about the Euclidean formulation of the problem, and then apply boundary conditions where the coordinates degenerate, one might think it makes no difference which signature we think about the static problem in. The mechanics of solving it will be identical in both. Whilst true, there is one important advantage to thinking about the static problem from the Lorentzian perspective, namely that one can consider multiple Killing horizons with respect to $\partial / \partial t$ with *different* surface gravities, each of which is individually preserved by Ricci flow or Newton’s method. Formally we could not previously consider this in the Riemannian case, as only one horizon boundary can be removed and made smooth by a choice of periodic time, the remaining horizons becoming conical singularities. By employing the boundary conditions above in the static adapted coordinates of course we may now treat these conical singularities in the Euclidean context, but the motivation to consider Euclidean signature is somewhat diminished.
Stationary vacuum spacetimes with globally timelike Killing vector {#sec:stationary}
==================================================================
We now wish to consider using the methods above to find stationary vacuum solutions. However, now the spacetimes must be considered as Lorentzian from the outset as there is generally no real Euclidean section to a solution. In this section we begin by considering the case of stationary spacetimes with globally timelike Killing vector (i.e. no horizons or ergo-regions may exist) and will argue that the Harmonic Einstein equation is elliptic. Of course we are ultimately interested in black hole spacetimes which violate such a condition and in the following section §\[sec:blackholes\] we consider more general stationary spacetimes which allow horizons and ergo-regions.
Consider the most general stationary metric with Killing vector $T = \partial / \partial t$, which we may write using coordinates adapted to the stationary isometry as, $$\begin{aligned}
\label{eq:stationarygeneral}
g = - N(x) \left( dt + A_i(x) dx^i \right)^2 + h_{ij}(x) dx^i dx^j\end{aligned}$$ Now under our assumption that $T$ is globally timelike we have $N > 0$ and we further assume that the function $N$ is bounded. Physically this implies our spacetime has no Killing horizons, and also no ergo-region. Since $\det{ g_{\mu\nu}} = - N \det h_{ij}$ we see that provided the metric $g$ is Lorentzian and smooth, so that $\det g_{\mu\nu} < 0$ and bounded, this implies that $\det{h_{ij}} > 0$. We may then regard this metric as a smooth fibration of time over a base manifold $\mathcal{M}$ so that $(\mathcal{M},h)$ is a smooth Riemannian manifold with Euclidean signature metric $h_{ij}$.
Whilst one might imagine that for a stationary spacetime the natural way to think about the metric is using the ADM ansatz, it is worth noting that the metric above is not of this form, but rather should be thought of as a Kaluza-Klein ansatz over time.[^3] Hence the base manifold $(\mathcal{M},h)$ is *not* the submanifold obtained by taking a constant time $t$ slice of the spacetime. Rather it is the geometry one obtains by performing a Kaluza-Klein reduction in the time direction. As a result of this the second order derivative terms acting on the metric components $g_{\mu\nu}$ in the stationary Harmonic Einstein equation then go as, $$\begin{aligned}
{R}^H_{\mu\nu} \sim -\frac{1}{2} g^{\alpha\beta} \partial_ \alpha \partial_\beta {g}_{\mu\nu} + \ldots = -\frac{1}{2} h^{ij} \partial_i \partial_j {g}_{\mu\nu} + \ldots\end{aligned}$$ and we see whilst the metric $g_{\mu\nu}$ is indeed Lorenztian, since there is no dependence on the coordinate $t$, it is actually the metric $h_{ij}$ that controls the character. This immediately implies that the Harmonic Einstein equation $R^H_{\mu\nu} = 0$ is elliptic since $h$ is smooth and Euclidean signature. Thus the stationary problem reduces to an elliptic problem on the Riemannian base manifold $\mathcal{M}$.
However, we also require that $R^H_{\mu\nu}$ is a tensor that is symmetric with respect to the stationary isometry $T$. Without this, Ricci-DeTurck flow and Newton’s method will not consistently truncate to the class of stationary metrics . In order that $R^H_{\mu\nu}$ preserves the symmetry $T$, we choose the reference metric $\bar{g}$ to also be a smooth Lorentzian metric which is stationary with respect to the vector field $T$, so that, $$\begin{aligned}
\bar{g} = - \bar{N}(x) \left( dt + \bar{A}_i(x) dx^i \right)^2 + \bar{h}_{ij}(x) dx^i dx^j \end{aligned}$$ where we also assume here that $T$ is globally timelike and bounded with respect to $\bar{g}$ so that $\bar{N} > 0$ and bounded. Then $\bar{h}_{ij}$ gives a second Riemannian metric on the same manifold $\mathcal{M}$. Since $R^H_{\mu\nu}$ preserves the stationary symmetry, the Ricci-DeTurck flow can be consistently truncated to a parabolic flow on the space of Lorentzian stationary metrics. Since this flow remains diffeomorphic to Ricci flow (subject at least to the normal component of $\xi$ vanishing on any boundaries), we arrive at the interesting result that we may apply Ricci flow to Lorentzian stationary spacetimes. Likewise the Newton method will preserve the symmetry.
In a situation where the solution we wish to find has a stationary Killing vector that is globally timelike and bounded then nearby to that solution the character of the Harmonic Einstein equation will be elliptic. Subject to imposing suitable boundary conditions on any boundaries or asymptotic regions, one may use the Lorentzian stationary Ricci-DeTurck flow or Newton method to solve for the solution. One must start with an initial guess that has globally timelike bounded stationary Killing field $T$, and then provided that guess is sufficiently good the subsequent Ricci-DeTurck flow or Newton iterations preserve that $T$ is globally timelike.
We recall that for a solution to the Harmonic Einstein equation to be Ricci flat we require $\xi = 0$. Thus we must ensure that our boundary or asymptotic conditions are compatible with vanishing $\xi$. Let us briefly consider the case of a Dirichlet boundary. The situation is entirely analogous to the static Euclidean case discussed above. Consider taking coordinates adapted to the boundary, so that, $$\begin{aligned}
ds^2 = - N(x) \left( dt + A_r(x) dr + A_a(x) dx^a \right)^2 + V dr^2 + V_a dr dx^a + h_{ab}(x) dx^a dx^b\end{aligned}$$ so that $x^i = ( r, x^a )$ and the boundary is at $r = 0$. Then fixing the induced metric would specify Dirichlet conditions for $N$, $A_a$ and $h_{ab}$. Requiring that $\xi^t$, $\xi^r$ and $\xi^a$ vanish then provides conditions for $V$, $V_a$, and $A_r$. Thus as in the static case we have precisely fixed both the induced metric and $\xi^\mu = 0$ on the boundary. Following an analogous argument as in [@PFLuciettiTW] we expect these boundary conditions are well posed for the elliptic Harmonic Einstein equation. Provided on other boundaries conditions are set consistent with $\xi$ vanishing we expect solutions where $\xi$ vanishes globally may be found (presuming they exist). Without the maximum principle of the static case [@PFLuciettiTW] we cannot rule out solitons with non-vanishing $\xi$, so in a practical context one simply has to test a solution found to see if it is Ricci flat or a soliton.[^4]
For completeness we now explicitly give the stationary Ricci-DeTurck flow equations (which of course give the Harmonic Einstein equations at a fixed point $\partial / \partial \lambda = 0$); $$\begin{aligned}
\frac{\partial N}{ \partial \lambda} &= \overbrace{\hat{\nabla}^{i}( \partial_{i} N)}^{{}= h^{km} \partial_{k} \partial_{m} N + \dots} - \frac{1}{N} ( \partial^{i} N)( \partial_{i} N)
- \frac{N^2}{2} F^{ij} F_{ij} + \hat{\xi}^{k} \partial_{k} N \\
& + \frac{1}{2 N} \bar{h}_{(-1)}^{km} ( \partial_{m} \bar{N})( \partial_{k} N) + \bar{h}_{(-1)}^{km} \bar{N} (A^{i} - \bar{A}^{i}) \bar{F}_{jm} \partial_{k} N + \frac{1}{2} (A_{i} - \bar{A}_{i})^2 \bar{h}_{(-1)}^{km} ( \partial_{m} N)( \partial_{k} N) \end{aligned}$$ $$\begin{aligned}
\frac{\partial A_{i}}{ \partial \lambda} &= \overbrace{\hat{\nabla}^{k} F_{ik} + \hat{\nabla}_{i}(A_{k} \hat{\xi}^{k})+ \hat{\nabla_{i}}(\hat{\nabla}^{p} A_{p} - \bar{\hat{\nabla}}^{p} \bar{A}_{p})}^{{}= h^{km} \partial_{k} \partial_{m} A_{i} + \dots}- \frac{1}{N} F_{ij} \partial^{j}N + \hat{ \xi^{k}}F_{ki}+ \frac{1}{2N}\bar{h}^{km}_{(-1)}F_{ki} \partial_{m} \bar{N} \\ & - \bar{h}^{km}_{(-1)} \bar{N} (A^{j} - \bar{A}^{j})F_{ik}\bar{F}_{jm} + \frac{1}{2} \bar{h}^{km}_{(-1)}(A_{p} - \bar{A}_{p})^2 F_{ki} \partial_{m} \bar{N}+ \hat{\nabla}_{i}\left( \left(\frac{1}{2N}\bar{h}^{mp}_{(-1)}(A_{m} - \bar{A}_{m})\right. \right. \\ & \left. \left. + \frac{1}{ \bar{N}}(A^{p} - \bar{A}^{p}) + \frac{1}{2} \bar{h}^{kp}_{(-1)}(A_{m} - \bar{A}_{m})^2 (A_{k} - \bar{A}_{k})\right) \partial_{p} \bar{N}\right) + \hat{\nabla_{i}}(\bar{h}^{km}_{(-1)} \bar{N} (A^{j} - \bar{A}^{j})(A_{k} - \bar{A}_{k}) \bar{F}_{jm}) \end{aligned}$$ $$\begin{aligned}
\frac{\partial h_{ij}}{ \partial \lambda} &= \overbrace{-2 \hat{R}_{ij} + 2\hat{\nabla}_{(i} \hat{\xi}_{j)}}^{{}= h^{km} \partial_{k} \partial_{m} h_{ij} + \dots} + \frac{1}{2 N^2}(\partial_{i} N)( \partial_{j} N) - N F_{j}^{k} F_{ki} \\ & \left(\frac{1}{2}h_{ik} \hat{\nabla}_{j}(\frac{1}{N} \bar{h}^{km}_{(-1)} \partial_{m} \bar{N}) + h_{ik} \hat{\nabla}_{j}(\bar{h}^{km}_{(-1)} \bar{N} \bar{F}_{qm} (A^{q} - \bar{A}^{q})) \right. \\ & \left. + \frac{1}{2} h_{ik} \hat{\nabla}_{j}(\bar{h}^{km}_{(-1)}(A^{p} - \bar{A}^{p})^2 \partial_{m} \bar{N}) + (i \leftrightarrow j) \right)\end{aligned}$$ where indices are contracted and covariant derivatives $\hat{\nabla}$ are with respect to the base metric $h_{ij}$, and we have defined the antisymmetric tensors $F_{ij} \equiv \partial_i A_j - \partial_j A_i$ and $\bar{F}_{ij} \equiv \partial_i \bar{A}_j - \partial_j \bar{A}_i$. Again $\hat{\xi}^i$ is the DeTurck vector of the base metric defined as before in .
We reiterate that the assumption the reference metric is also stationary with respect to $T$ is responsible for $R^H_{\mu\nu}$ preserving the stationary symmetry and hence the consistent truncation of the Ricci-DeTurck flow to the class that we see above. However the further assumption that $\bar{g}$ has globally timelike bounded Killing vector $T$, so that $\bar{N} > 0$ and bounded, ensures that the equations are regular and there are no singular terms arising from vanishing or diverging $\bar{N}$ in the above. We note that the components of the DeTurck tensor $R^H_{\mu\nu}$ may be derived from these DeTurck flow equations using the relations, $$\begin{aligned}
\frac{\partial N}{ \partial \lambda} &=& -2 R^H_{tt} \\
\frac{\partial A_{i}}{ \partial \lambda} &=& - \frac{2}{N} (R^H_{it} - R^H_{tt} A_{i}) \\
\frac{\partial h_{ij}}{ \partial \lambda} &=& -2(R^H_{ij} + R^H_{tt} A_{i} A_{j} - R_{i t} A_{j} - R_{j t} A_{i}) \end{aligned}$$
Stationary black holes and the Harmonic Einstein equation {#sec:blackholes}
=========================================================
We now proceed to consider the case of Ricci flat non-extremal stationary black holes. In the context of the discussion above now the norm of $T$ will vanish either at the horizon itself, assuming that $T$ is a globally timelike Killing vector (such as for certain Kerr-AdS black holes [@Caldarelli:1999xj]), or outside the horizon at the boundary of the ergo-region. Since we are interested in the exterior of the horizon, in the first case we may treat the system described above for globally timelike $T$ and now regard the horizon as a boundary of the base manifold $\mathcal{M}$ where suitable boundary conditions are required. However in the latter, more general case, outside the horizon but inside the ergo-region we have the norm of $T > 0$ and hence $\det{h_{ij}} < 0$. Now the base manifold in the previous section fails to be Riemannian and then our argument above that the Harmonic Einstein equation is elliptic fails.
In order to make progress we must use the Rigidity property of stationary black holes, proved in $D> 4$ by Ishibashi, Hollands and Wald [@Hollands:2006rj] and by Moncrief and Isenberg [@Moncrief:2008mr] for various asymptotics including asymptotic flatness. Assume there exists a stationary Killing vector $T$. Then the Rigidity theorem states that for a rotating non-extremal Killing horizon with topology $\mathbb{R} \times \Sigma$, for $\Sigma$ compact, there exists a Killing vector $K$ that commutes with $T$, which is normal to the horizon. Furthermore there exist some number $N \ge 1$ of commuting Killing vectors $R_a$, which also commute with $T$ and generate closed orbits with period $2 \pi$ and $K$ may be written in terms of these as, $$\begin{aligned}
K = T + \Omega^a R_a\end{aligned}$$ for some constants $\Omega^a$. This result is physically significant as it ensures that the rotation of the horizon is generated by an isometry of the spacetime. Were this not the case one would expect gravitational radiation to be emitted from the region near the horizon and this would presumably violate the assumption of stationarity.
Motivated by the Rigidity theorem we assume that our stationary spacetime, with stationary Killing vector $T$, has $N$ additional Killing vectors $R_a$ for $a = 1,\ldots,N$, which commute amongst themselves and with $T$ and generate rotational or translational isometries. In the former case they generate closed orbits which we take to have period $2 \pi$, and may have axes of rotational symmetry where $R_a$ vanishes. In the latter case they generate non-compact orbits. Let us take there to be a number of disconnected horizon components, $\mathcal{H}_1 , \ldots , \mathcal{H}_k$. Then Rigidity implies that each component is a Killing horizon with Killing vector given by a linear combination of the isometries $T$ and $R_a$ so that, $K_{\mathcal{H}_m} = T + \Omega_{\mathcal{H}_m}^a R_a$ for some constants $\Omega_{\mathcal{H}_m}^a$ (which may be different for each component).
As a consequence of these assumptions we may write the metric adapting coordinates to the isometries, so that using coordinates $y^A = \{ t, y^a \}$, $$\begin{aligned}
\label{eq:bhansatz}
d{s}^2 = {g}_{\mu\nu} dX^\mu dX^\nu = {G}_{AB}(x) \left( dy^A + {A}^A_i(x) dx^i \right) \left( dy^B +{A}^B_j(x) dx^j \right) +{h}_{ij}(x) dx^i dx^j\end{aligned}$$ where $T = \partial / \partial t$ and $R_a = \partial / \partial y^a$. In analogy with the stationary case in the previous section we see that the geometry may be thought of as a fibration of the Killing vector directions over a base manifold $\mathcal{M}$ with metric $h_{ij}$. This base manifold $\mathcal{M}$ is the orbit space of the full Lorentzian spacetime with respect to the isometries $T, R_a$. If $R_a$ generates a compact orbit, then since we have chosen to normalise the period to $2\pi$, then the coordinate $y^a$ is periodic with $y^a \sim y^a + 2 \pi$. Again we note that the metric $h_{ij}$ is not that induced on a constant $y^A$ submanifold of the full spacetime, but rather is the metric one would obtain by performing a ‘Kaluza-Klein’ reduction over these Killing directions.
The full spacetime is Lorentzian and so exterior to the horizons $\det g_{\mu\nu} = \det G_{AB} \det h_{ij} < 0$. On physical boundaries or asymptotic regions $T$ is timelike, $R_a$ are spacelike and hence the fiber metric $G_{AB}$ is Lorentzian there, and consequently the base metric is Euclidean. At a horizon $\mathcal{H}_m$, since the norm of $K_{\mathcal{H}_m} = T + \Omega_{\mathcal{H}_m}^a R_a$ vanishes, then $\det G_{AB} = 0$. In addition at axes of symmetry associated to the fixed action of a compact $R_a$, $\det G_{AB}$ will vanish.
Following the uniqueness theorem treatment of 4$D$ stationary black holes as an elliptic problem on the two dimensional Riemannian orbit space bounded by the horizon and axes of symmetry (as for example discussed in [@Carter]) and its generalization to $D$ dimensional metrics with $(D-2)$ commuting Killing vectors [@Morisawa:2004tc; @Hollands:2007aj] which is treated in the same manner, we make the follow key assumption;
- We assume that the orbit space base manifold $(\mathcal{M},h)$ is a smooth Riemannian manifold with boundaries given by the horizons and axes of symmetry of the $R_a$ that generate rotational isometries.
A consequence of $\det h_{ij} > 0$ everywhere on $\mathcal{M}$ (including the horizon and axis boundaries), is that $\det G_{AB} \ge 0$ everywhere on $\mathcal{M}$ with it vanishing only at the horizon or axis boundaries of $\mathcal{M}$. We note that $\det g_{\mu\nu} = 0$ at the horizons and axes, as one would expect since the chart breaks down there. As Harmark has discussed [@Harmark:2009dh], the structure of $\mathcal{M}$ together with the data $\Omega_{\mathcal{H}_m}^a$ at the horizon boundaries, and the data of which $R_a$ vanishes at the axis boundaries defines a ‘rod structure’ for stationary spacetime and has been conjectured to classify higher dimensional black holes.
Ellipticity
-----------
We note that we have not considered the stationary Killing field $T$ to be timelike. In the presence of horizons it will become null on the horizon or be spacelike if the horizon is surrounded by an ergo-region. We reiterate that in the previous section §\[sec:stationary\] it was precisely where $T$ failed to be timelike that ellipticity would break down, since the base metric would fail to be Riemannian. The crucial observation is that for our class of stationary spacetimes , $$\begin{aligned}
R^H_{AB} & = & -\frac{1}{2} g^{\alpha\beta} \partial_ \alpha \partial_\beta {g}_{AB} + \ldots = -\frac{1}{2} h^{mn} \partial_m\partial_n {G}_{AB} + \ldots \nonumber \\
R^H_{ij} & = & -\frac{1}{2} g^{\alpha\beta} \partial_ \alpha \partial_\beta {g}_{\mu\nu} + \ldots = -\frac{1}{2} h^{mn} \partial_m \partial_n {h}_{ij} + \ldots \end{aligned}$$ where again the $\ldots$ represent lower than second order derivative terms. We see the equations have character determined solely by the metric $h_{ij}$, and by our assumption above that the base $\mathcal{M}$ is Riemannian, this is indeed elliptic.
In analogy with the previous section \[sec:stationary\], in order to ensure that $R^H_{\mu\nu}$ shares the symmetries of $g$ we choose the reference metric $\bar{g}$ so that $T, R_a$ are again Killing with respect to it, and obey precisely the same assumptions as above for $g$. Thus we may write, $$\begin{aligned}
\label{eq:bhref}
d{s}^2 = \bar{g}_{\mu\nu} dX^\mu dX^\nu = \bar{G}_{AB}(x) \left( dy^A + \bar{A}^A_i(x) dx^i \right) \left( dy^B + \bar{A}^B_j(x) dx^j \right) + \bar{h}_{ij}(x) dx^i dx^j\end{aligned}$$ and we further assume that $(\mathcal{M},\bar{h})$ is a smooth Riemannian manifold. Then the Ricci-DeTurck flow and Newton’s method consistently truncate to the Lorentzian stationary spacetimes of the form .
We must impose suitable boundary conditions. In addition to the boundaries that define the asymptotics we must also now treat the boundaries at the horizons and axes of symmetries of the rotational Killing vectors. We will shortly discuss these boundary conditions explicitly. Using the Ricci-DeTurck flow or Newton method if we start from initial data in our stationary class, then for small flow times or updates we expect to remain in this class. In particular we expect $(\mathcal{M},h)$ to remain a Riemannian manifold. Provided this condition holds for the solution of interest, and our initial guess is sufficiently close to this, then we might hope to reach this solution.
We will now explicitly give the Ricci-DeTurck flow equations, from which the components of $R^H_{\mu\nu}$ can be deduced using, $$\begin{aligned}
\frac{\partial G_{AB}}{ \partial \lambda} &=& -2 R^H_{AB} \nonumber \\
\frac{\partial A_{j}^{C}}{ \partial \lambda} &=& -2 G^{AC}(R^H_{jA} - R^H_{AB} A^{B}_{j}) \nonumber \\
\frac{\partial h_{ij}}{ \partial \lambda} &=& -2(R^H_{ij} + R^H_{AB} A_{i}^{A} A_{j}^{B} - R^H_{iA} A^{A}_{j} - R^H_{jA} A^{A}_{i}) \end{aligned}$$ Contracting indices and taking covariant derivatives $\hat{\nabla}$ with respect to the base metric $h_{ij}$, we find, $$\begin{aligned}
\frac{\partial G_{AB}}{ \partial \lambda} &=& \overbrace{\hat{\nabla}^{i}( \partial_{i} G_{AB})}^{{}= h^{mp} \partial_{m} \partial_{p} G_{AB} + \dots} - G^{CD}( \partial^{i} G_{AD})( \partial_{i} G_{CB}) + \frac{1}{2} \bar{h}_{(-1)}^{km} G^{CD}( \partial_{m} \bar{G}_{CD})( \partial_{k} G_{AB}) \nonumber \\
&& -\frac{1}{2} G_{BE} G_{AF} F^{Eij} F_{ij}^{F} + \hat{\xi}^{k} \partial_{k} G_{AB} + \bar{h}_{(-1)}^{km} \bar{G}_{CD} (A^{Di} - \bar{A}^{Di}) \bar{F}^{C}_{im} \partial_{k} G_{AB} \nonumber\\
&& + \frac{1}{2} \bar{h}^{km}_{(-1)} (A^{Ci} A^{D}_{i} + \bar{A}^{Ci} \bar{A}^{D}_{i}- 2 A^{Ci} \bar{A}^{D}_{i})( \partial_{m} \bar{G}_{CD})( \partial_{k} G_{AB}) \nonumber \end{aligned}$$ $$\begin{aligned}
\frac{\partial A_{i}^{C}}{ \partial \lambda} &=& \overbrace{- \hat{\nabla}^{k} F_{ik}^{C} + \hat{\nabla}_{i} (A^{C}_{k} \hat{\xi}^{k}) + \hat{\nabla}_{i}(\hat{\nabla}^{p} A^{C}_{p} - \bar{\hat{\nabla}}^{p} \bar{A^{C}_{p}})}^{{}=h^{mp} \partial_{m} \partial_{p} A^{C}_{i} + \dots} - G^{AC} F_{ij}^{B} \partial^{j} G_{AB} + \hat{\xi}^{k} F_{ki} ^{C} \nonumber \\
&& + \frac{1}{2} \bar{h}^{km}_{(-1)} G^{DE} F_{ki}^{C} \partial_{m} \bar{G}_{DE} - \bar{h}^{km}_{(-1)} \bar{G}_{DE} (A^{jD} - \bar{A}^{jD}) F_{ik}^{C} \bar{F}_{jm}^{E} \nonumber \\
&& + \frac{1}{2} \bar{h}^{km}_{(-1)} (A^{pD} A^{E}_{p} + \bar{A}^{pD} \bar{A}^{E}_{p}- 2 A^{pD} \bar{A}^{E}_{p}) F_{ki}^{C} \partial_{m} \bar{G}_{DE} \nonumber \\
& & + \hat{\nabla}_{i} \left( \left(\frac{1}{2}\bar{h}^{mp}_{(-1)} G^{DE} (A^{C}_{m}- \bar{A}^{C}_{m}) + \bar{G}^{CE}(A^{pD}- \bar{A}^{pD}) \right. \right. \nonumber \\
&& \left. \left. + \frac{1}{2} \bar{h}^{kp}_{(-1)} (A^{mD} A^{E}_{m} + \bar{A}^{mD} \bar{A}^{E}_{m} - 2 A^{mD} \bar{A}^{E}_{m})(A^{C}_{k} - \bar{A}^{C}_{k})\right) \partial_{p} \bar{G}_{DE} \right) \nonumber \\
& & + \hat{\nabla}_{i}(\bar{h}^{km}_{(-1)} \bar{G}_{DE} (A^{jE} - \bar{A}^{jE})(A^{C}_{k} - \bar{A}^{C}_{k}) \bar{F}_{jm}^{D}) \nonumber \end{aligned}$$ $$\begin{aligned}
\frac{\partial h_{ij}}{ \partial \lambda} &=& \overbrace{-2 \hat{R}_{ij} +2 \hat{\nabla}_{(i} \hat{\xi}_{j)}}^{{}= h^{mp} \partial_{m} \partial_{p} h_{ij} + \dots} + \frac{1}{2} G^{AB} G^{CD} ( \partial_{i} G_{CB})( \partial_{j} G_{AD}) - G_{AB} F_{j}^{Ak} F^{B}_{ki} \nonumber \\
& & + \left (\frac{1}{2} h_{ik} \hat{ \nabla}_{j} (G^{AB} \bar{h}^{km}_{(-1)} \partial_{m} \bar{G}_{AB}) + h_{ik} \hat{ \nabla_{j}}( \bar{h}^{km}_{(-1)} \bar{G}_{AB} \bar{F}^{B}_{qm} (A^{qA} - \bar{A}^{qA})) \right. \nonumber \\
& & \left. + \frac{1}{2} h_{ik} \hat{\nabla}_{j}( \bar{h}^{km}_{(-1)} (A^{pA} A^{B}_{p} + \bar{A}^{pA} \bar{A}^{B}_{p} - 2 A^{pA} \bar{A}^{B}_{p}) \partial_{m} \bar{G}_{AB}) + (i \leftrightarrow j ) \right)\end{aligned}$$ where as before the base DeTurck vector field $\hat{\xi}^i$ is defined as in and we analogously define $F^A_{ij} \equiv \partial_i A^A_j - \partial_j A^A_i$ and similarly, $\bar{F}^A_{ij} \equiv \partial_i \bar{A}^A_j - \partial_j \bar{A}^A_i$. In the appendix \[app:equations\] to this paper we present useful intermediate results that lead to these expressions.
Reduced stationary case
-----------------------
We now make a simple observation, namely that if we require our stationary metric to have invariance under the discrete symmetry, $$\begin{aligned}
t \rightarrow -t \, , \quad y^a \rightarrow -y^a\end{aligned}$$ then this allows for a consistent truncation of the Harmonic Einstein equation in the sense that the Ricci-DeTurck tensor $R^H_{\mu\nu}$ is also invariant. This symmetry implies that all the $A^A_i$ vanish, and we explicitly see from the above equations that $R^H_{Ai}$ vanishes, as required for the invariance of $R^H_{\mu\nu}$. Consequently the Ricci-DeTurck flow and Newton methods consistently truncate. We term this the ‘reduced stationary’ case.
In this case the Ricci-DeTurck tensor considerably simplifies, and has non-zero components, $$\begin{aligned}
R^H_{AB} &= & - \frac{1}{2} \hat{\nabla}^{i}( \partial_{i} G_{AB}) +\frac{1}{2} G^{CD}( \partial^{i} G_{AD})( \partial_{i} G_{CB}) - \frac{1}{4} \bar{h}^{km} G^{CD}( \partial_{m} \bar{G}_{CD})( \partial_{k} G_{AB}) - \frac{1}{2} \hat{\xi}^{k} \partial_{k} G_{AB} \nonumber \\
R^H_{ij} &= & \hat{R}_{ij} - \hat{ \nabla}_{(i} \hat{\xi}_{j)} - \frac{1}{4} G^{AB} G^{CD} ( \partial_{i} G_{CB})( \partial_{j} G_{AD}) - \frac{1}{2} h_{k(i} \hat{ \nabla}_{j)} (G^{AB} \bar{h}^{km} \partial_{m} \bar{G}_{AB}) \end{aligned}$$ with $\xi^A = 0$. For a Dirichlet boundary one imposes the induced metric fixing $G_{AB}$ and the tangential components of $h_{ij}$. The remaining components of $h_{ij}$ are then determined by requiring $\xi^i = 0$.
We note that in 4$D$ the ‘circularity’ theorem [@WaldBook] implies all stationary vacuum solutions may be put in this reduced form, given a condition at a single point which is satisfied for asymptotically flat space. More generally all higher dimensional analytic solutions known to us are of this form.
Boundary conditions for the Killing horizons and axes of symmetry {#sec:bc}
-----------------------------------------------------------------
We now explicitly give the boundary conditions for the components of our stationary spacetime metric and reference metric at the Killing horizons or rotational symmetry axes. Recall that for the 4$D$ uniqueness theorems the horizon and symmetry axis play the role of boundaries for the Riemannian orbit space, in a very similar manner to our higher dimensional (and cohomogeneity) case here. Therefore we may regard the results in this section on the metric behaviour at the horizons and axes as generalising the boundary conditions in that context (see for example [@Carter]). They are also consistent with the boundary conditions discussed by Harmark using particular coordinates on the base manifold [@Harmark:2009dh].
This is an analogous problem to deducing the smoothness condition for a spherically symmetric function in spherical polar coordinates at the origin. The function depends only on the radial coordinate $r$, but since it is smooth, meaning in Cartesian coordinates $x^i$ it is a smooth $C^\infty$ function of the $x^i$’s, then as $r^2 = x^i x^i$ it cannot be a smooth function of $r$, but rather is smooth in $r^2$. If we require that the function is only $C^2$ then we deduce that the function simply has the Neumann condition $\partial f / \partial r |_{r=0}= 0$. We may perform a similar analysis for a tensor, the only difference being that now the components transform as one moves between a chart which manifests smoothness but not the symmetry, and a chart which manifests symmetry but not the smoothness. The details of this are straightforward and are given in appendix \[app:bc\] for a smooth $(0,2)$ tensor at a Killing horizon or rotational symmetry axis. We shall now apply these results to our spacetime metric and reference metric.
Let us first assume that there is a single Killing horizon, or multiple horizons with common normal Killing vector $K = T + \Omega^a R_a$. It is then convenient to change coordinates as, $$\begin{aligned}
\label{eq:trans}
t \, , \; y^a \quad \rightarrow \quad \tilde{t} = t \, , \quad \tilde{y}^a = y^a - \Omega^a t\end{aligned}$$ so that $K = \partial / \partial \tilde{t}$ and $R_a = \partial / \partial \tilde{y}^a$. We note that if $R_a$ generates a compact orbit, then the coordinate $\tilde{y}^a$ is periodic with $\tilde{y}^a \sim \tilde{y}^a + 2 \pi$. Now consider a boundary, either due to the vanishing norm of $K$ or a compact $R_a$. We take base coordinates $x^i = ( r , x^{\tilde{i}} )$ adapted to the boundary so that it lies at $r = 0$, and decompose the base metric as, $$\begin{aligned}
\label{eq:hdecomp}
h_{ij} dx^i dx^j = N dr^2 + r \, N_{\tilde{i}} d r dx^{\tilde{i}} + h_{\tilde{i}\tilde{j}} dx^{\tilde{i}} dx^{\tilde{j}}\end{aligned}$$ and likewise for the reference metric where $N \rightarrow \bar{N}$, $N_{\tilde{i}} \to \bar{N}_{\tilde{i}}$ and $h_{\tilde{i}\tilde{j}}\to \bar{h}_{\tilde{i}\tilde{j}}$.\
[**Horizon:**]{} For a Killing horizon we write the following metric components as, $$\begin{aligned}
G_{\tilde{t} A} = - r^2 f_A \, , \quad A^{A}_{r} = r g^A \, , \end{aligned}$$ for $A = (\tilde{t}, \tilde{y}^a)$ and then let $X = \left\{ f_A \, , \; g^A \, , \; G_{\tilde{y}^a \tilde{y}^b} \, , \;
A^{A}_{\tilde{i}}\, , \;
N \, , \; N_{\tilde{i}} \, , \; h_{\tilde{i}\tilde{j}} \right\}
$ be the set of functions describing our metric. Let $\bar{X}$ be the analogous set describing the reference metric. Then the results of appendix \[app:bc\] imply that for the metric and reference metric to be smooth we require the following behaviour; the functions $X$ and $\bar{X}$ must be smooth functions of $r^2$ and $x^{\tilde{i}}$ at $r=0$, and furthermore obey the regularity conditions, $$\begin{aligned}
\label{eq:horizon}
\left( f_{\tilde{t}} - \kappa^2 N \right) |_{r=0} = 0 \, , \quad \left( \bar{f}_{\tilde{t}} - \kappa^2 \bar{N} \right) |_{r=0} = 0 \end{aligned}$$ where $\kappa$ is constant and gives the surface gravity with respect to $T$. We note that since both the metric and reference metric are smooth with respect to the same vector field $K$, which is Killing for both, it is the same constant $\kappa$ that must enter the conditions above for both of them.\
[**Axis:**]{} Consider an axis associated to a vanishing compact $R_a$. Without loss of generality choose this to be $R_N$. Then we choose to write, $$\begin{aligned}
G_{\tilde{y}^N A} = r^2 f_A \, , \quad A^{A}_{r} = r g^A \, , \end{aligned}$$ and let $Y = \left\{ f_A \, , \; g^A \, , \; G_{\tilde{t}\tilde{t}} \, , \; G_{\tilde{t}\tilde{y}^{\tilde{a}}} \, , \; G_{\tilde{y}^{\tilde{a}}\tilde{y}^{\tilde{b}}} \, , \; A^{A}_{\tilde{i}}\, , \;
N \, , \; N_{\tilde{i}} \, , \; h_{\tilde{i}\tilde{j}} \right\}
$ be the set of functions describing our metric (where $\tilde{a} = 1,\ldots,N-1$). Let $\bar{Y}$ be the set of functions that analogously describe the reference metric. Appendix \[app:bc\] implies that for a smooth metric and reference metric we must have that the metric functions $Y$ and $\bar{Y}$ are smooth functions of $r^2$ and $x^{\tilde{i}}$ at $r=0$, and in addition we require, $$\begin{aligned}
\label{eq:axis}
\left( f_{\tilde{y}^N} - N \right) |_{r=0} = 0 \, , \quad \left( \bar{f}_{\tilde{y}^N} - \bar{N} \right) |_{r=0} = 0\end{aligned}$$ Of course we obtain analogous conditions for an axis with respect to a different $R_a$.
We see that if we have a single Killing vector $K = T + \Omega^a R_a$ normal to all horizons, we may use the coordinates $(\tilde{t}, \tilde{y}^a)$ and obtain rather simple boundary conditions. Of course one can write these conditions in the original $(t, y^a)$ coordinates. The axis conditions take exactly the same form, with replacements $\tilde{t} \rightarrow t$ and $\tilde{y}^a \rightarrow y^a$. If we have multiple Killing horizons with different normals, then one must deduce the boundary conditions for each taking coordinates as in with $\Omega^a$ appropriate to each horizon.\
[**A horizon meeting an axis:**]{} It is straightforward to check that the boundary conditions at the meeting of a horizon with an axis, or two axes, are compatible with each other. Here we will explicitly check this for the metric in the former case, for an axis of $R_N$, noting that one obtains similar results for the other cases and for the reference metric. Take coordinates on the base such that the horizon is at $r = 0$ and the axis at $\theta = 0$. Then we write the base metric as, $$\begin{aligned}
\label{eq:join1}
h_{ij} dx^i dx^j = N dr^2 + M d\theta^2 + r \theta A dr d\theta + r \, B_{\tilde{i}} d r dx^{\tilde{i}} + \theta \, C_{\tilde{i}} d \theta dx^{\tilde{i}} + h_{\tilde{i}\tilde{j}} dx^{\tilde{i}} dx^{\tilde{j}}\end{aligned}$$ where now $\tilde{i} = 1, \ldots, D-3$. Then writing, $$\begin{aligned}
\label{eq:join2}
G_{\tilde{t} \tilde{t} } &=& - r^2 f \, , \quad G_{\tilde{t} \tilde{y}^{\tilde{a}} } = r^2 f_{\tilde{a}} \, , \quad G_{\tilde{y}^N \tilde{y}^N} = \theta^2 g \, , \quad G_{\tilde{y}^N \tilde{y}^{\tilde{a}}} = \theta^2 g_{\tilde{a}} \quad G_{\tilde{t} \tilde{y}^N } = r^2 \theta^2 k \, , \nonumber \\
A^{A}_{r} &=& r \, p^A \; , \quad A^{A}_{\theta} = r \, q^A \, , \end{aligned}$$ our arguments from appendix \[app:bc\] applied to the horizon $r=0$ and to the axis $\theta = 0$ then imply that the set of functions characterising the metric found in the equations and above, $N, M, \ldots , k, p^A, q^A$, together with the remaining components $G_{\tilde{y}^a\tilde{y}^b}$ and $A_{\tilde{i}}^A$, must all be smooth functions in $r^2$, $\theta^2$ and $x^{\tilde{i}}$ near the meeting point $r = \theta = 0$. Furthermore regularity requires, $$\begin{aligned}
\left( f - \kappa^2 N \right) |_{r=0} = 0 \; , \quad \mathrm{and} \quad \left( g - M \right) |_{\theta=0} = 0\end{aligned}$$ We see that the conditions from each boundary give rise to a consistent set of behaviours above. In particular it implies that two boundaries (a horizon and axis, or two axes) meet in the base at right-angles.\
A very important point is that having introduced boundary conditions for the metric we must check that the conditions that this implies for the vector $\xi^\mu$ are compatible with ensuring the elliptic problem in equation is well posed with trivial solution. To investigate this we must consider our choice of reference metric , which also is required to be regular and hence is subject to the same boundary conditions above for its components. Then one can explicitly check that, $$\begin{aligned}
\xi^r |_{r = 0} = 0 \, , \quad \partial_r \xi^{\tilde{i}} |_{r = 0} = 0 \, , \quad \partial_r \xi^A |_{r = 0}= 0\end{aligned}$$ both at a horizon and axis of symmetry, which is indeed consistent with well posedness of the associated linear problem and with a trivial solution.
One may consider the Ricci-DeTurck flow or Newton method operating on the metric $g$ near a horizon. With the choice of reference metric above that has the same isometry $K$ and is also regular at the boundaries, the Ricci-DeTurck tensor will be symmetric under $K$. Consider the Ricci-DeTurck tensor in Cartesian coordinates. Since the metric and reference metric components will be smooth at the fixed point, then the Ricci-DeTurck tensor will be regular there. Thus in our adapted coordinates it will also obey the same regularity conditions as the metric and reference metric. In particular, Ricci-DeTurck flow and Newton’s method will preserve regularity, and therefore we have the very elegant result that they will leave the surface gravity constant. The same is true for an axis of symmetry where again regularity is preserved.
We note that we may view the above smoothness conditions in the weaker sense of requiring the Cartesian form to only be $C^2$. In this case instead of finding our various functions above are smooth in $r^2$ and $x^{\tilde{i}}$, we have instead only that these functions obey Neumann boundary conditions, ie. vanishing gradient. In addition we also have the regularity conditions $f_{\tilde{t}} = \kappa ^2 N$ for horizons, and $f_{\tilde{y}^N} = N$ for axes as before. Common folklore is that solutions to elliptic problems are analytic, and hence one might expect that only imposing such $C^2$ boundary conditions one would certainly find the stronger smooth behaviour at the horizons and axes, and presumably the much stronger analytic behaviour.
As a final comment we note that for these $C^\infty$ and $C^2$ conditions, the elliptic system has Neumann boundary conditions for the various functions above, and in addition to this, also the constraints, $f_{\tilde{t}} = \kappa ^2 N$ for horizons, and $f_{\tilde{y}^N} = N$ for axes. One might be concerned that these latter conditions should not be imposed in addition to the Neumann conditions, as this is ‘too much data’ for an elliptic problem. However, we emphasize here that this ‘fictitious’ boundary should be viewed as a regular singular point of the equations, due to the singular terms arising from the vanishing norm of the Killing vector, and hence a usual counting of boundary conditions does not apply. Instead we reiterate that the regularity conditions will be preserved by the Ricci-DeTurck flow and Newton method, and thus it is better to think of these conditions not as boundary conditions, but rather as a restriction of the problem to the class of regular metrics, and using the Ricci-DeTurck flow and Newton method which act within that class.
An example: Kerr
----------------
It is instructive to consider the simple example of the Kerr solution in light of our discussion above. The Kerr solution is in reduced stationary form, so that $A^A_i = 0$. In the conventional Boyer-Lindquist coordinates the Kerr metric takes the form, $$\begin{aligned}
ds^2 = G_{tt} dt^2 + 2 G_{t\phi} dt d\phi + G_{\phi\phi} d\phi^2 + h_{rr} dr^2 + h_{\theta\theta} d\theta^2\end{aligned}$$ with fiber metric, $$\begin{aligned}
&&G_{tt} = - \frac{ \left( \Delta - a^2 \sin^2 \theta \right) }{ \Sigma } \, , \quad
G_{\phi\phi} = \sin^2 \theta \frac{ \left( ( r^2 + a^2 )^2 - \Delta a^2 \sin^2\theta \right) }{ \Sigma} \, , \nonumber \\
&&G_{t\phi} = - a \sin^2\theta \frac{ \left( r^2 + a^2 - \Delta \right) }{ \Sigma} \end{aligned}$$ and base, $$\begin{aligned}
&& h_{rr} = \frac{\Sigma}{\Delta} \, , \quad h_{\theta\theta} = \Sigma\end{aligned}$$ where the functions $\Delta, \Sigma$ are defined as $\Delta = r^2 + a^2 - 2 M r$ and $\Sigma = r^2 + a^2 \cos^2\theta$. The stationary Killing vector $T = {\frac{\partial }{\partial t}}$ and the rotational Killing vector is $R = {\frac{\partial }{\partial \phi}}$.
The base manifold $\mathcal{M}$ has coordinates $r, \theta$ upon which the metric components depend explicitly. The outer horizon is a boundary of $\mathcal{M}$ and is located at $\Delta = 0$ where $r \equiv r_h = M + \sqrt{M^2 - a^2}$, and the remaining boundaries are from the axes of rotation at $\theta = 0,\pi$. One finds, $$\begin{aligned}
\det{G_{AB}} = -(a^2 + r(r-2M)) \sin^2 \theta\end{aligned}$$ which vanishes at these boundaries. Everywhere in the exterior of the black hole, $r_h < r$ and $0 < \theta < \pi$ we have that $G_{AB}$ has Lorentzian signature and $h_{ij}$ is Euclidean and smooth. The Killing field $K = T + \Omega R$ is tangent to the horizon and timelike near there, where the angular velocity of the horizon is given as, $\Omega = \frac{a}{a^2+r_h^2}$.
Whilst the $\theta$ coordinate is a regular coordinate on the base manifold at the rotation axes, the radial $r$ coordinate is not at the horizon (since $\Delta \to 0$ there). We therefore define a new radial coordinate, $\rho$, such that $d \rho = dr / \sqrt{\Delta}$, and $\rho = 0$ at the horizon, giving, $$\begin{aligned}
r = M + \sqrt{M^2 - a^2} \cosh \rho\end{aligned}$$ so that the components of the base metric $h_{ij}$ are smooth at the horizon boundary. In particular in these coordinates the determinant of the base metric, $$\begin{aligned}
h_{ij} dx^i dx^j = \frac{\Sigma}{\Delta} dr^2 + \Sigma d\theta^2 = \Sigma \left( d\rho^2 + d\theta^2 \right) \quad \implies \quad \det{h_{ij}} = \Sigma^2 \ge r_h^2\end{aligned}$$ and thus we see that since $r_h > 0$ the base is indeed a smooth Riemannian manifold everywhere on and in the exterior of the horizon. Since we have only one horizon it is convenient to use the coordinates discussed above, $$\begin{aligned}
\tilde{t} = t \, , \quad \tilde{\phi} = \phi - \Omega t\end{aligned}$$ and then we may confirm that near the horizon, $\rho = 0$, we have, $$\begin{aligned}
G_{\tilde{t}\tilde{t}} = -\kappa^2 \left( h_{\rho \rho}|_{\rho=0} \right) \rho^2 + \mathcal{O}( \rho^4) \end{aligned}$$ $$\begin{aligned}
G_{\tilde{t}\tilde{\phi}}= \mathcal{O}(\rho^2) \; , \quad
G_{\tilde{\phi} \tilde{\phi}} = \mathcal{O}(1)\end{aligned}$$ $$\begin{aligned}
h_{\rho \rho} = h_{ \theta \theta} = ( r_h^2 + a^2 \cos^2 \theta) + \mathcal{O}(\rho^2)\end{aligned}$$ in accord with our boundary behaviour above, where $\kappa$ is the surface gravity of the Kerr solution, $$\begin{aligned}
\kappa^2 = \frac{M^2 - a^2}{4M^2 r_{h}^2} \, .\end{aligned}$$ At the axis of symmetry $\theta = 0$ we have, $$\begin{aligned}
G_{\tilde{t} \tilde{t}} =\mathcal{O} (1) \; , \quad
G_{\tilde {t} \tilde{\phi}} = \mathcal{O}(\theta^2)\end{aligned}$$ $$\begin{aligned}
G_{\tilde{\phi} \tilde{\phi}} = \left( h_{\theta\theta} |_{\theta = 0} \right) \theta^2 + \mathcal{O}(\theta^4) \end{aligned}$$ $$\begin{aligned}
h_{\rho \rho} = h_{ \theta \theta} = (a^2 + (M + \sqrt{M^2 - a^2} \cosh \rho)^2) + \mathcal{O} (\theta^2)\end{aligned}$$ which again agrees with our calculation of axis boundary behaviour. Likewise the same agreement is seen for the axis $\theta = \pi$.
Example application: 4$D$ rotating black holes in a cavity {#sec:example}
==========================================================
In the above we have set out a numerical framework to find general stationary vacuum black holes. We now use the example of 4$D$ rotating black holes in a cavity as a toy example to demonstrate the methods discussed may be applied straightforwardly in practice.
We choose to impose Anderson’s boundary conditions at the cavity wall [@Anderson1], where we fix the conformal class of the induced metric, and also the trace of the extrinsic curvature of the boundary. A canonical choice of cavity is such that the induced metric on the 3 dimensional boundary is conformal to the product of time with a round 2-sphere, $$\begin{aligned}
ds^2_{B} & = & - dt^2 + d \theta^2 + \sin^2{\theta} \, d\phi^2\end{aligned}$$ where $\phi \sim \phi + 2 \pi$. We must then specify the trace of the extrinsic curvature of the boundary which we choose to be constant. If one takes the Schwarzschild solution and cuts it off at finite radius, then the extrinsic curvature of the boundary that is introduced is constant and positive. Hence in the rotating case we also take this trace of extrinsic curvature to be positive. Let us denote this positive constant $\alpha$. By a global scaling we may choose $\alpha$ to take any positive value. For later convenience we choose $\alpha = \sqrt{2}$.
We consider the spacetime to be in the reduced stationary class, with a single Killing horizon with spherical topology which is rotating in the $\phi$ direction, so that $R = \partial / \partial \phi$ is Killing, and $K = T + \Omega R$ is normal to the horizon. As discussed above it is convenient to use the coordinates, $$\begin{aligned}
\tilde{t} = t \, , \quad \tilde{\phi} = \phi - \Omega t\end{aligned}$$ so that the boundary behaviour at the horizon takes a simple form. Now $K = \partial / \partial \tilde{t}$ and $R = \partial / \partial \tilde{\phi}$. We explicitly write an ansatz for the reduced stationary metric as, $$\begin{aligned}
\label{eq:ansatz}
ds^2 &=& g_{\mu\nu} dx^\mu dx^\nu = G_{AB} dy^A dy^B + h_{ij} dx^i dx^j \\
&=& \left( \begin{matrix} - r^2 V B & r^2 \sin^2{\theta} \, W B \\ r^2 \sin^2{\theta} \, W B & \sin^2{\theta} \, S B \end{matrix} \right)_{AB} dy^A dy^B + \left( \begin{matrix} A + \frac{1}{B} r^2 \sin^2{\theta} F^2 & r \sin{\theta} \, F \\ r \sin{\theta} \, F & B \end{matrix} \right)_{ij} dx^i dx^j \nonumber\end{aligned}$$ where $y^A = ( \tilde{t}, \tilde{\phi} )$ and $x^i = ( r, \theta )$ and we take the horizon to be located at $r = 0$ where $K$ is null, and the axis of symmetry to be at $\theta = 0, \pi$ where $R$ vanishes. Finally we must specify the coordinate position of the cavity, and choose this to be at $r=1$. Having chosen this the base $\mathcal{M}$ is then covered by a single chart with domain $0 \le r \le 1$ and $0 \le \theta \le \pi$. However, we further assume the reflection symmetry $\theta \rightarrow \pi - \theta$, to reduce the domain to $0 \le \theta \le \pi/2$.
By assumption $h_{ij}$ is a Riemannian metric on our coordinate domain. Since $K$ is normal to the horizon, we have $V > 0$ in the neighbourhood of the exterior of the horizon. For the metric to be smooth at the horizon, our boundary conditions imply that near $r = 0$ the metric functions $\{ V, S, W, A, B, F \}$ are smooth in $r^2$ and $\theta$ (or for the metric only being $C^2$ we require Neumann boundary conditions in $r$, and $C^2$ in $\theta$), with the additional requirement that, $$\begin{aligned}
\label{eq:exbc1}
\left( V B - \kappa^2 A \right)|_{r = 0} = 0\end{aligned}$$ which specifies the surface gravity. At the axis $\theta = 0$ we require the functions $\{ V, S, W, A, B, F \}$ are smooth in $\theta^2$ and $r$ (again for a $C^2$ metric we require Neumann conditions in $\theta$ and $C^2$ in $r$), with the additional requirement that, $$\begin{aligned}
\label{eq:exbc2}
S |_{\theta = 0} = 1\end{aligned}$$ and likewise for the axis at $\theta = \pi $. The reflection symmetry $\theta \rightarrow \pi - \theta$ imposes that $F$ is odd about $\theta = \pi/2$ and the other metric functions are even there. The induced metric at the cavity boundary $r=1$ in the $\tilde{t}, \tilde{\phi}$ coordinates is conformal to, $$\begin{aligned}
\label{eq:boundary}
ds^2_{B} & = & ( - 1 + \Omega^2 \sin^2{\theta} ) d\tilde{t}^2 + 2 \Omega \sin^2{\theta} d \tilde{t} d \tilde{\phi} + \sin^2{\theta} d \tilde{\phi}^2 + d \theta^2 \, ,\end{aligned}$$ which implies that at $r = 1$ we have Dirichlet conditions, $$\begin{aligned}
V = (1 - \Omega^2 \sin^2{\theta} ) \, , \quad W = \Omega \, , \quad S = 1 \end{aligned}$$ where then the function $B$ gives the conformal factor. The projector onto the induced metric on the cavity boundary is $L_{\mu\nu} \equiv g_{\mu\nu} - n_\mu n_\nu$ where, $$\begin{aligned}
n = \frac{1}{\sqrt{A}} \left( \frac{\partial}{\partial r} - \frac{r \sin{\theta} F}{B} \frac{\partial}{\partial \theta} \right)\end{aligned}$$ is the outer unit normal to the boundary. The trace of the extrinsic curvature, $\alpha$, is then given as $\alpha \equiv L^{\mu\nu} \nabla_\mu n_\nu$. The remaining three metric functions $B, A, F$ at $r=1$ have boundary conditions determined from simultaneously requiring $\xi^r = \xi^\theta = 0$ together with imposing the extrinsic curvature has constant trace $\alpha = \sqrt{2}$. These take a somewhat complicated form but in essence are coupled oblique boundary conditions. We emphasize that Anderson has proven that these give a regular elliptic system [@Anderson1], ensuring that they give well posed boundary conditions for the numerical problem. Indeed we have encountered no problems with the cavity boundary in this toy example.
We see that the physical moduli of a black hole with such boundary conditions, the surface gravity $\kappa$ and angular rotation $\Omega$, are directly imposed in these boundary conditions. We see that for $\Omega \ge 1$ the horizon Killing field $K$ is no longer timelike. This implies that the stationary Killing vector $T$ develops an ergo-region for $\Omega \ge 1$. We note that for $\Omega < 1$ the Killing vector $T$ is timelike near the boundary and horizon, although in principle it might become spacelike for some intermediate region. However in the results we now present we see no evidence of such exotic behaviour.
For the reference metric we must choose the same form as in , with the same boundary conditions. We make a simple explicit choice, $$\begin{aligned}
\label{eq:ref}
\bar{ds}^2 = \bar{g}_{\mu\nu} dx^\mu dx^\nu &=& \left( \begin{matrix} - r^2 \left( 1 - \Omega^2 r^2 \sin^2{\theta} \right) & \Omega r^2 \sin^2{\theta} \\ \Omega r^2 \sin^2{\theta} & \sin^2{\theta} \end{matrix} \right)_{AB} dy^A dy^B + \left( \begin{matrix} \frac{1}{\kappa^2} & 0 \\ 0 & 1\end{matrix} \right)_{ij} dx^i dx^j \end{aligned}$$ which we note indeed satisfies the smoothness conditions at the horizon and axis, in addition to the requirements and above. The constants $\kappa$ and $\Omega$ entering the expression above give the surface gravity and angular rotation. One may compute that the trace of the extrinsic curvature is equal to $\kappa$ for this reference metric. We note that whilst we will fix the trace of the extrinsic curvature for the metric, $\alpha = \sqrt{2}$, the same quantity for the reference metric is not required to be fixed or equal to $\alpha$.
Numerical results
-----------------
We have used Newton’s method to solve the Harmonic Einstein equations in this example, and have also simulated the Lorentzian stationary Ricci-DeTurck flow. Our aim is not to perform high precision numerics, but rather to check that the method behaves as expected. We use simple second order finite difference to represent the Harmonic Einstein equations and boundary conditions numerically. We have used various resolutions up to $160 \times 80$ in the radial and angular directions respectively, and have checked the convergence of the results is consistent with second order scaling.
Whilst in principle smoothness in $r^2$ at the horizon should be preserved by Ricci flow or Newton’s method, numerical accuracy and stability is improved if in addition to taking smooth initial data, one also imposes Neumann conditions explicitly on the metric functions at the horizon. We take analogous conditions in $\theta$ at the axis boundaries. Likewise whilst the conditions and in principle are again preserved by Ricci flow or Newton’s method, we also impose these explicitly on the metric functions as boundary conditions to improve accuracy.
![Plot of the log (base 10) of the maximum value of $\phi$ over our domain for a typical solution ($\Omega = 0.6$ and $\kappa^2 = 2$) against log (base 10) of the number of radial lattice points $N$. The number of angular lattice points is $N/2$. The figure shows data for resolutions $20\times10$, $30\times15$, $40\times20$, $60\times30$, $90\times45$, $120\times60$ and $160\times80$. We see that $\phi$, and hence the vector field $\xi$ are consistent with vanishing in the continuum limit, as we would expect for a Ricci flat solution. The same behaviour is seen for all solutions of the Harmonic Einstein equation we have obtained. []{data-label="fig:phi"}](phimax.pdf){width="14cm"}
We have found solutions using the Newton method starting with the reference metric as an initial guess. For all these solutions we confirm that the base metric is smooth and Riemannian, although we note that presumably if there was a solution where this was not the case we should not expect to have found it. We have checked that the solutions found to the Harmonic Einstein equation are indeed Ricci flat, rather than solitons, by ensuring that $\xi^\mu$ is small and consistent with vanishing in the continuum limit. We note that for a reduced stationary metric $\xi^A = 0$, and hence the scalar $\phi = g_{\mu\nu} \xi^\mu \xi^\nu = h_{ij} \xi^i \xi^j > 0$ for $h_{ij}$ being Riemannian. Hence vanishing $\phi$ implies vanishing $\xi^\mu$. Figure \[fig:phi\] shows the maximum value of the scalar $\phi$ over our domain for a typical solution, with $\kappa = 2$ and $\Omega = 0.6$ plotted against resolution. We observe similar behaviour for the other values of $\kappa$ and $\Omega$.
![Plot of $x = \sqrt{ \mathcal{A}_h / \mathcal{A}_b }$ against surface gravity $\kappa^2$, where $\mathcal{A}_h$ is the horizon area. Data is shown for numerical solutions with $\Omega = 0.2, 0.6, 1$ and $1.2$. For $\Omega = 0$ the analytic curve is given as a dotted line (the numerical data is omitted as by eye it is indistinguishable from this curve). For $\Omega <1$ there is a minimum value of $\kappa$ which divides the branch of solutions into small (S) and large (L) black holes. The solutions very close to the minimum are difficult to obtain numerically. []{data-label="fig:plotx"}](plotx.pdf){width="14cm"}
![Plots of $R_{eq}$ (red), the equatorial radius, and $\sqrt{\mathcal{A}_h / 4 \pi}$ (blue) against surface gravity $\kappa^2$. Data is shown for $\Omega = 0.2, 0.6$ and $1.2$, with the analytic curve for $\Omega = 0$ drawn as a dotted line. For non zero $\Omega$ and away from the small $x$ limit the horizon becomes deformed from the round sphere to become prolate.[]{data-label="fig:plotRA"}](plotRA.pdf){width="14cm"}
Let us define the equatorial radius $R_{eq}$, horizon area $\mathcal{A}_h$, and cavity boundary area $\mathcal{A}_b$, $$\begin{aligned}
R_{eq} = \sqrt{S B} |_{r = 0 \, , \theta = 0} \, , \quad \mathcal{A}_h = 4\pi \int_{0}^{\pi/2} d\theta \,\left. \sqrt{S} B \sin{\theta} \right|_{r=0} \, , \quad \mathcal{A}_b = 4\pi \int_{0}^{\pi/2} d\theta \,\left. \sqrt{S} B \sin{\theta} \right|_{r=1} \; .\end{aligned}$$ Let us further define the quantity, $$\begin{aligned}
x \equiv \sqrt{ \frac{ \mathcal{A}_h }{ \mathcal{A}_b } } \; ,\end{aligned}$$ which we may intuitively think of as giving the ratio of the size of the horizon compared to that of the boundary.
We fix $\Omega$ and then varying $\kappa^2$ to scan through the moduli space of solutions. We have obtained solutions for $\Omega = 0, 0.2, 0.4, 0.6, 0.8, 1.$ and $1.2$, although for graphical clarity we have not presented all this data in the figures that follow. In figure we plot the quantity $x$ against $\kappa^2$ for fixed $\Omega = 0.2, 0.6, 1$. Similarly in figure we plot the quantities $R_{eq}$ and $\sqrt{A / 4\pi}$ against $\kappa^2$ for the same solutions. For $\Omega = 0$ the solutions are static, and therefore with spherical cavity boundary conditions will simply reproduce the Schwarzschild solution. For Schwarzschild one finds, $$\begin{aligned}
\kappa^2 = \frac{1}{4 x^2 \left( 1 - x \right) } \; , \quad R_{eq} = \frac{1}{2 \alpha} \frac{x \left( 4 - 3 x \right)}{ \sqrt{ 1 - x }} \end{aligned}$$ and we observe agreement with the data as we expect. We have not plotted the data points we have computed for $\Omega = 0$ as these lie on the analytic $\Omega = 0$ curves given in the figures. As we turn on the angular rotation, fixing $\Omega$, we see deviation from this Schwarzschild behaviour, again as we should expect. We note that in the small black hole limit, $x \to 0$, for fixed $\Omega$, we expect to recover the Schwarzschild behaviour and we see this is the case.
For static solutions, so $\Omega = 0$, there is a minimum surface gravity at $x = 2/3$, and so the solutions divide into the small ($x<2/3$) and large ($x > 2/3$) black holes in an analogous manner to York’s construction where one fixes the induced metric on the cavity wall [@York:1986it] (rather than the conformal class and $\alpha$). The AdS-Kerr solutions of Carter [@Carter:1968ks] give an indication of what to expect in the rotating case, where in the usual manner we think of AdS as being a box. Fixing the AdS length $\ell = 1$, then for $\Omega < 1$ there are small and large solutions and a minimum $\kappa$, and for $\Omega > 1$ there is only one branch of solutions which terminates in an extremal solution [@Caldarelli:1999xj]. In addition AdS-Kerr admits a globally timelike Killing vector for $\Omega < 1$, and has an ergo-region for $\Omega > 1$.
We recall that for $\Omega > 1$ we must develop an ergo-region for our solutions. Then in analogy with AdS-Kerr we might expect that for $0 < \Omega < 1$ we have small and large black holes with a minimum surface gravity, whereas for $\Omega > 1$ there is no minimum surface gravity and instead there is an extremal limit where $\kappa \to 0$. This is indeed borne out by our crude numerical results. It would be interesting to confirm with greater accuracy that this transition from minimum surface gravity to extremal limit at fixed $\Omega$ does indeed occur precisely at $\Omega = 1$. Here we have observed that $\Omega = 0.8$ has a minimum surface gravity, and for $\Omega = 1$ we have not found a minimum. We also confirm that for all our solutions with $\Omega < 1$ the vector field $K$ is globally timelike outside the horizon i.e. $G_{\tilde{t}\tilde{t}} < 0$.
We briefly comment on finding the solutions using Newton’s method. We begin by finding a solution for some $\kappa$, starting with the reference metric as an initial guess. For $\Omega < 0.6$ starting with $\kappa^2 = 2$ this yields a small black hole solution with boundary trace of extrinsic curvature $\alpha = \sqrt{2}$ as required. We then move along the branch of solutions by perturbing the solution and reference metric to yield a good approximation to a solution with nearby $\kappa$ and the same $\Omega$ and $\alpha$. Using this method we may quickly scan along a branch of solutions to the minimum value of $\kappa$. We could try to find an initial guess in the basin of attraction of a large solution and then scan along this branch. However the approach we take is to extrapolate our small black hole solutions near the minimum $\kappa$ to gain a good guess for a large solution, with $\kappa$ just greater than the minimum. Using this one finds a large solution, and can then scan along the large branch. The Newton method struggles to find solutions very close to the minimum value of $\kappa$. At the minimum there is a normalisable zero mode of the linearised Harmonic Einstein equation, and near to it there is a low lying mode that renders the linear operator that must be inverted in Newton’s method rather ill conditioned. This is why in figures \[fig:plotx\] and \[fig:plotRA\] for $\Omega = 0.2$ and $0.6$ the section of the curves connecting the small and large branches are missing. In order to find black holes with $\Omega > 0.6$ we have found that extrapolating these solutions together with their reference metrics for some fixed $\kappa$, say $\kappa^2 = 2.5$, on the small black hole branch then allows an initial guess for a higher value of $\Omega$ to be found. Having found this new solution branch, one can proceed to iterate this method to find larger $\Omega$ solutions.
![ Figure showing the horizon and cavity boundary embeddings for rotating black holes with $\Omega = 1$. This branch actually extends to lower $\kappa$ although the horizon cannot be globally embedded in $\mathbb{R}^3$ isometrically for $\kappa^2 < 0.7$. []{data-label="fig:embed10"}](embed10.png){width="9cm"}
We may isometrically embed the horizon and cavity boundary of these solutions into the Euclidean space $\mathbb{R}^3$ as surfaces of revolution. Taking a polar slice through these embeddings then gives curves in two dimensions representing the geometry of the horizon and boundary. In figures \[fig:embed00\], \[fig:embed06\] and \[fig:embed10\] we plot the upper quadrant (i.e. $0 \le \theta \le \pi/2$) of these two dimensional embedding curves against $\kappa^2$ for $\Omega = 0, 0.6$ and $1$. Since for $\Omega = 0$ the solutions are simply Schwarzschild, figure \[fig:embed00\] is included simply for comparison. For the solutions found with $\Omega < 1$ there is both a small and a large branch. Interestingly for $\Omega = 0.6$ we see that for the large solutions with sufficient $\kappa$, the geometry of the horizon is such that its vertical extent in the embedding is actually greater than that of the cavity indicating the geometry of the solution is rather exotic. For $\Omega = 1$ we find that the horizon may not be embedded isometrically into $\mathbb{R}^3$ for $\kappa^2 < 0.7$ in an analogous manner to the Kerr solution sufficiently near extremality [@Gibbons:2009qe].
We conclude our discussion of these solutions by studying certain stationary Ricci-DeTurck flows. From previous work [@HeadrickTW; @KitchenHeadrickTW] we know that for static black holes in a spherical cavity with fixed induced metric the small black hole is always unstable having a Euclidean negative mode [@GPY; @Gregory:2001bd] whereas the large black hole is stable. There are two flows for a small black hole generated by this unstable mode. Perturbing the small black hole by the negative mode in one sense generates a flow that asymptotes to the large black hole. By reversing the sign of this perturbation one then generates a second flow that at finite time develops a singularity where the horizon shrinks to zero size. In [@HeadrickTW] it was argued that by an appropriate surgery on the manifold one may continue the flow to flat space.
We note that whilst for fixed induced metric the static spherically symmetric problem (continued to Euclidean signature) is well posed [@Holzegel:2007zz], in our cohomogeneity two example we must take Anderson’s boundary conditions to obtain a regular elliptic system. Interestingly we find that for these boundary conditions and $\Omega < 1$ the large black holes now have one mode of instability for Ricci-DeTurck flow, and the small black holes tested have two. We emphasize that since we are considering the flow in the space of stationary Lorentzian metrics these are *not* Euclidean negative modes, but are rather negative eigenvalue eigenmodes of the Lichnerowicz operator restricted to our class of stationary spacetimes. This implies that to find the large black holes one must tune a one parameter family of initial geometries to reach these unstable fixed points. For the small black holes presumably one must tune two parameters although we have not tried this.
We now give an example of the Ricci flow algorithm to find a large black hole by tuning a one parameter family of data. We present results for $\Omega = 0.6$ and $\kappa^2 = 2$, although we expect to see qualitatively similar results for other $\Omega < 1$ and $\kappa$ greater than the minimum surface gravity. Let us take a one parameter family of initial data to be, $$\begin{aligned}
\label{eq:initdata}
{ds}^2 &=& \left( \begin{matrix} - r^2 \left( 1 - \Omega^2 r^2 \sin^2{\theta} \right) \left(1 + \delta f \right) & \Omega r^2 \sin^2{\theta} \\ \Omega r^2 \sin^2{\theta} & \sin^2{\theta} \end{matrix} \right)_{AB} dy^A dy^B + \left( \begin{matrix} \frac{1}{\kappa^2} \left(1 + \delta f \right) & 0 \\ 0 & 1 + \delta \left(r f \right)^2\end{matrix} \right)_{ij} dx^i dx^j \nonumber\end{aligned}$$ with $f = \left(1 - r^2\right)$ and where $\delta$ is the parameter we may tune. We choose $\kappa^2 = 2$ and $\Omega = 0.6$, and then the trace of the extrinsic curvature $\alpha = \sqrt{2}$ as required, independent of the value of $\delta$. Likewise the constants $\kappa$ and $\Omega$ entering the expression above give the surface gravity and angular velocity independent of $\delta$. The above metrics also satisfy all our other remaining boundary requirements discussed above. The reference metric is as before in equation .
We find that we may reach the critical point for $\delta \simeq 2.9$. In figure \[fig:flows\] we depict embeddings of the horizon and boundary for 3 values of $\delta$, one for slightly less than the critical value ($\delta = 2.8$), one for slightly greater ($\delta = 3.0$), and one close to the critical value ($\delta = 2.9$). We see that for $\delta$ smaller than the critical value the horizon and cavity boundary appear to collapse to a singularity at finite flow time. For $\delta$ greater than the critical value the horizon and boundary appear to expand forever under the Ricci flow. For an appropriately tuned flow we may approach the large rotating black hole solution with $\Omega = 0.6$ and $\kappa^2 = 2$. This is the same solution found by the Newton method. Indeed we see the horizon and cavity embeddings in the tuned flow in this figure match those presented in figure \[fig:embed06\] for $\kappa^2 = 2$.
Summary
=======
In this paper we have discussed applying the Harmonic Einstein equation with DeTurck term to the case of Lorentzian static and stationary spacetimes, where Ricci flow or Newton’s method acting on these spacetimes can be used as algorithms to solve this equation and yield solutions to the static or stationary vacuum Einstein equations.
Previously static spacetimes have been treated by continuing to periodic Euclidean time, which can remove boundaries associated to non-extremal horizons [@KitchenHeadrickTW]. The Harmonic Einstein equation is then solved as an elliptic boundary value problem on a Riemannian manifold, for example, using Ricci flow or Newton’s method.
We have firstly shown how to view this procedure for static spacetimes from an entirely Lorentzian point of view. It is straightforward to see, with a suitable choice of reference metric, that the Lorentzian Harmonic Einstein equation consistently truncates to the static case and gives an elliptic system. Whilst the horizon must be treated as a boundary now in Lorentzian signature, by considering the Euclidean continuation, it is easy to deduce Lorentzian boundary conditions for regularity at the horizon. Ricci flow on the space of static spacetimes, or Newton’s method can then be applied to solve for static solutions of the vacuum Einstein equations.
Having treated the static case from a manifestly Lorentzian perspective we then considered the stationary case for globally timelike stationary Killing vector. As in the static Lorentzian case it is straightforward to see that for a suitable choice of reference metric, the Harmonic Einstein equation consistently truncates to such stationary spacetimes, again giving an elliptic system. However, we learn that in general, if the stationary Killing field fails to be timelike, such as at horizons or in ergo-regions, it is unclear the system will remain elliptic.
Intuitively we may say that the challenge to ellipticity in the Lorentzian setting is spatial gradients in the direction of motion of the horizon. However for a spacetime to be stationary there can be no such gradients as is formalised in the Rigidity theorems. If there were, radiation would be emitted in contradiction to stationarity. Motivated by Rigidity we consider a broad class of stationary spacetimes which are given by a fibration of the orbits of the stationary Killing vector, together with orbits of other mutually commuting Killing vectors that generate rotation or translation and also commute with the stationary field. We then require that the Killing vector normal to a rotating horizon is some linear combination of these vector fields. This is the same class considered by Harmark in order to classify higher dimensional black holes [@Harmark:2009dh]. Further motivated by the stationary uniqueness theorems we make the key assumption that the base of this fibration is a Riemannian manifold with boundaries at the horizon and any axes of the commuting rotational symmetries. It is then straightforward to see that for a reference metric of the same form, the Harmonic Einstein equation consistently truncates to this class of stationary spacetimes. We provide the necessary boundary conditions at the Killing horizons or axes of symmetry and demonstrate they are consistent with obtaining vacuum solutions of the Einstein equations rather than only Ricci solitons. The Ricci-DeTurck flow is then parabolic on this class of Lorentzian stationary spacetimes, and gives an explicit algorithm to solve it, as does Newton’s method.
In order to demonstrate that these methods may be applied in practice we have considered a very simple example where analytic solutions are not known, namely a 4$D$ rotating black hole in a spherical cavity. We have used Anderson’s boundary conditions at the cavity wall, noting that the Dirichlet conditions (i.e. fixed induced metric) often discussed in the static spherically symmetric case are not well posed. Solutions were found using Newton’s method. Ricci flows of these Lorentzian stationary black hole spacetimes were also performed, and their use demonstrated to construct solutions. We believe these are the first example of Lorentzian Ricci flows that have been studied.
Of course this toy example is only of cohomogeneity two and is 4 dimensional. The main purpose of this paper is precisely to give an elegant geometric framework that applies beyond these cases, and it will be interesting in future work to test these methods in the many physically interesting situations where stationary black holes have been conjectured to exist.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank Pau Figueras and James Lucietti for many illuminating and invaluable discussions. We would also like to thank Michael Anderson, Barak Kol, Luis Lehner and Eric Woolgar for useful discussions and correspondence. Furthermore we would like to thank Matthew Headrick for many important discussions and initial collaboration on this work. AA and SK are supported by STFC studentships. TW is supported by an STFC advanced fellowship and Halliday award.
Boundary conditions for horizons and axes {#app:bc}
=========================================
Our elliptic problem is posed on a base $\mathcal{M}$ which has boundaries associated to Killing horizons and axes of symmetry. We must explicitly give boundary conditions that ensure regularity for these. In order to derive the necessary conditions, we perform a similar calculation to that in section \[sec:static\] where we used the smooth Euclidean continuation of a static black hole to deduce boundary conditions for the Lorentzian problem.
We are interested primarily in regularity of the metric. However we have argued that in the static Lorentzian case, as is clear from the Euclidean picture, that the Ricci-DeTurck tensor also shares the same regularity. Therefore we perform the analysis below for a general $(0,2)$ tensor $J$ which one can take to be the metric or Ricci-DeTurck tensor.
The procedure is best illustrated by a simple example. Consider a smooth $(0,2)$ tensor, $J$, which is symmetric with respect to a vector field $R$ that generates $U(1)$ orbits with period $2\pi$ and has fixed action at some point $p$. We will then derive regularity conditions for the components of $J$ in a chart with ‘polar’ coordinates $(r, \alpha)$ adapted to the symmetry so that $R = {\frac{\partial }{\partial \alpha}}$, and hence the components do not depend on $\alpha$.
We take $\alpha$ to have period $2 \pi$, and the fixed point of $R$ is $r=0$. To proceed the tensor $J$ is first written in ‘Cartesian coordinates’ $(a,b)$ that do not manifest the $U(1)$ symmetry, but where $J$ has components which are $ \mathbb{C}^{ \infty}$ smooth everywhere, including the fixed point; $$\begin{aligned}
\label{eq:cart}
J = N(a,b) da^2 + M(a,b) db^2 + K(a,b) da db\end{aligned}$$ We then introduce the ‘polar coordinates’ $r, \alpha$ that make explicit the $U(1)$ symmetry as, $$\begin{aligned}
a = r \sin \alpha \, , \;b = r \cos \alpha\end{aligned}$$ Let us write $J$ in this polar chart as, $$\begin{aligned}
\label{eq:polar}
ds^2 = r^2 A(r) d\alpha^2 + B(r) dr^2 + r^3 \, C(r) dr \, d\alpha\end{aligned}$$ where the metric functions $A,B,C$ now depend only on $r$. The symmetry conditions,$ \frac{ \partial A}{ \partial \alpha} = \frac{ \partial B}{ \partial \alpha}=\frac{ \partial C}{ \partial \alpha} =0$ then translate into conditions on the original metric functions $N, M$ and $K$ and evaluation of these gives the regularity conditions, $N= M$ and $K=0$ at the fixed point $a = b = 0$. When expressed in terms of the polar metric functions these conditions become, $(A - B)|_{r=0} =0$. Consider the function $A$ which can be given in terms of the Cartesian components as, $$\begin{aligned}
A = \frac{1}{a^2 + b^2} \left( a^2 M + b^2 N - a b K \right)\end{aligned}$$ As a consequence of the above regularity conditions, we see that $A$ is a smooth function of $a,b$. Since it is only a function of $r$, and $r^2 = a^2 + b^2$, it follows that $A$ is a smooth function of $r^2$. Similarly one finds $B, C$ are also smooth in $r^2$.
Thus we conclude that if we write the tensor $J$ is the polar chart as in , we may think of $r = 0$ as a boundary, where we impose that $A,B,C$ are smooth in $r^2$ there, and $(A - B)|_{r=0} = 0$.
Regularity and smoothness at a Killing horizon
----------------------------------------------
A Killing horizon implies the existence of a normal Killing field $K$, whose isometry group is $\mathbb{R}$, with a fixed point at the bifurcation surface, and whose orbits close on the future and past horizons. We consider a smooth $(0,2)$ tensor that is symmetric under $K$, in a chart adapted to the symmetry which covers the exterior of the Killing horizon. The fixed point may be regarded as a boundary of this chart, and we determine regularity conditions on the tensor components there.
We begin in smooth Cartesian coordinates with a $(0,2)$ tensor, $J$, written in components as, $$\begin{aligned}
\label{eq:Carthoriz}
J &=& N da^2 + M db^2 + U da db + Q_{ {i}} da d {x}^{ {i}} + R_{ {i}} db d {x}^{ {i}} + T_{ {i} {j}} d {x}^{ {i}} d {x}^{ {j}}\end{aligned}$$ where $ {i}=1,...,D-2$. We take a Killing horizon with respect to the Killing vector $K$ to be located at $a = b$ and $a = -b$ with bifurcation surface $a = b =0$. Since $J$ is smooth at the horizon these component functions are $C^\infty$ in the neighbourhood of the horizon. The horizon Killing symmetry is not manifest in these coordinates and in analogy with the toy example, we now change to hyperbolic coordinates, $$\begin{aligned}
a=r \sinh \kappa t \, , \quad b=r \cosh \kappa t\end{aligned}$$ so that $K = \partial / \partial t$ and $r = 0$ is the bifurcation surface, with $\kappa$ a constant related to the normalization of $K$ and giving the surface gravity. We write the metric in this polar form as, $$\begin{aligned}
\label{eq:Polarhoriz}
J &=& - r^2 A d t^2 + B dr^2 + r^3 C dr d t + r F_{ {i}} dr d {x}^{ {i}} + r^2 G_{ {i}} d t d {x}^{ {i}} + T_{ {i} {j}} d {x}^{ {i}} d {x}^{ {j}}\end{aligned}$$ where the component functions are independent of $t$. Repeating the analysis outlined in the toy example one arrives at the conclusion that the functions $A, B, C, F_{ {i}}, G_{ {i}}, T_{ {i} {j}}$ depend smoothly on $r^2$ and ${x}^{{i}}$, together with the regularity condition, $$\begin{aligned}
A|_{r=0} &=& \kappa^2 B|_{r=0}\end{aligned}$$ Thus we see explicitly that the regularity in the chart \[eq:Polarhoriz\] depends on the normalization of $K$, and hence the surface gravity. If we take the tensor $J$ to be the metric, we deduce the regularity conditions on the metric at the Killing horizon. Taking $J$ to be the Ricci-DeTurck tensor we see the behaviour it will exhibit if it shares the symmetry and is regular.
Axis of rotation
----------------
We now consider a $(0,2)$ tensor which is symmetric under a Killing field $R$ which generates rotation about an axis with period $2 \pi$. In a chart which manifests the symmetry the axis is fixed under the $U(1)$ action, and may be regarded as a boundary for the chart. We determine the regularity conditions for the components in this chart there. This case is very close to the toy example before.
We begin with a Cartesian line element of the form $$\begin{aligned}
\label{eq:Cartaxis}
J &=& N da^2 + M db^2 + U da db + Q_{ {i}} da d {x}^{ {i}} + R_{ {i}} db d {x}^{ {i}} + T_{ {i} {j}} d {x}^{ {i}} d {x}^{ {j}}\end{aligned}$$ and the component functions depend smoothly on $a,b, {x}^{ {i}}$ in the neighbourhood of the axis which we take to be $a = b = 0$. We now change to polar coordinates defined by, $$\begin{aligned}
a=r \sin \alpha \, , \quad b= r \cos \alpha\end{aligned}$$ where $ \alpha$ has period $ 2 \pi$ and $R = \partial / \partial \alpha$ and $r = 0$ is the axis. In these coordinates we write the tensor as, $$\begin{aligned}
\label{eq:Polaraxis}
J &=& r^2 A d \alpha^2 + B dr^2 + r^3 C dr d \alpha + r F_{ {i}} dr d {x}^{ {i}} + r^2 G_{ {i}} d \alpha d {x}^{ {i}} + T_{ {i} {j}} d {x}^{ {i}} d {x}^{ {j}}\end{aligned}$$ and the symmetry is manifest so the metric functions are independent of $\alpha$. Repeating the analysis outlined in the toy example, one finds the metric functions $A, B, C, F_{ {i}}, G_{ {i}}, T_{ {i} {j}}$ are *smooth* functions of $ r^2$ and ${x}^{ {i}}$, together with the regularity condition, $$\begin{aligned}
A|_{ r=0} &=& B|_{ r=0}\end{aligned}$$
Connection Components and Flow Equations {#app:equations}
========================================
In this appendix, we give the connection components of the metric \[eq:bhansatz\] together with the components of the Ricci tensor and $\xi$ vector. The Christoffel symbols are given by, $$\begin{aligned}
\Gamma^i_{jk} &=& \hat{ \Gamma}^i_{jk} + \frac{1}{2}h^{im}A_{Aj}F_{km}^{A}+ \frac{1}{2}h^{im}A_{Ak}F_{jm}^{A}- \frac{1}{2}h^{im}A^{A}_{k}A^{C}_{j} \partial_{m} G_{AC} \nonumber \\
\Gamma^i_{AB}&=& - \frac{1}{2} h^{ij}\partial_{j} G_{AB} \nonumber\\
\Gamma^A_{Bi} &=& -\frac{1}{2} A^{Aj} G_{BC} F_{ij}^{C} + \frac{1}{2} A^{Aj}A_{i}^{C} \partial_{j} G_{BC} + \frac{1}{2}G^{AC} \partial_{i} G_{BC} \nonumber\\
\Gamma^A_{ij} &=& -A^{A}_{m} \hat{ \Gamma}^m_{ij} + \frac{1}{2}A^{Ak}A_{Bi}F_{kj}^{B} + \frac{1}{2} A^{Ak}A_{Bj}F_{ki}^{B} + \partial_{(j} A_{i)}^{A} + G^{AB} A_{(i}^{C} \partial_{j)} G_{BC} + \frac{1}{2} A^{Ak} A^{B}_{j} A^{D}_{i} \partial_{k} G_{DB} \nonumber\\
\Gamma^A_{BC}&=& \frac{1}{2} A^{Ai} \partial_{i} G_{BC} \nonumber\\
\Gamma^i_{jA} &=& - \frac{1}{2}h^{ik} A_{j}^{B} \partial_{k} G_{AB} + \frac{1}{2}h^{ik} G_{AB} F_{jk}^{B} \end{aligned}$$ where $ \hat{ \Gamma}^i_{jk}$ is the Christoffel connection of the ’submetric’ $h_{ij}$ and $F^{A}_{ij} \equiv \partial_i A_{j}^{A}- \partial_j A_{i}^{A} = \hat{\nabla}_{i} A_{j}^{A} - \hat{\nabla}_{j} A_{i}^{A} $. The covariant derivative in the latter equation, $ \hat{ \nabla}_{i}$, is defined with respect to the connection $\hat{ \Gamma}^i_{jk}$ of $h_{ij}$ (and is metric compatible with respect to $h_{ij}$). Using these results one finds for the decomposition of the Ricci tensor, $$\begin{aligned}
R_{AB}&= \frac{1}{2} h^{ij} \hat {\nabla}_{i} (\partial_{j} G_{AB})- \frac{1}{4} G^{CD} h^{ip} (\partial_{p} G_{AB})( \partial_{i} G_{CD}) \\
& + \frac{1}{2} h^{ij} G^{CD} ( \partial_{j} G_{CB})( \partial_{i} G_{AD})+ \frac{1}{4}h^{mi} h^{jp} G_{BE} G_{AF} F_{mj}^{E} F_{ip}^{F}\end{aligned}$$ $$\begin{aligned}
R_{iA}-R_{AB} A^{B}_{i}&= \frac{1}{2} h^{jk} G_{AB} \hat{ \nabla}_{j} F_{ik}^{B}+ \frac{1}{2} h^{jk} F_{ij}^{B} \partial_{k} G_{AB} + \frac{1}{4} h^{jm} G^{CD} G_{AB} F_{im}^{B} \partial_{j} G_{CD} \end{aligned}$$ $$\begin{aligned}
R_{ij}+R_{AB}A^{A}_{i} A^{B}_{j}- R_{Ai} A^{A}_{j} - R_{Aj} A^{A}_{i} &= \hat{R}_{ij}- \frac{1}{2} G^{CB} \hat{ \nabla}_{j}( \partial_{i} G_{CB}) \\ &
+ \frac{1}{4} G^{CD} G^{BA} (\partial_{j} G_{DA})( \partial_{i} G_{CB}) + \frac{1}{2} h^{km} G_{AB} F_{jm}^{A} F_{ki}^{B} \end{aligned}$$ where $ \hat{R}_{ij}$ is the Ricci tensor computed with respect to $ \hat{ \Gamma}^i_{jk}$. The DeTurck vector $\xi^{ \mu} = g^{ \lambda \nu}( \Gamma^{ \mu}_{ \lambda \nu}- \bar{ \Gamma}^{ \mu}_{ \lambda \nu})$ decomposes as, $$\begin{aligned}
\xi^{k}&=& \hat{ \xi}^{k} - \frac{1}{2} G^{AB}h^{km} \partial_{m} G_{AB} + \frac{1}{2} G^{AB} \bar{h}^{km}_{(-1)} \partial_{m} \bar{G}_{AB} + h^{ij} \bar{h}^{km}_{(-1)} \bar{G}_{AB} (A_{i}^{A}-\bar{A}_{i}^{A}) \bar{F}_{jm}^{B} \nonumber\\
& & + \frac{1}{2} h^{ij} \bar{h}^{km}_{(-1)} (A^{A}_{j} A^{B}_{i} + \bar{A}^{A}_{j} \bar{A}^{B}_{i} - 2 A^{A}_{j} \bar{A}^{B}_{i}) \partial_{m} \bar{G}_{AB} \nonumber\\
\xi^{C}&=& \frac{1}{2}G^{AB} A^{Cj} \partial_{j} G_{AB} - \frac{1}{2}G^{AB} \bar{A}^{Cj} \partial_{j} \bar{G}_{AB}+ h^{ij} ( \hat{\nabla}_{i} A_{j}^{C} - \bar{ \hat{\nabla}}_{i} \bar{A}_{j}^{C}) \nonumber \\
& & - \frac{1}{2} h^{ij} \bar{A}^{Ck} (A^{A}_{j} A^{B}_{i} + \bar{A}^{A}_{j} \bar{A}^{B}_{i} - 2 A^{A}_{j} \bar{A}^{B}_{i}) \partial_{k} \bar{G}_{AB} + h^{ij} \bar{A}^{Ck} \bar{G}_{AB}(A^{A}_{i} - \bar{A}^{A}_{i}) \bar{F}^{B}_{kj} \nonumber\\
&& + h^{ij} \bar{G}^{CB} (A^{A}_{i} - \bar{A}^{A}_{i}) \partial_{j} \bar{G}_{AB} \end{aligned}$$ where $ \hat{ \xi}^{k} = h^{ij}( \hat{ \Gamma}^{k}_{ij}- \bar{ \hat{ \Gamma}}^{k}_{ij})$ and as usual, an overbar indicates that the quantity in question is evaluated in the reference metric.
As discussed in the main text, the flow equations for the various metric components of interest decompose as, $$\begin{aligned}
\frac{\partial G_{AB}}{ \partial \lambda} &=& -2 R_{AB} + 2 \nabla_{(A} \xi_{B)} \\
\frac{\partial A_{j}^{C}}{ \partial \lambda} &=& -2 G^{AC}(R_{jA} - R_{AB} A^{B}_{j}) + 2 G^{AC}( \nabla_{(A} \xi_{j)} - \nabla_{(A} \xi_{B)} A^{B}_{j} ) \\
\frac{\partial h_{ij}}{ \partial \lambda} &=& -2(R_{ij} + R_{AB} A_{i}^{A} A_{j}^{B} - R_{iA} A^{A}_{j} - R_{jA} A^{A}_{i}) \\ && + 2 ( \nabla_{(i} \xi_{j)} + \nabla_{(A} \xi_{B)} A_{i}^{A} A_{j}^{B} - \nabla_{(B} \xi_{i)} A_{j}^{B} - \nabla_{(B} \xi_{j)} A_{i}^{B} )\end{aligned}$$
This form is particularly useful as the linear combinations of the components of $ \nabla_{( \mu} \xi_{ \nu )}$ that arise take a relatively simple form. Explicitly one finds that, $$\begin{aligned}
2 \nabla_{(A} \xi_{B)} &= \hat{\xi}^{k} \partial_{k} G_{AB} - \frac{1}{2} G^{CD} (\partial^{k}G_{CD}) (\partial_{k} G_{AB}) + \frac{1}{2} G^{CD} \bar{h}^{km}_{(-1)} (\partial_{m} \bar{G}_{CD})(\partial_{k} G_{AB}) \\ & +\bar{h}^{km}_{(-1)} \bar{G}_{CD} \bar{F}^{D}_{jm} (A^{jC} - \bar{A}^{jC}) \partial_{k} G_{AB} \\ & + \frac{1}{2} \bar{h}^{km}_{(-1)} (A^{iC} A_{i}^{D} + \bar{A}^{iC} \bar{A}_{i}^{D} - 2 A^{iC} \bar{A}_{i}^{D}) (\partial_{m} \bar{G}_{CD})( \partial_{k} G_{AB})\end{aligned}$$ $$\begin{aligned}
2 (\nabla_{(i} \xi_{A)} &- \nabla_{(A} \xi_{B)} A^{B}_{i}) = G_{AC} \hat{\nabla}_{i}(A_{k}^{C} \hat{\xi}^{k}) + G_{AC} \hat{\nabla}_{i}(\hat{\nabla}^{p} A^{C}_{p} - \bar{\hat{\nabla}}^{p} \bar{A^{C}_{p}}) \\ & + G_{AC} \hat{\nabla}_{i} \left( \left(\frac{1}{2}\bar{h}^{mp}_{(-1)} G^{DE} (A^{C}_{m}- \bar{A}^{C}_{m}) + \bar{G}^{CE}(A^{pD}- \bar{A}^{pD}) \right. \right. \nonumber \\
& \left. \left. + \frac{1}{2} \bar{h}^{kp}_{(-1)} (A^{mD} A^{E}_{m} + \bar{A}^{mD} \bar{A}^{E}_{m} - 2 A^{mD} \bar{A}^{E}_{m})(A^{C}_{k} - \bar{A}^{C}_{k})\right) \partial_{p} \bar{G}_{DE} \right) \nonumber \\
& + G_{AC} \hat{\nabla}_{i}(\bar{h}^{km}_{(-1)} \bar{G}_{DE} (A^{jE} - \bar{A}^{jE})(A^{C}_{k} - \bar{A}^{C}_{k}) \bar{F}_{jm}^{D}) \nonumber \\ & + G_{AC} \hat{\xi}^{k} F_{ki} ^{C} - \frac{1}{2} G_{AC} G^{DE} F_{ki} ^{C} ( \partial^{k} G_{DE}) \nonumber \\ & + \frac{1}{2} \bar{h}^{km}_{(-1)}G_{AC} G^{DE} F_{ki}^{C} \partial_{m} \bar{G}_{DE} \nonumber \\ & - \bar{h}^{km}_{(-1)} G_{AC} \bar{G}_{DE} (A^{jD} - \bar{A}^{jD}) F_{ik}^{C} \bar{F}_{jm}^{E} \nonumber \\
& + \frac{1}{2} \bar{h}^{km}_{(-1)}G_{AC} (A^{pD} A^{E}_{p} + \bar{A}^{pD} \bar{A}^{E}_{p}- 2 A^{pD} \bar{A}^{E}_{p}) F_{ki}^{C} \partial_{m} \bar{G}_{DE}\end{aligned}$$ $$\begin{aligned}
2 ( \nabla_{(i} \xi_{j)} &+ \nabla_{(A} \xi_{B)} A_{i}^{A} A_{j}^{B} - \nabla_{(B} \xi_{i)} A_{j}^{B} - \nabla_{(B} \xi_{j)} A_{i}^{B} ) = 2 \hat{\nabla}_{(i} \hat{\xi}_{j)} + G^{AB} G^{CD} ( \partial_{i} G_{CB})( \partial_{j} G_{AD}) \nonumber \\
& - G^{AB} \hat{\nabla}_{i} ( \partial_{j} G_{AB}) + \left ( \frac{1}{2} h_{ik} \hat{ \nabla}_{j} (G^{AB} \bar{h}^{km}_{(-1)} \partial_{m} \bar{G}_{AB}) + h_{ik} \hat{ \nabla}_{j}( \bar{h}^{km}_{(-1)} \bar{G}_{AB} \bar{F}^{B}_{qm} (A^{qA} - \bar{A}^{qA})) \right. \nonumber \\
& \left. + \frac{1}{2} h_{ik} \hat{\nabla}_{j}( \bar{h}^{km}_{(-1)} (A^{pA} A^{B}_{p} + \bar{A}^{pA} \bar{A}^{B}_{p} - 2 A^{pA} \bar{A}^{B}_{p}) \partial_{m} \bar{G}_{AB}) + (i \leftrightarrow j ) \right)\end{aligned}$$ where we note that in these latter three expressions, all ’A term’ base indices have been contracted with the base metric $h_{ij}$ as appropriate. Using these results, one arrives at the flow equations in the main body of the paper, contracted in the same manner.
[^1]: Note that in the previous work [@KitchenHeadrickTW; @PFLuciettiTW] the ‘Harmonic Einstein’ equation with DeTurck gauge fixing was referred to as the ‘Einstein-DeTurck’ equation. We will use the ‘Harmonic Einstein’ terminology here to avoid potential confusion with the ‘Ricci-DeTurck’ flow.
[^2]: There may be special points in moduli space of Ricci flat solutions where a normalizable zero mode of the linearisation of the Harmonic Einstein equation exists, and this lifts to the non-linear equation to generate a deformation to a soliton solution. If a branch of solitons meets a branch of Ricci flat solutions this would be the case.
[^3]: Interestingly in the seemingly unrelated context of effective field theory used in the context of GR the utility of such a Kaluza-Klein ansatz has been emphasized over the ADM one [@KolCLEFT].
[^4]: In the static case we have the inequality $\nabla^2 \phi + \xi^\mu \partial_\mu \phi = \nabla_{\mu} \xi_\nu \nabla^\mu \xi^\nu > 0$ where $\phi = | \xi |^2$. However in the stationary case this inequality does not appear to hold as in Lorentzian signature with only stationary symmetry $\nabla_{\mu} \xi_\nu \nabla^\mu \xi^\nu$ is not of definite sign.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we introduce Stochastic Gradient Langevin Boosting (SGLB) — a powerful and efficient machine learning framework, which may deal with a wide range of loss functions and has provable generalization guarantees. The method is based on a special form of Langevin Diffusion equation specifically designed for gradient boosting. This allows us to guarantee the *global* convergence, while standard gradient boosting algorithms can guarantee only local optimum, which is a problem for multimodal loss functions. To illustrate the advantages of SGLB, we apply it to a classification task with 0-1 loss function, which is known to be multimodal, and to a standard Logistic regression task that is convex. The algorithm is implemented as a part of the CatBoost gradient boosting library and outperforms classic gradient boosting methods.'
bibliography:
- 'example\_paper.bib'
---
Introduction
============
Gradient boosting is a powerful machine-learning method that iteratively combines weak models to obtain more accurate ones [@friedman2001]. Nowadays, this technique remains the primary method for web search, recommendation systems, weather forecasting, and many other problems with complex dependencies and heterogeneous data. Combined with decision trees, gradient boosting underlies such widely-used software libraries like, e.g., XGBoost [@Chen:2016], LightGBM [@LightGBM], and CatBoost [@catboost].
For convex loss functions and under some regularity assumptions, stochastic gradient boosting (SGB) converges to the optimal solution. However, even local optima cannot be guaranteed for general losses. We fill this gap and build a *globally* convergent gradient boosting algorithm for convex and non-convex optimization with provable generalization guarantees. For this purpose, we combine gradient boosting with stochastic gradient Langevin dynamics (SGLD), which is a powerful iterative optimization algorithm [@DBLP:journals/corr/RaginskyRT17]. It turns out that gradient boosting can be easily modified to a globally convergent method: at each step, one has to shrink the currently built model and add a proper noise to stochastic gradient estimates.
We implemented the proposed algorithm as a part of the CatBoost gradient boosting library. Our experiments show that the obtained method outperforms classic SGB. Of particular importance is the ability to optimize globally and with provable guarantees such non-convex losses as 0-1 loss, which was previously claimed to be a challenge [@nguyen2013algorithms]. CatBoost team implemented our method as a part of the official library [@CatBoost_code], which means that SGLB is already acknowledged as a new SGB benchmark.
The rest of the paper is organized as follows. In the next section, we briefly discuss the related research on gradient boosting convergence and 0-1 loss optimization. Then, in Section \[sec:background\], we give the necessary background on gradient boosting and stochastic gradient Langevin dynamics. The proposed SGLB algorithm is described and analyzed in Section \[sec:sglb\], our particular CatBoost implementation is given in Section \[sec:catboost\]. The experiments comparing SGLB with SGB are discussed in Section \[sec:experiments\]. Section \[sec:conclusion\] concludes the paper and outlines promising directions for future research.
Related work {#sec:related}
============
Convergence of gradient boosting
--------------------------------
There are several theoretical attempts to study convergence properties of SGB like, e.g., Boulevard [@boulevard], AnyBoost [@anyboost], or gradient boosting in general $L_2$ setting [@doi:10.1198/016214503000125; @zhang2005; @2008arXiv0804.2752B]. These works consider a general boosting algorithm, but under restrictive assumptions like exact greediness of the weak learners’ selection algorithm [@anyboost], Structure Value Isolation properties [@boulevard], and, most importantly, all of them exploit convexity.
Even though convexity may seem a natural assumption, many practical tasks involve non-convex losses like 0-1 loss optimization [@nguyen2013algorithms], regret minimization in non-convex games [@DBLP:journals/corr/abs-1708-00075], Learning-to-Select-With-Order [@vorobev2019learning]. Thus, existing frameworks fail to efficiently solve these tasks as they strongly rely on convexity that is absent.
Many practical implementations of boosting like XGBoost [@Chen:2016], LightGBM [@LightGBM], and CatBoost [@catboost] use constant learning rate as in practice it outperforms dynamically decreasing ones. However, existing works on the convergence of boosting algorithms assume decreasing learning rates [@boulevard; @zhang2005], thus leaving an open question: if we assume a constant one $\epsilon > 0$, can convergence be guaranteed if we let $\epsilon \rightarrow 0_+$?
0-1 loss optimization
---------------------
For binary classification problems, convex loss functions are usually used since they can be efficiently optimized. However, as pointed out by @nguyen2013algorithms, such losses are sensitive to outliers. On the other hand, 0-1 loss (the fraction of errors) is more robust and more interpretable, but harder to optimize. @nguyen2013algorithms propose smoothing 0-1 loss with sigmoid function and show that an iteratively unrelaxed coordinate descent approach for gradient optimization of this smoothed loss outperforms optimization of convex upper bounds of the original 0-1 loss. In the current paper, we use a smoothed 0-1 loss as an example of a multimodal function and show that the SGLB algorithm achieves superior performance for this loss.
Background {#sec:background}
==========
General setup {#sec:setup}
-------------
For convenience, we list the notation frequently used in the paper in Table \[tab:notation\]. Assume that we are given some distribution $\mathcal{D}$ on $\mathcal{X}\times \mathcal{Y}$, where $\mathcal{X}$ is a feature space (typically $\mathbb{R}^k$) and $\mathcal{Y}$ is a target space (typically $\mathbb{R}$ for regression or $\{0, 1\}$ for classification). We are also given a loss function $L(z, y):\mathcal{Z}\times\mathcal{Y}\rightarrow \mathbb{R}$, where $\mathcal{Z}$ is a space of predictions (typically $\mathbb{R}$ or $\{0,1\}$). Our ultimate goal is to solve the following problem: $$\label{true-target}
\mathcal{L}(f|\mathcal{D}) := \mathbb{E}_{\mathcal{D}} L(f(x), y) \rightarrow \min_{f\in\mathcal{F}}$$ with respect to a family of functions $\mathcal{F} \subset \{f:\mathcal{X}\rightarrow\mathcal{Z}\}$.
In practice, the distribution $\mathcal{D}$ is unknown and we are given i.i.d. samples $(x_1, y_1), \ldots, (x_N, y_N) \sim \mathcal{D}$ denoted as $\mathcal{D}_N$, so the expectation in is replaced by the empirical average: $\mathcal{L}_N(f) := \mathcal{L}(f|\mathcal{D}_N) = \frac{1}{N}\sum_{i=1}^n L(f(x_i), y_i).$ Typically, to improve generalization, one adds a strongly convex regularization function $\Omega(f|\mathcal{D}) = \mathbb{E}_{x\sim\mathcal{D}}\Omega(f(x))$ with parameter $\gamma > 0$: $$\label{empirical-target-1}
\mathcal{L}(f|\mathcal{D}_N) + \gamma\,\Omega(f|\mathcal{D}_N) \rightarrow \min_{f\in \mathcal{F}}\,.$$ We consider only $\mathcal{F}$ corresponding to finite ensembles of weak learners $\mathcal{H} := \{h^s(x|\theta^s):\mathcal{X}\times \mathbb{R}^{m_s}\rightarrow \mathbb{R}|s\in S\}$, where $S$ is some index set that may be infinite and $h^s$ depends linearly on $\theta^s$. The latter condition is not only typical for practical implementations like CatBoost, XGBoost, and LightGBM, but is also crucial for our analysis as it allows us to claim the convergence.
Due to the linear dependence of $h^s$ on $\theta^s$, if $|S| < \infty$, we can represent any ensemble in $\mathcal{H}$ as a linear model $f(x|\theta) = \langle\phi(x), \theta\rangle_2$ for some feature map $\phi(x):\mathcal{X}\rightarrow \mathbb{R}^m$, where $\theta\in \mathbb{R}^m$ encodes all $\theta^s\in\mathbb{R}^{m_s}\,\forall s\in S$. If $|S| = \infty$ such representation also exists since ensembles are finite, so the infinite-dimensional vector $\theta$ would have only finite number of non-zero components, so $\langle \phi(x), \theta\rangle$ exists, linear, and correctly defined.
0.15in
Variable Description
----------------------------------------------------------------- -------------------------------------------------------
$x \in \mathcal{X}$ Features, typically from $\mathbb{R}^k$
$y \in \mathcal{Y}$ Target, typically from $\mathbb{R}$ or $\{0, 1\}$
$z \in \mathcal{Z}$ Prediction, typically from $\mathbb{R}$
$L(z,y):\mathcal{Z}\times\mathcal{Y}\rightarrow\mathbb{R}$ Loss function
$h^s(x|\theta^s) \in \mathcal{H}$ Weak learner
$f(x|\theta):\mathcal{X}\times \Theta \rightarrow \mathcal{Z} $ Model parametrized by $\theta$
$\mathcal{D}$ Data distribution over $\mathcal{X}\times\mathcal{Y}$
$\mathcal{D}_N=\{(x_i, y_i)\}_{i=1}^N$ I.i.d. samples from $\mathcal{D}$
$\mathcal{L}(\theta|\mathcal{D})$ Expected loss w.r.t. $\mathcal{D}$
$F(t)$ Markov Process $\big(f(x_i|\theta(t))\big)_{i=1}^N$
$P$ Symmetric preconditioner matrix
$\Gamma= \sqrt{P^{-1}}$ Regularization matrix
$\mathcal{L}_N({F},\gamma)$ Implicitly regularized loss
$t\in [0, +\infty)$ Continuous time
$\tau\in \mathbb{Z}_+$ Discrete time
$\widehat{F}_\tau$ Predictions’ Markov Chain
$\epsilon > 0$ Learning rate
$\beta > 0$ Diffusion temperature
$\gamma > 0$ Regularization parameter
$I_m \in \mathbb{R}^{m\times m}$ Identity matrix
$\mathbb{0}_m \in \mathbb{R}^{m}$ Zero vector
: Notation.[]{data-label="tab:notation"}
-0.1in
Stochastic Gradient Boosting {#sec:sgb}
----------------------------
A typical stochastic gradient boosting (SGB) algorithm [@friedman2002stochastic] is a recursive procedure that can be characterized by a triplet $\mathcal{B} := \big(\mathcal{H},\mathcal{S}, \mathcal{P}\big)$, where $\mathcal{H}$ is a set of weak learners, $\mathcal{S}$ is a sampling strategy that takes ensemble predictions $\widehat{F}_\tau:=\big(f(x_i|\widehat{\theta}_\tau)\big)_{i=1}^N$, where $\widehat{\theta}_\tau$ encodes ensemble’s weak learners parameters, and returns a weak learner index $s_\tau$, and $\mathcal{P}$ is a procedure that takes the weak learner $h^{s_\tau}$ and stochastic gradient estimates $\widehat{g}_\tau \in \mathbb{R}^N$, such that $\mathbb{E}\widehat{g}_\tau = \big(\frac{\partial}{\partial f}L(f(x_i|\widehat{\theta}_\tau), y_i)\big)_{i=1}^N$, and returns the optimal parameters $\theta^{s_\tau}_*$ according to the following problem: $$\begin{split}\label{eq:gb-wl-estimation}
&\text{minimize }\|W_{s_\tau}\theta^{s_\tau}\|_2^2\,\,\,\,\, \text{ subject to}\\
&\theta^{s_\tau}\in \operatorname*{arg\,min}_{\theta^{s_\tau}\in\mathbb{R}^{m_s}}\frac{1}{N}\sum_{i=1}^N (-\epsilon\widehat{g}_{\tau,i} - h^{s_\tau}(x_i|\theta^{s_\tau}))^2\,,
\end{split}$$ where $\epsilon > 0$ is a learning rate and $W_{s_\tau}$ is a matrix of weights typically picked as $I_{m_s}$, but in general we only assume that $\mathrm{rk } W_{s_\tau} = m_{s_\tau}$ in order to make the solution of to be unique. After obtaining $\theta^{s_\tau}_*$, the algorithm updates the ensemble as $f(\cdot|\widehat{\theta}_{\tau+1}) := f(\cdot|\widehat{\theta}_{\tau}) + h^{s_\tau}(\cdot|\theta^{s_\tau}_*)$ and, informally speaking, $\widehat{\theta}_{\tau+1} := (\widehat{\theta}_{\tau}, \theta_*^{s_\tau})$ so that $\widehat{\theta}_{\tau+1}$ encodes all $\theta^{s_\tau}_*$ from the previous iterations.
dataset $\mathcal{D}_N$, learning rate $\epsilon > 0$ $\tau = 0$ $f(\cdot|\widehat{\theta}_{0}) = 0$ estimate $\widehat{g}_\tau\in \mathbb{R}^N$ on $\mathcal{D}_N$ using $f(\cdot|\theta_{\tau})$ sample $s_\tau \sim \mathcal{S}(-\epsilon\widehat{g}_\tau)$ estimate $\theta^{s_\tau}_* = \mathcal{P}_\tau(-\epsilon \widehat{g}_\tau)$ update $f(\cdot|\widehat{\theta}_{\tau + 1}) = f(\cdot|\widehat{\theta}_{\tau}) + h^{s_\tau}(\cdot|\theta^{s_\tau}_*)$ update $\tau = \tau + 1$ $f(\cdot|\widehat{\theta}_\tau)$
Under the convexity of $L(z, y)$ by $z$ and some regularity assumptions on the triplet [@boulevard], one can show that the ensemble converges to the optimal one with respect to the closure of the set of all possible finite ensembles, so one can construct a converging SGB algorithm for convex losses. However, many problems require the minimization of non-convex losses. In such cases, SGB cannot guarantee convergence to the optimal ensemble for the same reasons as stochastic gradient descent (SGD) — it gives only first order stationarity guarantee [@first-order-non-convex] that in the case of non-convex losses may mean not only local minima points but also saddles, which prevents effective optimization.
The goal of our paper is to fill this gap: we build a globally convergent gradient boosting algorithm for convex and non-convex optimization with provable (under some assumptions on $\mathcal{H}$) generalization gap bounds. Our analysis gives an insight into how the choice of $\mathcal{B}$ and its properties affects optimization and generalization.
Stochastic Gradient Langevin Dynamics
-------------------------------------
The core idea of our algorithm is to combine SGB described above with stochastic gradient Langevin dynamics, which we briefly introduce here.
Assume that we are given a function $U(\theta):\mathbb{R}^m \rightarrow \mathbb{R}$ such that $U(\theta) \rightarrow \infty$ as $\|\theta\|_2\rightarrow\infty$. The SGLD algorithm [@GelfandGAA; @Welling2011BayesianLV; @DBLP:journals/corr/RaginskyRT17; @Erdogdu:2018:GNO:3327546.3327636] aims at finding the global minimum of $U(\theta)$ and it updates $\widehat{\theta}_\tau$ as: $$\widehat{\theta}_{\tau + 1} = \widehat{\theta}_\tau - \epsilon P\widehat{\nabla}U(\widehat{\theta}_\tau) + \mathcal{N}(\mathbb{0}_m, 2\epsilon\beta^{-1} P),$$ where $\widehat{\nabla}U(\theta)$ is an unbiased stochastic gradient estimate (i.e., $\mathbb{E}\widehat{\nabla}U(\theta) = \nabla U(\theta)$), $\epsilon > 0$ is a learning rate, $\beta > 0$ is a diffusion temperature and $P \in \mathbb{R}^{m\times m}$ is a symmetric positive definite (PD) preconditioner matrix. Then, under mild assumptions [@DBLP:journals/corr/RaginskyRT17; @Erdogdu:2018:GNO:3327546.3327636], the chain $\widehat{\theta}_\tau$ converges in distribution to a random variable with density $p_{\beta}(\theta) \propto \exp(-\beta U(\theta))$ as $\epsilon\tau \rightarrow +\infty$, $\epsilon \rightarrow 0_+$. Moreover, according to [@DBLP:journals/corr/RaginskyRT17; @Erdogdu:2018:GNO:3327546.3327636], we have $\mathbb{E}_{\theta\sim p_\beta(\theta)}U(\theta) - \min_{\theta\in\mathbb{R}^m} U(\theta) = \mathcal{O}\left(\frac{m}{\beta}\log\frac{\beta}{m}\right)$, so the distribution $p_\beta(\theta)$ concentrates around the global optimum of $U(\theta)$.
The trick is to show that the interpolated process $\theta_\epsilon(t) := \widehat{\theta}_{[\epsilon^{-1}\tau]}$ weakly converges to the solution of the following *associated* Langevin Dynamics (LD) Stochastic Differential Equation (SDE) as $\epsilon \rightarrow 0_+$: $$\mathrm{d}\theta(t) = -P\nabla U(\theta(t))\mathrm{d}t + \sqrt{2\beta^{-1}P}\mathrm{d}W(t),$$ where $W(t)$ is a standard Wiener process. The measure $p_\beta(\theta)$ is a so-called invariant measure of the SDE, $\theta(t)$ is a solution of the SDE typically defined by the Ito integral.
The convergence of $\theta(t)$ to the invariant measure $p_\beta(\theta)$ in Wasserstein-2 metric is known to be exponential under mild conditions on Lipschitz smoothness of $U(\theta)$, i.e., for $t \gg 1$ we have $\theta(t) \sim p_{\beta}(\theta)$ with an error of order $\exp(-\Omega(t\beta^{-1}))$ (see @DBLP:journals/corr/RaginskyRT17 for the details).
Note that the convergence in Wasserstein-2 metric is equivalent to the weak convergence plus the convergence of the second moments [@DBLP:journals/corr/RaginskyRT17]. Henceforth, we only need to fix a large enough horizon $T \gg 1$, so that $\exp(-\Omega(T\beta^{-1}))$ is negligible, and use SGLD with small enough $\epsilon > 0$ with $\epsilon^{-1}T$ iterations to obtain, informally speaking, $\widehat{\theta}_{\epsilon^{-1}T} \sim \theta(T) \sim p_\beta(\theta)$. Then integration suboptimality bound applies giving us that $\widehat{\theta}_{\epsilon^{-1}T}$ should be an almost-minimizer of the function.
Stochastic Gradient Langevin Boosting {#sec:sglb}
=====================================
We further assume that the loss $L(z, y)$ is Lipschitz continuous and Lipschitz smooth by the variable $z$ and that $\inf_z L(z, y) > -\infty\,\forall y$. Since $\mathcal{L}_N({F})$ is a sum of several $L(\cdot, \cdot)$, it necessarily inherits all these properties. Also, from the stochastic gradient estimates $\widehat{g}_\tau$ we require $\|\widehat{g}_\tau - \mathbb{E}\widehat{g}_\tau\|_2 = \mathcal{O}(1)$ with probability one.
Before we proceed with the optimization of $\mathcal{L}_N(f(\cdot|\theta))$, we show how a possibly infinite-dimensional (if $|S| = \infty$) problem of boosting convergence can be reduced to the finite-dimensional one.
Induced Linear Structure
------------------------
Linear dependence of the weak learners on $\theta^s$ is important for boosting especially if $|S| = \infty$. To see that, we define the set: $$V_{\mathcal{B}} := \Big\{F=\big(f(x_i|\theta)\big)_{i=1}^N\Big|\forall \text{ finite ensembles }\theta\Big\}\subset \mathbb{R}^N\,.$$ This set encodes all possible predictions of all possible finite ensembles.
\[lemma=linearity\] The set $V_{\mathcal{B}}$ is a linear subspace in $\mathbb{R}^N$ for arbitrary weak learners set $\mathcal{H}$.
The proof can be obtained by noting that every finite ensemble has a finite-dimensional parameters vector, and the sum of any finite ensembles is again a finite ensemble. Due to the finite-dimensional nature of $\mathbb{R}^N$, every increasing sequence of subspaces should stabilize, and henceforth the whole image can be spanned by some *finite* subset of the weak learners. The latter immediately implies the following statement.
\[statement=consistency\] A possibly infinite-dimensional problem of learning an almost optimal finite ensemble can be *linearly* reduced to the finite-dimensional problem of finding optimal predictions ${F} = \big(f(x_i|\theta)\big)_{i=1}^N \in V_\mathcal{B}\subset \mathbb{R}^N$ in the subspace.
The space $V_\mathcal{B}$ and its dimension $\dim V_\mathcal{B}$ are invariant for $\mathcal{H}$ and, hence, for $\mathcal{B}$. Moreover, $\mathcal{H}\subset \mathcal{H}'$ for $\mathcal{H}\ne \mathcal{H}'$ does not imply $\dim V_\mathcal{B} < \dim V_{\mathcal{B}'}$, so the choice of the larger weak learners set does not necessarily imply an increased capacity.
Recall that $\mathcal{P}$ takes unbiased estimates of the gradients $\widehat{g}_\tau \in \mathbb{R}^{N}$ and an index $s_\tau \in S$ sampled according to the strategy $\mathcal{S}$ and estimates $\theta^{s_\tau}_*$ according to . After that, we add $h(x|\theta^{s_\tau}_*)$ to the ensemble. We denote such operation that takes $-\epsilon\widehat{g}_\tau$ and maps to optimal $\theta^{s_\tau}_*$ as $\mathcal{P}_{\tau}:\mathbb{R}^{N}\rightarrow \mathbb{R}^{m_{s_\tau}}$.
$\mathcal{P}_\tau$ is a linear operator.
Indeed, let $H_{s_\tau}: \mathbb{R}^{m_s} \to \mathbb{R}^N$ denote a linear operator converting $\theta^{s_\tau}$ to $(h^{s_{\tau}}(x_i|\theta^{s_\tau}))_{i=1}^N$. Then, we can rewrite as $$\begin{gathered}
\theta_*^{s_\tau} = \lim_{\delta \rightarrow 0} \operatorname*{arg\,min}_{\theta^{s_\tau}}\|-\epsilon\widehat{g}_{\tau,i} - H_{s_\tau}\theta^{s_\tau}\|_2^2 \\ + \delta^2 \langle W_{s_\tau}^T W_{s_\tau}\theta^{s_\tau}, \theta^{s_\tau} \rangle\,.\end{gathered}$$ Taking the derivative of the inner expression, we obtain: $$\left(H_{s_\tau}^T H_{s_\tau} + \delta^2 W_{s_\tau}^T W_{s_\tau} \right) \theta^{s_\tau} - \epsilon H_{s_\tau}^T \widehat{g}_{\tau,i} = 0$$ So, $\mathcal{P}_\tau$ can be defined as $\lim_{\delta\rightarrow 0} (H^T_{s_\tau} H_{s_\tau} + \delta^2 W^T_{s_\tau} W_{s_\tau})^{-1}H_{s_\tau}^T$. Such limit is well defined and is known as the weighted pseudo-inverse of the matrix [@Gulliksson2000].
This is a fundamental property of the gradient boosting that allows us to transform it into SGLB by simple injection of the noise directly into $\widehat{g}_\tau$ to obtain a globally convergent method as $\epsilon\rightarrow 0_+$.
Now consider $P_\tau := H_{s_\tau} \mathcal{P}_\tau$. Such operation, first, estimates the optimal $\theta_*^{s_\tau}$, and then converts them to the predictions of the weak learner. We have $P_\tau v \in V_\mathcal{B} \,\, \forall v\in \mathbb{R}^N$. The following lemma characterizes the structure of $P_\tau$.
\[lemma=projector\] $P_\tau$ is an orthoprojector on the image of the weak learner $h^{s_\tau}$, i.e., $P_\tau^2 = P_\tau$ and $\mathrm{im }P_\tau = \mathrm{im }H_\tau$.
The matrix $P_\tau$ is symmetric since $P_\tau = \lim_{\delta\rightarrow 0} H_{s_\tau}(H^T_{s_\tau} H_{s_\tau} + \delta^2 W^T_{s_\tau} W_{s_\tau})^{-1}H_{s_\tau}^T$, so it remains to prove that such operation is a projector.
Clearly, $\mathrm{im} P_\tau \subset \mathrm{im }H_\tau$. Observe that if $H_{s_\tau}\theta^{s_\tau} = v$, then $P_\tau v = v$, since the problem has an exact solution for $\operatorname*{arg\,min}$ subproblem, so $\mathrm{im} P_\tau = \mathrm{im }H_\tau$. Then, for an arbitrary $v\in \mathbb{R}^N$ one obtains immediately that $P_\tau (P_\tau v) = P_\tau v$ since $P_\tau v \in \mathrm{im } H_{s_\tau}$.
SGLB Algorithm
--------------
We are going to inject Gaussian noise directly into SGB gradients estimation procedure $\mathcal{P}$ to obtain a global convergence by showing that after a proper time interpolation, the process weakly converges to the Langevin dynamics. Therefore, for a large enough time horizon, we should obtain an almost-minimizer of the loss.
Before we inject the noise, we add regularization to the loss for two reasons: a) we *did not* assume that $\mathcal{L}_N(F) \rightarrow \infty$ as $\|F\|_2 \rightarrow \infty$, but we need to ensure this property for Langevin dynamics to guarantee the existence of the invariant measure; b) regularization is known to improve generalization. To ensure that loss diverges at infinity, we note that we assumed Lipschitz continuity, which in turn implies at most linear growth at infinity. Thus, regularizer must grow faster than linearly.
Natural regularization that grows faster than linearly is $L_2$ regularization, i.e., to add $\frac{\gamma}{2}\|F\|_2^2$ to $\mathcal{L}_N(F)$ and then proceed with optimization of $\mathcal{L}_N(F, \gamma)$. Instead of fitting a weak learner on $-\epsilon\widehat{\nabla}_F \mathcal{L}_N(\widehat{F}_\tau, \gamma)$ observe that $-\epsilon\nabla_F\big(\frac{\gamma}{2}\|\widehat{F}_\tau\|_2^2\big) = -\epsilon\gamma \widehat{F}_\tau$, thus we can make the exact step for $\frac{\gamma}{2}\|F\|_2^2$ by shrinking the predictions $(1 - \gamma \epsilon )\widehat{F}_\tau$. Note that due to linearity, shrinkage of the ensemble parameters $(1 - \gamma\epsilon)\widehat{\theta}_\tau$ and formal shrinkage of the predictions $(1 - \gamma \epsilon) \widehat{F}_\tau$ is the same operation.
After doing shrinkage, we estimate the gradients $\widehat{g}_\tau$ of $\mathcal{L}_N((1 - \gamma \epsilon)\widehat{F}_\tau)$ and instead of approximating $-\epsilon\widehat{g}_\tau$ by a weak learner, we approximate $-\epsilon\widehat{g}_\tau + \mathcal{N}(\mathbb{0}_{N}, 2\epsilon\beta^{-1} I_{N})$: $$\begin{split} &\text{minimize }\|W_{s_\tau}\theta^{s_\tau}\|_2^2\,\,\,\,\, \text{ subject to}\\
&\theta^{s_\tau}\in \operatorname*{arg\,min}_{\theta^{s_\tau}\in\mathbb{R}^{m_s}}\frac{1}{N}\sum_{i=1}^N (-\epsilon\widehat{g}_{\tau,i}+\sqrt{2\epsilon\beta^{-1}}\zeta_i \\ & \hspace{110pt} -
h^{s_\tau}(x_i|\theta^{s_\tau}))^2\,,
\end{split}$$ where $\zeta_i \sim \mathcal{N}(0, 1)$. We will prove that such an estimate weakly converges to the solution of the Langevin dynamics equation in $V_{\mathcal{B}}$ and therefore converges globally. Algorithm \[alg:sglb\] summarizes the proposed method.
dataset $\mathcal{D}_N$, learning rate $\epsilon > 0$, $\beta > 0$ - diffusion temperature, $\gamma > 0$ - regularization $\tau = 0$\
$f(\cdot|\theta_{0}) = 0$\
$f(\cdot|\theta_{\tau+\frac{1}{2}}) = (1 - \gamma \epsilon) f(\cdot|\theta_\tau)$\
$\widehat{g}_\tau\in \mathbb{R}^N$ on $\mathcal{D}_N$ using $f(\cdot|\theta_{\tau+\frac{1}{2}})$\
$\zeta, \zeta' \sim \mathcal{N}(\mathbb{0}_N, I_N)$\
$s_\tau \sim \mathcal{S}(-\epsilon\widehat{g}_\tau+\sqrt{2\epsilon \beta^{-1}}\zeta)$\
$\theta^{s_\tau}_* = \mathcal{P}_\tau(-\epsilon \widehat{g}_\tau+\sqrt{2\epsilon\beta^{-1}}\zeta')$\
$f(\cdot|\theta_{\tau + 1}) = f(\cdot|\theta_{\tau+\frac{1}{2}}) + h^{s_\tau}(\cdot|\theta^{s_\tau}_*)$\
$\tau = \tau + 1$\
$f(\cdot|\theta_\tau)$
Observe that $\widehat{F}_\tau$ conforms the following update equation: $$\widehat{F}_{\tau + 1} = (1 - \gamma\epsilon)\widehat{F}_\tau + \epsilon P_\tau \widehat{\nabla}_{F}\mathcal{L}((1-\gamma\epsilon)\widehat{F}_\tau) + \sqrt{2\epsilon\beta^{-1}P_\tau}\zeta.$$ Here we exploit the fact that $P_\tau$ is a projector: $\sqrt{P_\tau} \equiv P_\tau$. The equation clearly mimics the SGLD update with the only difference that $P_\tau$ is not a constant preconditioner but a random projector. So, SGLB update can be seen as SGLD on random subspaces in $V_\mathcal{B}$.
Note that $\widehat{g}_\tau$ approximates the gradients of already shrunken predictions $(1 - \gamma \epsilon)\widehat{F}_\tau$, but due to Lipschitz smoothness of $L(z, y)$ with respect to $z$, we have $\mathbb{E}(\widehat{g}_\tau|\widehat{F}_\tau) = \nabla_{F} \mathcal{L}(\widehat{F}_\tau) + \mathcal{O}(\gamma\epsilon)$, i.e., shrinkage introduces a bias of order $\mathcal{O}(\gamma\epsilon)$ which is negligible as $\epsilon \rightarrow 0_+$.
We use $-\epsilon\widehat{g}_{\tau} + \sqrt{2\epsilon\beta^{-1}}\zeta$ not only for estimation of optimal parameters of a weak learner, but also inside a weak learners’ selection algorithm, i.e., $s \sim \mathcal{S}(-\epsilon\widehat{g}_{\tau} + \sqrt{2\epsilon\beta^{-1}}\zeta)$, where $\mathcal{S}(g)$ for $g \in \mathbb{R}^{N}$ is some predefined distribution over the weak learners. We require the following properties of $\mathcal{S}(g)$:
- **Continuity at Zero**: $\mathbb{P}(\mathcal{S}(g+\zeta) \ne \mathcal{S}(\zeta)) = \mathcal{O}(\|g\|_2)$ as $\|g\|_2 \rightarrow 0$ for $\zeta \sim \mathcal{N}(\mathbb{0}_N, I_N)$. Note that $\mathcal{S}$ can have an inner noise source different from added $\zeta$.
- **Non-degeneracy**: $\mathbb{E}\|P_s {F}\|_2^2 \ge m\|{F}\|_2^2\,\,\forall {F}\in V_\mathcal{B}$ for some $m > 0$, where $P_s$ is a random projector corresponding to the weak learner $s \sim \mathcal{S}(\zeta)$ and $\zeta \sim \mathcal{N}(\mathbb{0}_N, I_N)$.
- **Zero-order Positive Homogeneity**: $\forall \lambda > 0\,\,\forall g \in \mathbb{R}^{N}$ we have $\mathcal{S}(\lambda g) \equiv \mathcal{S}(g)$.
- **Structure Noise Isolation**: The noise $\zeta\sim \mathcal{N}(\mathbb{0}_N, I_N)$ used for weak learners selection and the noise $\zeta' \sim \mathcal{N}(\mathbb{0}_N, I_N)$ used for weak learner parameters estimation are *independent*.
Recall that we require $\|\widehat{g}_\tau - \nabla_F \mathcal{L}_N((1-\gamma\epsilon)\widehat{F}_\tau)\|_2 = \mathcal{O}(1)$ with probability one. Also, $\nabla_F \mathcal{L}_N(F)$ is uniformly bounded due to Lipschitz continuity, so $\widehat{g}_\tau$ is uniformly bounded with probability one. With Zero-order Positive Homogeneity and Continuity at Zero, we get $\mathcal{S}(-\epsilon\widehat{g}_{\tau} + \sqrt{2\epsilon\beta^{-1}}\zeta) \rightarrow \mathcal{S}(\zeta) =: \mathcal{S}_{\infty}$ with the rate at least $\mathcal{O}(\sqrt{\epsilon})$, i.e., the sampling converges to its stationary as $\epsilon\rightarrow 0_+$.
SGLB and Langevin Equation
--------------------------
Let us define an *implicit* limiting preconditioner matrix of the boosting algorithm $P_{\infty} := \mathbb{E}_{s\sim \mathcal{S}_\infty}H_{s}\mathcal{P}_s:\mathbb{R}^{N}\rightarrow\mathbb{R}^{N}$. Such expectation exists since each term is a projector and hence is uniformly bounded by 1 using the spectral norm. Since $\mathcal{S}(\cdot) = \mathcal{S}_\infty + \mathcal{O}(\sqrt{\epsilon})$, we obtain $\mathbb{E}_{s\sim \mathcal{S}}H_{s}\mathcal{P}_s = P_{\infty} + \mathcal{O}(\sqrt{\epsilon})$. Using Structure Noise Isolation, we get $$\epsilon^{-1}\mathbb{E}(\widehat{F}_{\tau+1}-\widehat{F}_\tau|\widehat{F}_\tau) = -\gamma \widehat{F}_\tau - P_{\infty}\nabla_{f}\mathcal{L}_N(\widehat{F}_\tau)+ \mathcal{O}(\sqrt{\epsilon}),$$ $$\epsilon^{-1}\mathrm{Var}(\widehat{F}_{\tau+1}|\widehat{F}_\tau) = 2\beta^{-1} P_{\infty} + \mathcal{O}(\sqrt{\epsilon}).$$ Under the assumptions that we placed on $L(\cdot, \cdot)$, we will show that $\widehat{F}_\tau$ weakly converges to the solution of the following SDE as $\epsilon \rightarrow 0_+$: $$\begin{gathered}
\label{eq:sde}
\mathrm{d} F(t) = - \gamma F(t)\mathrm{d}t - P_{\infty}\nabla_{F}\mathcal{L}_N(F(t))\mathrm{d}t \\ + \sqrt{2\beta^{-1}P_{\infty}}\mathrm{d}W(t).\end{gathered}$$ But first we need to deal with the problem of possibly singular operator $P_{\infty}$ due to the fact that generally $V_{\mathcal{B}} \ne \mathbb{R}^N$. Let us show why we can assume $P_{\infty} > 0$. The trick is to observe that $V_{\mathcal{B}} = \mathrm{im }P_{\infty} = \mathrm{coim }P_{\infty}$ due to Non-degeneracy of the sampling $\mathcal{S}$. Henceforth, we can easily factorize $V_{\mathcal{B}}\oplus \ker P_{\infty} = \mathbb{R}^N$ and assume that actually we live in $V_{\mathcal{B}}$ and *there* formally $P_{\infty} > 0$ as an operator from $V_{\mathcal{B}}$ to $V_{\mathcal{B}}$.
In order to obtain a more classic form of the equation, define an implicitly regularized function: $$\mathcal{L}_N({F}, \gamma) := \mathcal{L}_N({F}) + \frac{\gamma}{2}\|\Gamma_\infty{F}\|_2^2,$$ where $\Gamma_\infty := \sqrt{P^{-1}_\infty} > 0$. So, becomes $$\mathrm{d}F(t) = -P_\infty\nabla_{F}\mathcal{L}_N(F(t), \gamma)\mathrm{d}t + \sqrt{2\beta^{-1}}\sqrt{P_\infty}\mathrm{d}W(t).$$
Note that we assumed Lipschitz smoothness of $\mathcal{L}_N({F})$, but we added a shrinkage step that in turn is equivalent to adding $\frac{\gamma}{2}\|\Gamma_\infty{F}\|_2^2$ to the loss, so the Lipschitz smoothness is still preserved for $\mathcal{L}_N({F}, \gamma)$ for any $\gamma \ge 0$. Henceforth, Theorem 1, p. 446 of @gikhman1996introduction applies ensuring the weak convergence of ${F}_{\epsilon}(t) := \widehat{F}_{\big[\epsilon^{-1}t\big]}$ to the Langevin Dynamics as $\epsilon\rightarrow 0_+$ for $t \in [0, T]$ $\forall T > 0$. Note that this theorem is proved in dimension one, but it remains valid in an arbitrary dimension [@kushner1974].
In the next subsection, we study the properties of the limiting Langevin equation. We call it *Predictions’ Space* Langevin Dynamics equation as it describes the evolution of $F(t)$ in the space $V_\mathcal{B}$. Clearly, this allows us to claim the global convergence in $V_\mathcal{B}$ and then, tautologically, for $f(\cdot|\widehat{\theta}_\tau)$.
Predictions’ Space Langevin Dynamics
------------------------------------
We consider now the following Langevin-like SDE for $P > 0, P^T = P$ in $\mathbb{R}^d$ (here we assume that we live in $V_{\mathcal{B}} = \mathbb{R}^d$): $$\begin{gathered}
\label{eq:sde-1}
\mathrm{d}F(t) = -\gamma F(t)\mathrm{d}t - P\nabla_{F}\mathcal{L}_N(F(t) )\mathrm{d}{t} \\ + \sqrt{2\beta^{-1}}\sqrt{P}\mathrm{d}W(t).\end{gathered}$$ The classic form of the equation can be obtained using a so-called implicitly regularized loss function: $$\mathcal{L}_N({F}, \gamma) := \mathcal{L}_N({F}) + \frac{\gamma}{2}\|\Gamma{F}\|_2^2,$$ where $\Gamma := \sqrt{P^{-1}} > 0$ plays a role of a regularization matrix that is defined implicitly by $\mathcal{S}$ and $\mathcal{P}$ in the case of boosting. Due to the well-known properties of $L_2$-regularization, for small enough $\gamma > 0$, the minimization of $\mathcal{L}_N({F}, \gamma)$ leads to an almost minimization of $\mathcal{L}_N({F})$ with an error of order $\mathcal{O}(\delta_\Gamma(\gamma))$ for some function $\delta_\Gamma(\gamma)$ depending on the choice of $\mathcal{L}_N$[^1] and $\Gamma$, which is negligible as $\gamma \rightarrow 0_+$.
Clearly, the choice of $\mathcal{P}$ and $\mathcal{S}$ uniquely defines the preconditioner and henceforth the regularization $\Gamma$. The error $\delta_\Gamma(\gamma)$ heavily depends on $\Gamma = \sqrt{P^{-1}}$ and therefore affects the effectiveness of the algorithm, so a poor choice of $\mathcal{B}$ may drastically slow down the convergence.
Using $\mathcal{L}_N({F}, \gamma)$, we rewrite Equation as: $$\mathrm{d}F(t) = -P\nabla_{F}\mathcal{L}_N(F(t), \gamma)\mathrm{d}t + \sqrt{2\beta^{-1}}\sqrt{P}\mathrm{d}W(t).$$ Then, the results of @Erdogdu:2018:GNO:3327546.3327636 [@DBLP:journals/corr/RaginskyRT17] apply ensuring that $\mathcal{L}_N(\widehat{F}_\tau, \gamma)$ converges to an almost-minimizer of $\mathcal{L}_N({F}, \gamma)$ with an error of order $\mathcal{O}\Big(\frac{d}{\beta}\log\frac{\beta}{d}\Big)$, where $d = \dim V_{\mathcal{B}}$.
Henceforth, due to the weak convergence, we obtain the desired result: $\mathcal{L}_N(\widehat{F}_\tau)\rightarrow \inf_F$ among all possible predictions obtained by ensembles of weak learners from $\mathcal{H}$. Hence, we can formulate the following theorem.
Consider a Markov chain $\widehat{F}_{\tau}$ generated by Gradient Boosting algorithm $\mathcal{B}$. Then, the following bound holds almost surely: $$\lim\mathbb{E}\mathcal{L}_N(\widehat{F}_\tau) - \inf_{{F}\in V_{\mathcal{B}}}\mathcal{L}_N({F}) = \mathcal{O}\Big(\delta_\Gamma(\gamma) + \frac{d}{\beta}\log\frac{\beta}{d}\Big)$$ for $\epsilon\rightarrow 0_+, \epsilon\tau\rightarrow +\infty$, where $d = \dim V_{\mathcal{B}}$ and $\delta_{\Gamma}(\gamma)$ encodes the error from the regularization and is of order $o(1)$ as $\gamma \rightarrow 0_+$.
Generalization gap {#Generalization-Section}
------------------
We reduced the problem of convergence of a general boosting to the problem of convergence of *predictions* $\widehat{F}_\tau := \big(f(x_i|\widehat{\theta}_\tau)\big)_{i=1}^N$ in the space of all possible predictions $V_\mathcal{B}$ on the dataset $\mathcal{D}_N$ formed by finite ensembles. Assume that $|S| < \infty$ and we define a *design* matrix $\Phi := \big[\phi(x_1),\ldots, \phi(x_N)\big]^T \in \mathbb{R}^{N\times m}$, so we can write ${F} = \Phi \theta$ and $\mathcal{L}_N({F}, \gamma) = \mathcal{L}_N(\Phi\theta) + \frac{\gamma}{2}\|\Gamma \Phi \theta\|_2^2$. Note that $V_\mathcal{B}$ can be obtained as the image of $\Phi$.
Now we know how to solve the problem , but the crucial question is: how well the optimization of translates into ? The answer can be given explicitly by considering the uniform spectral gap parameter $\lambda_* \ge 0$ for the distribution $p_\beta(\theta) := \frac{\exp(-\beta \mathcal{L}_N(\Phi\theta, \gamma))}{\int_{\mathbb{R}^m}\small{\exp(-\beta \mathcal{L}_N(\Phi\theta, \gamma))}\mathrm{d}\theta}$ (for the definition of a uniform spectral gap see @DBLP:journals/corr/RaginskyRT17).
In order to make this definition fully correct (i.e., to ensure that $p_\beta^*(\theta)$ is a completely and correctly defined distribution) we need, first, to assume that $\ker \Phi = \bigoplus_{s\in S} \ker h^{s}$, where $\ker h^s := \{\theta^s \in \mathbb{R}^{m_s}: h(x_i|\theta^s) = 0\forall x_i\in \mathcal{D}_N \}\subset \mathbb{R}^{m_s}$. This condition is equivalent to the assumption that $(h^{s_1}(x_i|\theta^{s_1}))_{i=1}^N, \ldots, (h^{s_k}(x_i|\theta^{s_k}))_{i=1}^N$ is linearly independent for any choice of different weak learners indexes $s_1, \ldots, s_k \in S$ and an arbitrary choice of $\theta^{s_j}\in \mathbb{R}^{m_{s_j}}$ such that the vector $(h^{s_j}(x_i|\theta^{s_j}))_{i=1}^N$ is not zero $\forall j$. Then, $\ker \Phi$ has right structure, i.e., we have “basis weak learners” for $\mathcal{B}$.
Second, we use the factorization trick to factorize $\mathbb{R}^{m_s} = \ker h^s \oplus (\ker h^s)^{\perp} $ and hence w.l.o.g. we can assume $\ker h^s = \{\mathbb{0}_{m_s}\}$. The latter implies that w.l.o.g. $\ker \Phi = \{\mathbb{0}_m\}$, so the distribution $p_\beta(\theta)$ is a correctly defined distribution on $\mathbb{R}^m$. Moreover, observe that in that case $m = d = \dim V_{\mathcal{B}}$ and thus $d$ is *independent* from $N$ for large enough $N \gg 1$.
In practice, the condition $|S| < \infty$ is satisfied since boosting algorithms consider a finite number of possible trees on the dataset $\mathcal{D}_N$. Note that the above condition on kernel structure is somewhat restrictive since it implies $\sum_{s \in S} m_s - \dim\ker h^s = \dim V_\mathcal{B}$ and is hard to verify in practice. Such condition *can be* satisfied if we assume that each feature vector $x\in \mathbb{R}^k$ has independent components and $N \gg 1$. However, such restrictive condition allows us to deduce a generalization gap bound which gives an insight on how the choice of $\mathcal{S}$ and $\mathcal{P}$ affects the generalization.
Having $p_\beta(\theta) \propto \exp(-\beta \mathcal{L}_N(\Phi\theta, \gamma))$ from @DBLP:journals/corr/RaginskyRT17, we can transfer a bound $\propto \frac{(\beta+d)^2}{\lambda_* N}$ for the generalization gap where $\lambda_* \ge 0$ is a uniform spectral gap parameter. Since we added $L_2$-regularizer to the loss which is Lipschitz smooth and continuous, we necessarily obtain dissipativity and thus $\lambda_* > 0$ [@DBLP:journals/corr/RaginskyRT17].
The generalization gap $$\big|\mathbb{E}_{\theta\sim p_\beta(\theta)} \mathcal{L}(f(\cdot|\theta)) - \mathbb{E}_{\theta\sim p_\beta(\theta)}\mathcal{L}_N(f(\cdot|\theta))\big|$$ can be bounded by: $$\mathcal{O}\Big(\frac{(\beta+d)^2}{\lambda_* N}\Big),$$ where $d = \dim V_{\mathcal{B}}$ is independent from $N$ for large enough $N \gg 1$.
Note that $1/\lambda_*$ is not dimension-free and in general may depend on $d$, $\beta$, $\Gamma$ and $\gamma$. Moreover, the dependence can be exponential in $d$ [@DBLP:journals/corr/RaginskyRT17]. However, in the presence of convexity, we can get a dimension-free bound for $1/\lambda_*$. To see that, we need to bound $1/\lambda_* \le c_{P}(p_\beta)$, where $c_{P}$ is the Poincare constant for $p_\beta(\theta)$. If $L(\cdot, \cdot)$ is convex, then $\beta\mathcal{L}_N(\theta, \gamma)$ must be strongly convex with a constant $\ge \frac{\kappa \beta\gamma}{2}$, where $\kappa := \lambda_{\min}(\Phi\Gamma^2 \Phi^T)$ is the smallest eigenvalue of $\Phi\Gamma^2 \Phi^T > 0$. Then, $p^{*}_{\beta}$ is strongly log-concave, so by transferring the Poincare inequality for strongly log-concave distribution from [@Milman2007arXiv0712.4092M], we obtain a dimension free-bound $1/\lambda_*\le \frac{1}{\kappa\gamma\beta}$.
Together with integration suboptimality, we derive that our algorithm has the following performance in $\epsilon \rightarrow 0_+, \epsilon \tau \rightarrow +\infty$ limit: $$\begin{gathered}
\big|\mathbb{E}_{\theta\sim p_\beta(\theta)} \mathcal{L}(f(\cdot|\theta)) - \mathcal{L}_*\big| \\ = \mathcal{O}\Big(\delta_\Gamma(\gamma) + \frac{d}{\beta}\log\frac{\beta}{d} + \frac{(\beta+d)^2}{\lambda_* N}\Big),\end{gathered}$$ where $\delta_\Gamma(\gamma)$ encodes an error from regularization that is negligible in the $\gamma \rightarrow 0_+$ limit. However, $1/\lambda_*$ is also dependent on $\gamma$ (e.g., in the convex case the dependence is of order $\mathcal{O}(1/\gamma)$) and thus the optimal $\gamma_*$ must be strictly greater than zero $\gamma_* > 0$.
Moreover, note that here we use the expected loss on $\mathcal{D}$ but not $\mathcal{D}_N$, so by varying parameters and choosing $N \gg 1$ this bound can be made arbitrary small and therefore our algorithm reaches the ultimate goal stated in Section \[sec:sgb\].
The above bound completely answers the question how the choice of the triplet $\mathcal{B} = (\mathcal{H}, \mathcal{S}, \mathcal{P})$ affects the optimization quality. Note that our analysis, unfortunately, gives no insights on the speed of the convergence of the algorithm and we leave this for the future work.
CatBoost Implementation {#sec:catboost}
=======================
We implemented SGLB as a part of CatBoost gradient boosting library, which was shown to provide state-of-the-art results on a large number of datasets [@catboost]. Now we specify the particular triplet $\mathcal{B}= (\mathcal{H}, \mathcal{S}, \mathcal{P})$ such that all assumption stated before are satisfied and therefore the implementation must converge globally for a wide range of functions, not only for convex ones. First, note that $W_s$ is taken to be the identity matrix $I_{m_s}$ for any $s\in S$, so $\mathcal{P}$ is fully specified. Now we are going to describe the weak learners set $\mathcal{H}$ used by CatBoost.
For each numerical feature, CatBoost chooses between a finite number of splits $\mathbb{1}_{\{x_i \le c_{ij}\}}$, where $\{c_{ij}\}_{j=1}^{d_i}$ are some constants typically picked as quantiles of $x_i$ estimated on $\mathcal{D}_N$ and $d_i$ is bounded by a hyperparameter $\textit{border-count}$. So, the set of weak learners $\mathcal{H}$ consists of all non-trivial binary oblivious trees with splits $\mathbb{1}_{\{x_i \le c_{ij}\}}$ and with depth bounded by a hyperparameter $\textit{depth}$. This set is finite, $|S| < \infty$. We take $\theta^s \in \mathbb{R}^{m_s}$ as a vector of leaf values of the obtained tree.
Now we are going to describe $\mathcal{S}$. Assume that we are given a vector $g \in \mathbb{R}^N$ and already built a tree up to a depth $j$ with remaining (not used) binary splits $b_1, \ldots b_p$. Each split divides the vector $g$ in two components $g'\in \mathbb{R}^{N'}, g'' \in \mathbb{R}^{N''}$. So if we are in a leaf with corresponding components of $\widetilde{g} \in \mathbb{R}^{\widetilde{N}}$, we need to decide which split $b_i$ to apply or to keep the node as a leaf. CatBoost divides $\widetilde{g} = (\widetilde{g}', \widetilde{g}'')$ for each split $b_i$ and then calculates the following statistics: $$s_i := \sqrt{\text{Var}(\widetilde{g}') + \text{Var}(\widetilde{g'})},$$ where $\text{Var}(\cdot)$ is the variance of components from the component-wise mean. Denote also $\sigma := \sqrt{\text{Var}(g)}$. Then, CatBoost evaluates: $$s_i' := \mathcal{N}\left(s_i,\left(\frac{\rho \sigma}{1 + N^{\epsilon \tau}}\right)^2\right),$$ where $\rho \ge 0$ is a hyperparameter defined by the $\textit{random-strength}$ parameter. After obtaining $s_i'$, CatBoost selects a split with a highest $s_i'$ value and adds it to the tree and then proceeds recursively until a stopping criteria met.
Note that due to $\epsilon\tau \rightarrow \infty$, we can assume that actually $\rho = 0$ in the limit. Thus, the stationarity of sampling is preserved. So, $\mathcal{S}$ is fully specified and, henceforth, such CatBoost implementation $\mathcal{B}$ must converge globally for a large class of losses as $\epsilon \rightarrow 0_+, \epsilon \tau \rightarrow \infty$ under mild conditions.
Direct Accuracy Optimization {#sec:direct01}
============================
To show the power of SGLB for non-convex multimodal optimization, we select $\text{Accuracy}$ for the direct optimization by our framework: $$\text{Accuracy} = \frac{1}{N}\sum_{i=1}^N \mathbb{1}_{\{y_i f(x_i|\theta) > 0\}},$$ where $y_i \in \{0, 1\}$. The corresponding loss $1-\text{Accuracy}$ is called 0-1 loss.
In order to make the above function Lipschitz smooth and Lipschitz continious, we approximate 0-1 by $1-\sigma(\varsigma^{-1}(2y_i-1) f(x_i|\theta)) :=1 - (1 + \exp(-\varsigma^{-1} (2y_i-1) f(x_i|\theta)))^{-1}$, where $\varsigma > 0$ is a hyperparameter. We then minimize: $$\mathcal{L}_N(f) := 1-\frac{1}{N}\sum_{i=1}^N \sigma(\varsigma^{-1} (2y_i-1) f(x_i|\theta)).$$ Such smoothing is known as Smooth Loss Approximation (SLA) in the literature [@nguyen2013algorithms]. If $f(x_i|\theta) \ne 0\,\,\forall i$, then as $\varsigma \rightarrow 0_+$ we have $\sigma(\varsigma^{-1} y_i f(x_i|\theta)) \rightarrow \mathbb{1}_{\{y_i f(x_i|\theta) > 0\}}$, so that $\mathcal{L}_N(f) \rightarrow 1-\mathrm{Accuracy}$ as expected for an approximation.
To apply SGLB, we need to ensure Lipschitz smoothness and continuity. Observe that the gradient is uniformly bounded due to $\big|\frac{\mathrm{d}}{\mathrm{d}z}\sigma(z)\big| = (1 - \sigma(z))\sigma(z) \le 1$, which in turn implies Lipschitz continuity. Moreover, $\big|\frac{\mathrm{d}^2}{\mathrm{d}z^2}\sigma(z)\big| = |1 - 2\sigma(z)| \le 3$ which implies the uniform boundness of the Hessian and hence Lipschitz smoothness.
Note that similarly one can prove Lipschitz smoothness and continuity of Logistic regression loss $L(z, y) = -y \log \sigma(z) - (1 - y) \log (1 - \sigma(z))$ as $\frac{\mathrm{d}}{\mathrm{d}z}L(z, y) = -y + \sigma(z)$ and $\frac{\mathrm{d^2}}{\mathrm{d}z^2}L(z, y) = \sigma(z) (1 - \sigma(z))$.
Experiments {#sec:experiments}
===========
In this section, we illustrate that the theoretically grounded SGLB algorithm has a superior performance in various practical tasks. As the baseline approach, we consider the CatBoost SGB library, which is known to achieve state-of-the-art results across a wide variety of data sets [@catboost].
As mentioned above, we consider the logistic regression task with loss $L(z, y) = -y \log \sigma(z) - (1 - y) \log (1 - \sigma(z)), y \in \{0, 1\}$ (Logloss), and the classification task with 0-1 loss $L(z, y) = 1-\mathbb{1}_{\{z (2y-1) > 0\}}, y\in \{0, 1\}$. In the latter case we optimize SLA $L(z, y) = 1 - \sigma(\varsigma^{-1}z(2y-1))$ with $\varsigma = 10^{-1}$. Both functions are Lipschitz smooth and continuous and thus SGLB can be applied.
The datasets are described in Table \[tab:datasets\]. We split each dataset into train ($\mathcal{D}_N$), validation, and test sets in proportion 80/10/10. We tune the parameters on a validation set using 50 iterations of random search and select the best iteration on the validation, the details are given in Appendix. For all algorithms, the maximal number of trees is set to 1000.
The results are shown in Table \[table:results\]. We use bold font to highlight significance for two-tailed $z$-test with p-value $< 0.05$. We note that SGB baseline is strengthened with Minimal Variance Sampling [@ibragimov2019minimal] and leaves regularization, whilst SGLB is not, but MVS *can be* incorporated into SGLB without any modifications if sampling is performed prior to adding noise to gradient estimates (we leave such investigation for the future work).
Nevertheless, we see that in *convex* setting SGLB performs comparable to SGB, while for the *non-convex* task (for 0-1 loss optimization), SGLB performance is *superior* to SGB, which clearly shows the necessity of non-convex optimization methods in machine learning.
Method Dataset Optimization Loss Result
-------- --------- -------------- --------- -----------
SGB Adult SLA 0-1 0.130
SGLB Adult SLA 0-1 **0.127**
SGB Amazon SLA 0-1 0.050
SGLB Amazon SLA 0-1 **0.049**
SGB Epsilon SLA 0-1 [0.118]{}
SGLB Epsilon SLA 0-1 0.118
SGB Higgs SLA 0-1 [0.260]{}
SGLB Higgs SLA 0-1 **0.251**
SGB Adult Logloss Logloss 0.273
SGLB Adult Logloss Logloss 0.273
SGB Amazon Logloss Logloss 0.148
SGLB Amazon Logloss Logloss **0.145**
SGB Epsilon Logloss Logloss **0.273**
SGLB Epsilon Logloss Logloss 0.275
SGB Higgs Logloss Logloss **0.479**
SGLB Higgs Logloss Logloss [0.482]{}
: Experimental results[]{data-label="Results"}
\[table:results\]
Conclusion & Future Work {#sec:conclusion}
========================
Our experiments clearly demonstrate that the theoretically grounded SGLB algorithm also shows very promising experimental results. However, there are plenty directions for future research which can potentially further improve the performance.
Recall that our generalization gap estimate relies on restrictive assumption on linear independence of weak learners, thus a promising direction is how to modify the algorithm so that some form of Langevin diffusion is still preserved in the limit with good provable generalization gap guarantees.
Another idea is to incorporate momentum into the boosting so that there is the Hamiltonian dynamics [@gao2018global] in the limit instead of the ordinary Langevin dynamics. There are several theoretical attempts to incorporate momentum into the boosting like HistoricalGBM [@GCAI-2018:Historical_Gradient_Boosting_Machine], so the question is: if we use HistoricalGMB approach or its modification, would that be enough to claim the Hamiltonian dynamics equation in the limit?
Finally, our research does not investigate the rates of convergence, thus deducing the rates is another promising direction as it would allow to better understand the trade-offs between the parameters of the algorithm.
Dataset \# Examples \# Features
-------------------- ------------- -------------
Adult [@adult] 48842 15
Amazon [@amazon] 32769 9
Epsilon [@epsilon] 500K 2000
Higgs [@higgs] 11M 28
Parameter tuning
================
For all algorithms, we use the default value 64 for the parameter *border-count* and the default value 0 for *random-strength* ($\rho \ge 0$).
For SGB, we tune *learning-rate* ($\epsilon > 0$), *depth* (the maximal tree depth) and the regularization parameter *l2-leaf-reg*. Moreover, we consider Minimal Variance Sampling (MVS) by setting *bootstrap-type=MVS* as it improves the results for SGB [@ibragimov2019minimal].
Whilst for SGLB we consider *learning-rate*, *depth*, *model-shrink-rate* ($\gamma \ge 0$) and *diffusion-temperature* ($\beta > 0$). We *do not* consider MVS for SGLB by setting *bootstrap-type=No*.
For all methods, we consider *leaf-estimation-method=Gradient* as our main purpose is to compare first order optimization and option *use-best-model=True*.
For tuning, we use the random search (50 samples) with the following distributions:
- For *learning-rate* log-uniform distribution over $[10^{-5}, 1]$.
- For *l2-leaf-reg* log-uniform distribution over $[10^{-1}, 10^1]$ for SGB and *l2-leaf-reg=0* for SGLB.
- For *depth* uniform distribution over $\{6, 7, 8, 9, 10\}$.
- For *model-shrink-rate* log-uniform distribution over $[10^{-5}, 10^{-2}]$ for SGLB.
- For *diffusion-temperature* log-uniform distribution over $[10^2, 10^5]$ for SGLB.
[^1]: In the case when $\min_F \mathcal{L}_N(F)$ exists one can show that $\delta_\Gamma(\gamma) = \mathcal{O}(\lambda_{\max}(\Gamma^2) \gamma)$ with $\mathcal{O}(\cdot)$ depending on $\mathcal{L}_N$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents data obtained during the NICMOS Guaranteed Time Observations of a portion of the Hubble Deep Field. The data are in a catalog format similar to the publication of the original WFPC2 Hubble Deep Field Program ([@will96]). The catalog contains 342 objects in a $49.1 \times 48.4\arcsec$ subfield of the total observed field, 235 of which are considered coincident with objects in the WFPC2 catalog. The $3\sigma$ signal to noise level is at an aperture AB magnitude of approximately 28.8 at 1.6 microns. The catalog sources, listed in order of right ascension, are selected to satisfy a limiting signal to noise criterion of greater than or equal to 2.5. This introduces a few false detections into the catalog and users should take careful note of the completeness and reliability levels for the catalog discussed in Sections \[comp\] and \[rel\]. The catalog also contains a test parameter indicating the results of half catalog tests and the degree of coincidence with the original WFPC2 catalog.'
author:
- 'Rodger I. Thompson'
- 'Lisa J. Storrie-Lombardi and Ray J. Weymann'
- 'Marcia J. Rieke, Glenn Schneider, Elizabeth Stobie and Dyer Lytle'
title: |
NICMOS OBSERVATIONS OF THE HUBBLE DEEP FIELD: OBSERVATIONS,\
DATA REDUCTION, AND GALAXY PHOTOMETRY
---
Introduction
============
Deep observations with NICMOS, devoted to understanding the nature of galaxy formation and evolution along with information on cosmological parameters, have always been an important aspect of the NICMOS Instrument Development Team program since its inception in 1984. After the WFPC2 Hubble Deep Field (HDF) program in late 1995 it became obvious that observations at longer wavelengths would greatly enhance the value of the existing data in addition to satisfying the original intent of deep observations. The smaller field of view of the NICMOS instrument made it necessary to choose between deep observations of a portion of the HDF and a survey of the entire HDF at a brighter limiting magnitude. With the advent of the HDF there existed a large disparity between the depth of the HDF and the depth of observations at near infrared wavelengths ([@con97]). Since the majority of objects that might be at redshifts unobservable with the WFPC2 were expected to be relatively faint, the IDT decided to conduct a limited spatial survey to the faintest possible magnitude. The results of a NICMOS General Observer HDF survey program ([@dic97]) will provide coverage over the entire HDF.
The NICMOS HDF program consists of 127 orbits out of a total of 553 orbits for the entire GTO program. Table \[hdf\] shows the distribution of orbits between the two filters and two grisms. An additional two orbits were dedicated to confirmation of guide star acquisition. Although the bulk of the orbits are dedicated to imaging, the large comoving volume for line observations available to the grisms is very appealing. The small number of grism orbits shown in Table \[hdf\] were intended as test cases to see if more GTO orbits should be transferred to this program. In order to expedite the delivery of image data to the community at large we have not concentrated on the reduction of the grism data and it is not presented in this publication. The grism data will be published in subsequent papers. As discussed below the grism observations, however, significantly influenced the choice of the field for NICMOS imaging observations.
As with the [@will96] paper the purpose of this publication is a presentation of the data and analysis techniques rather than a discussion of the scientific content of the data. Future papers will discuss several implications of the new data. In the following we present the rationale for the observation methods, the methods for image production and source extraction, the catalog, and a discussion of the quality of the data in terms of signal to noise, completeness and reliability. Note that all magnitudes quoted in this paper are in the AB system.
FIELD SELECTION
===============
The decision to devote part of the observational time to grism observations limited field choices to regions of the HDF that are not dominated by large bright foreground galaxies. The slitless dispersed spectra of these galaxies would overlap large areas of the field of view and reduce the number of spectral observations of fainter galaxies. Although some information on high redshift objects was available at the time of field selection, no effort was made to bias the field position to include the largest number of high redshift sources.
The Space Telescope Science Institute decision to schedule the NICMOS HDF observations during the camera 3 campaign in January of 1998 determined the acceptable range of roll orientations. This time period was not part of the Continuous Viewing Zone (CVZ) opportunity period, however, it did offer a larger fraction of truly dark observing time. Given these constraints a field located roughly at the center of the WFPC2 chip 4 field offered the best observational opportunities. The J2000.0 center position is $12^{\rm h}$ $36^{\rm m}$ $45.129^{\rm s}$, $+62\arcdeg$ $12\arcmin$ $15.55\arcsec$. There is a relatively bright star ($AB_H$ of approximately 22.1 magnitude) near the center that provides an excellent fiducial location for the grism observations. The final orient of $261.851 \arcdeg$ is the result of fine tuning to obtain the best possible guide star orientation.
FILTER AND CAMERA SELECTION
===========================
Camera Selection
----------------
All of the NICMOS HDF data in this paper are from camera 3. The wide field format of camera 3 made it the obvious choice for HDF observations. The campaign also utilized cameras 1 and 2 with the same integration parameters as camera 3 but they were not in focus during operation. Parallel observation of these cameras with camera 3 prevented the occurence of the faint artifacts, termed bars, which occur when the autoflush and imaging output timing patterns overlap. Similar integration parameters for all cameras prevents the parallel cameras from defaulting into the autoflush pattern. The data from these cameras may be useful for background characterization but are not analyzed in this publication.
Filter Selection
----------------
The observations employed two imaging filters for camera 3, F110W and F160W centered at 1.1 and $1.6 \micron$ respectively. By careful design, the F160W spans the lowest background spectral region available to NICMOS. This is the minimum between the scattered zodiacal emission that decreases with wavelength and the thermal emission from the warm HST mirrors that increases with wavelength. Both of these emissions were lower than expected prior to the HDF observations. The second filter covers a shorter wavelength over a rather broad bandwidth. This filter provides a second color between the F814W WFPC2 filter and the NICMOS F160W filter. The addition of this filter provides an important discriminator between high redshift star-forming galaxies which will have blue infrared colors and lower redshift galaxies with large amounts of dust extinction which will have red infrared colors. Although very useful for all objects the extra filter is particularly important for objects detected only in the NICMOS bands. Figure \[Thompson.fig1.ps\] displays the normalized detectivity for the NICMOS filters. These plots include all of the color dependent terms of the detectivity, including the detector quantum efficiency. The rapidly changing indices of refraction for most optical coating materials in the 1 micron region account for the complicated shape of the F110W filter.
OBSERVATIONAL STRATEGY
======================
The preflight decision to devote 49 orbits to each imaging filter and 27 orbits to grism spectroscopy set the parameters for the observing strategy. The 49 orbits are based on a 7 by 7 grid of positions and the grism orbits employ 3 different roll angle positions to help remove confusion from overlapping spectra. There are 3 separate integrations in each orbit to insure that any problems encountered did not necessarily compromise all of the data during the orbit. Before the observations it appeared that the sky background would be the dominant noise source after about 900 seconds of integration. The lower than expected sky brightness, however, reduced the sky noise below the read noise for this integration time.
Detector Sampling Sequence
--------------------------
Since there are no bright sources in the HDF the logarithmic NICMOS sampling sequences, designed to handle high dynamic range images, are inappropriate. Those sequences add amplifier glow noise with several short time integrations near the beginning of the exposures. Good cosmic ray rejection, on the other hand, requires a sufficient number of samples to establish an accurate signal after any cosmic ray affected samples are rejected. The HDF integrations utilize the SPARS64 sampling sequence that has evenly spaced 64 second sample times after the first 3 short integration samples. The integrations have a NSAMP value of 17 which produces a total integration time of 896 seconds. Three of these integrations fill an orbit. The total number of integrations in each filter is then 147 integrations of 896 seconds in each filter for a total of $ 1.31712 \times 10^{5}$ seconds or 36.5 hours of observing time per filter.
Field Coverage
--------------
Several factors influenced the choice of field size. The basic purpose of the NICMOS HDF program is deep observations. This requirement favors very small or no dithers away from the field center. Accurate background subtraction requires many offsets large enough to insure that most pixels have the majority of their observations off of detectable objects. Spatial cosmic ray rejection and image resolution enhancement require at least one pixel and fractional pixel offsets respectively.
The pattern of observation positions on the sky is a combination of a small three point dither pattern during each orbit and a larger $7 \times 7$ raster pattern that covered the 49 orbits. The dither pattern is a 3 position spiral dither with a step size of 0.408 arc seconds, roughly 2 camera 3 pixels. The x and y spacings of the orbit to orbit raster are 0.918 and 1.523 arc seconds respectively which are 4.5 and 7.5 camera 3 pixels. The interorbit moves were accomplished with target offsets from the original center position. The basic purpose of the raster was to move the field of view sufficiently that any single pixel had the majority of its integrations with no observable source present.
POINTING ACCURACY
-----------------
Due to the paucity of bright stars in the HDF region and the roll constraints during the observational time period we were not able to utilize two FGS guide stars. This situation led to the possibility of roll errors in position about the location of the single guide star. Real time frequent updates of the gyro bias levels by HST Missions Operations Support Engineering Systems (MOSES) mitigated this problem. Data provided by the MOSES team ([@conr98]) indicated that the positional errors for all orbits used in this paper were less than 0.2 NICMOS camera 3 pixels. Subsequent analysis discussed in Section \[mos\] confirmed this data. Our absolute positions assume that the central star in our field (WFPC2 4-454) has the position stated in the published catalog ([@will96]).
DATA REDUCTION
==============
Data reduction procedures utilized the Interactive Data Language (IDL) software environment for most of the basic data analysis. KFOCAS ([@ads96]), a derivative of FOCAS ([@jt81], [@val82]) provided the source detections listed in the catalog. In order to provide a cross check on the images and catalog presented in this paper we have deliberately reduced and analyzed the data in two separate and independent ways. Specifically in addition to the IDL and KFOCAS procedures we utilized an independent IRAF based image processing algorithm and an alternative source extraction program, SExtractor ([@Ber96]). The IDL and KFOCAS reduction procedures are described in detail as they produced the bulk of the information on the sources listed in the catalog. Descriptions of the IRAF and SExtractor reductions are provided when they differ substantially from the IDL and KFOCAS reductions.
IDL Image Reduction
-------------------
Each 896 second SPARS64 integration produced an individual image. A set of 55 SPARS64 dark integrations of the same duration as the HDF exposures provided the required dark frames for the analysis. These dark exposures occurred just prior and coincident with the HDF exposures. The data reduction procedures produce a completely processed image for each integration. Section \[mos\] describes the combination of the images into the final mosaic.
### Dark Frames
Dark frame reductions begin with the division of each 17 sample SPARS64 darks into 16 first differences. A first difference is simply the difference between a readout and the previous readout. The first differences are then combined via a sigma clipping mean to produce a final super frame that is free of cosmic rays to the 3 sigma level of the 55 combined observations. Although for most observations a simple median of first differences would suffice, observations at the sensitivity level of the HDF required sigma clipped means to avoid digitization noise. The average camera 3 dark current is 0.2 electrons per second. In each 64 second first difference about 12.8 electrons accumulate. The detector gain for camera 3 is 6.5 electrons per ADU for an average of 2 ADUs per first difference. If medians are taken of these observations there would still be 50% noise even for an infinite number of integrations. The first differences are then recombined to produce a ramp dark for subtraction from the imaging integrations. The ramp dark is a sequence of summations ranging from just the initial first difference, the initial and second first difference, to the total of all of the first differences.
### Image Frames
Analysis of the image integrations starts with the production of a set of 16 ramp readouts for the 17 samples from each SPARS64 integration. Subtraction of the super dark ramp from each integration produces a set of dark current corrected but not sky corrected integrations. A set of standard linearity corrections are next applied to pixels that have exceeded the linear signal response region but have not saturated. In practice only bad “hot" pixels receive a correction due to the low signal levels in the HDF observations.
After correction for linearity the integrations are corrected for cosmic rays by fitting a linear function to the ramp values. The slope of this function is the signal rate in ADU per second. Cosmic rays produce an instantaneous discontinuity in the signal function. Subtraction of the fitted function from the signal produces an output that has a distinctive S shape if a cosmic ray is present. In one readout the difference between the fit and the signal transitions from negative to positive. Detection of a transition greater than expected from noise indicates the presence of a cosmic ray. The offending first difference is then removed from the ramp and a new fit is calculated. The new fit is again checked for cosmic rays and detected cosmic rays are removed in a similar matter. Any fits still beyond the expected noise are declared bad and flagged as bad. If the cosmic ray produces saturation, only the readouts before saturation are used in the final fit. All detected cosmic rays are recorded in the data quality image extension. This procedure is unique to the NICMOS instrument on HST due to the ability to nondestructively read out the detector during the integration. Further cosmic ray removal can occur if necessary during the image mosaic construction in the standard manner.
Before the flat field can be applied all of the quadrant biases in the individual images must be removed. If there is a bias level in the image, the flat field function will produce variations in the bias which will remain in the data. The removal procedure can be iterative but in practice one iteration is sufficient. The first step produces a median of all of the cosmic ray corrected images. This median is then subtracted from each individual image. If there are no quadrant biases the median value in each detector quadrant should be zero since the dark current and sky are removed by the subtraction. The sources do not dominate the image so they contribute very little to the median. The second step measures the median of each quadrant in each image and records the value. The final step then subtracts these quadrant bias levels from each cosmic ray corrected image. The flat field correction was then applied to each image.
The bad pixels are marked from a bad pixel mask determined from the previous observing history with the NICMOS camera 3. The bad pixel mask contains both low response pixels and hot pixels defined as pixels with an excessive dark current. In each image the bad pixels were replaced by the median of the total image. The drizzle mosaic process does not use pixels marked as bad in the final image construction.
The median of all of the final individual images determines the level of the sky emission. The raster and dither of the large number of images reduces the source contribution of the median to a value less than the expected noise level. Inspection of the median image did not reveal any source contributions. Subtraction of this sky level from each image completes the analysis of the individual images. The median sky level in the F160W filter is 0.55 electrons per second which is lower than the original estimates prior to the observations. This is not surprising as one of the selection criteria for the HDF was a low zodiacal background.
IRAF Image Reduction
--------------------
For comparison with the IDL procedures images were reduced in an independent pipeline using NICRED 1.5 ([@mcl97], [@leh98]) and modified IRAF scripts developed to reduce Camera 3 images taken in parallel mode ([@yan98]). Two median sky-dark frames were produced, one from the first exposure in each orbit and one from combining the second and third exposures in each orbit, to minimize the effect of any pedestal in the first exposure. These were used as dark frames along with the same flats utilized in the IDL reductions. These flats, observed on Dec. 23, 1997 are identical to the flats used in the IDL reduction. The residual bias levels in the individual quadrants were removed by fitting a Gaussian to a histogram of the pixels in each quadrant, and subtracting the peak value. A new bad pixel mask was created from the exposures. The images were inspected and any remaining cosmic ray hits or satellite trails were individually masked.
Mosaic Techniques {#mos}
-----------------
Both of the data reduction procedures utilize modified versions of the drizzle software developed for the reduction of the WFPC2 HDF images. The drizzled pixel size in each case is $\sim 0.1 \arcsec$, one half of the original NICMOS Camera 3 pixel size. The drizzle parameter PIXFRACT is 0.6 in the drizzling of the IDL reductions while it was set to 0.65 in the drizzling of the IRAF results.
### IDL Image Mosaic
The first task of mosaic production is an accurate determination of the relative offsets between the individual integrations. We compared offset information from the world coordinates in the header files, shifts computed from the IRAF/STSDAS Dither package, offsets from the IRAF imcentroid package of five individual objects in the field and finally individual inspection via interactive IDL tools. The NICMOS geometric distortions have been determined to be negligible so no geometric distortion corrections were made.
In general, the agreement among the four methods was quite good. For the F160W images the discrepancy between the world coordinate shifts and the IRAF generated shifts fell between -0.2 and 0.4 pixels in x with a mean of 0.15 (RMS=0.03) and a range of -0.5 to 0.2 pixels in y with a mean of -0.15 (RMS=0.03). The internal difference between the shifts determined by IRAF procedures averaged -0.02 pixels.
For the F110W filter the difference between the world coordinate shifts and those determined by IRAF procedures varied from -2.0 to 0.7 pixels in x and -0.4 and 2.7 pixels in y with means of 0.1 (RMS=0.14) and 0.7 (RMS=0.36) respectively. The mean internal difference between the IRAF procedures was -0.12 for x and 0.04 in y. The large excursions of 2.0 and 2.7 pixels in x and y were seen in only two images. The IRAF Dither and imcentroid positions agreed to 0.1 pixels in these images. Visual inspection of the images confirmed the IRAF positions. In both the F160W and the F110W filter the rotation angle varied by less than 0.005 degrees due to the excellent effort of the MOSES group in limiting the roll during the single guide star observations.
The data were drizzled using Drizzle Version 1.2 February, 1998 ([@fru97]) with image offsets derived from the mean of the IRAF procedures since in cases where the IRAF positions differed from the world coordinates interactive inspection via the IDL tools confirmed the IRAF positions. As discussed above no geometric distortion correction or image rotation was required. High cosmic ray persistence noise levels after transit of the Southern Atlantic Anomaly required removal of 28 F160W integrations and 36 F110W integrations from the final mosaic image. A comparison of a straight combination of the drizzled frames and a combination averaged with $3\sigma$ clipping showed no differences indicating that the IDL cosmic ray removal techniques were effective.
### IRAF Image Mosaic
The IRAF reduction images were drizzled (Drizzle Version 1.1, Fruchter et al. 1997) with offsets determined from the centroid of the central star (NICMOS 249) in each frame. No rotation or geometric distortion corrections were necessary. Due to persistence of cosmic rays encountered in SAA passages, 25 F160W images and 37 F110W images were removed from the final mosaic. Which frames to remove was determined independently by inspection which leads to the slight difference from the number not included in the IDL image mosaic. The drizzled frames were averaged with 3$\sigma$ clipping to remove any residual low level cosmic rays.
THE IMAGES
==========
Figures \[Thompson.fig2.ps\] and \[Thompson.fig3.ps\] show the F110W and F160W images respectively, produced by the IDL image reduction and drizzle procedures described in the preceding section. The raster pattern of observations produces much lower signal to noise areas in the image at the edges where the number of overlapping integrations are greatly reduced. These areas are deleted from the image even though many strong sources are evident in these regions. The area covered by the images is a $49.19 \times 48.53 \arcsec$ ($481 \times 476$ pixels) rectangle. Figure \[Thompson.fig4.ps\] is a color composite of the two infrared images and the F606W WFPC2 image. The WFPC2 image has been rotated and resampled to fit the orientation and pixel size of the NICMOS images. The red, green, and blue colors represent the F160W, F110W, and F606W intensities. As with the original WFPC2 color image, the stretch and color curves have been manipulated to show faint objects while preserving the detail of features in the brighter objects. This image should not be used for quantitative purposes. Figures \[Thompson.fig2.ps\] and \[Thompson.fig3.ps\] are also stretched to show the best range of features. The very high dynamic range of the image can not be displayed in a linear intensity image.
SOURCE DETECTION AND PHOTOMETRY
===============================
Since the original WFPC2 HDF catalog ([@will96]) utilized KFOCAS to generate its listings our primary catalog listings also utilize KFOCAS to provide consistency. We also provide a description of the SExtractor source extraction process. The main difference between KFOCAS and FOCAS is the utilization of a supplemental image by KFOCAS that specifies the relative detectivity at each point in the image. This is important for the NICMOS HDF images where there are significant variations of quantum efficiency and total integration time over the image area.
Estimation of the input sigma for KFOCAS {#sig}
----------------------------------------
KFOCAS uses a constant $1\sigma$ level that is either determined from the first few lines of the image or is input manually by the user. Since the first few lines of the NICMOS image have much lower signal to noise than rest of the image we estimated the $1\sigma$ value manually from a histogram of the signal levels in all of the pixels. Figure \[Thompson.fig5.ps\] gives the histogram of the pixel values in the two dithered images shown in figures \[Thompson.fig2.ps\] and \[Thompson.fig3.ps\]. Only the pixels from the area covered by the images are used in this histogram. The histogram peaks at zero signal as expected for sky subtracted observations. The long extensions of the histograms toward positive values due to the sources in the field are cut off in this figure. The flat fielding process multiplies the true noise value in the image by the value of the flat field. This process raises the noise level in low quantum efficiency areas and lowers it in high efficiency areas. Since the median efficiency of the area is set to one this process should not appreciably alter the width of the curve.
As learned in the production of the WFPC2 images, the drizzling process produces a correlated image and hence correlated noise ([@will96], [@fru98]). There is approximately a factor of two reduction in the apparent noise as a result of the drizzle process for a factor of two reduction in linear pixel size. The numbers given in figure 5 should therefore be multiplied by a factor of 2 to determine the true $1\sigma$ value of the noise. This gives the noise figures of $1.22 \times 10^{-9}$ Jy for the F160W filter and $1.54 \times 10^{-9}$ Jy for the F110W filters. These are the powers that produce a signal equal to a $1\sigma$ noise in a single pixel. Use of these levels resulted in KFOCAS missing a large number of real sources easily identified by eye. We therefore dropped the $1\sigma$ estimates to a very low value of $5.5 \times 10^{-10}$ Jy or $2.0 \times 10^{-4}$ ADUs per second for the F160W filter and $3.5 \times 10^{-10}$ Jy or $2.3 \times 10^{-4}$ ADUs per second for the F110W filter. The number of $\sigma$ for the detection limit was then varied until all known real sources were detected without excessive over selection. It is of course the product of the chosen $1\sigma$ noise level and the number of sigma for detection parameter that determines the signal value a pixel must have to be considered a potential source. A known real source is an object easily seen by eye in the NICMOS images that is exactly coincident with an observed source in the WFPC2 HDF images. A more rigorous discussion of the completeness and reliability of the selected sources occurs in sections \[comp\] and \[rel\]. The catalog listings are limited to sources with signal to noise ratios that exceed or equal 2.5. This discards some sources that by many tests appear to be real but eliminates a large number of sources that have a significant chance of being false.
KFOCAS Reduction
----------------
We prepared the drizzled images for the KFOCAS procedure by multiplying the signal in ADUs per second by $10^{5}$ and subtracting the minimum value, a negative number, from the multiplied image. This produced an image that had no negative values and where all of the significant values were well represented in the integer arithmetic used by KFOCAS. The zero point magnitudes of the modified images are 35.3 for the F160W image and 35.186 for the F110W image. The source extraction utilized the standard KFOCAS procedures of the series KDETECT, KSKY, KEVALUATE, KSPLIT and RESOLUTION. Our drizzled pixels have 6.25 times the area of the WFPC2 drizzled pixels, therefore, we set the minimum area for detection in pixels to 2, to avoid missing very compact galaxies. The parameters for the KFOCAS reduction are listed in Table \[kfoc\]. The point spread function (PSF) matrix for smoothing the data is a $3 \times 3$ matrix that mimics the PSF of the central star in the drizzled data. This is much more sharply peaked than the Gaussian function used in the SExtractor analysis.
### Preparation of the detectivity image {#qualim}
Each pixel, $pix_{i,j}$, in the final image has a quality $Q_{i,j}$ value associated with it. The quality value is the square root of the sum of the squares of the total efficiency of each pixel in the individual image that contributes to the final image. Due to the raster and dither pattern a pixel in the final image has contributions from many different individual image pixels. $$Q_{i,j} = \sqrt{\sum_{k=1}^n (eff_{k})^2}$$ Here n is the number of pixels contributing to the final image pixel, $pix_{i,j}$, and the efficiency $eff_{k}$ of each contributing pixel is measured by the inverse of its multiplicative flat field value. Figures \[Thompson.fig6.ps\] and \[Thompson.fig7.ps\] show the detectivity functions over the image areas used in the catalog.
SExtractor Reduction
--------------------
As a cross check on the KFOCAS detections we utilized an alternative source extraction system, SExtractor, on data reduced via the IRAF reduction procedure rather than the IDL based procedure.
### Galaxy Detection
We performed object detection on the F160W images and photometry on the F160W and F110W images using SExtractor version 2.0.7 ([@Ber96]). The final F160W and F110W images reduced in the IRAF pipeline showed low frequency structure in the background in the X-direction. We created background frames in SExtractor using a $64 \times 32$ mesh which were subtracted from the reduced frames, producing cosmetically more uniform backgrounds. We measure $1\sigma$ noise levels from histograms of all the pixels in the frames of $2.0 \times 10^{-4}$ ADUs per second in the F160W image and $2.4 \times
10^{-4}$ ADUs per second in the F110W frame, consistent with the values shown in Fig. \[Thompson.fig5.ps\]. The amplitude of the fluctuations in the background varies by up to 50 % across the image due to variations in the quantum efficiency of the detector and the dither pattern used. Thus we used the option in SExtractor (WEIGHT\_TYPE) which accepts a user supplied variance map (for which we used the ‘detectivity’ function described in Section \[qualim\]). SExtractor robustly scales the weight map to the appropriate absolute level by comparing the weight map to an internal, low resolution, absolute variance map built from the science image itself. In contrast to the KFOCAS source extraction, all object detection was done on the F160W image and magnitudes were measured to the corresponding isophotes on the F110W image.
After experimenting with different values of SExtractor parameters, we adopted the values given in Table \[t\_sextparam\]. Aside from the determination of the local variance itself, the three most critical parameters which affect the detection of very faint isolated sources are FILTER\_NAME, DETECT\_MINAREA and DETECT\_THRESH. The FILTER\_NAME parameter describes the smoothing kernel which is applied to the image and for this a Gaussian with a full width half-maximum of 2.0 (drizzled) pixels was used over a 3 x 3 pixel grid. We tested various combinations of DETECT\_MINAREA, the minimum number of contiguous pixels above a level which is the product of the parameter DETECT\_THRESH times the local RMS fluctuation in the background. Our final choices for these parameters are DETECT\_MINAREA = 2 pixels and DETECT\_THRESH = 2.15$\sigma$. This choice for DETECT\_MINAREA (the same value as for the equivalent KFOCAS parameter) favors slightly the detection of the most compact sources and the final choice of 2.15 for DETECT\_THRESH was dictated by an attempt to strike a judicious balance between completeness and reliability. We tested an alternative set of parameters with DETECT\_MINAREA set to 3 pixels and with DETECT\_THRESH to a lower value in order to detect the same number of sources as with DETECT\_MINAREA =2 and DETECT\_THRESH=2.15. Estimating the number of false detection rates and completeness as discussed in sections \[comp\] and \[rel\], we found that these two sets of parameters behaved very similarly. For the reason stated above, we selected DETECT\_MINAREA=2. Sections \[comp\] and \[rel\] contain a more quantitative discussion of the completeness and reliability of the detected sources.
Although like the KFOCAS reductions we have deliberately erred on the side of extracting a fairly large estimated fraction of false detections (at the faintest levels) and a completeness level which is only of order 50%, the catalog listings only contains sources with signal to noise ratios greater than 2.5.
Comparison of the IDL-KFOCAS and the IRAF-SExtractor Photometry {#kscomp}
---------------------------------------------------------------
At this point our analysis contains four components, the IDL and IRAF reduced images and the KFOCAS and SExtractor source extractions and photometry. In the spirit of independent cross checks it is useful to compare these results and to see if any differences lie primarily in the images or in the source extraction procedures. We will discuss the differences in completeness and reliability between the two methods after sections \[comp\] and \[rel\]. Data presented so far have been for either the IDL-KFOCAS or IRAF-SExtractor procedures. The third panel in fig. \[Thompson.fig8.ps\] shows a comparison between the aperture magnitudes found by KFOCAS in the IDL image and the aperture magnitudes found by SExtractor in the IRAF images for all objects detected in common. The delta magnitudes are KFOCAS magnitude minus SExtractor magnitude. In both cases the diameter of the aperture is 0.6.
As expected the correlation is very good at bright magnitudes and gets worse at the faint end. The SExtractor magnitudes are equal to the KFOCAS magnitudes up until 28th magnitude where the SExtractor magnitudes become significantly brighter than the KFOCAS magnitudes. These differences could be either to differences in the input images or in the magnitudes extracted by KFOCAS and SExtractor. To check this we ran SExtractor on both the IDL reduced and the IRAF reduced images. The first panel in fig. \[Thompson.fig8.ps\] shows the result of this test which shows a uniform scatter about the zero level at the fainter magnitudes. The slight offset at the bright end is probably due to the IRAF procedure IMCOMBINE clipping some pixels in the brightest galaxies in the IRAF image. A follow up test comparing the KFOCAS reductions of the IDL image with SExtractor reductions of the IDL images is shown in the middle panel of fig. \[Thompson.fig8.ps\]. This plot is essentially identical to the last panel except for about a 0.07 magnitude offset at the brighter end. This set of tests shows that the differences between the KFOCAS/IDL magnitudes and the SExtractor/IRAF magnitudes shown in the last panel of fig. \[Thompson.fig8.ps\] are entirely due to the differences between the KFOCAS and SExtractor algorithms, not from any differences between the IDL and IRAF image production procedures. The origin of this difference is not clear but the reader should be aware that these two standard procedures do produce differences at the very faintest levels.
THE CATALOG {#cat-ref}
===========
Table \[cat\] contains the catalog of sources from the KFOCAS source extraction from the F160W and F110W images. This catalog only contains sources with signal to noise ratios greater than 2.5. We anticipate a future publication describing the sources with less reliable detections. The catalog contains 342 objects, some of which are components of a larger object. The catalog contains 235 objects in common with the WFPC2 catalog. 221 objects have detections in both filters, 56 objects have only a detection in the F160W filter, 53 have only a detection in the F110W filter, and none have detections only in SExtractor. The objects are arranged by right ascension which sometimes separates different components of the same object in the catalog. The data and numbering in the catalog have the priorities, in order, of F160W KFOCAS, F110W KFOCAS, and SExtractor. This means that all KFOCAS objects detected in both filters use the KFOCAS F160W RA, DEC, x, and y positions. The magnitudes come from the KFOCAS F160W and F110W extractions. Positional coincidence is relative to the F160W positions. Objects that have F110W KFOCAS detections but not KFOCAS F160W detections use the F110W KFOCAS positions and magnitudes. The catalog columns contain the following parameters.
**ID:** This is a running number for each object. The numbers after the decimal point indicate the level of splitting by KFOCAS up to three levels of daughter objects. Since the list is arranged by right ascension daughter objects can appear separately from the parent objects. No object is repeated. Numbers of 900 or higher are split F110W objects that are not coincident with any F160W split even though some of the components are in common.
**WFPC:** The WFPC column lists the nearest WFPC2 source from the [@will96] catalog. This is not necessarily the same object, just the nearest.
**s:** The s value is the separation in arc seconds between the NICMOS and the nearest WFPC2 object as listed in the WFPC column. A large value of separation indicates that the NICMOS and WFPC2 object are probably not associated.
**x,y:** These columns give the x and y values of the centroid of the source in the F160W or F110W image. If the object is detected in both images the x and y values refer to the F160W image. Objects detected only by SExtractor have the values determined by SExtractor. This order of precedence holds for all of the subsequent values.
**RA,DEC:** These columns give the right ascension and declination of the centroid of the source. Only the minutes and seconds are listed. The hour of right ascension is 12 hours for all sources and all sources have 62 degrees of declination. The source positions assume that the central star, NICMOS 145 and the WFPC2 object 4-454 have the same position and that the measured plate scales of the NICMOS camera 3 are correct. In this sense all positions are relative to the position of the WFPC2 4-454 object.
$\mathbf{t_{160}}$,$\mathbf{i_{160}}$,$\mathbf{a_{160}}$,$\mathbf{t_{110}}$,$\mathbf{i_{110}}$,$\mathbf{a_{110}}$: These are the total, isophotal and aperture magnitudes found by KFOCAS in the F160W and F110W images. The aperture diameter for the aperture magnitudes is $0.6\arcsec$. The total and isophotal magnitudes are as described in [@will96]. A value of 99.99 indicates that the object was not detected in that filter. The F160W and F110W objects are considered to be in common if they lie within $0.25\arcsec$ of each other. If the last digit in the tests parameter (see below) is 3, the magnitudes are from the SExtractor procedure.
**S/N:** The signal to noise value quoted in the catalog is calculated by the same technique used in the optical HDF. The value of the signal to noise is $S/N = L_{i} / \sigma (L_{i})$ where the variance $[\Gamma \sigma (L_{i})]^{2}$ is, $$[\Gamma \sigma(L_{i})]^{2} = \Gamma N_{obj} + 1.9^{2} \Gamma^{2} \sigma_{sky}^{2} A_{obj} + 1.9^{2} \Gamma^{2} \sigma_{sky}^{2} A_{obj}^{2} / A_{sky}$$ as quoted in its correct form by ([@poz98]). For NICMOS the value of $\Gamma$ in electrons per ADU is 6.5. $L_{i}$ is the sky subtracted number of counts, $\sigma_{sky}$ is the sky sigma in ADUs, and the object and sky areas $A_{sky}$ and $A_{obj}$ are the areas in pixels returned by KFOCAS. This equation reformulated in terms of count rates in ADUs per second is given by $$\sigma_{tot}^{2} = rate_{obj} / (\Gamma t) + 1.9^{2} \sigma_{sky rate}^{2} A_{obj} + 1.9^2 \sigma_{sky rate}^{2} A_{obj}^{2}/A_{sky}$$ Here $rate_{obj}$ is the sky subtracted source count rate in ADUs/sec, $\sigma_{sky rate}$ is the sky sigma value in ADUs/sec, and t is the integration time in seconds. The final signal to noise is $rate_{obj} / \sigma_{tot}$. The measured value of the sky sigma is $2.2 \times 10^{-4}$ ADUs per second for the F160W image and $2.5 \times 10^{-4}$ for F110W as shown in Fig. 2. The factor of 1.9 in each of the equations is the estimated value of the noise correlation discussed earlier. Figure \[Thompson.fig9.ps\] shows a plot of the signal to noise ratios calculated by this method for the KFOCAS determined sources.
**A:** This is the isophotal area of the source in square arcseconds as determined from the value returned by KFOCAS.
$\mathbf{r_{1}}$: This is the half light radius returned by KFOCAS.
**tests:** This parameter indicates which of the various reliability tests the source passed. A source that passed all tests has a value of 22111, one that passed no tests has a value of 00000. The first number is 0,1,or 2 if the source was detected in none, one, or both of the F160W half catalogs. For an explanation of the half catalogs see Section \[half\]. The second number is the same test in the F110W catalog. The third number is 0 or 1 depending on whether the source was found in both NICMOS bands of the full image extractions. The fourth number is 0 or 1 depending on whether the source is detected in the WFPC2 catalog. The last number is 0 or 1 depending on whether the source was found in the independent Sextractor catalog. Note the discussions in sections \[krel\] and \[srel\] on the differences in half catalog detection probabilities between KFOCAS and SExtractor. In all cases a common detection means that the source centroids lie within 0.25 arc seconds of each other. No color or magnitude tests are applied as part of the common object association.
Completeness {#comp}
============
Our calculated $50\%$ completeness levels are KFOCAS AB aperture magnitudes of 28.7 and 28.2 for point objects and extended sources in the F160W filter and KFOCAS AB magnitudes of 28.6 and 28.0 respectively in the F110W filter. These limits are based on a technique of adding sources to the image at various flux levels and running the source extraction programs to see what percentage of the added sources are recovered. These numbers are based on the KFOCAS reductions.
KFOCAS Completeness {#kcomp}
-------------------
The test for the KFOCAS program established a regular grid of 49 positions in a 7 by 7 pattern that evenly covered the image area utilized in the catalog extraction. Sources are placed at these positions to create a new image which contains the original image plus the added sources. KFOCAS then creates a new catalog utilizing exactly the same parameters used in the final catalog preparation. An automatic program checks the new catalog to see what percentage of the added sources are recovered by KFOCAS. The sources are sequentially dimmed in half magnitude steps from their original magnitudes of about 21 to the final magnitude of 32. The added sources are real sources extracted from the final image. The NICMOS source 145, the central star, represents a point object and the elliptical galaxy, NICMOS 212, is the extended object. These are WFPC2 sources $4-454$ and $4-471$ respectively. Alteration of the position of the grid confirmed that the completeness limits were consistent at different grid positions. The extended source completeness limit will of course be brighter for more extended objects than NICMOS 212 but this galaxy is one of the largest in the image.
The completeness value at a given magnitude is for the magnitude of the input source not the magnitude at which the source is recovered. This value is approximately equal to the recovered magnitude for sources brighter than 26.5. At the very faintest levels the recovered magnitudes are 0.3 to 0.5 magnitudes fainter than the input magnitudes. The measured completeness limits are fit by $$(1 - {\frac{m - m_{b}}{m_{a} - m_{b}}})^n \sqrt{\frac{m_{b}}{m}}$$
In this function $m_{a}$ is the magnitude at which the completeness goes to zero, $m_{b}$ is the magnitude where the completeness is $100\%$ and the power index n is adjusted to fit the data. The function is purely empirical, simply designed to fit the data well. This function smooths the curves and provides the interpolation from the observed magnitudes to the magnitudes used in table \[cr\]. As expected from the nonuniform quality map of the image, the completeness limit is not uniform over the image. The completeness is somewhat lower in the right hand half of the image. The completeness limits in table \[cr\] should be considered as an average across the image.
SExtractor completeness
-----------------------
For the SExtractor catalog we used an analysis similar to the one described in Yan et al. (1998). We selected a compact galaxy (4-289 in the Williams catalog), representative of the majority of the objects in our field, and dimmed it in half-magnitude increments from 25.5 to 29. The dimmed galaxy was dropped at random positions 10,000 times, superposed on the full image and two images with half the exposure time of the full image. (The “half” images are discussed in section \[half\].) We ran SExtractor at the position where the galaxy was dropped at each iteration to determine whether or not the dimmed galaxy was recovered and at what magnitude. The use of random positions in the simulations allows us to include completeness corrections arising from non-detections and magnitude errors caused by crowding and spatially dependent errors in the sky subtraction and flat field correction. As discussed in section \[kcomp\], the completeness values for recovery of the images are again an area-average of the completeness since the sensitivity across the NICMOS images is not uniform. These same experiments also give us the matrix relating the input and output aperture magnitudes for those galaxies which are recovered. The input magnitude and mean recovered magnitude agrees to within 0.07 magnitudes in the bins through 28. At 28.5 the mean of the recovered galaxy magnitudes begins to brighten (28.21) and by an input magnitude of 29 it brightens substantially (28.11). The galaxies that land on negatives noise regions are lost completely. The SExtractor catalog becomes 50% complete at AB$\approx$28.3 for the compact galaxies in our survey.
To test differences in detectability of the point spread function due to the source landing at different locations within the 0.2" pixels, we ran incompleteness tests using the central star (NICMOS 145, AB=22.1) taken from 5 different individual exposures, all taken at different dither positions. These have only 1/120 the exposure time of the final image and therefore have lower signal-to-noise ratios. We found no substantial differences in detection rates, other than as the star is dimmed to AB$=$28, it is missed most frequently in the upper right hand quadrant of the detector, the least sensitive region of the image. In the final combined image, we found no obvious location dependence. We detect the star 95% of the time as it is dimmed to AB$=$27, and then the detection rate drops rapidly to 50% complete at AB$=$28.5.
Source Reliability Tests {#rel}
========================
Even though the listed catalog does not contain objects with signal to noise ratios less than 2.5 there can be false detection still in the catalog. As indicated in Section \[cat-ref\] the catalog indicates the degree of coincidence between the various subcatalogs that make up the total catalog. This data is provided as an aid in discerning the reality of the sources. Any statistical study of these results should utilize the test flag indices of Table \[cat\] carefully along with the completeness and reliability results discussed here and in Section \[comp\] and summarized in Table \[cr\]. See section \[crt\] for a discussion of this table.
Half Data Reductions {#half}
--------------------
Our primary test of the reliability of observed sources utilizes two independent images formed from subsets of the integrations in each filter. The two images contain the even and odd numbered integrations from a sequential numbering of the integrations after removal of images with excess cosmic ray persistence. Since there are three images per orbit this technique insures that orbits are mixed between the groups and that each group has an equal mix of images observed at different times during the orbit. The width of the histograms of pixel values in the half images are a factor of $\sqrt{2}$ wider than the full image histograms. This is a good indication that the width of the histogram in Fig. \[Thompson.fig5.ps\] is due to noise rather than faint sources.
These half data reductions are the primary tests as they represent truly independent sets of data that measure the same quantity. Although useful, the coincidence between the KFOCAS and SExtractor catalogs are not measurements of two independent data sets. The coincidence between the objects detected in the various NICMOS and WFPC2 filter sets are again useful but they are not measuring the same quantity. As with the completeness tests, the checks on the KFOCAS and Sextractor image catalogs are carried out independently. With slight modifications, however, the logic of the tests is essentially identical.
### Logic of the half catalog tests {#logic}
Our goal is a measurement of the probability that a detected source with a given magnitude range is real. To facilitate the comparison between KFOCAS and SExtractor tests we both utilize aperture magnitudes in a $0.6\arcsec$ diameter aperture. We start by grouping sources into half magnitude bins centered on integer and half magnitudes. Our analysis method then utilizes the statistics of objects detected in both, only one, or neither, of the independent half catalogs for each aperture magnitude bin. We consider objects as being present in both catalogs if their centroids are within 0.25 arc seconds (2.5 drizzled pixels) of each other.
From the completeness studies described in Section \[comp\] we determined the probability $P_{A,B}(j,k)$ that an object in a magnitude bin j is recovered in a bin k where A or B refers to one of the half catalogs. The completeness is then C$_{\rm A}(j) = {\displaystyle\sum_k} {\rm P_A}(j,k)$ and similarly for image B. Let N$_{\rm R}(j)$ be the number of real objects whose true magnitudes lie within bin j. In addition, due to noise fluctuations (both Poisson and non-random), there will be some false objects detected in bin $j$. Let f$_{\rm A}(j)$ be the probability that any object in image A in bin $j$ is a false detection and let N$_{\rm A}(j)$ be the number of objects found in bin $j$ on image A. Then $$\label{sext_equ1}
{\rm N_A}(j) = \sum_k {\rm N_R}(k) \times {\rm P_A}(k,j) + {\rm f_A}(j) \times {\rm N_A}(j)$$ and $$\label{sext_equ2}
{\rm N_B}(j) = \sum_k {\rm N_R}(k) \times {\rm P_B}(k,j) + {\rm f_B}(j) \times {\rm N_B}(j).$$
We then count the number of objects $N_{AB}(i,j)$ which appear in common in both half images A and B, (ie, which agree in position to within 0.25" of each other) and which have measured magnitudes in A which place them in bin i, and measured magnitudes in B which place them in bin j. Then
$$\label{sext_equ3}
{\rm N_{AB}}(i,j) = \sum_k {\rm N_R}(k)\times {\rm P_A}(k,i) \times {\rm P_B}(k,j)$$
Strictly speaking, we should add to equation \[sext\_equ3\] a term which represents the number of times that a false detection in both half images will be coincident to within 0.25" and land in the two magnitude bins in question. In practice this number is small compared to one, and we neglect it.
If $N_{bins}$ is the number of magnitude bins, then equations \[sext\_equ1\] and \[sext\_equ2\] hold for the $N_{bins}$ values of j, and in equation \[sext\_equ3\] for the $ N_{bins} \times N_{bins}$ combinations of i and j. The unknown quantities are the $N_{bins}$ values for the number of real sources with true flux placing them in bin $N_R(k)$ along with the $N_{bins}$ estimates for the probabilities $f_A(j)$ and $f_B(j)$ that a given source in bin j is not a real source. Obviously, the system is over–determined. This is to be expected since equations \[sext\_equ1\], \[sext\_equ2\] and \[sext\_equ3\] are just discrete representations of integral equations describing the observed number count distribution from which we are trying to recover the true distribution, taking into account losses, false sources and errors between the true and measured magnitudes due to noise.
A simplification of the preceding equations which is useful for illustrative purposes and actual calculations in some cases comes from ignoring the cross terms and letting the completeness in any bin be equal to C independent of which half catalog is addressed. This eliminates the cross terms in equations \[sext\_equ1\], \[sext\_equ2\], and \[sext\_equ3\].
In that case, we obtain for each bin,
$$\label{simp1}
N_{R} = N_{AB}/C^2$$
and $$\label{simp2}
f_{A} = 1 - N_{AB}/(C \times N_{A})$$ with a similar expression for $f_B$.
### Negative Image Tests {#negim}
As described below, we are limited in the applicability of the full formalism described above due to small number statistics in the observed number of objects, which can lead to reliability estimates greater than one. An alternative procedure is to multiply the final images by -1.0 and search for “detections” of objects in these negative images. This assumes that the noise properties of the images are the same for negative excursions as for positive ones. This is not in general true since, for example, cosmic rays which are not completely removed have no counterpart in the negative image. In the case of the NICMOS HDF images, however, not only are the cosmic rays removed fairly effectively within each frame as a consequence of the non-destructive readout, but the very large number of dithered frames making up our final images reduces any residual cosmic rays by a further larger factor. Unfortunately, we have found that this method does not appear to be well-suited for the KFOCAS extractions for reasons associated with edge effects near the large negative “holes” in the counts produced by the bright real sources in the negative images. However, this method does seem to yield useful results for the SExtractor algorithm. We now discuss the particular tests actually carried out on the KFOCAS and SExtractor half catalogs.
KFOCAS Half Catalog Tests {#krel}
-------------------------
The KFOCAS analysis of each half image used the same parameters as the total image analysis. Since the images are in units of photon flux the half images have the same signal strength for true sources but have a higher noise. Unlike SExtractor, the $1\sigma$ noise level for KFOCAS is an input parameter. Retention of whole catalog input parameters results in an input $1\sigma$ noise value that is a $\sqrt{2}$ lower than the noise in the half catalog. The half image KFOCAS analysis then detects more sources since more random noise fluctuations appear above the detection threshold. Bright true sources should be detected in both images. Faint sources of course could be detected in only one or even neither of the half images. Each source is marked in the catalog test column as to whether it appeared in both, only one or none of the half images.
In practice for the KFOCAS source we utilize the simplified formalism described in equations \[simp1\] and \[simp2\] of Section \[logic\]. In particular we note that the reliability in either half catalog is $1 - f_{A}$ or $1 - f_{B}$ so we can say that the reliability r is
$$\label{simp3}
r_{A} = N_{AB}/(C \times N_{A})$$
Since all of the quantities on the right side of equation \[simp3\] are known r can be calculated using the values of C previously determined. However, when this is formally carried out the values of r for some magnitude bins become greater than 1 due to a low value of C for that bin or small number statistics. Since the true value of the completeness can never exceed one we can get a robust lower limit on r by setting C equal to 1 and noting that
$$\label{simp4}
r_{A} \geq N_{AB}/N_{A}$$
with again a similar equation for $r_{B}$. This equation depends solely on the ratio of the number of sources detected in both catalogs to the number detected in one of the half catalogs. The final reliability for a magnitude bin is just the average of $f_{A}$ and $f_{B}$. This reliability is appropriate for the half catalogs. The signal to noise in the whole catalog is a factor of $\sqrt{2}$ higher. This corresponds the objects in the half catalog that are the same factor brighter. This is an offset of 0.376 magnitudes, therefore, the calculated reliability numbers in the half catalog are appropriate for sources that are 0.376 magnitudes fainter in the whole catalog.
Table \[cr\] contains the results of these tests under the section marked KFOCAS. The completeness values listed in the table are the values found from the analysis in Section \[kcomp\]. The reliability numbers are the numbers from the above calculation adjusted to the appropriate aperture magnitude for the total catalog. These values are in general lower than those calculated using the measured values for completeness in equation \[simp3\].
Sextractor Half Catalog Tests {#srel}
-----------------------------
For both the whole image and half images, we use the SExtractor parameters given in table \[t\_sextparam\] but determine the completeness independently for the half images using the same procedure described in \[comp\].
As noted above, the noise properties in the half images scale almost exactly as expected, so that the false detection rate at a given magnitude bin in the half images should be applied to a magnitude bin in the whole image fainter by $1/\sqrt{2}$ lower in the flux, or 0.376 magnitudes. We then use the mean of the two false detection rates determined from the half images for the estimate of the false detection rate at this slightly fainter magnitude.
As described in \[logic\], the system of equations \[sext\_equ1\]–\[sext\_equ3\] for ${\rm N_R}(k)$ is over-determined and we determined these values by a least squares fit to the observed values ${\rm N_{AB}}(i,j)$. Equations \[sext\_equ1\] and \[sext\_equ2\] then give the false detection rates for the two half images, and use the mean of these determinations for the whole image as explained above.
In practice, as already discussed in \[krel\], the small number of sources actually detected in common in the two half images result in uncertainties in the reliability estimates for magnitudes at which the completeness is near unity. We have also estimated the reliability of the detections by the negative image method described in \[negim\]. The objects which SExtractor finds using this technique do not seem to occur preferentially near the “holes” associated with the negative sources but occur in the higher signal to noise regions, as expected, so that SExtractor does not seem subject to the same degree to the edge effects described in \[negim\] for KFOCAS.
The Completeness and Reliability Table {#crt}
--------------------------------------
The completeness and reliability table, Table \[cr\], summarizes the results of our tests described above. Columns 2–8 refer to the KFOCAS reductions only and the last three refer to the results from SExtractor. As described in \[kscomp\], there is a systematic difference between aperture magnitudes measured by KFOCAS and SExtractor which becomes significant for objects fainter than $\approx 28.0$, as shown in Figure \[Thompson.fig8.ps\]. Thus, in the first column, the magnitude (Mag.) is the aperture AB magnitude measured by KFOCAS for the KFOCAS reductions and by SExtractor for the SExtractor reductions. The width of the magnitude bin is 0.5 magnitudes centered on the value in the magnitude column. The signal to noise ratio (S/N) is the average signal to noise ratio for all objects in the magnitude bin. The columns labeled $Cs_{16}$, $Cl_{16}$, $Cs_{11}$ and $Cl_{11}$ are the completeness numbers for the KFOCAS reduction listed in order for small and large objects in the F160W filter and the F110W filter, respectively. Next are the reliability numbers for the F160W and F110W filters where no discrimination has been made between small and large objects. Completeness and reliability SExtractor results for the F160W filter comprise the last three columns, where the column labeled $R_{co16}$ uses the full half image formalism, while the column labeled $R_{neg16}$ uses the negative image technique. It should be emphasized that the reliability estimates at the faint end of the table are subject to considerable uncertainty. However the results all seem to indicate a fairly low incidence of false detections $\sim 5-15 \%$ at magnitude $\sim27.5$ but this incidence rises steeply at fainter magnitudes, while the completeness is of order $\sim 80-90\%$ at magnitude 27.5, of order $\sim 70-75\%$ at magnitude 28.0 and falls rapidly beyond that point. KFOCAS appears to lose more objects due to merging with other objects than SExtractor particularly for bright objects. This is probably the cause of the less than $100\%$ completeness at bright magnitudes for the KFOCAS reductions. The low number of objects in the brighter bins limits the accuracy of the measurements and differences of $ \pm 5\%$ should not be considered significant.
Our discussion of the differences in source extractions in section \[kscomp\] clearly indicates that the IDL and IRAF images are essentially identical and that the magnitude differences in fig. \[Thompson.fig8.ps\] are solely due to differences between the two extraction programs, KFOCAS and SExtractor. There are also differences in the number of detections between the two programs. Running SExtractor on the IDL image with the same set of parameters used for the SExtractor analysis of the IRAF images we find 284 galaxies, somewhat less than the 356 found on the IRAF image. On the other hand there are a total of 350 objects selected by KFOCAS from the whole IDL F160W image, also more than those found by SExtractor on the IDL image. The total number of objects found in the F160W image by KFOCAS and in the IRAF image by SExtractor are nearly identical.
Inspection of table \[cr\] indicates the range of reliability and completeness measures returned by the two methods. In general the reliability and completeness measures from the KFOCAS analysis fall below those determined via the SExtractor analysis. This is particularly true when you consider the difference in faint magnitudes discussed in section \[kscomp\]. This indicates that at the faintest magnitudes the KFOCAS numbers should be compared with the SExtractor numbers for sources with SExtractor magnitudes nearly a magnitude brighter than the KFOCAS magnitude. Part of this difference in reliability is due to the KFOCAS numbers being lower limits on the reliability as discussed in section \[krel\]. Another part of the difference, however, is due to the uncertainty inherent in these calculations and users of this catalog should be aware of them. Our net philosophy is to be aggressive in identifying potential sources but to be relatively conservative in calculating their reliability and completeness.
GALAXY COUNTS
=============
As with the original optical catalog of [@will96] it is not our intention in this paper to discuss scientific results. A commonly used statistic, however, is the differential number count of galaxies. Fig. \[Thompson.fig10.ps\] presents this statistic for the region of the HDF covered by our catalog. The galaxy counts in number per magnitude per square degree have been divided into half magnitude bins. If the object radius is less than $0.3 \arcsec$, the aperture magnitude is used. If the object radius is greater than this the isophotal magnitude is used. The number counts for the F814W filter in the same area as determined from the optical catalog are also shown for comparison. The same division between aperture and isophotal magnitudes are used for the F814W data. The counts for the WFPC2 F814W objects includes all of the WFPC2 catalog objects in the NICMOS field, not just those in common with NICMOS objects. There are no aperture corrections applied to this data in order to facilitate comparison with the statistics presented in [@will96]. It should be noted that the area covered by the NICMOS image is very small. For a value of $H_{o} = 50$ and a value of $\Omega_{o} = 1$ the sides of the NICMOS image are on the order of 250 kpc for redshifts greater than 1. This is about 10 times smaller that the typical diameter of a region forming a single galaxy from Cold Dark Matter simulations ([@stz98]). Drawing any cosmological conclusions from this small sample may be very suspect. Following the discussion in [@will96] we have not eliminated split objects from the number counts. Except at very bright magnitudes we do not expect this to significantly affect the statistics.
CONCLUSIONS
===========
The NICMOS observations of the Hubble Deep Field add significant value to the existing data by providing improved wavelength coverage and access to objects that are either too heavily extincted or too highly redshifted to be visible in the original optical catalog. This paper is designed to be a reference source for the use of this data in various areas of research. Future papers will discuss various aspects of the significance of these data.
ACKNOWLEDGEMENTS
================
Many people contributed to the NICMOS observations of the Hubble Deep Field. The entire NICMOS Instrument Definition Team contributed to the success of NICMOS and participated in the decision to allocate a large fraction of the team’s guaranteed time to this effort. We wish to thank Mark Dickinson for help with the KFOCAS reduction techniques and Andy Fruchter for his aid in implementing the Drizzle image procedure. Andy Lubenow spent many hours refining our observation plan to handle the single guide star acquisition. Chris Conner and the Lockheed MOSES group went to extraordinary efforts to keep the gyro biases updated to ensure good pointing under single guide star tracking. Zolt Levay provided invaluable assistance in the preparation of the images for publication and Dr. Bertin provided the SExtractor software and quick response to inquiries. The schedulers at STScI diligently worked to minimize the impact of SAA crossings. LS-L thanks Lin Yan and Patrick McCarthy and RJW thanks David Koo for very useful discussions on incompleteness testing and galaxy surveys. This work was supported by NASA grant NAG 5-3043 and the observations were obtained with the NASA/ESA Hubble Space Telescope operated by the Space Telescope Science Institute managed by the Association of Universities for Research in Astronomy Inc. under NASA contract NAS5-26555.
Adelberger, K., Steidel, C.C. 1996, Private Communication to STScI.
Bershady, M.A., Lowenthal, J.D., and Koo, D.C. 1998, ApJ, in press
Bertin, E., and Arnouts, S. 1996, , 117, 393.
Conner,C. 1998, private communication from the Missions Operations Support Engineering Systems.
Connolly, A.J., Szalay, A.S., Dickinson, M., SubbaRao, M.U., and Brunner, R.J. 1997, , 486, L11.
Dickinson et al. 1997 HST Approved Cycle 7 Observing Plan.
Fruchter, A. 1998, private communication.
Fruchter, A.S. and Hook, R.N. 1997, in Applications of Digital Image Processing XX, Proc SPIE, Vol. 3164, ed. A. Tescher, in press.
Fruchter, A., Hook, R.N., Busko, C., and Mutchler, M. 1997, in 1997 HST Calibration Workshop, ed. S. Casertano et al., 518.
Jarvis, J.F. and Tyson, J.A. 1981, , 86, 476.
Lehar, J., Falco, E.E., Impey, C.D., Kochanek, C.S., McLeod, B.A., Munoz, J., and Rix, H.-W. 1998, in preparation.
McLeod, B. 1997, in 1997 HST Calibration Workshop, ed. S. Casertano et al., 281.
Pozzetti, L., Madau, P., Zamorani, G., Ferguson, H.C., and Bruzual, G. 1998, , in press.
Steinmetz, M. 1998, in Space Telescope Symposium on the Hubble Deep Field, in press.
Valdes, F. 1982, Faint Object Classification and Analysis System (KPNO Internal Publication)
Williams, R.E., Blacker, B. Dickinson, M., Dixon, W.V.D, Ferguson, H.C.,Fruchter, A.S., Giavalisco, M., Gilliland,R.L., Heyer, I. Katsanis, R., Levay, Z. Lucas, R.A.,McElroy, D.B., Petro, L., Postman, M. Adorf, H-M., and Hook, R.N. 1996, , 112, 1335.
Yan, L., McCarthy, P.J., Storrie-Lombardi, L.J. & Weymann,R.J. 1998, , 503, L19.
[cccccccc]{} $12^{\rm h} 36^{\rm m} 45.129^{\rm s}$ & $+62\arcdeg 12\arcmin 15.55\arcsec$ &$261.851\arcdeg$ &$57.87 \times 61.34\arcsec$ &49 orbits &49 orbits&12 orbits&15 orbits
[crr]{}
Magnitude zero point &35.3 &35.186Catalog magnitude limit &100. &100.Radius of fixed circular aperture &6.0 &6.0Sigma of sky &20 &22Sigma above sky for detection &2.5 &3.0Sigma below sky for detection &20.0 &20.0Minimum area for detection &2 &2Significance for evaluation and splitting &0.15 &0.15
[ll]{} DETECT\_MINAREA & 2\
DETECT\_THRESH & 2.15\
ANALYSIS\_THRESH& 2.15\
FILTER\_NAME & gauss\_2.0\_3x3.conv\
CLEAN & N\
MASK\_TYPE & CORRECT\
WEIGHT\_TYPE & MAP\_WEIGHT\
PHOT\_APERTURES & 6\
GAIN & 6.5\
PIXEL\_SCALE & 0.1\
SEEING\_FWHM & 0.2\
BACK\_SIZE & 64\
BACK\_FILTERSIZE& 3\
BACKPHOTO\_TYPE & LOCAL\
BACKPHOTO\_THICK& 24\
MAG\_ZEROPOINT & 22.80 (F160W)\
MAG\_ZEROPOINT & 22.68 (F110W)\
[ccccccccccccccccc]{} 1.00000 &4-851.0& 0.31& 560.40& 134.80 &36:40.81 &12:9.258& 28.56& 28.86& 28.54& 28.85& 29.17& 28.47& 2.553& 0.180& 0.163&20100 2.00000 &4-830.0& 0.10& 561.00& 152.00 &36:40.95 &12:10.67& 27.43& 27.76& 27.42& 27.87& 28.21& 27.92& 6.434& 0.220& 0.122&21111 3.00000 &4-810.0& 0.08& 579.80& 180.00 &36:40.98 &12:14.13& 28.18& 28.78& 28.08& 99.99& 99.99& 99.99& 3.502& 0.190& 0.084&10010 4.00000 &4-851.0& 2.45& 540.50& 127.00 &36:40.98 &12:7.445& 99.99& 99.99& 99.99& 28.86& 29.28& 28.63& 2.804& 1.142& 0.050&00000 5.00000 &4-822.0& 0.08& 550.08& 158.03 &36:41.13 &12:10.50& 24.50& 24.65& 25.33& 24.88& 25.04& 25.44& 22.66& 2.360& 0.277&22111 6.00000 &4-813.2& 1.48& 526.50& 135.00 &36:41.22 &12:7.239& 99.99& 99.99& 99.99& 28.80& 29.46& 28.41& 2.974& 0.968& 0.043&01000 7.00000 &4-766.0& 0.91& 574.50& 201.00 &36:41.23 &12:15.56& 99.99& 99.99& 99.99& 28.89& 29.42& 28.56& 2.746& 1.010& 0.062&00000 8.00000 &4-766.0& 0.88& 566.25& 197.75 &36:41.29 &12:14.73& 28.69& 29.36& 28.34& 99.99& 99.99& 99.99& 2.570& 0.140& 0.072&10000 9.00000 &4-766.0& 0.47& 568.66& 210.66 &36:41.37 &12:15.94& 28.69& 29.68& 28.48& 99.99& 99.99& 99.99& 2.861& 0.110& 0.085&00000 10.0000 &4-813.2& 0.21& 514.50& 141.00 &36:41.40 &12:6.992& 99.99& 99.99& 99.99& 28.94& 29.23& 28.55& 2.605& 1.200& 0.046&01010 11.0000 &4-790.0& 0.21& 539.33& 175.66 &36:41.41 &12:11.30& 28.73& 29.36& 28.52& 27.73& 28.16& 27.82& 2.857& 0.110& 0.069&22111 12.0000 &4-813.2& 0.15& 511.72& 139.27 &36:41.41 &12:6.649& 27.18& 27.44& 27.16& 27.26& 28.15& 27.13& 7.199& 0.270& 0.109&21111 13.0000 &4-767.0& 1.05& 551.75& 198.00 &36:41.47 &12:13.92& 99.99& 99.99& 99.99& 28.46& 28.97& 28.15& 2.640& 0.760& 0.102&01000 14.0000 &4-767.0& 0.08& 557.46& 207.10 &36:41.47 &12:14.97& 21.53& 21.56& 22.04& 22.10& 22.13& 22.55& 157.7& 7.990& 0.267&22111 15.0000 &4-813.0& 0.08& 506.15& 144.57 &36:41.53 &12:6.731& 25.94& 26.14& 26.18& 26.37& 26.61& 26.48& 12.36& 0.750& 0.162&21111 16.0000 &4-794.0& 0.02& 512.73& 155.56 &36:41.55 &12:8.022& 25.98& 26.29& 26.43& 26.39& 26.85& 26.67& 11.24& 0.830& 0.190&22111 17.0000 &4-739.0& 0.75& 562.00& 227.00 &36:41.59 &12:16.86& 28.35& 29.10& 28.31& 99.99& 99.99& 99.99& 4.062& 0.110& 0.054&10000 18.0000 &4-760.0& 0.35& 537.44& 194.77 &36:41.60 &12:12.71& 27.73& 28.19& 27.83& 29.17& 29.78& 28.87& 4.568& 0.250& 0.132&22101 19.0000 &4-802.0& 0.65& 493.00& 135.50 &36:41.61 &12:5.221& 99.99& 99.99& 99.99& 28.68& 29.56& 28.51& 2.843& 0.887& 0.062&00000 20.0000 &4-802.0& 0.06& 488.50& 130.62 &36:41.61 &12:4.507& 26.46& 26.97& 27.17& 27.34& 27.73& 27.40& 7.592& 0.740& 0.204&21110 21.0000 &4-739.0& 0.06& 555.60& 223.00 &36:41.63 &12:16.13& 28.08& 28.75& 27.99& 99.99& 99.99& 99.99& 4.206& 0.160& 0.093&20011 22.0000 &4-802.0& 0.98& 491.50& 140.50 &36:41.66 &12:5.495& 28.46& 28.97& 28.35& 99.99& 99.99& 99.99& 3.402& 0.120& 0.082&10000 23.0000 &4-706.0& 1.15& 569.00& 249.50 &36:41.71 &12:19.16& 99.99& 99.99& 99.99& 28.78& 29.37& 28.57& 2.582& 1.054& 0.040&01000 24.0000 &4-706.0& 1.06& 567.50& 248.00 &36:41.72 &12:18.94& 99.99& 99.99& 99.99& 28.86& 29.59& 28.39& 2.817& 0.862& 0.056&01000 25.0000 &4-683.0& 0.97& 572.75& 270.25 &36:41.84 &12:21.04& 28.55& 28.99& 28.58& 99.99& 99.99& 99.99& 2.750& 0.160& 0.078&10000 26.1200&4-795.112& 0.09& 489.40& 159.18 &36:41.85 &12:6.910& 25.70& 25.78& 25.72& 25.68& 25.74& 25.75& 18.94& 0.680& 0.160&22111 26.1000 &4-795.0& 0.13& 476.22& 150.64 &36:41.93 &12:5.303& 20.16& 20.21& 21.55& 20.52& 20.58& 21.90& 336.0& 17.63& 0.510&22111 26.0000 &4-795.0& 0.12& 474.67& 150.20 &36:41.93 &12:5.303& 20.16& 20.19& 21.55& 20.52& 20.58& 21.90& 272.2& 22.53& 0.534&22111 26.1100 &4-795.0& 0.13& 475.39& 150.33 &36:41.93 &12:5.289& 20.17& 20.24& 21.55& 20.52& 20.58& 21.90& 474.9& 11.49& 0.475&22111 27.0000 &4-665.0& 0.04& 567.34& 298.55 &36:42.16 &12:23.01& 25.83& 26.04& 26.16& 25.90& 26.19& 26.26& 13.39& 0.790& 0.170&21111 28.0000 &4-652.0& 0.04& 575.60& 314.80 &36:42.20 &12:24.84& 28.47& 28.87& 28.43& 28.72& 29.30& 28.16& 2.765& 0.180& 0.085&22110 26.2000 &4-757.0& 1.14& 454.33& 154.00 &36:42.21 &12:4.328& 27.61& 27.66& 26.93& 26.60& 26.93& 27.01& 5.781& 0.230& 0.110&11101 29.0000 &4-769.0& 0.18& 432.56& 132.29 &36:42.28 &12:1.238& 25.47& 25.74& 26.24& 25.82& 26.09& 26.34& 14.40& 1.200& 0.248&21111 30.0000 &4-671.2& 0.68& 537.84& 279.55 &36:42.32 &12:19.65& 27.88& 29.57& 28.25& 99.00& 99.00& 99.00& 5.800& 0.040& 0.000&00003 31.0000 &4-671.0& 0.05& 533.48& 274.67 &36:42.35 &12:18.99& 26.25& 26.33& 26.53& 26.33& 26.54& 26.51& 8.648& 0.850& 0.186&22111 32.0000 &4-690.0& 0.11& 502.42& 234.66 &36:42.36 &12:13.83& 26.62& 27.08& 27.03& 26.86& 27.18& 26.93& 6.986& 0.690& 0.161&22111 33.0000 &4-743.0& 0.19& 421.25& 138.25 &36:42.47 &12:1.018& 28.32& 29.00& 28.16& 29.34& 29.73& 29.46& 3.790& 0.130& 0.082&20111 34.0000 &4-715.0& 0.10& 443.50& 176.50 &36:42.54 &12:5.537& 99.99& 99.99& 99.99& 28.16& 28.79& 28.04& 3.694& 0.900& 0.082&02011 35.0000 &4-687.0& 0.07& 476.28& 223.50 &36:42.57 &12:11.33& 27.11& 27.42& 27.25& 28.20& 28.85& 28.06& 7.124& 0.310& 0.135&22111 36.0000 &4-619.0& 0.14& 556.62& 333.87 &36:42.59 &12:25.24& 26.93& 27.31& 27.12& 29.89& 29.85& 30.79& 7.998& 0.340& 0.148&21111 37.0000 &4-725.0& 0.15& 428.95& 168.74 &36:42.64 &12:3.985& 25.54& 25.74& 26.26& 25.98& 26.46& 26.70& 12.38& 1.360& 0.245&22111 38.0000 &4-636.0& 0.04& 528.16& 301.66 &36:42.64 &12:20.86& 26.22& 26.44& 26.49& 26.34& 26.56& 26.52& 10.11& 0.680& 0.154&22111 39.0000 &4-671.2& 2.40& 509.33& 280.26 &36:42.66 &12:17.96& 28.27& 30.24& 28.64& 27.87& 30.47& 28.33& 5.700& 0.030& 0.000&01003 40.0000 &4-636.0& 0.74& 522.00& 297.50 &36:42.68 &12:20.14& 29.04& 29.92& 28.97& 99.99& 99.99& 99.99& 2.540& 0.080& 0.050&10001 41.0000 &4-636.0& 0.62& 521.66& 300.33 &36:42.70 &12:20.36& 28.77& 29.42& 28.67& 99.99& 99.99& 99.99& 3.044& 0.090& 0.088&20001 42.0000 &4-716.0& 1.31& 405.50& 149.00 &36:42.75 &12:0.977& 99.99& 99.99& 99.99& 28.93& 29.63& 28.84& 2.631& 0.827& 0.047&01000 43.0000 &4-707.0& 0.19& 421.54& 173.00 &36:42.76 &12:3.861& 27.70& 28.15& 27.80& 27.55& 28.08& 27.59& 4.439& 0.270& 0.128&22111 44.0000 &4-561.0& 1.27& 578.00& 393.00 &36:42.86 &12:31.36& 28.24& 28.65& 28.52& 99.99& 99.99& 99.99& 3.551& 0.170& 0.130&20000 47.1000 &4-581.0& 0.03& 556.85& 365.35 &36:42.86 &12:27.86& 24.93& 25.09& 25.36& 25.48& 25.63& 25.67& 23.10& 1.370& 0.212&22111 45.0000 &4-716.0& 0.17& 392.66& 145.50 &36:42.86 &11:59.86& 28.51& 29.11& 28.63& 28.41& 28.96& 28.22& 2.752& 0.170& 0.116&20111 46.0000 &4-697.0& 0.08& 411.57& 174.40 &36:42.89 &12:3.353& 25.81& 25.94& 26.12& 26.66& 26.87& 26.74& 12.80& 0.860& 0.177&22111 47.0000 &4-581.12& 0.21& 551.59& 363.05 &36:42.89 &12:27.42& 24.45& 24.56& 25.36& 28.69& 29.17& 28.54& 20.69& 2.990& 0.658&11110 48.0000 &4-606.0& 0.06& 525.50& 328.00 &36:42.90 &12:22.85& 28.50& 29.23& 28.35& 99.99& 99.99& 99.99& 3.221& 0.130& 0.088&20011 49.0000 &4-656.0& 0.05& 486.95& 276.44 &36:42.90 &12:16.31& 20.06& 20.08& 21.89& 20.34& 20.36& 22.20& 299.5& 22.09& 0.618&22111 50.0000 &4-614.2& 1.59& 508.01& 306.80 &36:42.91 &12:20.03& 28.06& 28.74& 28.43& 27.99& 29.59& 28.45& 3.300& 0.090& 0.000&10003 47.2000 &4-581.2& 0.11& 545.75& 360.37 &36:42.95 &12:26.67& 25.55& 25.71& 25.98& 25.96& 26.01& 26.18& 15.73& 1.030& 0.202&22111 51.0000 &4-660.0& 0.17& 454.80& 240.43 &36:42.96 &12:11.38& 25.27& 25.37& 25.79& 25.92& 26.13& 26.29& 17.34& 1.190& 0.225&22111 52.0000 &4-554.2& 0.07& 545.00& 370.50 &36:43.05 &12:27.55& 99.99& 99.99& 99.99& 28.08& 28.52& 27.95& 3.360& 0.766& 0.087&01011 53.0000 &4-537.0& 0.33& 563.50& 398.00 &36:43.08 &12:30.91& 99.99& 99.99& 99.99& 28.60& 29.18& 28.12& 3.585& 1.253& 0.045&01000 54.0000 &4-664.0& 0.11& 422.28& 213.00 &36:43.10 &12:7.171& 28.11& 28.43& 27.98& 28.26& 28.76& 28.14& 3.844& 0.180& 0.086&22111 55.0000 &4-531.0& 0.09& 575.60& 418.36 &36:43.10 &12:33.26& 25.68& 25.88& 26.10& 26.02& 26.29& 26.43& 13.33& 1.000& 0.192&22110 56.0000 &4-537.0& 0.12& 562.71& 401.42 &36:43.10 &12:31.10& 28.52& 28.63& 28.47& 99.99& 99.99& 99.99& 2.726& 0.170& 0.108&20011 57.0000 &4-554.0& 0.03& 544.77& 377.70 &36:43.12 &12:28.12& 24.32& 24.38& 24.80& 24.39& 24.56& 24.84& 26.95& 2.370& 0.247&21111 58.0000 &4-695.0& 0.46& 381.00& 166.00 &36:43.18 &12:0.799& 28.59& 28.81& 28.79& 99.99& 99.99& 99.99& 2.745& 0.150& 0.131&10000 59.0000 &4-599.0& 0.04& 488.50& 310.70 &36:43.18 &12:19.17& 25.76& 26.00& 26.03& 26.66& 26.81& 26.67& 15.02& 0.710& 0.154&21111 60.0000 &4-541.0& 0.87& 534.00& 371.50 &36:43.19 &12:26.94& 99.99& 99.99& 99.99& 28.12& 28.75& 28.09& 3.394& 0.929& 0.097&00000 61.0000 &4-590.0& 0.09& 477.28& 308.59 &36:43.29 &12:18.30& 24.85& 24.90& 25.20& 25.27& 25.42& 25.62& 24.57& 1.250& 0.206&22111 62.0000 &4-531.0& 1.49& 565.50& 429.00 &36:43.32 &12:33.56& 99.99& 99.99& 99.99& 28.90& 29.54& 28.62& 2.718& 0.899& 0.045&00000 63.0000 &4-605.0& 0.12& 460.80& 290.90 &36:43.33 &12:15.86& 27.62& 27.99& 27.57& 28.33& 29.11& 28.23& 5.433& 0.220& 0.109&22111 64.0000 &4-677.0& 0.14& 368.90& 175.09 &36:43.40 &12:0.826& 27.91& 28.22& 27.97& 99.99& 99.99& 99.99& 3.487& 0.300& 0.124&20011 65.0000 &4-541.0& 1.26& 527.81& 391.54 &36:43.43 &12:28.21& 99.99& 99.99& 99.99& 27.28& 27.75& 27.47& 5.091& 0.852& 0.175&01000 66.0000 &4-541.0& 1.18& 525.00& 389.00 &36:43.45 &12:27.82& 99.99& 99.99& 99.99& 28.28& 28.77& 28.20& 3.865& 1.219& 0.066&00000 67.0000 &4-553.0& 0.01& 485.66& 338.33 &36:43.46 &12:21.26& 28.86& 29.29& 28.48& 29.15& 29.34& 28.84& 2.822& 0.090& 0.086&21110 68.0000 &4-565.0& 0.05& 459.04& 321.55 &36:43.62 &12:18.25& 22.19& 22.24& 23.06& 22.39& 22.45& 23.20& 106.5& 5.840& 0.344&22111 69.0000 &4-437.0& 2.44& 573.66& 474.66 &36:43.62 &12:37.74& 28.74& 29.30& 28.78& 99.99& 99.99& 99.99& 2.837& 0.110& 0.054&20000 70.0000 &4-669.0& 0.16& 330.14& 155.85 &36:43.68 &11:56.87& 28.11& 28.81& 28.11& 28.29& 28.97& 28.11& 3.355& 0.230& 0.096&20111 71.0000 &4-572.0& 0.04& 439.20& 303.01 &36:43.69 &12:15.50& 24.82& 25.01& 25.88& 25.05& 25.26& 26.01& 17.41& 2.310& 0.308&22111 72.0000 &4-598.0& 0.08& 411.70& 266.76 &36:43.69 &12:10.90& 27.15& 27.43& 27.40& 27.64& 28.13& 27.72& 5.961& 0.400& 0.165&21111 73.0000 &4-479.0& 0.60& 533.66& 430.66 &36:43.71 &12:31.75& 99.99& 99.99& 99.99& 28.52& 29.21& 28.82& 2.767& 0.814& 0.073&01000 74.0000 &4-513.0& 0.05& 493.00& 381.25 &36:43.75 &12:25.20& 28.62& 29.20& 28.41& 29.09& 29.30& 28.88& 2.876& 0.130& 0.076&21111 75.0000 &4-479.0& 0.04& 531.88& 436.91 &36:43.78 &12:32.11& 25.85& 26.04& 26.20& 25.98& 26.12& 26.14& 11.83& 0.940& 0.170&21111 76.0000 &4-525.0& 0.07& 473.90& 362.00 &36:43.80 &12:22.45& 27.51& 27.88& 27.58& 27.38& 27.64& 27.36& 5.136& 0.290& 0.120&22111 77.0000 &4-486.0& 0.07& 516.00& 420.00 &36:43.81 &12:29.74& 28.66& 29.46& 28.54& 27.86& 28.46& 27.91& 3.030& 0.110& 0.062&12111 78.0000 &4-513.0& 0.79& 486.16& 384.33 &36:43.86 &12:25.06& 99.99& 99.99& 99.99& 27.96& 28.37& 28.02& 3.356& 0.887& 0.106&01000 79.0000 &4-626.0& 0.13& 349.00& 205.75 &36:43.90 &12:2.103& 28.41& 28.91& 28.36& 28.28& 28.83& 28.20& 2.858& 0.190& 0.111&21111 80.0000 &4-444.0& 0.05& 545.33& 480.83 &36:44.00 &12:36.51& 27.43& 27.91& 27.74& 28.40& 29.15& 28.37& 5.650& 0.280& 0.171&22111 81.1000 &4-576.0& 0.18& 386.20& 267.80 &36:44.01 &12:9.450& 25.37& 25.98& 25.91& 25.58& 26.01& 26.06& 23.56& 0.670& 0.171&22111 81.0000 &4-576.0& 0.28& 384.10& 267.85 &36:44.02 &12:9.327& 25.02& 25.13& 25.91& 25.13& 25.30& 26.06& 18.04& 1.640& 0.319&22101 82.0000 &4-481.0& 0.13& 487.50& 407.00 &36:44.03 &12:26.96& 29.23& 29.81& 29.00& 99.99& 99.99& 99.99& 2.504& 0.060& 0.050&10010 83.0000 &4-500.0& 0.48& 464.00& 376.40 &36:44.04 &12:23.04& 28.56& 29.05& 28.43& 28.48& 29.03& 28.44& 2.694& 0.160& 0.104&11101 84.0000 &4-524.0& 2.69& 401.00& 292.50 &36:44.06 &12:12.40& 99.99& 99.99& 99.99& 28.68& 29.15& 28.72& 2.833& 1.294& 0.075&01000 85.0000 &4-494.0& 1.18& 473.00& 389.50 &36:44.06 &12:24.68& 99.99& 99.99& 99.99& 28.75& 29.60& 28.87& 2.664& 0.856& 0.051&01000 81.2000 &4-576.2& 0.06& 381.50& 268.00 &36:44.06 &12:9.148& 26.40& 27.01& 26.44& 26.31& 26.74& 26.51& 16.23& 0.250& 0.114&21111 86.0000 &4-500.0& 0.19& 462.70& 380.35 &36:44.09 &12:23.26& 26.97& 27.45& 27.43& 27.92& 28.34& 27.79& 7.103& 0.390& 0.201&22111 87.0000 &4-653.0& 0.19& 284.09& 143.00 &36:44.11 &11:53.01& 27.78& 28.17& 27.85& 99.99& 99.99& 99.99& 4.064& 0.280& 0.115&20011 88.0000 &4-430.0& 0.05& 559.01& 519.49 &36:44.18 &12:40.44& 22.88& 22.98& 24.42& 23.04& 23.11& 24.57& 50.78& 6.790& 0.465&21111 89.0000 &4-501.0& 0.03& 435.50& 360.50 &36:44.24 &12:19.99& 28.30& 28.96& 28.13& 99.99& 99.99& 99.99& 4.039& 0.120& 0.090&20011 90.0000 &4-580.0& 0.11& 331.27& 231.90 &36:44.33 &12:3.147& 28.05& 28.41& 28.13& 28.08& 28.48& 28.20& 3.460& 0.240& 0.130&21111 91.0000 &4-524.0& 0.19& 394.46& 317.46 &36:44.34 &12:13.98& 27.85& 28.13& 27.98& 99.99& 99.99& 99.99& 3.658& 0.300& 0.148&20011 92.0000 &4-499.0& 0.07& 408.85& 346.28 &36:44.42 &12:17.19& 28.02& 28.55& 28.07& 28.82& 29.40& 28.67& 3.958& 0.200& 0.112&21111 93.0000 &4-452.0& 0.96& 467.12& 425.87 &36:44.44 &12:27.23& 28.38& 28.52& 28.31& 99.99& 99.99& 99.99& 2.658& 0.220& 0.118&20001 94.0000 &4-627.0& 0.20& 263.65& 158.76 &36:44.48 &11:53.05& 26.05& 26.31& 26.66& 26.29& 26.58& 26.77& 9.286& 1.030& 0.216&22111 95.0000 &4-438.0& 0.04& 483.82& 453.37 &36:44.48 &12:30.50& 25.85& 26.19& 26.65& 26.25& 26.55& 26.96& 10.96& 1.080& 0.258&22111 96.0000 &4-400.0& 0.51& 513.66& 500.66 &36:44.55 &12:36.18& 28.81& 29.52& 28.82& 28.51& 29.42& 28.17& 2.658& 0.110& 0.077&11101 97.0000 &4-505.0& 0.31& 390.77& 337.94 &36:44.55 &12:15.42& 23.05& 23.09& 24.49& 24.00& 24.09& 25.48& 48.38& 5.780& 0.484&22101 98.0000 &4-527.0& 0.05& 356.03& 292.25 &36:44.57 &12:9.560& 26.52& 26.77& 26.86& 26.65& 26.97& 26.89& 7.206& 0.760& 0.170&22111 99.0000 &4-400.2& 1.17& 507.21& 495.52 &36:44.57 &12:35.36& 27.79& 29.04& 28.16& 99.00& 99.00& 99.00& 3.000& 0.060& 0.000&00003 100.000 &4-472.0& 0.15& 418.56& 378.25 &36:44.59 &12:20.39& 27.11& 27.42& 27.29& 28.24& 28.93& 27.99& 6.743& 0.340& 0.165&22111 101.000 &4-632.0& 0.29& 229.33& 129.66 &36:44.63 &11:48.56& 28.76& 29.43& 28.76& 28.97& 29.75& 28.62& 2.775& 0.110& 0.090&11101 102.000 &4-445.0& 0.05& 457.23& 435.27 &36:44.63 &12:27.44& 22.57& 22.62& 23.39& 22.98& 23.04& 23.65& 80.24& 5.240& 0.329&21111 900.000 &4-625.2& 0.18& 243.12& 148.18 &36:44.64 &11:50.96& 99.99& 99.99& 99.99& 26.75& 27.00& 26.75& 5.934& 1.053& 0.142&01011 103.000 &4-625.0& 0.09& 238.06& 145.36 &36:44.66 &11:50.32& 24.33& 24.46& 25.34& 24.72& 24.84& 25.55& 24.01& 2.830& 0.402&21111 104.000 &4-409.0& 0.65& 497.50& 495.00 &36:44.69 &12:34.78& 99.99& 99.99& 99.99& 28.77& 28.85& 28.79& 3.064& 1.710& 0.083&00000 105.000 &4-377.0& 0.85& 545.50& 560.00 &36:44.70 &12:43.00& 99.99& 99.99& 99.99& 28.86& 29.42& 28.77& 2.807& 1.006& 0.044&01000 106.000 &4-603.0& 1.19& 252.80& 170.00 &36:44.71 &11:53.30& 28.53& 29.27& 28.43& 28.64& 29.15& 28.55& 2.591& 0.180& 0.080&12101 107.000 &4-455.0& 1.38& 402.00& 370.50 &36:44.71 &12:18.74& 28.40& 28.91& 28.23& 29.59& 29.87& 29.71& 3.152& 0.160& 0.095&21101 108.000 &4-446.0& 0.24& 432.33& 413.66 &36:44.73 &12:24.11& 28.85& 29.49& 28.70& 99.99& 99.99& 99.99& 2.564& 0.110& 0.078&20010 109.000 &4-579.0& 0.14& 272.39& 199.69 &36:44.73 &11:56.91& 24.97& 25.04& 25.30& 24.98& 25.06& 25.23& 22.51& 1.200& 0.196&22111 110.000 &4-603.0& 0.23& 256.12& 178.75 &36:44.74 &11:54.24& 25.73& 25.92& 26.35& 26.51& 26.81& 27.01& 10.34& 1.380& 0.234&22111 111.000 &4-587.2& 0.08& 260.16& 186.60 &36:44.75 &11:55.08& 28.12& 29.01& 28.49& 29.53& 30.23& 29.99& 3.200& 0.070& 0.000&10013 112.000 &4-340.0& 0.10& 567.00& 598.50 &36:44.78 &12:47.45& 99.99& 99.99& 99.99& 28.31& 29.36& 28.03& 3.978& 1.066& 0.045&02010 113.000 &4-351.0& 0.94& 539.50& 565.00 &36:44.80 &12:43.00& 29.14& 29.73& 28.82& 99.99& 99.99& 99.99& 2.725& 0.060& 0.051&10001 114.000 &4-558.0& 0.11& 286.32& 228.51 &36:44.82 &12:0.098& 22.63& 22.66& 24.12& 22.78& 22.85& 24.31& 50.99& 9.250& 0.513&22111 115.000 &4-422.0& 0.66& 461.00& 461.50 &36:44.83 &12:29.81& 99.99& 99.99& 99.99& 28.78& 29.43& 28.72& 2.599& 0.995& 0.054&00000 116.000 &4-509.0& 0.09& 329.54& 287.77 &36:44.84 &12:7.569& 26.50& 26.64& 26.78& 26.62& 26.85& 26.95& 6.880& 0.850& 0.178&22111 901.000 &4-378.0& 1.28& 518.30& 544.90 &36:44.89 &12:40.10& 99.99& 99.99& 99.99& 25.80& 26.71& 25.90& 23.01& 2.364& 0.119&00000 117.000 &4-455.0& 0.04& 395.33& 382.19 &36:44.89 &12:19.29& 26.56& 26.89& 26.85& 26.87& 27.08& 26.89& 7.612& 0.660& 0.146&22111 118.000 &4-601.0& 0.28& 222.44& 153.11 &36:44.91 &11:50.07& 27.86& 28.24& 27.92& 28.65& 29.42& 28.44& 4.049& 0.250& 0.127&21101 119.000 &4-351.0& 0.12& 531.10& 567.21 &36:44.92 &12:42.68& 26.62& 26.90& 26.83& 28.15& 28.69& 28.04& 9.022& 0.450& 0.156&22111 120.000 &4-423.0& 0.09& 435.14& 441.65 &36:44.94 &12:26.57& 25.37& 25.53& 26.00& 25.66& 25.93& 26.29& 14.34& 1.390& 0.248&21111 121.000 &4-329.0& 1.61& 540.50& 582.00 &36:44.94 &12:44.44& 29.22& 29.74& 28.78& 99.99& 99.99& 99.99& 2.511& 0.060& 0.052&10001 122.000 &4-431.0& 0.77& 407.66& 406.33 &36:44.96 &12:22.01& 28.79& 29.53& 28.54& 99.99& 99.99& 99.99& 3.003& 0.090& 0.078&10001 123.000 &4-607.0& 0.21& 210.25& 144.37 &36:44.98 &11:48.60& 28.08& 28.60& 28.10& 99.99& 99.99& 99.99& 3.614& 0.210& 0.106&20011 124.000 &4-378.0& 0.55& 509.33& 546.02 &36:44.99 &12:39.67& 23.04& 23.14& 24.86& 23.32& 23.48& 25.24& 40.35& 7.560& 0.573&21101 125.000 &4-596.0& 0.19& 210.30& 151.07 &36:45.04 &11:49.16& 26.68& 26.92& 27.10& 26.61& 26.78& 27.00& 6.300& 0.740& 0.189&21111 126.000 &4-515.0& 0.11& 291.00& 259.25 &36:45.04 &12:2.900& 28.67& 29.47& 28.48& 99.99& 99.99& 99.99& 2.768& 0.130& 0.070&20011 127.000 &4-543.0& 0.10& 261.79& 221.36 &36:45.05 &11:58.01& 25.08& 25.18& 25.81& 25.25& 25.35& 25.92& 17.37& 1.580& 0.272&22111 141.120 &4-571.0& 0.18& 237.26& 190.86 &36:45.06 &11:54.04& 23.87& 23.98& 24.95& 24.04& 24.16& 25.18& 48.44& 2.500& 0.335&21111 128.000 &4-431.2& 0.04& 401.46& 412.21 &36:45.07 &12:22.09& 27.47& 28.33& 27.84& 28.38& 29.05& 28.84& 4.400& 0.130& 0.000&20013 129.000 &4-342.0& 1.96& 487.50& 531.00 &36:45.12 &12:37.09& 99.99& 99.99& 99.99& 28.93& 29.69& 28.48& 2.644& 0.788& 0.049&00000 130.000 &4-573.0& 0.21& 210.38& 162.66 &36:45.14 &11:50.11& 26.85& 27.05& 27.01& 28.43& 29.04& 28.46& 5.945& 0.630& 0.146&22111 131.000 &4-530.0& 0.17& 265.54& 236.90 &36:45.14 &11:59.53& 27.69& 28.29& 27.93& 27.67& 28.09& 27.72& 4.457& 0.280& 0.158&22111 132.000 &4-502.0& 0.31& 300.24& 283.82 &36:45.14 &12:5.454& 24.75& 24.82& 25.75& 25.06& 25.18& 25.96& 18.27& 2.370& 0.318&22101 133.000 &4-329.0& 0.10& 528.70& 591.95 &36:45.17 &12:44.52& 26.23& 26.45& 26.38& 26.69& 26.97& 26.68& 10.61& 0.630& 0.140&22111 134.000 &4-456.0& 1.03& 341.00& 350.00 &36:45.25 &12:13.35& 28.34& 28.78& 28.26& 99.99& 99.99& 99.99& 2.920& 0.200& 0.098&20001 135.000 &4-407.0& 0.12& 412.89& 447.69 &36:45.26 &12:25.69& 25.16& 25.29& 25.86& 25.25& 25.43& 25.94& 16.56& 1.510& 0.261&22111 136.000 &4-593.0& 0.30& 173.22& 129.51 &36:45.28 &11:45.13& 26.20& 26.35& 26.48& 26.07& 26.24& 26.46& 9.240& 0.820& 0.178&21101 141.112 &4-555.11& 0.18& 213.18& 184.93 &36:45.29 &11:52.05& 23.54& 23.92& 24.37& 24.04& 24.49& 24.96& 106.5& 1.080& 0.273&22111 137.000 &4-342.0& 0.08& 486.66& 551.11 &36:45.30 &12:38.64& 27.62& 27.96& 27.57& 27.14& 27.64& 27.09& 5.417& 0.220& 0.105&21111 138.000 &4-498.0& 0.11& 278.72& 274.80 &36:45.32 &12:3.422& 25.99& 26.10& 26.24& 26.44& 26.67& 26.68& 10.51& 0.920& 0.170&22111 141.100 &4-555.0& 0.57& 222.83& 197.86 &36:45.33 &11:53.94& 20.84& 20.92& 22.04& 21.41& 21.48& 22.59& 170.5& 19.86& 0.786&22100 141.111 &4-555.0& 0.17& 224.45& 204.37 &36:45.34 &11:54.37& 21.20& 21.41& 22.04& 21.80& 22.01& 22.59& 403.9& 4.100& 0.296&22111 139.000 &4-395.0& 0.04& 414.87& 460.04 &36:45.34 &12:26.82& 26.39& 26.73& 26.73& 26.15& 26.29& 26.33& 8.580& 0.700& 0.161&22111 141.110 &4-555.0& 0.59& 219.17& 199.24 &36:45.34 &11:53.94& 20.91& 21.02& 22.04& 21.51& 21.63& 22.59& 259.6& 11.01& 0.693&22100 140.000 &4-368.0& 0.04& 455.06& 514.43 &36:45.35 &12:33.74& 24.16& 24.30& 24.98& 24.80& 24.95& 25.44& 28.98& 2.650& 0.283&22111 141.000 &4-555.0& 0.66& 219.83& 199.84 &36:45.36 &11:53.91& 20.71& 20.74& 22.04& 21.51& 21.63& 22.59& 135.4& 29.60& 0.998&22100 141.112 &4-555.1& 0.56& 213.08& 191.80 &36:45.37 &11:52.66& 22.50& 22.71& 23.81& 23.10& 23.31& 24.41& 143.1& 3.290& 0.643&21100 142.000 &4-379.0& 0.01& 438.50& 493.00 &36:45.37 &12:30.99& 99.99& 99.99& 99.99& 28.95& 29.32& 28.47& 2.596& 1.100& 0.049&01010 143.000 &4-407.2& 0.71& 407.18& 453.94 &36:45.37 &12:25.84& 28.83& 30.17& 29.20& 29.04& 31.81& 29.50& 2.800& 0.030& 0.000&01003 144.000 &4-487.0& 0.85& 265.50& 263.50 &36:45.38 &12:1.719& 99.99& 99.99& 99.99& 28.09& 28.86& 28.01& 3.036& 0.565& 0.114&01000 145.000 &4-454.0& 0.01& 332.74& 358.70 &36:45.41 &12:13.54& 22.07& 22.10& 22.35& 21.43& 21.45& 21.68& 119.0& 5.920& 0.206&22111 146.000 &4-467.0& 0.04& 299.50& 313.00 &36:45.41 &12:7.803& 28.57& 29.11& 28.49& 29.42& 30.10& 29.03& 2.698& 0.160& 0.097&21111 141.112 &4-555.1& 0.15& 212.64& 196.90 &36:45.41 &11:53.02& 23.02& 23.41& 23.81& 23.68& 24.14& 24.41& 163.4& 1.160& 0.225&21111 147.000 &4-384.2& 0.04& 418.76& 475.17 &36:45.42 &12:28.29& 27.67& 28.51& 28.04& 27.83& 28.84& 28.29& 3.100& 0.100& 0.000&10013 148.000 &4-557.0& 0.23& 186.45& 163.79 &36:45.42 &11:48.72& 24.45& 24.53& 25.04& 25.16& 25.34& 25.63& 22.87& 2.550& 0.253&22111 149.000 &4-330.0& 0.10& 486.00& 565.00 &36:45.43 &12:39.71& 28.23& 28.45& 28.40& 99.99& 99.99& 99.99& 3.444& 0.180& 0.089&20011 150.000 &4-487.0& 0.32& 263.63& 268.88 &36:45.45 &12:2.021& 26.39& 26.75& 27.70& 27.21& 27.58& 27.76& 6.179& 1.200& 0.452&22101 151.000 &4-582.2& 0.50& 153.04& 125.15 &36:45.47 &11:43.53& 27.73& 28.62& 28.10& 28.71& 29.84& 29.17& 2.700& 0.090& 0.000&00003 152.000 &4-344.0& 0.50& 456.00& 529.50 &36:45.47 &12:34.99& 28.97& 29.63& 28.73& 99.99& 99.99& 99.99& 2.690& 0.080& 0.048&10000 153.000 &4-424.0& 0.07& 346.00& 385.00 &36:45.49 &12:16.49& 28.32& 28.87& 28.19& 28.28& 28.70& 28.18& 3.278& 0.170& 0.069&21111 154.000 &4-487.0& 0.08& 260.75& 270.25 &36:45.50 &12:1.994& 99.99& 99.99& 99.99& 28.09& 28.42& 28.04& 2.760& 0.590& 0.098&01011 155.000 &4-344.0& 0.06& 451.68& 528.13 &36:45.51 &12:34.62& 26.72& 26.99& 26.95& 26.66& 26.99& 26.94& 6.447& 0.670& 0.155&22111 156.000 &4-307.0& 0.91& 493.80& 584.40 &36:45.51 &12:41.81& 99.99& 99.99& 99.99& 27.86& 28.55& 27.80& 4.102& 0.898& 0.100&01000 157.000 &4-330.2& 0.73& 478.52& 569.07 &36:45.54 &12:39.59& 27.25& 27.93& 27.62& 32.21& 99.00& 32.67& 7.700& 0.170& 0.000&10003 158.000 &4-323.0& 1.43& 460.75& 545.25 &36:45.55 &12:36.57& 28.57& 29.38& 28.53& 99.99& 99.99& 99.99& 2.690& 0.160& 0.072&20001 159.000 &4-345.0& 0.03& 442.70& 522.20 &36:45.56 &12:33.57& 26.30& 26.50& 26.50& 26.29& 26.64& 26.51& 9.460& 0.680& 0.149&22111 160.000 &4-323.0& 1.15& 457.66& 542.66 &36:45.57 &12:36.21& 99.99& 99.99& 99.99& 28.51& 29.24& 28.50& 2.806& 0.795& 0.084&01000 161.000 &4-330.2& 1.22& 480.06& 576.14 &36:45.59 &12:40.26& 27.47& 28.30& 27.84& 29.12& 30.32& 29.58& 4.600& 0.110& 0.000&10003 162.000 &4-582.2& 0.77& 149.09& 133.67 &36:45.59 &11:43.99& 27.52& 27.65& 27.89& 27.86& 28.56& 28.32& 7.900& 0.220& 0.000&01003 163.000 &4-389.0& 0.09& 389.35& 456.44 &36:45.61 &12:24.95& 26.02& 26.25& 26.43& 26.60& 26.84& 26.88& 10.85& 0.830& 0.182&21111 164.000 &4-307.0& 0.84& 482.88& 582.88 &36:45.63 &12:41.05& 99.99& 99.99& 99.99& 27.23& 27.72& 27.70& 4.670& 1.067& 0.178&01000 165.000 &4-563.0& 0.21& 140.98& 126.03 &36:45.63 &11:42.86& 25.98& 26.13& 26.62& 26.73& 27.03& 26.94& 9.138& 1.170& 0.305&21111 166.000 &4-307.0& 0.16& 488.13& 592.80 &36:45.65 &12:42.05& 22.24& 22.29& 22.97& 23.37& 23.49& 24.04& 87.28& 7.230& 0.329&22111 167.000 &4-551.0& 0.20& 156.00& 148.25 &36:45.65 &11:45.63& 99.99& 99.99& 99.99& 28.29& 28.92& 28.05& 3.133& 0.795& 0.063&01011 141.200 &4-516.0& 0.15& 203.76& 213.36 &36:45.66 &11:53.83& 23.09& 23.17& 23.65& 23.60& 23.68& 24.08& 101.4& 2.400& 0.243&22111 168.000 &4-563.2& 1.27& 147.31& 139.45 &36:45.66 &11:44.34& 27.56& 28.20& 27.93& 28.50& 30.31& 28.96& 4.100& 0.140& 0.000&10003 169.000 &4-371.2& 0.51& 393.15& 470.94 &36:45.68 &12:26.38& 29.02& 30.73& 29.39& 99.00& 99.00& 99.00& 3.700& 0.020& 0.000&00003 170.000 &4-371.2& 0.34& 395.51& 478.78 &36:45.72 &12:27.16& 28.25& 29.22& 28.62& 28.05& 29.10& 28.51& 14.70& 0.070& 0.000&21003 171.000 &4-323.0& 0.08& 447.50& 547.50 &36:45.72 &12:35.94& 28.78& 29.14& 28.64& 99.99& 99.99& 99.99& 2.593& 0.120& 0.076&10011 172.000 &4-497.0& 0.16& 219.01& 243.37 &36:45.74 &11:57.18& 25.17& 25.27& 25.61& 25.69& 25.89& 26.09& 17.98& 1.310& 0.210&21111 173.000 &4-307.0& 0.96& 484.80& 599.60 &36:45.75 &12:42.50& 99.99& 99.99& 99.99& 27.72& 28.35& 27.71& 4.763& 1.079& 0.087&00000 174.000 &4-478.0& 0.18& 227.55& 259.77 &36:45.78 &11:59.05& 27.96& 28.38& 27.97& 28.08& 28.57& 28.02& 3.761& 0.240& 0.121&22111 175.000 &4-419.0& 0.10& 313.16& 375.16 &36:45.79 &12:13.69& 27.55& 27.90& 27.61& 27.92& 28.37& 27.81& 5.131& 0.270& 0.120&22111 176.000 &4-387.0& 0.06& 365.31& 445.42 &36:45.79 &12:22.62& 26.92& 27.26& 27.23& 27.16& 27.54& 27.42& 5.658& 0.620& 0.163&22111 177.000 &4-520.0& 0.17& 175.89& 192.10 &36:45.80 &11:50.38& 25.14& 25.23& 25.52& 25.28& 25.38& 25.56& 18.24& 1.340& 0.199&22111 178.000 &4-307.0& 1.18& 479.25& 599.75 &36:45.81 &12:42.14& 28.33& 29.16& 28.26& 99.99& 99.99& 99.99& 3.509& 0.140& 0.123&00000 179.000 &4-305.0& 0.97& 458.25& 571.25 &36:45.82 &12:38.57& 99.99& 99.99& 99.99& 28.49& 28.69& 28.23& 2.861& 0.987& 0.091&01000 180.000 &4-497.0& 1.22& 208.66& 240.33 &36:45.83 &11:56.33& 28.70& 28.64& 28.87& 99.99& 99.99& 99.99& 3.269& 0.090& 0.100&10000 181.000 &4-305.0& 1.40& 449.00& 561.50 &36:45.84 &12:37.21& 99.99& 99.99& 99.99& 28.69& 29.61& 28.59& 2.810& 0.846& 0.052&00000 182.000 &4-552.0& 0.33& 126.30& 130.80 &36:45.85 &11:42.39& 99.99& 99.99& 99.99& 27.92& 28.28& 27.84& 2.937& 0.573& 0.138&01000 183.000 &4-305.0& 0.62& 451.50& 569.50 &36:45.88 &12:38.02& 99.99& 99.99& 99.99& 27.98& 28.59& 27.66& 3.669& 0.718& 0.121&00000 184.000 &4-473.0& 0.08& 215.83& 256.95 &36:45.90 &11:58.12& 26.80& 27.01& 27.01& 26.48& 26.70& 26.77& 5.773& 0.720& 0.155&21111 185.000 &4-337.2& 0.85& 397.48& 503.75 &36:45.92 &12:29.30& 28.17& 29.48& 28.54& 99.00& 99.00& 99.00& 11.60& 0.060& 0.000&10003 186.000 &4-305.0& 0.09& 449.50& 574.54 &36:45.94 &12:38.31& 25.31& 25.53& 26.02& 25.55& 25.82& 26.10& 15.55& 1.360& 0.249&22111 187.000 &4-326.0& 0.03& 409.20& 521.53 &36:45.95 &12:31.49& 26.97& 27.30& 27.04& 27.01& 27.45& 27.14& 7.849& 0.330& 0.133&22111 188.000 &4-514.0& 0.29& 161.04& 189.65 &36:45.95 &11:49.31& 25.82& 25.89& 26.22& 25.82& 26.01& 26.24& 10.80& 1.140& 0.211&22101 189.000 &4-460.0& 0.22& 230.81& 284.67 &36:45.96 &12:1.307& 23.27& 23.34& 24.24& 23.50& 23.55& 24.41& 49.73& 4.100& 0.348&22111 190.000 &4-391.0& 0.07& 325.00& 415.00 &36:46.00 &12:17.67& 28.71& 29.45& 28.46& 28.35& 29.03& 28.18& 3.240& 0.090& 0.065&21111 191.000 &4-353.0& 0.02& 359.12& 464.87 &36:46.04 &12:23.82& 27.98& 28.42& 27.89& 28.65& 28.91& 28.49& 4.201& 0.190& 0.100&22111 192.000 &4-331.2& 0.64& 379.15& 494.44 &36:46.05 &12:27.44& 28.51& 30.13& 28.88& 30.32& 99.00& 30.78& 6.400& 0.040& 0.000&00003 193.000 &4-331.2& 0.78& 372.66& 489.76 &36:46.09 &12:26.65& 28.00& 29.07& 28.37& 27.69& 28.80& 28.15& 16.80& 0.070& 0.000&11003 194.000 &4-288.0& 0.24& 437.00& 574.50 &36:46.09 &12:37.53& 99.99& 99.99& 99.99& 28.44& 29.31& 28.31& 3.552& 1.118& 0.042&00010 902.000 &4-550.0& 0.54& 103.80& 132.53 &36:46.13 &11:41.17& 99.99& 99.99& 99.99& 23.19& 24.13& 23.71& 204.7& 16.06& 0.150&01000 195.000 &4-331.0& 0.09& 375.00& 497.50 &36:46.14 &12:27.44& 28.89& 29.78& 28.74& 28.94& 29.69& 29.04& 2.933& 0.080& 0.047&22111 196.000 &4-285.0& 0.05& 429.46& 572.66 &36:46.16 &12:36.88& 27.19& 27.51& 27.38& 99.99& 99.99& 99.99& 6.405& 0.330& 0.134&20011 197.000 &4-299.0& 0.17& 416.12& 555.75 &36:46.16 &12:34.67& 27.83& 28.29& 28.03& 28.08& 28.60& 27.89& 4.438& 0.220& 0.155&22111 198.200 &4-522.0& 0.21& 122.47& 163.02 &36:46.16 &11:44.75& 25.09& 25.12& 25.48& 25.32& 25.44& 25.66& 24.42& 1.170& 0.215&22111 198.100 &4-550.0& 0.61& 105.23& 142.07 &36:46.19 &11:41.86& 19.84& 19.87& 22.01& 20.38& 20.45& 22.69& 292.8& 27.98& 0.834&22101 198.000 &4-550.0& 0.63& 106.77& 143.04 &36:46.19 &11:41.88& 19.83& 19.85& 22.01& 20.38& 20.45& 22.69& 269.0& 31.00& 0.862&22101 199.000 &4-294.0& 0.21& 417.40& 560.20 &36:46.19 &12:35.13& 28.25& 29.00& 28.18& 99.99& 99.99& 99.99& 3.718& 0.150& 0.114&20011 201.100 &4-322.0& 0.02& 377.44& 508.20 &36:46.20 &12:28.45& 23.72& 24.38& 24.41& 24.13& 24.67& 24.66& 108.9& 0.690& 0.160&22111 200.000 &4-465.0& 0.21& 181.00& 245.50 &36:46.20 &11:55.04& 29.04& 30.01& 28.93& 99.99& 99.99& 99.99& 2.556& 0.080& 0.046&20011 201.000 &4-322.0& 0.35& 378.54& 512.71 &36:46.21 &12:28.77& 23.32& 23.40& 24.41& 23.60& 23.70& 24.66& 44.67& 4.490& 0.466&22100 201.200 &4-322.1& 0.10& 379.72& 514.44 &36:46.23 &12:29.10& 24.60& 25.25& 25.09& 24.63& 25.17& 25.26& 53.74& 0.590& 0.170&22111 202.000 &4-300.0& 0.03& 403.80& 550.05 &36:46.26 &12:33.49& 24.58& 24.69& 25.17& 24.86& 24.97& 25.26& 24.70& 1.870& 0.238&22111 203.000 &4-341.0& 0.03& 340.50& 466.50 &36:46.27 &12:22.81& 27.41& 27.75& 27.54& 28.65& 28.97& 28.65& 5.139& 0.340& 0.158&22111 204.000 &4-448.0& 0.13& 203.91& 284.88 &36:46.27 &11:59.65& 24.15& 24.28& 25.49& 24.38& 24.53& 25.64& 23.95& 3.600& 0.414&22111 205.000 &4-262.0& 2.35& 436.83& 600.00 &36:46.31 &12:39.56& 27.76& 28.42& 27.72& 28.34& 29.34& 28.05& 5.325& 0.180& 0.120&12100 206.000 &4-322.1& 0.99& 379.00& 525.00 &36:46.33 &12:29.92& 28.92& 29.41& 28.72& 99.99& 99.99& 99.99& 2.634& 0.090& 0.075&10000 207.000 &4-327.0& 0.04& 347.32& 485.77 &36:46.36 &12:24.77& 25.50& 25.66& 25.97& 25.95& 26.13& 26.33& 14.62& 1.110& 0.210&22111 208.000 &4-372.0& 0.03& 292.80& 416.45 &36:46.39 &12:15.80& 26.18& 26.43& 26.69& 26.33& 26.58& 26.76& 8.814& 0.910& 0.202&22111 209.000 &4-418.0& 0.34& 225.20& 328.88 &36:46.41 &12:4.507& 24.19& 24.36& 25.79& 24.45& 24.64& 25.93& 18.89& 4.700& 0.475&21101 210.000 &4-489.0& 0.20& 116.93& 186.06 &36:46.43 &11:46.29& 27.28& 27.59& 27.40& 99.99& 99.99& 99.99& 5.689& 0.350& 0.151&20011 211.000 &4-303.0& 0.04& 359.17& 515.47 &36:46.48 &12:27.93& 26.91& 27.21& 27.12& 27.95& 28.35& 27.85& 7.756& 0.370& 0.139&22111 212.000 &4-471.0& 0.18& 140.16& 227.69 &36:46.52 &11:51.17& 20.89& 20.91& 21.59& 21.31& 21.34& 21.94& 181.9& 15.49& 0.358&22111 213.000 &4-488.0& 0.26& 105.41& 183.80 &36:46.55 &11:45.39& 25.85& 26.00& 26.20& 26.30& 26.75& 26.62& 11.92& 0.900& 0.191&22101 214.000 &4-416.0& 0.07& 208.98& 322.61 &36:46.55 &12:3.023& 24.14& 24.20& 24.71& 24.22& 24.33& 24.76& 29.82& 2.530& 0.258&22111 215.000 &4-350.0& 0.02& 280.30& 422.50 &36:46.59 &12:15.53& 26.42& 26.58& 26.65& 26.75& 26.92& 26.88& 7.997& 0.740& 0.160&22111 216.000 &4-434.0& 0.15& 171.40& 277.01 &36:46.59 &11:57.02& 25.52& 25.63& 26.01& 25.54& 25.71& 25.96& 14.16& 1.140& 0.218&21111 217.000 &4-262.0& 1.11& 405.60& 591.70 &36:46.60 &12:36.98& 27.41& 27.93& 27.48& 28.57& 29.28& 28.45& 5.676& 0.270& 0.112&22100 218.000 &4-412.0& 0.18& 198.42& 315.00 &36:46.61 &12:1.774& 28.29& 29.00& 28.33& 99.99& 99.99& 99.99& 2.843& 0.230& 0.118&20011 219.000 &4-280.0& 0.07& 362.00& 539.86 &36:46.66 &12:30.09& 27.16& 27.45& 27.22& 27.14& 27.55& 27.29& 6.704& 0.320& 0.140&22111 220.000 &4-475.0& 0.26& 102.78& 193.91 &36:46.66 &11:46.06& 26.38& 26.73& 26.72& 28.06& 28.76& 27.48& 8.220& 0.730& 0.159&22101 221.000 &4-411.0& 0.12& 188.71& 309.14 &36:46.67 &12:0.703& 27.90& 28.48& 27.80& 27.75& 28.22& 27.76& 4.679& 0.180& 0.100&22111 222.000 &4-442.0& 0.20& 147.91& 255.89 &36:46.67 &11:53.86& 24.75& 24.86& 25.35& 24.94& 25.03& 25.45& 20.65& 1.890& 0.242&21111 223.000 &4-488.0& 1.38& 94.666& 185.38 &36:46.70 &11:44.92& 99.99& 99.99& 99.99& 25.72& 26.73& 26.41& 11.76& 1.135& 0.267&01000 224.000 &4-475.2& 1.16& 94.890& 189.46 &36:46.71 &11:45.21& 25.24& 25.58& 26.36& 25.39& 25.64& 26.38& 24.50& 0.810& 0.000&22003 225.000 &4-336.0& 0.02& 276.10& 432.86 &36:46.73 &12:16.12& 26.14& 26.43& 26.73& 27.59& 28.02& 27.66& 8.661& 0.980& 0.207&22111 226.000 &4-334.0& 0.04& 284.27& 444.55 &36:46.73 &12:17.59& 26.68& 27.07& 27.00& 27.02& 27.22& 27.06& 7.052& 0.620& 0.148&21111 227.000 &4-241.2& 0.07& 397.14& 599.88 &36:46.76 &12:37.12& 21.80& 21.80& 22.06& 22.25& 22.25& 22.56& 222.9& 2.550& 0.000&22013 228.000 &4-388.2& 0.06& 219.00& 361.25 &36:46.77 &12:6.800& 28.69& 29.43& 28.32& 99.99& 99.99& 99.99& 2.702& 0.130& 0.100&20011 229.000 &4-241.2& 1.84& 378.75& 575.75 &36:46.78 &12:34.08& 99.99& 99.99& 99.99& 28.57& 29.13& 28.45& 2.533& 0.655& 0.109&01000 230.000 &4-432.0& 0.12& 139.00& 258.29 &36:46.81 &11:53.50& 27.43& 27.82& 27.68& 27.85& 28.52& 28.29& 4.724& 0.370& 0.148&21111 231.000 &4-382.2& 0.10& 210.20& 355.00 &36:46.82 &12:5.743& 27.76& 28.09& 27.75& 29.24& 29.93& 28.72& 4.405& 0.250& 0.101&21111 232.000 &4-360.0& 0.07& 231.00& 383.50 &36:46.82 &12:9.341& 28.10& 28.49& 28.07& 27.96& 28.60& 27.91& 3.227& 0.250& 0.111&21111 233.000 &4-241.2& 1.28& 382.60& 586.60 &36:46.84 &12:35.20& 99.99& 99.99& 99.99& 28.09& 28.70& 28.14& 3.062& 0.784& 0.130&01001 234.000 &4-216.0& 2.25& 362.33& 565.66 &36:46.88 &12:32.20& 28.82& 29.35& 28.72& 99.99& 99.99& 99.99& 2.623& 0.110& 0.075&10000 235.000 &4-289.0& 0.06& 323.43& 519.69 &36:46.93 &12:26.09& 25.05& 25.16& 25.62& 25.36& 25.50& 25.77& 19.06& 1.420& 0.238&22111 236.000 &4-295.0& 0.19& 305.70& 497.63 &36:46.94 &12:23.18& 26.44& 26.78& 27.09& 26.33& 26.74& 27.18& 7.534& 0.810& 0.233&22111 237.000 &4-382.0& 0.08& 199.65& 356.86 &36:46.95 &12:5.276& 23.92& 24.03& 24.85& 24.06& 24.13& 24.86& 31.74& 3.210& 0.319&22111 238.000 &4-415.0& 0.18& 143.34& 288.00 &36:47.01 &11:56.19& 25.63& 25.83& 26.33& 25.70& 25.91& 26.38& 12.36& 1.220& 0.250&22111 239.000 &4-266.0& 1.40& 329.50& 541.00 &36:47.06 &12:28.23& 99.99& 99.99& 99.99& 28.89& 29.65& 28.68& 2.730& 0.816& 0.055&01000 240.000 &4-332.0& 0.13& 235.07& 419.31 &36:47.09 &12:12.48& 24.90& 25.01& 25.55& 24.86& 24.99& 25.41& 20.56& 1.550& 0.252&22111 241.000 &4-335.0& 0.02& 205.44& 393.66 &36:47.21 &12:8.599& 27.75& 28.23& 27.76& 27.52& 27.89& 27.54& 4.587& 0.240& 0.116&22111 242.000 &4-332.2& 0.00& 227.60& 425.96 &36:47.23 &12:12.58& 26.26& 26.41& 26.49& 25.95& 26.20& 26.22& 9.082& 0.750& 0.162&22111 243.000 &4-385.0& 0.11& 146.60& 318.12 &36:47.24 &11:58.84& 25.91& 26.09& 26.36& 26.08& 26.28& 26.44& 10.61& 1.020& 0.201&22111 244.000 &4-232.0& 0.08& 330.23& 569.21 &36:47.28 &12:30.72& 22.43& 22.49& 23.73& 22.57& 22.65& 23.85& 63.81& 9.030& 0.520&22111 244.100 &4-232.0& 0.06& 331.74& 570.95 &36:47.28 &12:30.79& 22.45& 22.58& 23.73& 22.60& 22.75& 23.85& 90.37& 5.700& 0.446&22111 244.200 &4-232.2& 0.10& 319.40& 558.55 &36:47.32 &12:28.99& 26.58& 26.71& 26.64& 26.86& 27.02& 26.57& 8.191& 0.690& 0.171&21111 245.000 &4-298.0& 0.11& 240.61& 455.05 &36:47.33 &12:15.78& 27.14& 27.46& 27.53& 27.62& 27.84& 27.60& 5.890& 0.410& 0.188&22111 246.000 &4-356.0& 0.14& 146.45& 338.63 &36:47.42 &12:0.524& 27.39& 27.72& 27.43& 27.58& 27.88& 27.55& 5.822& 0.280& 0.116&22111 247.000 &4-319.0& 0.10& 190.38& 402.11 &36:47.46 &12:8.379& 25.98& 26.19& 26.48& 26.03& 26.22& 26.49& 10.11& 1.000& 0.211&21111 248.000 &4-356.0& 0.96& 139.37& 335.25 &36:47.47 &11:59.79& 28.13& 28.49& 28.20& 28.10& 28.49& 28.06& 3.572& 0.200& 0.103&22101 249.000 &4-315.0& 0.11& 205.72& 425.93 &36:47.48 &12:11.23& 24.57& 24.68& 25.51& 24.86& 24.96& 25.65& 21.26& 2.470& 0.333&21111 250.000 &4-291.0& 0.04& 221.35& 449.21 &36:47.51 &12:14.09& 27.21& 27.53& 27.31& 27.61& 28.01& 27.60& 6.248& 0.330& 0.129&22111 251.000 &4-346.0& 0.09& 152.70& 358.04 &36:47.52 &12:2.502& 24.15& 24.22& 25.06& 24.16& 24.28& 25.13& 28.41& 2.770& 0.317&21111 252.000 &4-386.0& 0.16& 93.585& 287.07 &36:47.58 &11:53.06& 25.85& 26.06& 26.46& 25.79& 26.13& 26.31& 11.27& 1.020& 0.223&22111 253.000 &4-182.0& 0.22& 324.50& 599.00 &36:47.62 &12:32.65& 99.99& 99.99& 99.99& 28.22& 28.99& 28.13& 3.510& 0.747& 0.094&02010 254.000 &4-182.0& 0.50& 322.00& 599.00 &36:47.64 &12:32.42& 28.23& 28.53& 28.52& 99.99& 99.99& 99.99& 3.527& 0.170& 0.117&10000 255.000 &4-301.2& 0.13& 181.37& 414.66 &36:47.66 &12:8.833& 27.74& 28.74& 28.11& 27.48& 28.75& 27.94& 2.600& 0.100& 0.000&11013 256.000 &4-283.0& 0.09& 202.82& 442.41 &36:47.66 &12:12.37& 27.12& 27.24& 27.39& 28.04& 28.55& 28.05& 6.265& 0.380& 0.178&22111 257.000 &4-231.0& 0.11& 259.83& 519.41 &36:47.67 &12:22.15& 27.62& 27.91& 27.71& 27.75& 28.38& 27.60& 4.838& 0.270& 0.119&22111 258.000 &4-308.0& 0.10& 177.43& 409.50 &36:47.68 &12:8.187& 27.09& 27.40& 27.26& 27.66& 28.08& 27.73& 6.808& 0.350& 0.149&22111 259.000 &4-247.0& 0.04& 242.50& 497.50 &36:47.68 &12:19.32& 28.65& 29.35& 28.55& 27.65& 27.95& 27.67& 2.939& 0.120& 0.082&22111 260.000 &4-297.0& 1.65& 186.00& 436.00 &36:47.80 &12:10.85& 28.77& 29.59& 28.56& 99.99& 99.99& 99.99& 2.764& 0.110& 0.066&10001 261.000 &4-212.0& 0.16& 263.26& 542.34 &36:47.82 &12:24.22& 25.67& 25.86& 26.31& 25.88& 26.08& 26.33& 12.63& 1.100& 0.250&22111 262.000 &4-267.0& 0.17& 200.64& 458.52 &36:47.83 &12:13.59& 27.53& 28.07& 27.91& 99.99& 99.99& 99.99& 3.752& 0.480& 0.233&20011 263.000 &4-267.0& 0.65& 196.28& 454.85 &36:47.85 &12:13.02& 28.38& 28.94& 28.16& 99.99& 99.99& 99.99& 3.108& 0.170& 0.114&10000 264.000 &4-304.0& 0.10& 150.50& 395.00 &36:47.86 &12:5.358& 26.25& 26.40& 26.43& 26.53& 26.65& 26.65& 9.461& 0.730& 0.150&21111 265.000 &4-174.0& 0.05& 291.12& 586.12 &36:47.89 &12:29.52& 27.64& 28.21& 27.53& 27.57& 28.02& 27.55& 5.606& 0.200& 0.102&22111 266.000 &4-212.0& 2.51& 265.00& 565.00 &36:48.01 &12:26.20& 28.76& 29.67& 28.68& 99.99& 99.99& 99.99& 2.771& 0.110& 0.062&10001 267.000 &4-314.0& 0.16& 115.09& 364.36 &36:48.01 &12:0.689& 27.56& 27.92& 27.59& 27.79& 27.97& 27.85& 5.461& 0.240& 0.114&22111 268.000 &4-187.0& 0.12& 246.00& 549.50 &36:48.10 &12:23.78& 29.05& 29.77& 28.98& 99.99& 99.99& 99.99& 2.514& 0.080& 0.057&10010 277.120 &4-260.0& 0.08& 191.55& 480.69 &36:48.13 &12:14.86& 24.11& 24.46& 24.66& 24.13& 24.69& 24.79& 70.33& 0.900& 0.187&22111 269.000 &4-265.0& 0.13& 153.00& 430.50 &36:48.15 &12:8.434& 99.99& 99.99& 99.99& 28.81& 29.88& 28.67& 2.512& 0.660& 0.045&01011 270.000 &4-198.0& 0.86& 234.50& 544.00 &36:48.18 &12:22.62& 28.36& 29.31& 28.23& 99.99& 99.99& 99.99& 3.361& 0.150& 0.116&10000 271.000 &4-170.0& 0.78& 231.40& 543.00 &36:48.21 &12:22.35& 28.48& 29.04& 28.33& 99.99& 99.99& 99.99& 2.914& 0.160& 0.092&10000 277.100&4-260.111& 0.28& 180.49& 478.66 &36:48.24 &12:13.96& 21.66& 21.73& 23.96& 21.76& 22.05& 24.33& 94.11& 16.80& 0.822&22101 277.110&4-260.111& 0.21& 179.09& 478.29 &36:48.25 &12:13.85& 21.78& 22.13& 23.96& 21.89& 22.44& 24.33& 163.7& 7.180& 0.596&22111 272.000 &4-135.0& 1.42& 261.00& 586.50 &36:48.25 &12:27.75& 99.99& 99.99& 99.99& 28.74& 29.43& 28.89& 2.684& 1.000& 0.058&01000 273.000 &4-170.0& 0.09& 228.62& 550.62 &36:48.31 &12:22.79& 28.24& 28.55& 28.18& 27.40& 28.05& 27.34& 3.047& 0.220& 0.134&21111 274.000 &4-200.0& 0.14& 192.62& 509.51 &36:48.37 &12:17.25& 26.02& 26.48& 26.66& 26.16& 26.46& 26.50& 10.20& 0.930& 0.200&22111 275.000 &4-237.0& 0.05& 141.00& 446.14 &36:48.42 &12:8.942& 28.00& 28.72& 27.97& 27.93& 28.36& 27.97& 4.022& 0.200& 0.101&22111 276.000 &4-135.0& 0.08& 247.14& 588.71 &36:48.42 &12:27.04& 27.89& 28.28& 27.74& 27.89& 28.27& 27.75& 4.726& 0.180& 0.092&22111 277.000&4-260.111& 1.69& 175.38& 489.09 &36:48.42 &12:14.76& 20.99& 21.05& 22.92& 99.99& 99.99& 99.99& 97.33& 32.80& 2.054&20000 278.000 &4-228.0& 0.02& 133.33& 445.33 &36:48.50 &12:8.393& 28.63& 29.52& 28.61& 99.99& 99.99& 99.99& 2.985& 0.120& 0.086&10011 279.000 &4-122.0& 0.31& 239.09& 594.63 &36:48.57 &12:27.02& 26.53& 26.88& 26.96& 26.34& 26.68& 27.24& 9.547& 0.470& 0.193&22101 280.000 &4-257.0& 0.15& 101.98& 413.78 &36:48.59 &12:3.902& 25.11& 25.20& 25.47& 25.42& 25.50& 25.62& 20.92& 1.120& 0.198&22111 277.211 &4-186.0& 0.09& 170.11& 509.01 &36:48.63 &12:15.83& 22.04& 22.25& 22.92& 23.09& 23.19& 23.75& 221.6& 3.210& 0.297&22111 281.000 &4-229.0& 0.10& 122.44& 445.73 &36:48.63 &12:7.761& 25.41& 25.52& 25.75& 25.59& 25.76& 25.85& 16.67& 1.040& 0.189&22111 277.210 &4-186.0& 0.17& 170.45& 511.07 &36:48.63 &12:15.91& 21.96& 22.08& 22.92& 22.94& 23.03& 23.75& 157.0& 6.060& 0.393&22111 277.200 &4-186.0& 0.19& 167.16& 507.53 &36:48.64 &12:15.71& 21.83& 21.91& 22.92& 23.09& 23.19& 23.75& 115.6& 10.67& 0.558&22111 277.220 &4-260.2& 0.03& 158.51& 497.39 &36:48.66 &12:14.18& 24.26& 24.37& 25.23& 24.15& 24.44& 25.38& 38.16& 2.090& 0.316&21111 282.000 &4-111.0& 0.12& 228.00& 591.00 &36:48.68 &12:26.09& 99.99& 99.99& 99.99& 27.92& 28.55& 27.69& 3.804& 0.901& 0.103&01010 277.212 &4-169.0& 0.08& 171.23& 519.32 &36:48.71 &12:16.74& 24.79& 25.00& 25.26& 25.19& 25.29& 25.61& 36.20& 0.970& 0.213&22111 283.000 &4-154.0& 0.04& 182.16& 539.93 &36:48.76 &12:19.10& 25.29& 25.40& 25.97& 25.48& 25.64& 26.12& 16.37& 1.290& 0.260&22111 284.000 &4-183.0& 0.08& 135.00& 487.50 &36:48.85 &12:11.93& 28.71& 29.32& 28.38& 99.99& 99.99& 99.99& 2.784& 0.120& 0.082&20011 285.000 &4-128.0& 0.13& 182.09& 554.36 &36:48.89 &12:20.25& 27.85& 28.30& 28.07& 99.99& 99.99& 99.99& 3.647& 0.300& 0.148&20011 286.000 &4-199.0& 0.08& 113.28& 464.00 &36:48.91 &12:8.736& 99.99& 99.99& 99.99& 27.98& 28.42& 27.88& 3.431& 0.734& 0.099&02011 287.000 &4-148.0& 0.14& 160.00& 527.80 &36:48.92 &12:16.78& 99.99& 99.99& 99.99& 28.30& 28.81& 28.13& 2.761& 0.705& 0.083&01010 288.100 &4-203.0& 0.11& 107.58& 459.56 &36:48.92 &12:7.981& 25.11& 25.49& 25.84& 25.68& 25.81& 25.94& 23.16& 1.060& 0.208&22111 288.000 &4-203.0& 0.26& 107.33& 461.59 &36:48.94 &12:8.077& 24.86& 25.08& 25.84& 25.68& 25.81& 25.94& 15.47& 2.610& 0.369&22101 288.200 &4-203.12& 0.16& 103.06& 465.25 &36:49.03 &12:8.160& 26.61& 27.00& 26.82& 27.91& 28.54& 27.99& 11.42& 0.340& 0.138&22111 289.000 &4-120.0& 0.57& 178.11& 570.54 &36:49.06 &12:21.28& 21.84& 21.87& 23.29& 22.02& 22.05& 23.50& 111.5& 8.500& 0.464&22101 290.000 &4-99.0& 0.13& 159.08& 569.75 &36:49.29 &12:20.10& 27.52& 28.06& 27.75& 28.10& 28.69& 28.02& 4.911& 0.310& 0.170&22111 291.000 &4-119.0& 2.05& 102.00& 494.50 &36:49.30 &12:10.52& 99.99& 99.99& 99.99& 28.44& 29.49& 28.23& 3.552& 0.942& 0.051&00000 292.000 &4-131.0& 0.18& 123.72& 527.72 &36:49.34 &12:14.53& 24.56& 24.70& 25.43& 24.64& 24.76& 25.52& 22.78& 2.200& 0.291&22111 293.000 &4-74.0& 0.34& 163.08& 596.20 &36:49.47 &12:22.49& 26.59& 26.96& 27.21& 28.36& 28.97& 28.15& 6.479& 0.830& 0.223&22101 294.000 &4-119.0& 0.03& 99.500& 515.75 &36:49.51 &12:12.04& 27.79& 28.29& 27.84& 27.42& 27.74& 27.16& 4.707& 0.210& 0.121&22111 295.100 &4-85.0& 0.09& 146.10& 580.31 &36:49.52 &12:20.14& 23.87& 24.04& 24.66& 24.02& 24.13& 24.74& 55.24& 1.680& 0.278&22111 295.000 &4-85.0& 0.48& 141.71& 581.26 &36:49.58 &12:20.01& 23.40& 23.48& 24.66& 99.99& 99.99& 99.99& 40.72& 4.620& 0.591&20000 296.000 &4-109.0& 0.11& 98.397& 524.45 &36:49.61 &12:12.70& 24.37& 24.43& 24.75& 24.95& 25.01& 25.19& 34.17& 1.500& 0.209&21111 297.000 &4-100.0& 0.12& 106.00& 535.50 &36:49.61 &12:14.06& 29.01& 29.99& 28.84& 99.99& 99.99& 99.99& 2.617& 0.080& 0.046&20011 298.000 &4-95.0& 0.15& 112.50& 545.18 &36:49.62 &12:15.27& 27.05& 27.32& 27.17& 27.75& 28.01& 27.66& 7.161& 0.340& 0.138&22111 295.200 &4-85.2& 0.12& 135.60& 583.23 &36:49.68 &12:19.76& 24.54& 24.71& 25.10& 25.11& 25.24& 25.48& 35.91& 1.270& 0.222&21111 299.000 &4-50.2& 0.65& 98.680& 564.65 &36:49.94 &12:15.98& 27.56& 28.92& 27.93& 99.00& 30.37& 99.00& 2.500& 0.070& 0.000&10003 300.000 &4-50.0& 0.65& 101.50& 571.00 &36:49.97 &12:16.68& 29.13& 29.76& 28.76& 99.99& 99.99& 99.99& 2.695& 0.060& 0.059&00000 302.200 &4-50.0& 0.15& 96.083& 570.75 &36:50.03 &12:16.33& 27.14& 27.18& 26.67& 27.26& 27.45& 27.58& 7.245& 0.330& 0.140&12111 301.000 &4-33.2& 0.48& 95.890& 584.44 &36:50.15 &12:17.42& 21.76& 21.76& 22.70& 22.34& 22.34& 23.23& 301.0& 3.830& 0.000&22003 302.000 &4-33.0& 0.13& 94.419& 581.94 &36:50.17 &12:17.05& 20.97& 21.00& 22.18& 21.58& 21.68& 22.78& 200.0& 10.17& 0.473&22110 302.100 &4-33.0& 0.14& 93.381& 582.58 &36:50.17 &12:17.05& 20.97& 21.02& 22.18& 21.58& 21.68& 22.78& 222.0& 8.970& 0.459&22110 303.000 &4-33.0& 1.25& 93.600& 592.20 &36:50.26 &12:17.97& 99.99& 99.99& 99.99& 27.85& 28.38& 27.76& 4.187& 1.046& 0.086&00000
[cccccccccccc]{} 20.000& 283.8& 100.0& 100.0& 100.0& 100.0& 1.000& 1.000 & & & 20.500& 287.9& 100.0& 100.0& 100.0& 100.0& 1.000& 0.833 & & & 21.000& 166.5& 100.0& 100.0& 100.0& 100.0& 1.000& 0.750 & & & 21.500& 195.8& 100.0& 100.0& 100.0& 100.0& 1.000& 0.750 & & & 22.000& 114.2& 100.0& 98.85& 100.0& 100.0& 1.000& 0.800 & & & 22.500& 125.0& 100.0& 97.75& 100.0& 98.88& 1.000& 0.900 & & & 23.000& 79.12& 100.0& 96.67& 100.0& 97.79& 1.000& 0.875 & & & 23.500& 48.68& 98.93& 95.62& 98.93& 96.71& 1.000& 0.542 & & & 24.000& 58.21& 97.89& 94.53& 97.89& 95.58& 1.000& 0.916 & & & 24.500& 28.31& 96.87& 93.35& 96.87& 94.28& 1.000& 0.663 & & & 25.000& 22.32& 95.83& 91.91& 95.83& 92.63& 0.961& 0.732 & & & 25.500& 17.82& 94.66& 90.01& 94.66& 90.36& 1.000& 0.783 & 97.6 & & 26.000& 12.59& 93.14& 87.31& 93.14& 87.11& 0.973& 0.806 & 96.2 & & 26.500& 9.327& 90.89& 83.32& 90.89& 82.42& 0.957& 0.722 & 95.1 & & 27.000& 7.192& 87.28& 77.38& 87.28& 75.68& 0.977& 0.601 & 92.2 & & 27.500& 5.989& 81.38& 68.62& 81.38& 66.17& 0.848& 0.610 & 89.1 & & 28.000& 4.852& 71.85& 55.91& 71.85& 53.01& 0.681& 0.458 & 76.6 & 0.94 & 0.87 28.500& 3.580& 56.86& 37.82& 56.86& 35.17& 0.461& 0.205 & 41.5 & 0.68 & 0.50 29.000& 2.889& 33.96& 12.57& 33.96& 11.44& 0.247& 0.122 & 14.1 && 29.500& 2.260& 0.000& 0.000& 0.000& 0.000& 0.157& 0.075 &&& 30.000& 1.635& 0.000& 0.000& 0.000& 0.000& 0.123& 0.079 &&&
[ccccccccccccccccc]{} 501.000 &4-851.0& 0.53& 557.50& 139.00 &36:40.88 &12:9.423& 29.98& 29.85& 30.34& 99.99& 99.99& 99.99& 1.255& 0.060& 0.051&20000 502.000 &4-807.0& 0.15& 561.00& 170.50 &36:41.11 &12:12.19& 29.35& 30.28& 29.02& 28.28& 28.79& 28.19& 1.917& 0.080& 0.050&20111 503.000 &4-822.0& 1.01& 543.50& 151.00 &36:41.15 &12:9.547& 29.99& 29.89& 29.69& 99.99& 99.99& 99.99& 1.240& 0.060& 0.051&10000 504.000 &4-822.2& 1.47& 538.16& 149.94 &36:41.19 &12:9.121& 27.82& 30.29& 28.19& 99.00& 99.00& 99.00& 0.900& 0.020& 0.000&00003 505.000 &4-807.0& 1.71& 559.33& 188.16 &36:41.28 &12:13.54& 28.64& 28.82& 28.64& 99.99& 99.99& 99.99& 2.365& 0.180& 0.119&20000 506.000 &4-766.0& 1.00& 562.50& 197.00 &36:41.33 &12:14.49& 99.99& 99.99& 99.99& 29.14& 30.00& 29.14& 2.146& 0.592& 0.050&01000 507.000 &4-790.0& 1.00& 531.50& 170.00 &36:41.45 &12:10.35& 29.74& 29.46& 29.67& 99.99& 99.99& 99.99& 1.564& 0.060& 0.054&10000 508.000 &4-739.0& 0.58& 554.00& 218.00 &36:41.61 &12:15.63& 28.87& 29.35& 28.86& 99.99& 99.99& 99.99& 2.069& 0.150& 0.112&00000 509.000 &4-739.2& 1.43& 541.66& 220.44 &36:41.77 &12:15.06& 27.96& 30.28& 28.33& 31.83& 99.00& 32.29& 0.700& 0.020& 0.000&00003 510.000 &4-795.2& 0.91& 495.16& 167.43 &36:41.84 &12:7.912& 27.57& 29.67& 27.94& 99.00& 99.00& 99.00& 1.700& 0.040& 0.000&00003 511.000 &4-683.0& 0.63& 574.50& 275.00 &36:41.87 &12:21.53& 30.84& 29.52& 0.000& 99.99& 99.99& 99.99& 0.567& 0.060& 0.052&10000 512.000 &4-781.0& 1.09& 459.33& 130.33 &36:41.96 &12:2.749& 99.99& 99.99& 99.99& 29.06& 29.24& 29.15& 1.595& 0.794& 0.081&00000 513.000 &4-746.0& 0.45& 478.00& 167.50 &36:42.05 &12:6.869& 29.84& 30.50& 29.45& 99.99& 99.99& 99.99& 1.164& 0.080& 0.049&00000 514.000 &4-711.2& 0.09& 522.67& 230.44 &36:42.07 &12:14.72& 27.93& 28.93& 28.30& 27.37& 29.24& 27.83& 1.800& 0.060& 0.000&10013 515.000 &4-709.2& 0.17& 512.16& 225.43 &36:42.15 &12:13.68& 27.79& 29.56& 28.16& 30.35& 99.00& 30.81& 1.300& 0.040& 0.000&10013 516.000 &4-665.0& 0.42& 564.00& 295.50 &36:42.17 &12:22.60& 99.99& 99.99& 99.99& 28.82& 29.79& 28.28& 2.483& 0.719& 0.044&00000 517.000 &4-757.0& 0.13& 449.00& 144.50 &36:42.19 &12:3.229& 28.86& 29.11& 28.63& 99.99& 99.99& 99.99& 2.416& 0.120& 0.086&20010 518.000 &4-652.0& 0.53& 572.50& 311.00 &36:42.20 &12:24.33& 29.31& 29.94& 28.71& 99.99& 99.99& 99.99& 2.325& 0.060& 0.044&20000 519.000 &4-671.0& 0.76& 540.00& 277.50 &36:42.29 &12:19.61& 29.30& 29.88& 29.17& 99.99& 99.99& 99.99& 2.009& 0.080& 0.063&10000 520.000 &4-657.0& 1.34& 547.50& 294.00 &36:42.35 &12:21.42& 29.56& 30.03& 29.10& 99.99& 99.99& 99.99& 1.847& 0.060& 0.050&10000 521.000 &4-671.0& 0.47& 531.33& 278.66 &36:42.41 &12:19.22& 99.99& 99.99& 99.99& 28.79& 29.22& 28.63& 2.132& 0.805& 0.062&00000 522.000 &4-671.2& 1.07& 522.00& 275.63 &36:42.48 &12:18.36& 28.49& 29.95& 28.86& 29.28& 30.63& 29.74& 1.100& 0.030& 0.000&00003 523.000 &4-707.0& 0.88& 420.50& 166.00 &36:42.71 &12:3.229& 29.34& 30.26& 29.01& 99.99& 99.99& 99.99& 2.251& 0.060& 0.047&10001 524.000 &4-655.0& 1.75& 490.00& 258.50 &36:42.72 &12:15.05& 99.99& 99.99& 99.99& 29.51& 30.27& 29.14& 1.320& 0.462& 0.052&00000 525.000 &4-584.2& 0.20& 566.82& 365.77 &36:42.74 &12:28.44& 27.85& 28.85& 28.22& 29.01& 99.00& 29.47& 1.900& 0.060& 0.000&10013 526.000 &4-707.0& 0.43& 420.00& 170.50 &36:42.76 &12:3.573& 29.30& 30.26& 28.94& 27.55& 28.08& 27.59& 1.981& 0.080& 0.054&22100 527.000 &4-715.2& 1.60& 434.18& 191.49 &36:42.76 &12:6.141& 28.24& 29.66& 28.61& 99.00& 99.00& 99.00& 1.300& 0.040& 0.000&00003 528.000 &4-616.0& 0.11& 531.50& 322.00 &36:42.79 &12:22.77& 99.99& 99.99& 99.99& 29.41& 29.62& 29.00& 1.695& 0.838& 0.053&00011 529.000 &4-581.2& 1.03& 544.00& 349.50 &36:42.87 &12:25.73& 29.15& 29.85& 28.93& 99.99& 99.99& 99.99& 2.304& 0.080& 0.050&10000 530.000 &4-606.0& 0.51& 524.50& 323.00 &36:42.87 &12:22.42& 99.99& 99.99& 99.99& 29.30& 29.86& 29.30& 1.882& 0.672& 0.052&00000 531.000 &4-537.0& 1.50& 576.00& 398.50 &36:42.93 &12:31.68& 29.43& 29.96& 29.54& 99.99& 99.99& 99.99& 1.773& 0.080& 0.056&10000 532.000 &4-614.2& 0.69& 512.66& 317.47 &36:42.95 &12:21.19& 28.68& 30.40& 29.05& 99.00& 31.86& 99.00& 0.800& 0.020& 0.000&00003 533.000 &4-606.2& 1.42& 529.17& 342.43 &36:42.97 &12:24.24& 28.50& 29.60& 28.87& 29.15& 29.76& 29.61& 1.600& 0.040& 0.000&00003 534.000 &4-599.2& 1.57& 502.67& 307.66 &36:42.98 &12:19.79& 28.09& 29.68& 28.46& 28.31& 30.53& 28.77& 1.700& 0.040& 0.000&00003 535.000 &4-554.2& 0.47& 548.18& 372.94 &36:43.02 &12:27.89& 28.09& 28.70& 28.46& 29.03& 30.54& 29.49& 1.900& 0.060& 0.000&20003 536.000 &4-614.2& 1.07& 511.02& 323.61 &36:43.02 &12:21.59& 28.52& 29.87& 28.89& 29.01& 32.07& 29.47& 1.000& 0.030& 0.000&00003 537.000 &4-649.0& 0.43& 449.00& 239.50 &36:43.03 &12:11.00& 99.99& 99.99& 99.99& 28.93& 29.71& 28.84& 2.243& 0.768& 0.044&00000 538.000 &4-614.0& 1.21& 513.50& 326.00 &36:43.03 &12:22.00& 99.99& 99.99& 99.99& 29.10& 29.55& 28.75& 2.246& 0.892& 0.057&01000 539.000 &4-630.0& 0.06& 465.50& 269.00 &36:43.09 &12:14.36& 29.65& 30.10& 29.10& 99.99& 99.99& 99.99& 1.701& 0.060& 0.059&20010 540.000 &4-695.0& 0.55& 381.50& 168.00 &36:43.19 &12:1.005& 29.72& 29.92& 29.93& 99.99& 99.99& 99.99& 1.583& 0.060& 0.052&10000 541.000 &4-606.0& 2.44& 518.50& 351.00 &36:43.19 &12:24.33& 99.99& 99.99& 99.99& 29.41& 29.72& 29.18& 1.698& 0.764& 0.050&01000 542.000 &4-666.2& 0.07& 405.32& 202.27 &36:43.20 &12:5.248& 28.67& 29.96& 29.04& 29.19& 31.07& 29.65& 0.800& 0.030& 0.000&00013 543.000 &4-541.2& 2.61& 515.34& 361.24 &36:43.30 &12:24.92& 28.13& 29.74& 28.50& 29.02& 31.98& 29.48& 1.000& 0.030& 0.000&00003 544.000 &4-645.2& 0.23& 413.00& 224.62 &36:43.30 &12:7.542& 28.07& 30.11& 28.44& 28.22& 29.97& 28.68& 1.000& 0.030& 0.000&10013 545.000 &4-553.2& 2.22& 503.66& 350.42 &36:43.34 &12:23.33& 28.42& 30.30& 28.79& 99.00& 99.00& 99.00& 0.700& 0.020& 0.000&00003 546.000 &4-669.0& 2.80& 341.00& 131.50 &36:43.34 &11:55.56& 29.66& 29.78& 30.00& 99.99& 99.99& 99.99& 1.446& 0.080& 0.059&10000 547.000 &4-583.2& 0.49& 461.17& 309.41 &36:43.48 &12:17.38& 27.62& 29.41& 28.00& 28.35& 30.78& 28.81& 1.200& 0.040& 0.000&00003 548.000 &4-645.0& 1.45& 402.00& 233.50 &36:43.52 &12:7.583& 29.13& 29.61& 29.21& 99.99& 99.99& 99.99& 2.342& 0.080& 0.071&10000 549.000 &4-583.2& 0.51& 457.18& 308.94 &36:43.52 &12:17.11& 27.75& 30.18& 28.12& 27.64& 30.45& 28.10& 0.600& 0.020& 0.000&00003 550.000 &4-492.0& 1.46& 527.00& 411.50 &36:43.61 &12:29.73& 29.25& 29.54& 29.26& 99.99& 99.99& 99.99& 2.105& 0.080& 0.047&10000 551.000 &4-513.0& 1.56& 489.00& 366.50 &36:43.67 &12:23.78& 99.99& 99.99& 99.99& 29.11& 29.65& 28.95& 1.914& 0.816& 0.049&00000 552.000 &4-479.0& 1.73& 546.00& 445.50 &36:43.69 &12:33.71& 99.99& 99.99& 99.99& 28.95& 29.69& 28.65& 2.226& 0.787& 0.051&02000 553.000 &4-525.0& 0.92& 472.00& 352.50 &36:43.75 &12:21.61& 99.99& 99.99& 99.99& 28.90& 29.63& 28.85& 2.313& 0.827& 0.058&01000 554.000 &4-492.0& 0.10& 513.66& 412.33 &36:43.78 &12:29.02& 99.99& 99.99& 99.99& 28.76& 29.19& 28.45& 2.487& 0.832& 0.087&02010 555.000 &4-479.0& 1.20& 522.00& 429.50 &36:43.84 &12:30.94& 99.99& 99.99& 99.99& 29.16& 29.64& 28.90& 1.829& 0.822& 0.050&00000 556.000 &4-437.2& 0.97& 565.27& 489.74 &36:43.84 &12:38.44& 28.11& 29.11& 28.48& 99.00& 99.00& 99.00& 1.400& 0.050& 0.000&01003 557.000 &4-486.0& 0.20& 514.00& 421.50 &36:43.85 &12:29.76& 29.27& 29.82& 29.05& 27.86& 28.46& 27.91& 2.052& 0.080& 0.047&12110 558.000 &4-479.0& 0.53& 527.00& 438.50 &36:43.86 &12:31.97& 99.99& 99.99& 99.99& 29.22& 29.96& 29.38& 1.706& 0.611& 0.048&01001 559.000 &4-479.0& 0.94& 523.00& 433.50 &36:43.86 &12:31.32& 99.99& 99.99& 99.99& 29.16& 29.75& 28.91& 1.816& 0.746& 0.048&01000 560.000 &4-479.0& 0.67& 525.00& 436.50 &36:43.86 &12:31.69& 99.99& 99.99& 99.99& 29.20& 29.76& 28.90& 1.759& 0.737& 0.054&01000 561.000 &4-437.2& 0.10& 567.66& 498.44 &36:43.89 &12:39.29& 28.31& 30.20& 28.68& 29.42& 31.82& 29.88& 0.600& 0.020& 0.000&00013 562.000 &4-663.0& 0.19& 302.00& 143.50 &36:43.90 &11:54.15& 29.21& 30.04& 29.06& 28.26& 29.00& 28.14& 2.179& 0.080& 0.051&20110 563.000 &4-479.2& 1.32& 529.66& 450.43 &36:43.91 &12:33.05& 28.12& 30.19& 28.49& 99.00& 99.00& 99.00& 0.600& 0.020& 0.000&00003 564.000 &4-481.0& 0.58& 484.00& 401.50 &36:44.03 &12:26.34& 99.99& 99.99& 99.99& 29.19& 29.45& 29.41& 1.777& 0.977& 0.057&01000 565.000 &4-640.0& 0.15& 300.00& 163.50 &36:44.11 &11:55.69& 99.99& 99.99& 99.99& 29.46& 30.22& 29.22& 1.388& 0.483& 0.054&01010 566.000 &4-393.2& 1.76& 574.33& 542.33 &36:44.20 &12:43.28& 29.82& 30.20& 29.85& 99.99& 99.99& 99.99& 0.960& 0.120& 0.071&10000 567.000 &4-499.0& 1.61& 416.00& 332.50 &36:44.22 &12:16.50& 29.31& 29.68& 29.23& 99.99& 99.99& 99.99& 1.986& 0.080& 0.049&10000 568.000 &4-577.0& 0.12& 343.25& 239.25 &36:44.25 &12:4.465& 28.93& 29.64& 28.60& 28.85& 29.73& 28.65& 2.381& 0.110& 0.082&22111 569.000 &4-393.2& 1.06& 567.50& 549.00 &36:44.35 &12:43.45& 99.99& 99.99& 99.99& 30.01& 30.34& 29.34& 0.969& 0.431& 0.050&01000 570.000 &4-624.2& 0.22& 275.15& 164.43 &36:44.39 &11:54.20& 28.17& 29.87& 28.54& 28.42& 30.65& 28.88& 1.000& 0.030& 0.000&00013 571.000 &4-534.0& 0.06& 363.50& 289.00 &36:44.45 &12:9.753& 29.47& 30.17& 29.02& 28.70& 29.53& 28.64& 2.006& 0.060& 0.041&22111 572.000 &4-393.2& 0.93& 572.00& 568.50 &36:44.46 &12:45.28& 29.48& 30.09& 29.21& 99.99& 99.99& 99.99& 1.671& 0.080& 0.050&20000 573.000 &4-393.2& 0.66& 567.50& 564.00 &36:44.48 &12:44.67& 99.99& 99.99& 99.99& 28.99& 29.64& 28.43& 2.424& 0.823& 0.048&01000 574.000 &4-430.2& 2.36& 535.36& 523.28 &36:44.48 &12:39.34& 28.80& 29.75& 29.17& 99.00& 30.71& 99.00& 0.900& 0.030& 0.000&00003 575.000 &4-452.0& 0.46& 456.00& 421.50 &36:44.54 &12:26.26& 99.99& 99.99& 99.99& 29.18& 29.80& 29.29& 1.793& 0.710& 0.054&01000 576.000 &4-452.0& 0.71& 453.50& 422.00 &36:44.57 &12:26.13& 99.99& 99.99& 99.99& 29.39& 30.05& 29.08& 1.716& 0.563& 0.055&01000 577.000 &4-400.2& 0.08& 512.66& 506.41 &36:44.60 &12:36.57& 28.25& 30.15& 28.62& 27.50& 29.88& 27.96& 0.600& 0.020& 0.000&00013 578.000 &4-452.0& 0.94& 451.00& 421.50 &36:44.60 &12:25.94& 99.99& 99.99& 99.99& 28.99& 29.80& 29.14& 2.129& 0.712& 0.049&01000 579.000 &4-546.0& 0.06& 322.00& 251.50 &36:44.61 &12:4.163& 29.41& 30.42& 29.20& 29.01& 29.74& 28.85& 1.805& 0.080& 0.056&22111 580.000 &4-422.2& 1.05& 465.15& 446.94 &36:44.63 &12:28.82& 27.73& 30.07& 28.10& 30.19& 99.00& 30.65& 0.700& 0.020& 0.000&00003 581.000 &4-400.2& 1.40& 518.14& 518.43 &36:44.64 &12:37.89& 28.49& 29.55& 28.86& 99.00& 99.00& 99.00& 1.300& 0.040& 0.000&00003 582.000 &4-409.0& 0.98& 501.00& 496.50 &36:44.67 &12:35.11& 99.99& 99.99& 99.99& 29.10& 29.56& 28.81& 1.936& 0.883& 0.060&00000 583.000 &4-377.2& 0.52& 550.08& 565.31 &36:44.67 &12:43.65& 27.85& 29.35& 28.22& 27.71& 29.90& 28.17& 1.700& 0.040& 0.000&00003 584.000 &4-483.0& 0.87& 376.33& 332.33 &36:44.68 &12:14.07& 28.87& 29.88& 28.70& 99.99& 99.99& 99.99& 2.391& 0.120& 0.070&10001 585.000 &4-455.0& 1.42& 404.50& 371.00 &36:44.69 &12:18.99& 99.99& 99.99& 99.99& 29.14& 29.77& 28.82& 2.169& 0.730& 0.054&01001 586.000 &4-472.2& 0.56& 410.00& 378.50 &36:44.70 &12:19.92& 99.99& 99.99& 99.99& 28.98& 29.54& 28.96& 2.153& 0.902& 0.055&02001 587.000 &4-566.0& 0.31& 275.00& 204.50 &36:44.76 &11:57.50& 99.99& 99.99& 99.99& 28.79& 29.59& 28.56& 1.775& 0.430& 0.088&02000 588.000 &4-483.0& 0.11& 373.33& 339.33 &36:44.78 &12:14.44& 28.84& 29.77& 28.79& 99.99& 99.99& 99.99& 2.467& 0.120& 0.078&20011 589.000 &4-422.0& 2.48& 473.00& 476.50 &36:44.82 &12:31.76& 99.99& 99.99& 99.99& 29.14& 29.68& 29.29& 1.864& 0.794& 0.049&01000 590.000 &4-423.0& 0.64& 441.50& 437.00 &36:44.83 &12:26.57& 29.30& 29.86& 29.38& 99.99& 99.99& 99.99& 2.334& 0.060& 0.063&10000 591.000 &4-446.0& 1.45& 421.00& 417.50 &36:44.90 &12:23.74& 29.07& 30.02& 28.79& 99.99& 99.99& 99.99& 2.481& 0.080& 0.049&20000 592.000 &4-458.0& 0.04& 383.66& 371.66 &36:44.94 &12:17.71& 29.13& 29.92& 28.90& 27.75& 28.09& 27.78& 1.979& 0.110& 0.080&22111 593.000 &4-340.2& 1.94& 546.66& 595.40 &36:44.98 &12:45.91& 28.33& 29.93& 28.70& 31.66& 99.00& 32.12& 0.600& 0.020& 0.000&10003 594.000 &4-378.2& 1.86& 499.87& 533.32 &36:44.98 &12:37.98& 27.81& 29.31& 28.18& 28.03& 29.47& 28.49& 1.700& 0.040& 0.000&10003 595.000 &4-519.0& 0.14& 281.33& 251.33 &36:45.08 &12:1.664& 28.97& 29.73& 28.88& 99.99& 99.99& 99.99& 2.185& 0.120& 0.073&20011 596.000 &4-329.2& 0.85& 528.13& 583.41 &36:45.09 &12:43.79& 27.80& 28.83& 28.17& 28.08& 30.25& 28.54& 1.900& 0.060& 0.000&00003 597.000 &4-596.0& 0.69& 206.00& 154.50 &36:45.12 &11:49.16& 29.64& 30.21& 29.77& 99.99& 99.99& 99.99& 1.451& 0.080& 0.050&20000 598.000 &4-573.0& 0.51& 207.50& 162.00 &36:45.17 &11:49.90& 99.99& 99.99& 99.99& 29.66& 30.37& 29.04& 1.348& 0.421& 0.049&01000 599.000 &4-593.2& 0.40& 178.97& 132.62 &36:45.23 &11:45.73& 28.12& 29.81& 28.49& 27.92& 31.87& 28.38& 0.700& 0.030& 0.000&10003 600.000 &4-368.0& 0.94& 459.66& 522.33 &36:45.37 &12:34.63& 29.11& 29.66& 29.23& 99.99& 99.99& 99.99& 2.215& 0.090& 0.060&20000 601.000 &4-487.0& 0.96& 264.00& 261.50 &36:45.38 &12:1.472& 99.99& 99.99& 99.99& 29.70& 30.24& 29.64& 1.107& 0.473& 0.051&01000 602.000 &4-368.0& 0.65& 449.00& 511.50 &36:45.40 &12:33.15& 99.99& 99.99& 99.99& 29.48& 29.80& 29.33& 1.355& 0.713& 0.052&00000 603.000 &4-342.0& 0.75& 482.00& 556.50 &36:45.41 &12:38.83& 99.99& 99.99& 99.99& 28.92& 29.48& 29.16& 2.286& 0.951& 0.055&00000 604.000 &4-559.2& 0.21& 179.98& 154.42 &36:45.41 &11:47.56& 27.85& 29.37& 28.22& 27.82& 29.38& 28.28& 1.600& 0.050& 0.000&01013 605.000 &4-454.0& 0.01& 332.74& 358.70 &36:45.41 &12:13.54& 22.07& 22.10& 22.35& 21.43& 21.45& 21.68& 119.0& 5.920& 0.206&22111 606.000 &4-593.2& 1.28& 163.66& 136.45 &36:45.45 &11:45.10& 28.36& 30.40& 28.73& 29.23& 32.69& 29.69& 0.500& 0.020& 0.000&00003 607.000 &4-384.2& 0.54& 414.66& 472.43 &36:45.45 &12:27.81& 28.57& 30.41& 28.94& 30.15& 99.00& 30.61& 0.800& 0.020& 0.000&10003 608.000 &4-397.0& 0.13& 388.50& 442.00 &36:45.49 &12:23.74& 29.51& 30.24& 29.14& 99.99& 99.99& 99.99& 1.935& 0.060& 0.047&10011 609.000 &4-582.0& 0.17& 150.00& 126.50 &36:45.52 &11:43.46& 30.12& 30.39& 30.21& 99.99& 99.99& 99.99& 0.938& 0.080& 0.048&10010 610.000 &4-330.0& 1.27& 483.66& 576.33 &36:45.56 &12:40.54& 99.99& 99.99& 99.99& 28.83& 29.00& 28.49& 2.319& 0.987& 0.072&01000 611.000 &4-312.0& 0.82& 499.50& 599.00 &36:45.57 &12:43.31& 29.63& 30.14& 29.05& 99.99& 99.99& 99.99& 1.709& 0.060& 0.050&10000 612.000 &4-312.0& 0.70& 497.66& 596.33 &36:45.57 &12:43.04& 99.99& 99.99& 99.99& 28.77& 29.05& 28.53& 2.442& 0.942& 0.071&01000 613.000 &4-307.0& 1.02& 485.50& 581.00 &36:45.57 &12:40.99& 29.34& 30.04& 28.90& 99.99& 99.99& 99.99& 2.245& 0.060& 0.049&10000 614.000 &4-330.2& 1.04& 478.16& 572.94 &36:45.58 &12:39.88& 27.81& 28.94& 28.18& 30.44& 30.50& 30.90& 2.000& 0.060& 0.000&10003 615.000 &4-389.2& 0.72& 384.66& 450.45 &36:45.60 &12:24.18& 28.39& 30.65& 28.76& 99.00& 33.11& 99.00& 1.300& 0.020& 0.000&00003 616.000 &4-330.0& 1.60& 476.00& 577.50 &36:45.65 &12:40.13& 29.16& 30.10& 28.82& 99.99& 99.99& 99.99& 2.278& 0.080& 0.045&10001 617.000 &4-371.0& 0.04& 396.00& 474.50 &36:45.69 &12:26.85& 29.35& 30.22& 29.18& 99.99& 99.99& 99.99& 1.912& 0.080& 0.054&20011 618.000 &4-563.2& 0.75& 140.18& 135.94 &36:45.71 &11:43.61& 27.87& 30.36& 28.24& 99.00& 99.00& 99.00& 0.500& 0.020& 0.000&00003 619.000 &4-307.0& 0.88& 478.50& 589.00 &36:45.73 &12:41.21& 29.40& 29.80& 29.01& 99.99& 99.99& 99.99& 2.085& 0.060& 0.049&10000 620.000 &4-563.2& 1.19& 141.66& 140.19 &36:45.73 &11:44.05& 27.98& 29.63& 28.35& 27.75& 29.77& 28.21& 1.100& 0.040& 0.000&00003 621.000 &4-520.2& 1.45& 187.16& 202.94 &36:45.75 &11:51.95& 28.71& 30.57& 29.08& 99.00& 34.48& 99.00& 0.500& 0.020& 0.000&00003 622.000 &4-563.2& 1.13& 136.16& 138.44 &36:45.78 &11:43.57& 27.67& 29.24& 28.04& 28.70& 30.35& 29.16& 1.800& 0.050& 0.000&01003 623.000 &4-510.0& 0.76& 172.50& 186.00 &36:45.78 &11:49.68& 29.90& 30.50& 29.53& 99.99& 99.99& 99.99& 1.344& 0.060& 0.051&10000 624.000 &4-484.0& 0.14& 217.66& 250.66 &36:45.82 &11:57.70& 29.35& 30.09& 29.16& 99.99& 99.99& 99.99& 1.575& 0.110& 0.065&20011 625.000 &4-389.0& 1.54& 374.50& 461.00 &36:45.83 &12:24.47& 99.99& 99.99& 99.99& 29.42& 29.70& 28.97& 1.682& 0.781& 0.062&01000 626.000 &4-441.0& 0.07& 259.00& 315.50 &36:45.90 &12:5.509& 29.12& 30.13& 28.79& 99.99& 99.99& 99.99& 2.364& 0.080& 0.066&20011 627.000 &4-448.2& 1.65& 219.18& 276.94 &36:46.02 &11:59.93& 28.13& 30.45& 28.50& 28.14& 32.26& 28.60& 0.500& 0.020& 0.000&00003 628.000 &4-522.0& 0.53& 129.50& 162.00 &36:46.07 &11:45.10& 29.27& 30.15& 28.72& 99.99& 99.99& 99.99& 2.382& 0.060& 0.052&00000 629.000 &4-293.0& 0.02& 431.50& 568.00 &36:46.09 &12:36.64& 29.34& 29.92& 29.01& 28.94& 29.36& 28.60& 2.259& 0.060& 0.057&22111 630.000 &4-528.0& 0.44& 123.50& 155.00 &36:46.09 &11:44.19& 99.99& 99.99& 99.99& 29.76& 30.04& 29.13& 1.217& 0.570& 0.053&01000 631.000 &4-300.0& 1.10& 408.33& 540.33 &36:46.12 &12:32.96& 28.91& 29.47& 28.85& 99.99& 99.99& 99.99& 2.315& 0.120& 0.090&10000 632.000 &4-288.0& 0.94& 441.50& 585.00 &36:46.13 &12:38.67& 99.99& 99.99& 99.99& 29.28& 29.28& 29.29& 1.912& 1.142& 0.048&01000 633.000 &4-396.2& 0.12& 283.68& 379.63 &36:46.16 &12:12.23& 28.32& 30.17& 28.69& 28.87& 32.16& 29.33& 0.900& 0.020& 0.000&00013 634.000 &4-362.0& 0.08& 329.66& 442.83 &36:46.19 &12:20.20& 28.60& 28.90& 28.40& 28.86& 29.46& 28.67& 2.309& 0.200& 0.091&22111 635.000 &4-470.2& 2.10& 156.18& 213.44 &36:46.20 &11:50.91& 27.92& 29.93& 28.29& 28.57& 30.97& 29.03& 1.300& 0.010& 0.000&00003 636.000 &4-396.0& 0.50& 283.50& 384.00 &36:46.22 &12:12.63& 99.99& 99.99& 99.99& 29.28& 29.95& 28.87& 1.916& 0.619& 0.049&01000 637.000 &4-294.0& 0.08& 416.00& 562.50 &36:46.22 &12:35.24& 29.38& 30.19& 29.22& 99.99& 99.99& 99.99& 1.860& 0.080& 0.050&10010 638.000 &4-418.0& 1.61& 238.50& 326.00 &36:46.23 &12:5.125& 29.49& 30.30& 29.41& 99.99& 99.99& 99.99& 1.972& 0.060& 0.049&10000 639.000 &4-418.0& 0.83& 230.50& 323.00 &36:46.30 &12:4.383& 29.51& 30.36& 28.89& 99.99& 99.99& 99.99& 1.903& 0.060& 0.052&10000 640.000 &4-372.0& 0.50& 298.00& 415.50 &36:46.32 &12:16.04& 29.42& 30.26& 29.10& 99.99& 99.99& 99.99& 1.784& 0.080& 0.057&00000 641.000 &4-390.0& 0.26& 271.50& 384.00 &36:46.37 &12:11.89& 99.99& 99.99& 99.99& 29.60& 30.14& 29.51& 1.418& 0.520& 0.042&01000 642.000 &4-390.2& 0.00& 268.66& 383.45 &36:46.37 &12:11.63& 28.44& 30.55& 28.81& 28.07& 30.47& 28.53& 0.500& 0.020& 0.000&00013 643.000 &4-470.0& 0.62& 141.50& 212.50 &36:46.38 &11:49.95& 29.11& 29.91& 29.01& 99.99& 99.99& 99.99& 1.894& 0.120& 0.081&10000 644.000 &4-470.0& 0.36& 137.50& 212.00 &36:46.42 &11:49.66& 29.86& 30.73& 29.62& 99.99& 99.99& 99.99& 1.389& 0.060& 0.050&10000 645.000 &4-461.0& 1.38& 155.00& 236.50 &36:46.43 &11:52.72& 29.76& 30.60& 29.94& 99.99& 99.99& 99.99& 1.301& 0.080& 0.046&00000 646.000 &4-262.2& 0.08& 415.66& 590.43 &36:46.47 &12:37.47& 28.37& 30.05& 28.74& 28.24& 30.64& 28.70& 0.700& 0.020& 0.000&00013 647.000 &4-504.0& 0.93& 93.500& 159.00 &36:46.47 &11:42.64& 29.27& 29.80& 28.92& 99.99& 99.99& 99.99& 2.366& 0.060& 0.057&10000 648.000 &4-443.0& 0.16& 172.50& 265.00 &36:46.47 &11:56.11& 29.57& 30.39& 29.01& 28.20& 29.09& 28.19& 1.833& 0.060& 0.047&21111 649.000 &4-470.0& 0.51& 131.33& 215.16 &36:46.53 &11:49.57& 99.99& 99.99& 99.99& 28.84& 29.23& 28.72& 1.544& 0.401& 0.118&01000 650.000 &4-262.0& 0.93& 407.50& 588.00 &36:46.55 &12:36.83& 99.99& 99.99& 99.99& 29.19& 29.64& 28.85& 2.078& 0.819& 0.054&01000 651.000 &4-434.0& 0.44& 175.50& 281.00 &36:46.58 &11:57.59& 29.73& 30.39& 29.17& 99.99& 99.99& 99.99& 1.562& 0.060& 0.047&00000 652.000 &4-442.0& 0.85& 143.50& 250.00 &36:46.68 &11:53.12& 30.06& 30.58& 30.33& 99.99& 99.99& 99.99& 1.158& 0.060& 0.051&20000 653.000 &4-388.0& 0.06& 225.33& 359.33 &36:46.68 &12:7.034& 28.91& 29.67& 28.65& 99.99& 99.99& 99.99& 2.303& 0.120& 0.080&10011 654.000 &4-241.0& 1.36& 388.00& 588.50 &36:46.78 &12:35.69& 99.99& 99.99& 99.99& 29.12& 29.72& 28.82& 1.880& 0.766& 0.041&00000 655.000 &4-415.0& 1.20& 153.00& 280.00 &36:46.83 &11:56.15& 29.33& 29.50& 29.87& 99.99& 99.99& 99.99& 1.637& 0.110& 0.063&10000 656.000 &4-241.2& 1.31& 384.00& 592.50 &36:46.87 &12:35.76& 99.99& 99.99& 99.99& 29.58& 29.95& 31.04& 1.209& 0.618& 0.054&00000 657.000 &4-280.0& 1.76& 357.00& 556.50 &36:46.87 &12:31.17& 99.99& 99.99& 99.99& 29.53& 29.94& 29.44& 1.294& 0.621& 0.049&01000 658.000 &4-336.0& 1.44& 265.00& 440.50 &36:46.93 &12:16.11& 99.99& 99.99& 99.99& 29.22& 30.06& 29.06& 1.722& 0.556& 0.045&01000 659.000 &4-415.0& 0.93& 143.50& 279.00 &36:46.94 &11:55.52& 99.99& 99.99& 99.99& 29.29& 30.20& 29.20& 1.881& 0.493& 0.049&01000 660.000 &4-470.0& 3.52& 105.00& 230.50 &36:46.96 &11:49.18& 29.23& 30.06& 29.16& 99.99& 99.99& 99.99& 2.129& 0.080& 0.048&00000 661.000 &4-382.0& 1.03& 193.50& 350.00 &36:46.97 &12:4.314& 29.78& 30.41& 29.29& 99.99& 99.99& 99.99& 1.488& 0.060& 0.051&00000 662.000 &4-298.0& 2.26& 264.00& 451.50 &36:47.04 &12:16.94& 99.99& 99.99& 99.99& 29.26& 30.08& 29.00& 1.673& 0.550& 0.054&01000 663.000 &4-381.0& 0.68& 161.00& 326.50 &36:47.14 &12:0.428& 29.30& 30.37& 29.05& 99.99& 99.99& 99.99& 2.011& 0.080& 0.047&10001 664.000 &4-216.2& 0.04& 351.16& 585.43 &36:47.17 &12:33.12& 28.03& 29.17& 28.40& 27.92& 29.38& 28.38& 1.200& 0.040& 0.000&10013 665.000 &4-315.2& 1.68& 220.00& 409.50 &36:47.18 &12:10.78& 29.51& 30.09& 29.61& 28.46& 29.14& 28.37& 1.653& 0.080& 0.044&11101 666.000 &4-386.0& 3.29& 101.50& 255.00 &36:47.21 &11:50.95& 29.58& 29.68& 29.63& 99.99& 99.99& 99.99& 1.804& 0.060& 0.042&20000 667.000 &4-216.2& 0.74& 344.01& 582.62 &36:47.23 &12:32.46& 27.79& 30.03& 28.16& 32.90& 99.00& 33.36& 1.300& 0.030& 0.000&00003 668.000 &4-281.2& 1.27& 272.66& 491.44 &36:47.27 &12:20.67& 28.35& 30.54& 28.72& 99.00& 99.00& 99.00& 0.700& 0.020& 0.000&00003 669.000 &4-386.0& 3.28& 95.000& 254.50 &36:47.28 &11:50.51& 29.69& 30.02& 29.93& 99.99& 99.99& 99.99& 1.403& 0.080& 0.050&10001 670.000 &4-298.2& 0.06& 239.00& 452.50 &36:47.34 &12:15.50& 99.99& 99.99& 99.99& 29.22& 30.00& 28.79& 1.716& 0.589& 0.036&01010 671.000 &4-386.2& 2.22& 95.660& 265.43 &36:47.36 &11:51.44& 27.97& 30.56& 28.34& 28.12& 31.86& 28.58& 1.100& 0.020& 0.000&10003 672.000 &4-250.2& 1.14& 281.89& 515.12 &36:47.37 &12:23.17& 28.19& 29.47& 28.56& 99.00& 31.74& 99.00& 1.500& 0.050& 0.000&10003 673.000 &4-298.0& 1.12& 245.00& 465.50 &36:47.38 &12:16.92& 99.99& 99.99& 99.99& 28.97& 29.84& 28.68& 2.179& 0.681& 0.055&01000 674.000 &4-232.11& 1.40& 333.50& 584.00 &36:47.38 &12:31.98& 99.99& 99.99& 99.99& 29.35& 29.58& 29.23& 1.785& 0.869& 0.050&00000 675.000 &4-188.0& 0.33& 320.50& 592.00 &36:47.60 &12:31.83& 99.99& 99.99& 99.99& 29.06& 29.80& 28.58& 2.332& 0.713& 0.055&01000 676.000 &4-346.2& 1.64& 141.15& 369.94 &36:47.74 &12:2.735& 28.16& 30.70& 28.53& 30.20& 32.51& 30.66& 0.800& 0.020& 0.000&00003 677.000 &4-188.0& 1.24& 311.33& 599.33 &36:47.77 &12:31.86& 99.99& 99.99& 99.99& 28.83& 29.04& 28.66& 1.970& 0.950& 0.082&00000 678.000 &4-174.2& 2.34& 283.13& 564.44 &36:47.78 &12:27.26& 27.93& 29.49& 28.30& 27.60& 29.66& 28.06& 1.100& 0.040& 0.000&10003 679.000 &4-212.0& 0.50& 258.66& 539.33 &36:47.86 &12:23.73& 29.03& 29.38& 28.96& 99.99& 99.99& 99.99& 2.389& 0.090& 0.075&10000 680.000 &4-198.0& 1.22& 251.00& 531.50 &36:47.89 &12:22.66& 99.99& 99.99& 99.99& 28.96& 29.72& 28.96& 2.192& 0.766& 0.056&01000 681.000 &4-236.0& 0.45& 226.66& 499.66 &36:47.89 &12:18.54& 29.15& 29.88& 28.86& 28.40& 29.11& 28.32& 1.931& 0.110& 0.081&21100 682.000 &4-304.0& 0.75& 144.33& 396.33 &36:47.95 &12:5.083& 28.97& 29.83& 28.81& 99.99& 99.99& 99.99& 2.160& 0.120& 0.085&20000 683.000 &4-212.0& 1.24& 259.50& 551.00 &36:47.96 &12:24.76& 99.99& 99.99& 99.99& 29.38& 29.72& 29.10& 1.744& 0.765& 0.050&01000 684.000 &4-170.0& 0.89& 229.00& 541.50 &36:48.23 &12:22.07& 29.31& 30.08& 29.27& 99.99& 99.99& 99.99& 1.973& 0.080& 0.042&10000 685.000 &4-158.0& 0.35& 235.50& 562.00 &36:48.33 &12:24.15& 29.64& 29.95& 29.04& 99.99& 99.99& 99.99& 1.716& 0.060& 0.059&10000 686.000 &4-158.0& 0.11& 232.66& 561.66 &36:48.36 &12:23.95& 28.89& 29.62& 28.62& 99.99& 99.99& 99.99& 2.462& 0.110& 0.066&10010 687.000 &4-260.2& 1.67& 163.32& 470.60 &36:48.36 &12:12.28& 27.98& 30.07& 28.35& 29.74& 99.00& 30.20& 0.800& 0.030& 0.000&00003 688.000 &4-122.0& 0.79& 249.00& 597.50 &36:48.49 &12:27.90& 99.99& 99.99& 99.99& 29.62& 29.57& 29.99& 1.200& 0.880& 0.052&00000 689.000 &4-122.0& 0.33& 241.00& 599.50 &36:48.60 &12:27.59& 99.99& 99.99& 99.99& 29.23& 29.71& 29.32& 1.699& 0.768& 0.044&01000 690.000 &4-203.2& 0.24& 115.00& 455.00 &36:48.81 &12:8.105& 99.99& 99.99& 99.99& 28.98& 29.54& 28.82& 1.814& 0.601& 0.108&01010 691.000 &4-203.2& 0.20& 116.66& 459.66 &36:48.82 &12:8.544& 29.13& 29.85& 29.04& 99.99& 99.99& 99.99& 1.967& 0.110& 0.079&20010 692.000 &4-105.2& 1.07& 205.17& 581.44 &36:48.84 &12:23.87& 27.80& 29.46& 28.17& 28.52& 30.19& 28.98& 1.400& 0.040& 0.000&10003 693.000 &4-128.0& 0.48& 185.75& 556.00 &36:48.85 &12:20.62& 29.24& 29.54& 29.02& 29.72& 29.91& 29.77& 1.611& 0.130& 0.063&22101 694.000 &4-148.2& 0.32& 161.16& 526.94 &36:48.88 &12:16.74& 27.91& 30.40& 28.28& 27.47& 30.43& 27.93& 0.500& 0.020& 0.000&00003 695.000 &4-120.0& 0.74& 173.50& 559.00 &36:49.02 &12:20.12& 29.61& 30.30& 29.32& 99.99& 99.99& 99.99& 1.741& 0.060& 0.042&10000 696.000 &4-131.0& 1.95& 127.50& 511.00 &36:49.14 &12:13.39& 29.57& 30.13& 29.25& 99.99& 99.99& 99.99& 1.826& 0.060& 0.046&10000 697.000 &4-124.0& 0.07& 140.50& 534.00 &36:49.19 &12:16.05& 29.39& 30.02& 28.91& 99.99& 99.99& 99.99& 2.155& 0.060& 0.055&20011 698.000 &4-99.0& 1.53& 169.50& 579.00 &36:49.25 &12:21.49& 29.81& 30.26& 29.13& 99.99& 99.99& 99.99& 1.420& 0.060& 0.057&00000 699.000 &4-95.0& 2.30& 133.50& 555.00 &36:49.47 &12:17.37& 99.99& 99.99& 99.99& 29.14& 29.79& 29.04& 2.173& 0.716& 0.050&01000 700.000 &4-74.0& 1.57& 148.00& 596.50 &36:49.66 &12:21.64& 99.99& 99.99& 99.99& 29.05& 29.47& 29.17& 2.029& 0.958& 0.058&01000 701.000 &4-50.0& 1.35& 101.00& 556.50 &36:49.86 &12:15.52& 99.99& 99.99& 99.99& 28.96& 29.64& 28.68& 2.200& 0.820& 0.048&01000 702.000 &4-50.2& 1.75& 94.330& 552.61 &36:49.88 &12:14.73& 28.00& 29.68& 28.37& 99.00& 99.00& 99.00& 1.000& 0.030& 0.000&00003 703.000 &4-50.2& 1.04& 105.07& 571.55 &36:49.93 &12:16.93& 27.78& 29.29& 28.15& 30.77& 99.00& 31.23& 1.600& 0.050& 0.000&00003 704.000 &4-50.0& 0.78& 93.500& 562.00 &36:49.99 &12:15.45& 29.30& 29.51& 28.96& 99.99& 99.99& 99.99& 2.346& 0.060& 0.047&10000 705.000 &4-33.0& 2.13& 107.00& 596.50 &36:50.12 &12:19.07& 29.29& 29.93& 28.85& 99.99& 99.99& 99.99& 1.986& 0.080& 0.064&10000
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.